SQL Tuning Guide

4_sql-tuning-guide

4_PDF_12c_SQLTuningGuide_E85762-03

SQLTUNE_02_PDF_122_SQLTuningGuide_E85762-03

User Manual:

Open the PDF directly: View PDF PDF.
Page Count: 837

DownloadSQL Tuning Guide Sql-tuning-guide
Open PDF In BrowserView PDF
Oracle® Database
SQL Tuning Guide

12c Release 2 (12.2)
E85762-03
March 2018

Oracle Database SQL Tuning Guide, 12c Release 2 (12.2)
E85762-03
Copyright © 2013, 2018, Oracle and/or its affiliates. All rights reserved.
Primary Author: Lance Ashdown
Contributing Authors: Nigel Bayliss, Maria Colgan, Tom Kyte
Contributors: Hermann Baer, Ali Cakmak, Sunil Chakkappen, Immanuel Chan, Deba Chatterjee, Chris
Chiappa, Dinesh Das, Leonidas Galanis, William Endress, Marcus Fallen, Bruce Golbus, Katsumi Inoue,
Shantanu Joshi, Adam Kociubes, Keith Laker, Allison Lee, Sue Lee, David McDermid, Colin McGregor, Ajit
Mylavarapu, Ted Persky, Lei Sheng, Ekrem Soylemez, Hong Su, Murali Thiyagarajah, Randy Urbano, Sahil
Vazirani, Bharath Venkatakrishnan, Hailing Yu, John Zimmerman
This software and related documentation are provided under a license agreement containing restrictions on
use and disclosure and are protected by intellectual property laws. Except as expressly permitted in your
license agreement or allowed by law, you may not use, copy, reproduce, translate, broadcast, modify,
license, transmit, distribute, exhibit, perform, publish, or display any part, in any form, or by any means.
Reverse engineering, disassembly, or decompilation of this software, unless required by law for
interoperability, is prohibited.
The information contained herein is subject to change without notice and is not warranted to be error-free. If
you find any errors, please report them to us in writing.
If this is software or related documentation that is delivered to the U.S. Government or anyone licensing it on
behalf of the U.S. Government, then the following notice is applicable:
U.S. GOVERNMENT END USERS: Oracle programs, including any operating system, integrated software,
any programs installed on the hardware, and/or documentation, delivered to U.S. Government end users are
"commercial computer software" pursuant to the applicable Federal Acquisition Regulation and agencyspecific supplemental regulations. As such, use, duplication, disclosure, modification, and adaptation of the
programs, including any operating system, integrated software, any programs installed on the hardware,
and/or documentation, shall be subject to license terms and license restrictions applicable to the programs.
No other rights are granted to the U.S. Government.
This software or hardware is developed for general use in a variety of information management applications.
It is not developed or intended for use in any inherently dangerous applications, including applications that
may create a risk of personal injury. If you use this software or hardware in dangerous applications, then you
shall be responsible to take all appropriate fail-safe, backup, redundancy, and other measures to ensure its
safe use. Oracle Corporation and its affiliates disclaim any liability for any damages caused by use of this
software or hardware in dangerous applications.
Oracle and Java are registered trademarks of Oracle and/or its affiliates. Other names may be trademarks of
their respective owners.
Intel and Intel Xeon are trademarks or registered trademarks of Intel Corporation. All SPARC trademarks are
used under license and are trademarks or registered trademarks of SPARC International, Inc. AMD, Opteron,
the AMD logo, and the AMD Opteron logo are trademarks or registered trademarks of Advanced Micro
Devices. UNIX is a registered trademark of The Open Group.
This software or hardware and documentation may provide access to or information about content, products,
and services from third parties. Oracle Corporation and its affiliates are not responsible for and expressly
disclaim all warranties of any kind with respect to third-party content, products, and services unless otherwise
set forth in an applicable agreement between you and Oracle. Oracle Corporation and its affiliates will not be
responsible for any loss, costs, or damages incurred due to your access to or use of third-party content,
products, or services, except as set forth in an applicable agreement between you and Oracle.

Contents
Preface
Audience

xxv

Documentation Accessibility

xxv

Related Documents

xxvi

Conventions

xxvi

Changes in This Release for Oracle Database SQL Tuning Guide
Changes in Oracle Database 12c Release 2 (12.2.0.1)

xxvii

Changes in Oracle Database 12c Release 1 (12.1.0.2)

xxx

Changes in Oracle Database 12c Release 1 (12.1.0.1)

xxxi

Part I
1

SQL Performance Fundamentals

Introduction to SQL Tuning
1.1

About SQL Tuning

1-1

1.2

Purpose of SQL Tuning

1-1

1.3

Prerequisites for SQL Tuning

1-2

1.4

Tasks and Tools for SQL Tuning

1-2

1.4.1

SQL Tuning Tasks

1-3

1.4.2

SQL Tuning Tools

1-5

1.4.2.1

Automated SQL Tuning Tools

1-5

1.4.2.2

Manual SQL Tuning Tools

1-8

1.4.3

2

User Interfaces to SQL Tuning Tools

1-11

SQL Performance Methodology
2.1

2.2

Guidelines for Designing Your Application

2-1

2.1.1

Guideline for Data Modeling

2-1

2.1.2

Guideline for Writing Efficient Applications

2-2

Guidelines for Deploying Your Application

2-3

iii

Part II
3

2.2.1

Guideline for Deploying in a Test Environment

2-3

2.2.2

Guidelines for Application Rollout

2-4

Query Optimizer Fundamentals

SQL Processing
3.1

About SQL Processing
3.1.1

3.2

3.3

4

SQL Parsing

3-1
3-3

3.1.1.1

Syntax Check

3-3

3.1.1.2

Semantic Check

3-4

3.1.1.3

Shared Pool Check

3-4

3.1.2

SQL Optimization

3-6

3.1.3

SQL Row Source Generation

3-6

3.1.4

SQL Execution

3-7

How Oracle Database Processes DML

3-9

3.2.1

How Row Sets Are Fetched

3.2.2

Read Consistency

3-10

3.2.3

Data Changes

3-10

How Oracle Database Processes DDL

3-9

3-10

Query Optimizer Concepts
4.1

4.2

Introduction to the Query Optimizer

4-1

4.1.1

Purpose of the Query Optimizer

4-2

4.1.2

Cost-Based Optimization

4-2

4.1.3

Execution Plans

4-3

4.1.3.1

Query Blocks

4-3

4.1.3.2

Query Subplans

4-4

4.1.3.3

Analogy for the Optimizer

4-4

About Optimizer Components

4-5

4.2.1

Query Transformer

4-6

4.2.2

Estimator

4-7

4.2.2.1

Selectivity

4-9

4.2.2.2

Cardinality

4-10

4.2.2.3

Cost

4-10

Plan Generator

4-11

4.2.3
4.3

About Automatic Tuning Optimizer

4-13

4.4

About Adaptive Query Optimization

4-13

4.4.1

Adaptive Query Plans

4.4.1.1

About Adaptive Query Plans

4-14
4-15

iv

4.4.1.2

Purpose of Adaptive Query Plans

4-16

4.4.1.3

How Adaptive Query Plans Work

4-16

4.4.1.4

When Adaptive Query Plans Are Enabled

4-24

4.4.2

5

Adaptive Statistics

4-24

4.4.2.1

Dynamic Statistics

4-24

4.4.2.2

Automatic Reoptimization

4-25

4.4.2.3

SQL Plan Directives

4-28

4.4.2.4

When Adaptive Statistics Are Enabled

4-28

4.5

About Approximate Query Processing

4-29

4.6

About SQL Plan Management

4-31

4.7

About the Expression Statistics Store (ESS)

4-32

Query Transformations
5.1

OR Expansion

5-2

5.2

View Merging

5-3

5.2.1

Query Blocks in View Merging

5-4

5.2.2

Simple View Merging

5-5

5.2.3

Complex View Merging

5-7

5.3

Predicate Pushing

5-9

5.4

Subquery Unnesting

5-10

5.5

Query Rewrite with Materialized Views

5-11

5.6

Star Transformation

5-12

5.6.1

About Star Schemas

5-12

5.6.2

Purpose of Star Transformations

5-13

5.6.3

How Star Transformation Works

5-13

5.6.4

Controls for Star Transformation

5-14

5.6.5

Star Transformation: Scenario

5-14

5.6.6

Temporary Table Transformation: Scenario

5-17

5.7

In-Memory Aggregation (VECTOR GROUP BY)

5-19

5.8

Cursor-Duration Temporary Tables

5-19

5.8.1

Purpose of Cursor-Duration Temporary Tables

5-19

5.8.2

How Cursor-Duration Temporary Tables Work

5-20

5.8.3

Cursor-Duration Temporary Tables: Example

5-20

5.9

Table Expansion

5-21

5.9.1

Purpose of Table Expansion

5-22

5.9.2

How Table Expansion Works

5-22

5.9.3

Table Expansion: Scenario

5-22

5.9.4

Table Expansion and Star Transformation: Scenario

5-25

5.10

Join Factorization

5.10.1

Purpose of Join Factorization

5-27
5-27

v

Part III
6

How Join Factorization Works

5-27

5.10.3

Factorization and Join Orders: Scenario

5-28

5.10.4

Factorization of Outer Joins: Scenario

5-30

Query Execution Plans

Generating and Displaying Execution Plans
6.1

Introduction to Execution Plans

6-1

6.2

About Plan Generation and Display

6-1

6.3

6.4

7

5.10.2

6.2.1

About the Plan Explanation

6-2

6.2.2

Why Execution Plans Change

6-2

6.2.2.1

Different Schemas

6-3

6.2.2.2

Different Costs

6-3

6.2.3

Guideline for Minimizing Throw-Away

6-4

6.2.4

Guidelines for Evaluating Execution Plans Using EXPLAIN PLAN

6-4

6.2.5

Guidelines for Evaluating Plans Using the V$SQL_PLAN Views

6-5

6.2.6

EXPLAIN PLAN Restrictions

6-5

6.2.7

Guidelines for Creating PLAN_TABLE

6-6

Generating Execution Plans

6-6

6.3.1

Executing EXPLAIN PLAN for a Single Statement

6-7

6.3.2

Executing EXPLAIN PLAN Using a Statement ID

6-8

6.3.3

Directing EXPLAIN PLAN Output to a Nondefault Table

6-8

Displaying PLAN_TABLE Output
6.4.1

Displaying an Execution Plan: Example

6.4.2

Customizing PLAN_TABLE Output

6-8
6-9
6-10

Reading Execution Plans
7.1

Reading Execution Plans: Basic

7-1

7.2

Reading Execution Plans: Advanced

7-2

7.2.1

Reading Adaptive Query Plans

7-2

7.2.2

Viewing Parallel Execution with EXPLAIN PLAN

7-6

7.2.2.1

About EXPLAIN PLAN and Parallel Queries

7-7

7.2.2.2

Viewing Parallel Queries with EXPLAIN PLAN: Example

7-8

7.2.3

Viewing Bitmap Indexes with EXPLAIN PLAN

7.2.4

Viewing Result Cache with EXPLAIN PLAN

7-10

7.2.5

Viewing Partitioned Objects with EXPLAIN PLAN

7-10

7.2.5.1

Displaying Range and Hash Partitioning with EXPLAIN PLAN:
Examples

7-9

7-11

vi

7.2.5.2

Pruning Information with Composite Partitioned Objects:
Examples

7-13

7.2.5.3

Examples of Partial Partition-Wise Joins

7-14

7.2.5.4

Example of Full Partition-Wise Join

7-16

7.2.5.5

Examples of INLIST ITERATOR and EXPLAIN PLAN

7-17

7.2.5.6

Example of Domain Indexes and EXPLAIN PLAN

7-18

7.2.6
7.3

Part IV
8

PLAN_TABLE Columns

Execution Plan Reference

7-18
7-29

7.3.1

Execution Plan Views

7-30

7.3.2

PLAN_TABLE Columns

7-30

7.3.3

DBMS_XPLAN Program Units

7-39

SQL Operators: Access Paths and Joins

Optimizer Access Paths
8.1

Introduction to Access Paths

8-1

8.2

Table Access Paths

8-2

8.2.1

8-3

8.2.1.1

Row Storage in Data Blocks and Segments: A Primer

8-3

8.2.1.2

Importance of Rowids for Row Access

8-4

8.2.1.3

Direct Path Reads

8-5

8.2.2

Full Table Scans

8-6

8.2.2.1

When the Optimizer Considers a Full Table Scan

8-6

8.2.2.2

How a Full Table Scan Works

8-7

8.2.2.3

Full Table Scan: Example

8-8

8.2.3

Table Access by Rowid

8-9

8.2.3.1

When the Optimizer Chooses Table Access by Rowid

8.2.3.2

How Table Access by Rowid Works

8-10

8.2.3.3

Table Access by Rowid: Example

8-10

8.2.4

Sample Table Scans

8-9

8-10

8.2.4.1

When the Optimizer Chooses a Sample Table Scan

8-11

8.2.4.2

Sample Table Scans: Example

8-11

8.2.5

8.3

About Heap-Organized Table Access

In-Memory Table Scans

8-12

8.2.5.1

When the Optimizer Chooses an In-Memory Table Scan

8-12

8.2.5.2

In-Memory Query Controls

8-12

8.2.5.3

In-Memory Table Scans: Example

8-13

B-Tree Index Access Paths
8.3.1

About B-Tree Index Access

8-14
8-15

8.3.1.1

B-Tree Index Structure

8-15

8.3.1.2

How Index Storage Affects Index Scans

8-16

vii

8.3.1.3

Unique and Nonunique Indexes

8-17

8.3.1.4

B-Tree Indexes and Nulls

8-17

8.3.2

8-19

8.3.2.1

When the Optimizer Considers Index Unique Scans

8-19

8.3.2.2

How Index Unique Scans Work

8-20

8.3.2.3

Index Unique Scans: Example

8-21

8.3.3

Index Range Scans

8-22

8.3.3.1

When the Optimizer Considers Index Range Scans

8-22

8.3.3.2

How Index Range Scans Work

8-23

8.3.3.3

Index Range Scan: Example

8-24

8.3.3.4

Index Range Scan Descending: Example

8-25

8.3.4

Index Full Scans

8-26

8.3.4.1

When the Optimizer Considers Index Full Scans

8-26

8.3.4.2

How Index Full Scans Work

8-26

8.3.4.3

Index Full Scans: Example

8-27

8.3.5

Index Fast Full Scans

8-28

8.3.5.1

When the Optimizer Considers Index Fast Full Scans

8-28

8.3.5.2

How Index Fast Full Scans Work

8-28

8.3.5.3

Index Fast Full Scans: Example

8-29

8.3.6

Index Skip Scans

8-29

8.3.6.1

When the Optimizer Considers Index Skips Scans

8-29

8.3.6.2

How Index Skip Scans Work

8-30

8.3.6.3

Index Skip Scans: Example

8-30

8.3.7

8.4

Index Unique Scans

Index Join Scans

8-31

8.3.7.1

When the Optimizer Considers Index Join Scans

8-32

8.3.7.2

How Index Join Scans Work

8-32

8.3.7.3

Index Join Scans: Example

8-32

Bitmap Index Access Paths
8.4.1

About Bitmap Index Access

8-33
8-34

8.4.1.1

Differences Between Bitmap and B-Tree Indexes

8-34

8.4.1.2

Purpose of Bitmap Indexes

8-35

8.4.1.3

Bitmaps and Rowids

8-36

8.4.1.4

Bitmap Join Indexes

8-37

8.4.1.5

Bitmap Storage

8-38

Bitmap Conversion to Rowid

8-39

8.4.2

8.4.2.1

When the Optimizer Chooses Bitmap Conversion to Rowid

8-39

8.4.2.2

How Bitmap Conversion to Rowid Works

8-39

8.4.2.3

Bitmap Conversion to Rowid: Example

8-39

8.4.3

Bitmap Index Single Value

8-40

8.4.3.1

When the Optimizer Considers Bitmap Index Single Value

8-40

8.4.3.2

How Bitmap Index Single Value Works

8-40

viii

8.4.3.3
8.4.4

8-41

When the Optimizer Considers Bitmap Index Range Scans

8-42

8.4.4.2

How Bitmap Index Range Scans Work

8-42

8.4.4.3

Bitmap Index Range Scans: Example

8-42

Bitmap Merge

8-43

8.4.5.1

When the Optimizer Considers Bitmap Merge

8-43

8.4.5.2

How Bitmap Merge Works

8-43

8.4.5.3

Bitmap Merge: Example

8-44

Table Cluster Access Paths
8.5.1

Cluster Scans

8-44
8-45

8.5.1.1

When the Optimizer Considers Cluster Scans

8-45

8.5.1.2

How a Cluster Scan Works

8-45

8.5.1.3

Cluster Scans: Example

8-46

8.5.2

9

Bitmap Index Range Scans

8-41

8.4.4.1

8.4.5

8.5

Bitmap Index Single Value: Example

Hash Scans

8-47

8.5.2.1

When the Optimizer Considers a Hash Scan

8-47

8.5.2.2

How a Hash Scan Works

8-47

8.5.2.3

Hash Scans: Example

8-47

Joins
9.1

9.2

About Joins

9-1

9.1.1

Join Trees

9-2

9.1.2

How the Optimizer Executes Join Statements

9-3

9.1.3

How the Optimizer Chooses Execution Plans for Joins

9-4

Join Methods
9.2.1

Nested Loops Joins

9-5
9-6

9.2.1.1

When the Optimizer Considers Nested Loops Joins

9-7

9.2.1.2

How Nested Loops Joins Work

9-8

9.2.1.3

Nested Nested Loops

9-8

9.2.1.4

Current Implementation for Nested Loops Joins

9-11

9.2.1.5

Original Implementation for Nested Loops Joins

9-13

9.2.1.6

Nested Loops Controls

9-14

9.2.2

Hash Joins

9-16

9.2.2.1

When the Optimizer Considers Hash Joins

9-16

9.2.2.2

How Hash Joins Work

9-17

9.2.2.3

How Hash Joins Work When the Hash Table Does Not Fit in the
PGA

9-19

Hash Join Controls

9-20

9.2.2.4
9.2.3

Sort Merge Joins

9-20

9.2.3.1

When the Optimizer Considers Sort Merge Joins

9-21

9.2.3.2

How Sort Merge Joins Work

9-21

ix

9.2.3.3
9.2.4

9.3

Cartesian Joins

9-25

9.2.4.2

How Cartesian Joins Work

9-26

9.2.4.3

Cartesian Join Controls

9-27
9-27

Inner Joins

9-28

9.3.1.1

Equijoins

9-28

9.3.1.2

Nonequijoins

9-28

9.3.1.3

Band Joins

9-29

9.3.2

Outer Joins

9-32

9.3.2.1

Nested Loops Outer Joins

9-33

9.3.2.2

Hash Join Outer Joins

9-34

9.3.2.3

Sort Merge Outer Joins

9-35

9.3.2.4

Full Outer Joins

9-36

9.3.2.5

Multiple Tables on the Left of an Outer Join

9-37

9.3.3

Semijoins

9-37

9.3.3.1

When the Optimizer Considers Semijoins

9-38

9.3.3.2

How Semijoins Work

9-38

9.3.4

Antijoins

9-40

9.3.4.1

When the Optimizer Considers Antijoins

9-40

9.3.4.2

How Antijoins Work

9-41

9.3.4.3

How Antijoins Handle Nulls

9-42

Join Optimizations
9.4.1

Bloom Filters

9-44
9-45

9.4.1.1

Purpose of Bloom Filters

9-45

9.4.1.2

How Bloom Filters Work

9-46

9.4.1.3

Bloom Filter Controls

9-46

9.4.1.4

Bloom Filter Metadata

9-47

9.4.1.5

Bloom Filters: Scenario

9-47

9.4.2

Partition-Wise Joins

9-48

9.4.2.1

Purpose of Partition-Wise Joins

9-49

9.4.2.2

How Partition-Wise Joins Work

9-49

9.4.3

Part V

9-25

When the Optimizer Considers Cartesian Joins

Join Types

9.4

9-25

9.2.4.1

9.3.1

10

Sort Merge Join Controls

In-Memory Join Groups

9-52

Optimizer Statistics

Optimizer Statistics Concepts
10.1

Introduction to Optimizer Statistics

10-1

10.2

About Optimizer Statistics Types

10-3

x

10.2.1

Table Statistics

10-4

10.2.2

Column Statistics

10-4

10.2.3

Index Statistics

10-5

10.2.3.1

Types of Index Statistics

10-5

10.2.3.2

Index Clustering Factor

10-6

10.2.3.3

Effect of Index Clustering Factor on Cost: Example

10-9

10.2.4

Session-Specific Statistics for Global Temporary Tables

10.2.4.1
10.2.4.2

Shared and Session-Specific Statistics for Global Temporary
Tables

10-11

Effect of DBMS_STATS on Transaction-Specific Temporary
Tables

10-11

10.2.5

System Statistics

10-12

10.2.6

User-Defined Optimizer Statistics

10-12

10.3

How the Database Gathers Optimizer Statistics

10-13

10.3.1

DBMS_STATS Package

10-13

10.3.2

Supplemental Dynamic Statistics

10-14

10.3.3

Online Statistics Gathering for Bulk Loads

10-15

10.4

11

10-10

10.3.3.1

Purpose of Online Statistics Gathering for Bulk Loads

10-16

10.3.3.2

Global Statistics During Inserts into Empty Partitioned Tables

10-16

10.3.3.3

Index Statistics and Histograms During Bulk Loads

10-16

10.3.3.4

Restrictions for Online Statistics Gathering for Bulk Loads

10-17

10.3.3.5

Hints for Online Statistics Gathering for Bulk Loads

10-18

When the Database Gathers Optimizer Statistics

10-19

10.4.1

Sources for Optimizer Statistics

10-19

10.4.2

SQL Plan Directives

10-20

10.4.2.1

When the Database Creates SQL Plan Directives

10-20

10.4.2.2

How the Database Uses SQL Plan Directives

10-21

10.4.2.3

SQL Plan Directive Maintenance

10-22

10.4.2.4

How the Optimizer Uses SQL Plan Directives: Example

10-22

10.4.2.5

How the Optimizer Uses Extensions and SQL Plan Directives:
Example

10-27

10.4.3

When the Database Samples Data

10-30

10.4.4

How the Database Samples Data

10-32

Histograms
11.1

Purpose of Histograms

11-2

11.2

When Oracle Database Creates Histograms

11-2

11.3

How Oracle Database Chooses the Histogram Type

11-4

11.4

Cardinality Algorithms When Using Histograms

11-4

11.4.1

Endpoint Numbers and Values

11-5

11.4.2

Popular and Nonpopular Values

11-5

xi

11.4.3
11.5

Frequency Histograms

11-6

Criteria For Frequency Histograms

11-7

11.5.2

Generating a Frequency Histogram

11-7

Top Frequency Histograms

11-10

11.6.1

Criteria For Top Frequency Histograms

11-10

11.6.2

Generating a Top Frequency Histogram

11-11

11.7

Height-Balanced Histograms (Legacy)

11-14

11.7.1

Criteria for Height-Balanced Histograms

11-14

11.7.2

Generating a Height-Balanced Histogram

11-15

11.8

Hybrid Histograms

11-18

11.8.1

How Endpoint Repeat Counts Work

11-18

11.8.2

Criteria for Hybrid Histograms

11-20

11.8.3

Generating a Hybrid Histogram

11-21

Configuring Options for Optimizer Statistics Gathering
12.1

About Optimizer Statistics Collection

12-1

12.1.1

Purpose of Optimizer Statistics Collection

12-2

12.1.2

User Interfaces for Optimizer Statistics Management

12-2

12.2

12.1.2.1

Graphical Interface for Optimizer Statistics Management

12-2

12.1.2.2

Command-Line Interface for Optimizer Statistics Management

12-3

Setting Optimizer Statistics Preferences

12.2.1

About Optimizer Statistics Preferences

12-3
12-4

12.2.1.1

Purpose of Optimizer Statistics Preferences

12-4

12.2.1.2

DBMS_STATS Procedures for Setting Statistics Preferences

12-5

12.2.1.3

Statistics Preference Overrides

12-6

12.2.1.4

Setting Statistics Preferences: Example

12-7

12.2.2

Setting Global Optimizer Statistics Preferences Using Cloud Control

12-9

12.2.3

Setting Object-Level Optimizer Statistics Preferences Using Cloud
Control

12-9

12.2.4
12.3

Setting Optimizer Statistics Preferences from the Command Line

Configuring Options for Dynamic Statistics

12-10
12-11

12.3.1

About Dynamic Statistics Levels

12-12

12.3.2

Setting Dynamic Statistics Levels Manually

12-13

12.3.3

Disabling Dynamic Statistics

12-16

12.4

13

11-6

11.5.1
11.6

12

Bucket Compression

Managing SQL Plan Directives

12-16

Gathering Optimizer Statistics
13.1

Configuring Automatic Optimizer Statistics Collection

13.1.1

About Automatic Optimizer Statistics Collection

13-1
13-2

xii

13.1.2
13.1.3
13.2

Configuring Automatic Optimizer Statistics Collection Using Cloud
Control

13-2

Configuring Automatic Optimizer Statistics Collection from the
Command Line

13-4

Gathering Optimizer Statistics Manually

13-6

13.2.1

About Manual Statistics Collection with DBMS_STATS

13-7

13.2.2

Guidelines for Gathering Optimizer Statistics Manually

13-8

13.2.2.1

Guideline for Setting the Sample Size

13-9

13.2.2.2

Guideline for Gathering Statistics in Parallel

13-9

13.2.2.3

Guideline for Partitioned Objects

13-10

13.2.2.4

Guideline for Frequently Changing Objects

13-10

13.2.2.5

Guideline for External Tables

13-10

13.2.3

Determining When Optimizer Statistics Are Stale

13-11

13.2.4

Gathering Schema and Table Statistics

13-12

13.2.5

Gathering Statistics for Fixed Objects

13-13

13.2.6

Gathering Statistics for Volatile Tables Using Dynamic Statistics

13-14

13.2.7

Gathering Optimizer Statistics Concurrently

13-15

13.2.7.1

About Concurrent Statistics Gathering

13-15

13.2.7.2

Enabling Concurrent Statistics Gathering

13-18

13.2.7.3

Monitoring Statistics Gathering Operations

13-20

Gathering Incremental Statistics on Partitioned Objects

13-22

13.2.8

13.2.8.1

Purpose of Incremental Statistics

13-22

13.2.8.2

How DBMS_STATS Derives Global Statistics for Partitioned
tables

13-23

13.2.8.3

Gathering Statistics for a Partitioned Table: Basic Steps

13-27

13.2.8.4

Maintaining Incremental Statistics for Partition Maintenance
Operations

13-30

Maintaining Incremental Statistics for Tables with Stale or
Locked Partition Statistics

13-32

13.2.8.5
13.3

Gathering System Statistics Manually

13-34

13.3.1

About Gathering System Statistics with DBMS_STATS

13-35

13.3.2

Guidelines for Gathering System Statistics

13-37

13.3.3

Gathering Workload Statistics

13-37

13.3.3.1

About Workload Statistics

13-38

13.3.3.2

Starting and Stopping System Statistics Gathering

13-39

13.3.3.3

Gathering System Statistics During a Specified Interval

13-40

13.3.4

Gathering Noworkload Statistics

13-41

13.3.5

Deleting System Statistics

13-43

13.4

Running Statistics Gathering Functions in Reporting Mode

13-43

xiii

14

Managing Extended Statistics
14.1

Managing Column Group Statistics

14.1.1

About Statistics on Column Groups

14.1.1.1

Why Column Group Statistics Are Needed: Example

14-3

14.1.1.2

Automatic and Manual Column Group Statistics

14-5

14.1.1.3

User Interface for Column Group Statistics

14-5

Detecting Useful Column Groups for a Specific Workload

14-6

14.1.3

Creating Column Groups Detected During Workload Monitoring

14-9

14.1.4

Creating and Gathering Statistics on Column Groups Manually

14-11

14.1.5

Displaying Column Group Information

14-12

14.1.6

Dropping a Column Group

14-13

Managing Expression Statistics

14.2.1

About Expression Statistics

14.2.1.1

When Expression Statistics Are Useful: Example

14-13
14-14
14-15

14.2.2

Creating Expression Statistics

14-15

14.2.3

Displaying Expression Statistics

14-16

14.2.4

Dropping Expression Statistics

14-17

Controlling the Use of Optimizer Statistics
15.1

Locking and Unlocking Optimizer Statistics

15-1

15.1.1

Locking Statistics

15-1

15.1.2

Unlocking Statistics

15-2

15.2

Publishing Pending Optimizer Statistics

15-3

15.2.1

About Pending Optimizer Statistics

15-3

15.2.2

User Interfaces for Publishing Optimizer Statistics

15-5

15.2.3

Managing Published and Pending Statistics

15-6

15.3

16

14-2

14.1.2

14.2

15

14-1

Creating Artificial Optimizer Statistics for Testing

15-9

15.3.1

About Artificial Optimizer Statistics

15-9

15.3.2

Setting Artificial Optimizer Statistics for a Table

15-10

15.3.3

Setting Optimizer Statistics: Example

15-11

Managing Historical Optimizer Statistics
16.1

Restoring Optimizer Statistics

16-1

16.1.1

About Restore Operations for Optimizer Statistics

16-1

16.1.2

Guidelines for Restoring Optimizer Statistics

16-2

16.1.3

Restrictions for Restoring Optimizer Statistics

16-2

16.1.4

Restoring Optimizer Statistics Using DBMS_STATS

16-3

16.2

Managing Optimizer Statistics Retention

16.2.1

Obtaining Optimizer Statistics History

16-4
16-5

xiv

16.2.2

Changing the Optimizer Statistics Retention Period

16-5

16.2.3

Purging Optimizer Statistics

16-6

16.3

17

18

Reporting on Past Statistics Gathering Operations

Transporting Optimizer Statistics
17.1

About Transporting Optimizer Statistics

17-1

17.2

Transporting Optimizer Statistics to a Test Database: Tutorial

17-2

Analyzing Statistics Using Optimizer Statistics Advisor
18.1

About Optimizer Statistics Advisor

18.1.1

18-2

Problems with a Traditional Script-Based Approach

18-3

18.1.1.2

Advantages of Optimizer Statistics Advisor

18-4

Optimizer Statistics Advisor Concepts

18-4

18.1.2.1

Components of Optimizer Statistics Advisor

18-5

18.1.2.2

Operational Modes for Optimizer Statistics Advisor

18-9

18.1.3
18.2

Purpose of Optimizer Statistics Advisor

18-1

18.1.1.1
18.1.2

Command-Line Interface to Optimizer Statistics Advisor

Basic Tasks for Optimizer Statistics Advisor

18-9
18-11

18.2.1

Creating an Optimizer Statistics Advisor Task

18-14

18.2.2

Listing Optimizer Statistics Advisor Tasks

18-15

18.2.3

Creating Filters for an Optimizer Advisor Task

18-16

18.2.3.1

About Filters for Optimizer Statistics Advisor

18-16

18.2.3.2

Creating an Object Filter for an Optimizer Advisor Task

18-17

18.2.3.3

Creating a Rule Filter for an Optimizer Advisor Task

18-19

18.2.3.4

Creating an Operation Filter for an Optimizer Advisor Task

18-22

18.2.4

Executing an Optimizer Statistics Advisor Task

18-25

18.2.5

Generating a Report for an Optimizer Statistics Advisor Task

18-26

18.2.6

Implementing Optimizer Statistics Advisor Recommendations

18-30

18.2.6.1
18.2.6.2

Part VI
19

16-7

Implementing Actions Recommended by Optimizer Statistics
Advisor

18-30

Generating a Script Using Optimizer Statistics Advisor

18-32

Optimizer Controls

Influencing the Optimizer
19.1

Techniques for Influencing the Optimizer

19-1

19.2

Influencing the Optimizer with Initialization Parameters

19-3

19.2.1

About Optimizer Initialization Parameters

19-4

xv

19.2.2

Enabling Optimizer Features

19-8

19.2.3

Choosing an Optimizer Goal

19-9

19.2.4

Controlling Adaptive Optimization

19.3

Influencing the Optimizer with Hints

19.3.1

19-11
19-12

19.3.1.1

Types of Hints

19-13

19.3.1.2

Scope of Hints

19-14

19.3.1.3

Guidelines for Hints

19-14

Guidelines for Join Order Hints

19-15

19.3.2

20

About Optimizer Hints

19-10

Improving Real-World Performance Through Cursor Sharing
20.1

Overview of Cursor Sharing

20.1.1

About Cursors

20-1

20.1.1.1

Private and Shared SQL Areas

20-2

20.1.1.2

Parent and Child Cursors

20-4

20.1.2

About Cursors and Parsing

20.1.3

About Literals and Bind Variables

20-7
20-11

20.1.3.1

Literals and Cursors

20-11

20.1.3.2

Bind Variables and Cursors

20-12

20.1.3.3

Bind Variable Peeking

20-13

20.1.4

20.2

20-1

About the Life Cycle of Shared Cursors

20-16

20.1.4.1

Cursor Marked Invalid

20-16

20.1.4.2

Cursors Marked Rolling Invalid

20-18

CURSOR_SHARING and Bind Variable Substitution

20-20

20.2.1

CURSOR_SHARING Initialization Parameter

20-20

20.2.2

Parsing Behavior When CURSOR_SHARING = FORCE

20-21

20.3

Adaptive Cursor Sharing

20-23

20.3.1

Purpose of Adaptive Cursor Sharing

20-23

20.3.2

How Adaptive Cursor Sharing Works: Example

20-24

20.3.3

Bind-Sensitive Cursors

20-26

20.3.4

Bind-Aware Cursors

20-29

20.3.5

Cursor Merging

20-32

20.3.6

Adaptive Cursor Sharing Views

20-33

20.4

Real-World Performance Guidelines for Cursor Sharing

20.4.1

20-33

Develop Applications with Bind Variables for Security and Performance
20-34

20.4.2

Do Not Use CURSOR_SHARING = FORCE as a Permanent Fix

20-35

20.4.3

Establish Coding Conventions to Increase Cursor Reuse

20-36

20.4.4

Minimize Session-Level Changes to the Optimizer Environment

20-38

xvi

Part VII
21

Monitoring and Tracing SQL

Monitoring Database Operations
21.1

About Monitoring Database Operations

21.1.1

Simple Database Operation Use Cases

21-3

21.1.1.2

Composite Database Operation Use Cases

21-3

Database Operation Monitoring Concepts

21-4

21.1.2.1

About the Architecture of Real-Time SQL Monitoring

21-4

21.1.2.2

When the Database Monitors Operations

21-6

21.1.2.3

Attributes of Database Operations

21-7

21.1.3

User Interfaces for Database Operations Monitoring

21-8

21.1.3.1

Monitored SQL Executions Page in Cloud Control

21-9

21.1.3.2

DBMS_SQL_MONITOR Package

21-9

21.1.3.3

Views for Monitoring and Reporting on Database Operations

21.1.4

22

21-2

21.1.1.1
21.1.2

21.2

Purpose of Monitoring Database Operations

21-1

21-10

Basic Tasks in Database Operations Monitoring

21-12

Enabling and Disabling Monitoring of Database Operations

21-12

21.2.1

Enabling Monitoring of Database Operations at the System Level

21-13

21.2.2

Enabling and Disabling Monitoring of Database Operations at the
Statement Level

21-14

21.3

Creating a Database Operation

21-14

21.4

Monitoring SQL Executions Using Cloud Control

21-17

Gathering Diagnostic Data with SQL Test Case Builder
22.1

Purpose of SQL Test Case Builder

22-1

22.2

Concepts for SQL Test Case Builder

22-1

22.2.1

SQL Incidents

22-2

22.2.2

What SQL Test Case Builder Captures

22-2

22.2.3

Output of SQL Test Case Builder

22-3

22.3

User Interfaces for SQL Test Case Builder

22.3.1

22-4

22.3.1.1

Accessing the Incident Manager

22-5

22.3.1.2

Accessing the Support Workbench

22-6

22.3.2
22.4

Graphical Interface for SQL Test Case Builder

22-4

Command-Line Interface for SQL Test Case Builder

Running SQL Test Case Builder

22-6
22-7

xvii

23

Performing Application Tracing
23.1

Overview of End-to-End Application Tracing

23-1

23.1.1

Purpose of End-to-End Application Tracing

23-2

23.1.2

End-to-End Application Tracing in a Multitenant Environment

23-3

23.1.3

Tools for End-to-End Application Tracing

23-3

23.2

23.1.3.1

Overview of the SQL Trace Facility

23-4

23.1.3.2

Overview of TKPROF

23-5

Enabling Statistics Gathering for End-to-End Tracing

23-5

23.2.1

Enabling Statistics Gathering for a Client ID

23-5

23.2.2

Enabling Statistics Gathering for Services, Modules, and Actions

23-6

23.3

Enabling End-to-End Application Tracing

23-7

23.3.1

Enabling Tracing for a Client Identifier

23-7

23.3.2

Enabling Tracing for a Service, Module, and Action

23-8

23.3.3

Enabling Tracing for a Session

23-9

23.3.4

Enabling Tracing for the Instance or Database

23.4

Generating Output Files Using SQL Trace and TKPROF

23-10
23-11

23.4.1

Step 1: Setting Initialization Parameters for Trace File Management

23-12

23.4.2

Step 2: Enabling the SQL Trace Facility

23-13

23.4.3

Step 3: Generating Output Files with TKPROF

23-14

23.4.4

Step 4: Storing SQL Trace Facility Statistics

23-15

23.5

23.4.4.1

Generating the TKPROF Output SQL Script

23-16

23.4.4.2

Editing the TKPROF Output SQL Script

23-16

23.4.4.3

Querying the Output Table

23-16

Guidelines for Interpreting TKPROF Output

23-18

23.5.1

Guideline for Interpreting the Resolution of Statistics

23-18

23.5.2

Guideline for Recursive SQL Statements

23-18

23.5.3

Guideline for Deciding Which Statements to Tune

23-19

23.5.4

Guidelines for Avoiding Traps in TKPROF Interpretation

23-20

23.6.1

23.5.4.1

Guideline for Avoiding the Argument Trap

23-20

23.5.4.2

Guideline for Avoiding the Read Consistency Trap

23-20

23.5.4.3

Guideline for Avoiding the Schema Trap

23-21

23.5.4.4

Guideline for Avoiding the Time Trap

23-22

Application Tracing Utilities

23.6.1.1

TRCSESS

23-22
23-22

23.6.1.1.1

Purpose

23-23

23.6.1.1.2

Guidelines

23-23

23.6.1.1.3

Syntax

23-23

23.6.1.1.4

Options

23-23

23.6.1.1.5

Examples

23-24

23.6.1.2

TKPROF

23-24

xviii

23.7.1

Purpose

23-25

23.6.1.2.2

Guidelines

23-25

23.6.1.2.3

Syntax

23-25

23.6.1.2.4

Options

23-26

23.6.1.2.5

Output

23-28

23.6.1.2.6

Examples

23-31

Views for Application Tracing
Views Relevant for Trace Statistics

23-35

23.7.1.2

Views Related to Enabling Tracing

23-36

Automatic SQL Tuning

Managing SQL Tuning Sets
24.1

About SQL Tuning Sets

24-1

24.1.1

Purpose of SQL Tuning Sets

24-2

24.1.2

Concepts for SQL Tuning Sets

24-2

24.1.3

User Interfaces for SQL Tuning Sets

24-4

24.1.3.1

Accessing the SQL Tuning Sets Page in Cloud Control

24-4

24.1.3.2

Command-Line Interface to SQL Tuning Sets

24-5

24.1.4

Basic Tasks for SQL Tuning Sets

24-5

24.2

Creating a SQL Tuning Set

24-6

24.3

Loading a SQL Tuning Set

24-7

24.4

Displaying the Contents of a SQL Tuning Set

24-8

24.5

Modifying a SQL Tuning Set

24-10

24.6

Transporting a SQL Tuning Set

24-12

24.6.1

24.7

About Transporting SQL Tuning Sets

24-12

24.6.1.1

Basic Steps for Transporting SQL Tuning Sets

24-12

24.6.1.2

Basic Steps for Transporting SQL Tuning Sets When the
CON_DBID Values Differ

24-13

24.6.2

25

23-35

23.7.1.1

Part VIII
24

23.6.1.2.1

Transporting SQL Tuning Sets with DBMS_SQLTUNE

Dropping a SQL Tuning Set

24-14
24-16

Analyzing SQL with SQL Tuning Advisor
25.1

About SQL Tuning Advisor

25-1

25.1.1

Purpose of SQL Tuning Advisor

25-1

25.1.2

SQL Tuning Advisor Architecture

25-3

25.1.2.1

Input to SQL Tuning Advisor

25-4

25.1.2.2

Output of SQL Tuning Advisor

25-5

25.1.2.3

Automatic Tuning Optimizer Analyses

25-5

xix

25.1.3

25.2

SQL Tuning Advisor Operation

25.1.3.1

Automatic and On-Demand SQL Tuning

25-16

25.1.3.2

Local and Remote SQL Tuning

25-17

Managing the Automatic SQL Tuning Task

25-18

25.2.1

About the Automatic SQL Tuning Task
Purpose of Automatic SQL Tuning

25-19

25.2.1.2

Automatic SQL Tuning Concepts

25-20

25.2.1.3

Command-Line Interface to SQL Tuning Advisor

25-20

25.2.1.4

Basic Tasks for Automatic SQL Tuning

25-21

Enabling and Disabling the Automatic SQL Tuning Task

25.2.2.1
25.2.2.2
25.2.3

25.2.3.2
25.2.4

25-22

Enabling and Disabling the Automatic SQL Tuning Task from the
Command Line

25-23

25-25

Configuring the Automatic SQL Tuning Task Using the
Command Line

25-26

Viewing Automatic SQL Tuning Reports Using the Command
Line

Running SQL Tuning Advisor On Demand

25.3.1

25-24

Configuring the Automatic SQL Tuning Task Using Cloud
Control

Viewing Automatic SQL Tuning Reports

25.2.4.1

25-22

Enabling and Disabling the Automatic SQL Tuning Task Using
Cloud Control

Configuring the Automatic SQL Tuning Task

25.2.3.1

26

25-19

25.2.1.1

25.2.2

25.3

25-16

About On-Demand SQL Tuning

25-27
25-28
25-30
25-31

25.3.1.1

Purpose of On-Demand SQL Tuning

25-31

25.3.1.2

User Interfaces for On-Demand SQL Tuning

25-31

25.3.1.3

Basic Tasks in On-Demand SQL Tuning

25-33

25.3.2

Creating a SQL Tuning Task

25-34

25.3.3

Configuring a SQL Tuning Task

25-36

25.3.4

Executing a SQL Tuning Task

25-38

25.3.5

Monitoring a SQL Tuning Task

25-39

25.3.6

Displaying the Results of a SQL Tuning Task

25-40

Optimizing Access Paths with SQL Access Advisor
26.1

About SQL Access Advisor

26-1

26.1.1

Purpose of SQL Access Advisor

26-2

26.1.2

SQL Access Advisor Architecture

26-2

26.1.2.1

Input to SQL Access Advisor

26-4

26.1.2.2

Filter Options for SQL Access Advisor

26-4

26.1.2.3

SQL Access Advisor Recommendations

26-5

26.1.2.4

SQL Access Advisor Actions

26-6

26.1.2.5

SQL Access Advisor Repository

26-8

xx

26.1.3

User Interfaces for SQL Access Advisor

26.1.3.1
26.1.3.2
26.2

Command-Line Interface to SQL Tuning Sets

Optimizing Access Paths with SQL Access Advisor: Basic Tasks

26-9
26-10
26-10

26.2.1

Creating a SQL Tuning Set as Input for SQL Access Advisor

26-12

26.2.2

Populating a SQL Tuning Set with a User-Defined Workload

26-13

26.2.3

Creating and Configuring a SQL Access Advisor Task

26-15

26.2.4

Executing a SQL Access Advisor Task

26-17

26.2.5

Viewing SQL Access Advisor Task Results

26-18

26.2.6

Generating and Executing a Task Script

26-22

26.3

Performing a SQL Access Advisor Quick Tune

26-23

26.4

Using SQL Access Advisor: Advanced Tasks

26-24

26.4.1

Evaluating Existing Access Structures

26-25

26.4.2

Updating SQL Access Advisor Task Attributes

26-25

26.4.3

Creating and Using SQL Access Advisor Task Templates

26-26

26.4.4

Terminating SQL Access Advisor Task Execution

26-28

26.4.4.1

Interrupting SQL Access Advisor Tasks

26-29

26.4.4.2

Canceling SQL Access Advisor Tasks

26-29

26.4.5

Deleting SQL Access Advisor Tasks

26-30

26.4.6

Marking SQL Access Advisor Recommendations

26-31

26.4.7

Modifying SQL Access Advisor Recommendations

26-32

26.5

SQL Access Advisor Examples

26-33

26.6

SQL Access Advisor Reference

26-33

26.6.1

Action Attributes in the DBA_ADVISOR_ACTIONS View

26-34

26.6.2

Categories for SQL Access Advisor Task Parameters

26-35

26.6.3

SQL Access Advisor Constants

26-36

Part IX
27

Accessing the SQL Access Advisor: Initial Options Page Using
Cloud Control

26-8

SQL Controls: Profiles and Plan Baselines

Managing SQL Profiles
27.1

About SQL Profiles

27-1

27.1.1

Purpose of SQL Profiles

27-2

27.1.2

Concepts for SQL Profiles

27-2

27.1.2.1

SQL Profile Recommendations

27-4

27.1.2.2

SQL Profiles and SQL Plan Baselines

27-6

27.1.3

User Interfaces for SQL Profiles

27-6

27.1.4

Basic Tasks for SQL Profiles

27-6

27.2

Implementing a SQL Profile

27.2.1

About SQL Profile Implementation

27-7
27-8

xxi

27.2.2

28

27-9

27.3

Listing SQL Profiles

27.4

Altering a SQL Profile

27-10

27.5

Dropping a SQL Profile

27-11

27.6

Transporting a SQL Profile

27-12

27-9

Overview of SQL Plan Management
28.1

Purpose of SQL Plan Management

28-2

28.1.1

Benefits of SQL Plan Management

28-2

28.1.2

Differences Between SQL Plan Baselines and SQL Profiles

28-3

28.2

Plan Capture

28.2.1

Automatic Initial Plan Capture

28-4
28-4

28.2.1.1

Eligibility for Automatic Initial Plan Capture

28-5

28.2.1.2

Plan Matching for Automatic Initial Plan Capture

28-6

28.2.2

Manual Plan Capture

28-6

28.3

Plan Selection

28-8

28.4

Plan Evolution

28-9

28.4.1

Purpose of Plan Evolution

28-10

28.4.2

PL/SQL Subprograms for Plan Evolution

28-10

28.5

29

Implementing a SQL Profile

Storage Architecture for SQL Plan Management

28-11

28.5.1

SQL Management Base

28-11

28.5.2

SQL Statement Log

28-12

28.5.3

SQL Plan History

28-14

28.5.3.1

Enabled Plans

28-15

28.5.3.2

Accepted Plans

28-15

28.5.3.3

Fixed Plans

28-15

Managing SQL Plan Baselines
29.1

About Managing SQL Plan Baselines

29.1.1

29-2

29.1.1.1

Accessing the SQL Plan Baseline Page in Cloud Control

29-2

29.1.1.2

DBMS_SPM Package

29-3

29.1.2
29.2

User Interfaces for SQL Plan Management

29-1

Basic Tasks in SQL Plan Management

Configuring SQL Plan Management

29.2.1

Configuring the Capture and Use of SQL Plan Baselines

29.2.1.1

29-4
29-5
29-5

Enabling Automatic Initial Plan Capture for SQL Plan
Management

29-7

29.2.1.2

Configuring Filters for Automatic Plan Capture

29-8

29.2.1.3

Disabling All SQL Plan Baselines

29.2.2

Managing the SPM Evolve Advisor Task

29-10
29-10

xxii

29-11

29.2.2.2

Enabling and Disabling the SPM Evolve Advisor Task

29-11

29.2.2.3

Configuring the Automatic SPM Evolve Advisor Task

29-12

Displaying Plans in a SQL Plan Baseline

29-14

29.4

Loading SQL Plan Baselines

29-15

29.4.1

About Loading SQL Plan Baselines

29-16

29.4.2

Loading Plans from AWR

29-17

29.4.3

Loading Plans from the Shared SQL Area

29-19

29.4.4

Loading Plans from a SQL Tuning Set

29-21

29.4.5

Loading Plans from a Staging Table

29-23

Evolving SQL Plan Baselines Manually

29-26

29.5.1

About the DBMS_SPM Evolve Functions

29-26

29.5.2

Managing an Evolve Task

29-28

29.6

Dropping SQL Plan Baselines

29-35

29.7

Managing the SQL Management Base

29-37

29.7.1

About Managing the SMB

29-37

29.7.2

Changing the Disk Space Limit for the SMB

29-38

29.7.3

Changing the Plan Retention Policy in the SMB

29-39

Migrating Stored Outlines to SQL Plan Baselines
30.1

A

About the SPM Evolve Advisor Task

29.3

29.5

30

29.2.2.1

About Stored Outline Migration

30-1

30.1.1

Purpose of Stored Outline Migration

30-2

30.1.2

How Stored Outline Migration Works

30-3

30.1.2.1

Stages of Stored Outline Migration

30-3

30.1.2.2

Outline Categories and Baseline Modules

30-4

30.1.3

User Interface for Stored Outline Migration

30-5

30.1.4

Basic Steps in Stored Outline Migration

30-7

30.2

Preparing for Stored Outline Migration

30-7

30.3

Migrating Outlines to Utilize SQL Plan Management Features

30-9

30.4

Migrating Outlines to Preserve Stored Outline Behavior

30-10

30.5

Performing Follow-Up Tasks After Stored Outline Migration

30-11

Guidelines for Indexes and Table Clusters
A.1

Guidelines for Tuning Index Performance

A-1

A.1.1

Guidelines for Tuning the Logical Structure

A-2

A.1.2

Guidelines for Choosing Columns and Expressions to Index

A-3

A.1.3

Guidelines for Choosing Composite Indexes

A-4

A.1.3.1

Guidelines for Choosing Keys for Composite Indexes

A-5

A.1.3.2

Guidelines for Ordering Keys for Composite Indexes

A-5

xxiii

A.1.4

Guidelines for Writing SQL Statements That Use Indexes

A-6

A.1.5

Guidelines for Writing SQL Statements That Avoid Using Indexes

A-6

A.1.6

Guidelines for Re-Creating Indexes

A-6

A.1.7

Guidelines for Compacting Indexes

A-7

A.1.8

Guidelines for Using Nonunique Indexes to Enforce Uniqueness

A-7

A.1.9

Guidelines for Using Enabled Novalidated Constraints

A-8

A.2

Guidelines for Using Function-Based Indexes for Performance

A-9

A.3

Guidelines for Using Partitioned Indexes for Performance

A-10

A.4

Guidelines for Using Index-Organized Tables for Performance

A-10

A.5

Guidelines for Using Bitmap Indexes for Performance

A-11

A.6

Guidelines for Using Bitmap Join Indexes for Performance

A-12

A.7

Guidelines for Using Domain Indexes for Performance

A-12

A.8

Guidelines for Using Table Clusters

A-13

A.9

Guidelines for Using Hash Clusters for Performance

A-14

Glossary
Index

xxiv

Preface
This manual explains how to tune Oracle SQL.
This preface contains the following topics:
•

Audience

•

Documentation Accessibility

•

Related Documents

•

Conventions

Audience
This document is intended for database administrators and application developers who
perform the following tasks:
•

Generating and interpreting SQL execution plans

•

Managing optimizer statistics

•

Influencing the optimizer through initialization parameters or SQL hints

•

Controlling cursor sharing for SQL statements

•

Monitoring SQL execution

•

Performing application tracing

•

Managing SQL tuning sets

•

Using SQL Tuning Advisor or SQL Access Advisor

•

Managing SQL profiles

•

Managing SQL baselines

Documentation Accessibility
For information about Oracle's commitment to accessibility, visit the Oracle
Accessibility Program website at http://www.oracle.com/pls/topic/lookup?
ctx=acc&id=docacc.
Access to Oracle Support
Oracle customers that have purchased support have access to electronic support
through My Oracle Support. For information, visit http://www.oracle.com/pls/topic/
lookup?ctx=acc&id=info or visit http://www.oracle.com/pls/topic/lookup?ctx=acc&id=trs
if you are hearing impaired.

xxv

Preface

Related Documents
This manual assumes that you are familiar with Oracle Database Concepts. The
following books are frequently referenced:
•

Oracle Database Data Warehousing Guide

•

Oracle Database VLDB and Partitioning Guide

•

Oracle Database SQL Language Reference

•

Oracle Database Reference

Many examples in this book use the sample schemas, which are installed by default
when you select the Basic Installation option with an Oracle Database. See Oracle
Database Sample Schemas for information on how these schemas were created and
how you can use them.

Conventions
The following text conventions are used in this document:
Convention

Meaning

boldface

Boldface type indicates graphical user interface elements associated
with an action, or terms defined in text or the glossary.

italic

Italic type indicates book titles, emphasis, or placeholder variables for
which you supply particular values.

monospace

Monospace type indicates commands within a paragraph, URLs, code
in examples, text that appears on the screen, or text that you enter.

xxvi

Changes in This Release for Oracle
Database SQL Tuning Guide
This preface describes the most important changes in Oracle Database SQL Tuning
Guide.
This preface contains the following topics:
•

Changes in Oracle Database 12c Release 2 (12.2.0.1)
Oracle Database SQL Tuning Guide for Oracle Database 12c Release 2 (12.2.0.1)
has the following changes.

•

Changes in Oracle Database 12c Release 1 (12.1.0.2)
Oracle Database SQL Tuning Guide for Oracle Database 12c Release 1 (12.1.0.2)
has the following changes.

•

Changes in Oracle Database 12c Release 1 (12.1.0.1)
Oracle Database SQL Tuning Guide for Oracle Database 12c Release 1 (12.1)
has the following changes.

Changes in Oracle Database 12c Release 2 (12.2.0.1)
Oracle Database SQL Tuning Guide for Oracle Database 12c Release 2 (12.2.0.1)
has the following changes.
•

New Features
The following features are new in this release:

•

Desupported Features
The following features are desupported in Oracle Database 12c Release 2
(12.2.0.1).

•

Other Changes
This topic describes additional changes in the release.

New Features
The following features are new in this release:
•

Advisor enhancements
–

Optimizer Statistics Advisor
Optimizer Statistics Advisor is built-in diagnostic software that analyzes the
quality of statistics and statistics-related tasks. The advisor task runs
automatically in the maintenance window, but you can also run it on demand.
You can then view the advisor report. If the advisor makes recommendations,
then in some cases you can run system-generated scripts to implement them.
See "Analyzing Statistics Using Optimizer Statistics Advisor".

xxvii

Changes in This Release for Oracle Database SQL Tuning Guide

–

Active Data Guard Support for SQL Tuning Advisor
Using database links, you can tune a standby database workload on a primary
database.
See "Local and Remote SQL Tuning".

•

DBMS_STATS enhancements

–

DBMS_STATS preference for automatic column group statistics

If the DBMS_STATS preference AUTO_STAT_EXTENSIONS is set to ON (by default it is
OFF), then a SQL plan directive can automatically trigger the creation of column
group statistics based on usage of predicates in the workload.
See "Purpose of Optimizer Statistics Preferences".
–

DBMS_STATS support for external table scan rates and In-Memory column store
(IM column store) statistics

If the database uses an IM column store, then you can set the im_imcu_count
parameter to the number of IMCUs in the table or partition, and im_block_count
to the number of blocks. For an external table, scanrate specifies the rate at
which data is scanned in MB/second.
See "Guideline for External Tables".
–

DBMS_STATS statistics preference PREFERENCE_OVERRIDES_PARAMETER

The PREFERENCE_OVERRIDES_PARAMETER statistics preference determines whether,
when gathering optimizer statistics, to override the input value of a parameter
with the statistics preference. In this way, you control when the database
honors a parameter value passed to the statistics gathering procedures.
See "Statistics Preference Overrides".
–

Access to current statistics does not require FLUSH_DATABASE_MONITORING_INFO
You no longer need to ensure that view metadata is up-to-date by using
DBMS_STATS.FLUSH_DATABASE_MONITORING_INFO to save monitoring information to
disk. The statistics shown in DBA_TAB_STATISTICS and DBA_IND_STATISTICS come
from the same source as DBA_TAB_MODIFICATIONS, which means these views
show statistics obtained from disk and memory.
See "Determining When Optimizer Statistics Are Stale".

•

Separate controls for adaptive plans and adaptive statistics
The OPTIMIZER_ADAPTIVE_PLANS initialization parameter enables (default) or disables
adaptive plans. The OPTIMIZER_ADAPTIVE_STATISTICS initialization parameter enables
or disables (default) adaptive statistics.
See "When Adaptive Query Plans Are Enabled" and "When Adaptive Statistics Are
Enabled".

•

Join enhancements
–

Join groups
A join group is a user-created object that lists two columns that can be
meaningfully joined. In certain queries, join groups enable the database to
eliminate the performance overhead of decompressing and hashing column
values. Join groups require an IM column store.
See "In-Memory Join Groups".

xxviii

Changes in This Release for Oracle Database SQL Tuning Guide

–

Band join enhancements
A band join is a special type of nonequijoin in which key values in one data set
must fall within the specified range (“band”) of the second data set. When the
database detects a band join, the database evaluates the costs of band joins
more efficiently, avoiding unnecessary scans of rows that fall outside the
defined bands. In most cases, optimized performance is comparable to an
equijoin.
See "Band Joins".

•

Cursor management enhancements
–

Cursor-duration temporary tables
To materialize the intermediate results of a query, Oracle Database may
create a cursor-duration temporary table in memory during query compilation.
For complex operations such as WITH clause queries and star transformations,
this internal optimization, which enhances the materialization of intermediate
results from repetitively used subqueries, improves performance and
optimizes I/O.
See "Cursor-Duration Temporary Tables".

–

Fine-grained cursor invalidation
Starting in this release, you can specify deferred invalidation on DDL
statements. When shared SQL areas are marked rolling invalid, the database
assigns each one a randomly generated time period. A hard parse occurs only
if the query executes after the time period has expired. In this way, the
database can diffuse the performance overhead of hard parsing over time.
See "About the Life Cycle of Shared Cursors".

•

OR expansion enhancement

In previous releases, the optimizer used the CONCATENATION operator to perform the
OR expansion. Now the optimizer uses the UNION-ALL operator instead. This
enhancement provides several benefits, including enabling interaction among
various transformations, and avoiding the sharing of query structures.
See "OR Expansion".
•

•

SQL plan management enhancements
–

You can now capture plans from AWR. See "Manual Plan Capture".

–

In previous releases, automatic capture applied to all repeatable queries.
Starting in this release, you can create filters to capture only the plans for
statements that you choose. See "Eligibility for Automatic Initial Plan Capture".

Real-Time database operation monitoring enhancements
A session can start or stop a database operation in a different session by
specifying its session ID and serial number.
See "Creating a Database Operation".

•

Expression tracking
SQL statements commonly include expressions such as plus (+) or minus (-).
More complicated examples include PL/SQL functions or SQL functions such as
LTRIM and TO_NUMBER. The Expression Statistics Store (ESS) maintains usage
information about expressions identified during compilation and captured during
execution.

xxix

Changes in This Release for Oracle Database SQL Tuning Guide

See "About the Expression Statistics Store (ESS)".
•

Enhancements for application tracing in a multitenant environment
CDB administrators and PDB administrators can use new V$ views to access trace
data that is relevant for a specific PDB.
See "End-to-End Application Tracing in a Multitenant Environment".

Desupported Features
The following features are desupported in Oracle Database 12c Release 2 (12.2.0.1).
•

The OPTIMIZER_ADAPTIVE_FEATURES initialization parameter

See Also:
Oracle Database Upgrade Guide for a list of desupported features

Other Changes
This topic describes additional changes in the release.
•

New Real-World Performance content
In this release, the book incorporates information provided by the Real-World
Performance group, including the following:
–

"Improving Real-World Performance Through Cursor Sharing" explains how to
use bind variables and new features such as adaptive cursor sharing

Changes in Oracle Database 12c Release 1 (12.1.0.2)
Oracle Database SQL Tuning Guide for Oracle Database 12c Release 1 (12.1.0.2)
has the following changes.
•

New Features
The following features are new in this release.

New Features
The following features are new in this release.
•

In-Memory aggregation
This optimization minimizes the join and GROUP BY processing required for each row
when joining a single large table to multiple small tables, as in a star schema.
VECTOR GROUP BY aggregation uses the infrastructure related to parallel query (PQ)
processing, and blends it with CPU-efficient algorithms to maximize the
performance and effectiveness of the initial aggregation performed before
redistributing fact data.
See "In-Memory Aggregation (VECTOR GROUP BY)".

•

SQL Monitor support for adaptive query plans

xxx

Changes in This Release for Oracle Database SQL Tuning Guide

SQL Monitor supports adaptive query plans in the following ways:
–

Indicates whether a query plan is adaptive, and show its current status:
resolving or resolved.

–

Provides a list that enables you to select the current, full, or final query plans
See "Adaptive Query Plans" to learn more about adaptive query plans, and
"Monitoring SQL Executions Using Cloud Control" to learn more about SQL
Monitor.

Changes in Oracle Database 12c Release 1 (12.1.0.1)
Oracle Database SQL Tuning Guide for Oracle Database 12c Release 1 (12.1) has
the following changes.
•

New Features
The following features are new in this release.

•

Deprecated Features
The following features are deprecated in this release, and may be desupported in
a future release.

•

Desupported Features
Some features previously described in this document are desupported in Oracle
Database 12c.

•

Other Changes
The following are additional changes in the release.

New Features
The following features are new in this release.
•

Adaptive SQL Plan Management (SPM)
The SPM Evolve Advisor is a task infrastructure that enables you to schedule an
evolve task, rerun an evolve task, and generate persistent reports. The new
automatic evolve task, SYS_AUTO_SPM_EVOLVE_TASK, runs in the default maintenance
window. This task ranks all unaccepted plans and runs the evolve process for
them. If the task finds a new plan that performs better than existing plan, the task
automatically accepts the plan. You can also run evolution tasks manually using
the DBMS_SPM package.
See "Managing the SPM Evolve Advisor Task".

•

Adaptive query optimization
Adaptive query optimization is a set of capabilities that enable the optimizer to
make run-time adjustments to execution plans and discover additional information
that can lead to better statistics. The set of capabilities include:
–

Adaptive query plans
An adaptive query plan has built-in options that enable the final plan for a
statement to differ from the default plan. During the first execution, before a
specific subplan becomes active, the optimizer makes a final decision about
which option to use. The optimizer bases its choice on observations made
during the execution up to this point. The ability of the optimizer to adapt plans
can improve query performance.

xxxi

Changes in This Release for Oracle Database SQL Tuning Guide

See "Adaptive Query Plans".
–

Automatic reoptimization
When using automatic reoptimization, the optimizer monitors the initial
execution of a query. If the actual execution statistics vary significantly from
the original plan statistics, then the optimizer records the execution statistics
and uses them to choose a better plan the next time the statement executes.
The database uses information obtained during automatic reoptimization to
generate SQL plan directives automatically.
See "Automatic Reoptimization".

–

SQL plan directives
In releases earlier than Oracle Database 12c, the database stored compilation
and execution statistics in the shared SQL area, which is nonpersistent.
Starting in this release, the database can use a SQL plan directive, which is
additional information and instructions that the optimizer can use to generate a
more optimal plan. The database stores SQL plan directives persistently in the
SYSAUX tablespace. When generating an execution plan, the optimizer can use
SQL plan directives to obtain more information about the objects accessed in
the plan.
See "SQL Plan Directives".

–

Dynamic statistics enhancements
In releases earlier than Oracle Database 12c, Oracle Database only used
dynamic statistics (previously called dynamic sampling) when one or more of
the tables in a query did not have optimizer statistics. Starting in this release,
the optimizer automatically decides whether dynamic statistics are useful and
which dynamic statistics level to use for all SQL statements. Dynamic statistics
gathers are persistent and usable by other queries.
See "Supplemental Dynamic Statistics".

•

New types of histograms
This release introduces top frequency and hybrid histograms. If a column contains
more than 254 distinct values, and if the top 254 most frequent values occupy
more than 99% of the data, then the database creates a top frequency histogram
using the top 254 most frequent values. By ignoring the nonpopular values, which
are statistically insignificant, the database can produce a better quality histogram
for highly popular values. A hybrid histogram is an enhanced height-based
histogram that stores the exact frequency of each endpoint in the sample, and
ensures that a value is never stored in multiple buckets.
Also, regular frequency histograms have been enhanced. The optimizer computes
frequency histograms during NDV computation based on a full scan of the data
rather than a small sample (when AUTO_SAMPLING is used). The enhanced frequency
histograms ensure that even highly infrequent values are properly represented
with accurate bucket counts within a histogram.
See "Histograms ".

•

Monitoring database operations
Real-Time Database Operations Monitoring enables you to monitor long running
database tasks such as batch jobs, scheduler jobs, and Extraction,
Transformation, and Loading (ETL) jobs as a composite business operation. This
feature tracks the progress of SQL and PL/SQL queries associated with the

xxxii

Changes in This Release for Oracle Database SQL Tuning Guide

business operation being monitored. As a DBA or developer, you can define
business operations for monitoring by explicitly specifying the start and end of the
operation or implicitly with tags that identify the operation.
See "Monitoring Database Operations ".
•

Concurrent statistics gathering
You can concurrently gather optimizer statistics on multiple tables, table partitions,
or table subpartitions. By fully utilizing multiprocessor environments, the database
can reduce the overall time required to gather statistics. Oracle Scheduler and
Advanced Queuing create and manage jobs to gather statistics concurrently. The
scheduler decides how many jobs to execute concurrently, and how many to
queue based on available system resources and the value of the
JOB_QUEUE_PROCESSES initialization parameter.
See "Gathering Optimizer Statistics Concurrently".

•

Reporting mode for DBMS_STATS statistics gathering functions
You can run the DBMS_STATS functions in reporting mode. In this mode, the
optimizer does not actually gather statistics, but reports objects that would be
processed if you were to use a specified statistics gathering function.
See "Running Statistics Gathering Functions in Reporting Mode".

•

Reports on past statistics gathering operations
You can use DBMS_STATS functions to report on a specific statistics gathering
operation or on operations that occurred during a specified time.
See "Reporting on Past Statistics Gathering Operations".

•

Automatic column group creation
With column group statistics, the database gathers optimizer statistics on a group
of columns treated as a unit. Starting in Oracle Database 12c, the database
automatically determines which column groups are required in a specified
workload or SQL tuning set, and then creates the column groups. Thus, for any
specified workload, you no longer need to know which columns from each table
must be grouped.
See "Detecting Useful Column Groups for a Specific Workload".

•

Session-private statistics for global temporary tables
Starting in this release, global temporary tables have a different set of optimizer
statistics for each session. Session-specific statistics improve performance and
manageability of temporary tables because users no longer need to set statistics
for a global temporary table in each session or rely on dynamic statistics. The
possibility of errors in cardinality estimates for global temporary tables is lower,
ensuring that the optimizer has the necessary information to determine an optimal
execution plan.
See "Session-Specific Statistics for Global Temporary Tables".

•

SQL Test Case Builder enhancements
SQL Test Case Builder can capture and replay actions and events that enable you
to diagnose incidents that depend on certain dynamic and volatile factors. This
capability is especially useful for parallel query and automatic memory
management.
See Gathering Diagnostic Data with SQL Test Case Builder.

xxxiii

Changes in This Release for Oracle Database SQL Tuning Guide

•

Online statistics gathering for bulk loads
A bulk load is a CREATE TABLE AS SELECT or INSERT INTO ... SELECT operation. In
releases earlier than Oracle Database 12c, you needed to manually gather
statistics after a bulk load to avoid the possibility of a suboptimal execution plan
caused by stale statistics. Starting in this release, Oracle Database gathers
optimizer statistics automatically, which improves both performance and
manageability.
See "Online Statistics Gathering for Bulk Loads".

•

Reuse of synopses after partition maintenance operations
ALTER TABLE EXCHANGE is a common partition maintenance operation. During a

partition exchange, the statistics of the partition and the table are also exchanged.
A synopsis is a set of auxiliary statistics gathered on a partitioned table when the
INCREMENTAL value is set to true. In releases earlier than Oracle Database 12c, you
could not gather table-level synopses on a table. Thus, you could not gather tablelevel synopses on a table, exchange the table with a partition, and end up with
synopses on the partition. You had to explicitly gather optimizer statistics in
incremental mode to create the missing synopses. Starting in this release, you can
gather table-level synopses on a table. When you exchange this table with a
partition in an incremental mode table, the synopses are also exchanged.
See "Maintaining Incremental Statistics for Partition Maintenance Operations".
•

Automatic updates of global statistics for tables with stale or locked partition
statistics
Incremental statistics can automatically calculate global statistics for a partitioned
table even if the partition or subpartition statistics are stale and locked.
See "Maintaining Incremental Statistics for Tables with Stale or Locked Partition
Statistics".

•

Cube query performance enhancements
These enhancements minimize CPU and memory consumption and reduce I/O for
queries against cubes.
See Table 7-7 to learn about the CUBE JOIN operation.

Deprecated Features
The following features are deprecated in this release, and may be desupported in a
future release.
•

Stored outlines
See "Managing SQL Plan Baselines" for information about alternatives.

•

The SIMILAR value for the CURSOR_SHARING initialization parameter
This value is deprecated. Use FORCE instead.
See "Do Not Use CURSOR_SHARING = FORCE as a Permanent Fix".

Desupported Features
Some features previously described in this document are desupported in Oracle
Database 12c.

xxxiv

Changes in This Release for Oracle Database SQL Tuning Guide

See Oracle Database Upgrade Guide for a list of desupported features.

Other Changes
The following are additional changes in the release.
•

New tuning books
The Oracle Database 11g Oracle Database Performance Tuning Guide has been
divided into two books for Oracle Database 12c:
–

Oracle Database Performance Tuning Guide, which contains only topics that
pertain to tuning the database

–

Oracle Database SQL Tuning Guide, which contains only topics that pertain to
tuning SQL

xxxv

Part I
SQL Performance Fundamentals
SQL tuning is improving SQL statement performance to meet specific, measurable,
and achievable goals.
This part contains the following chapters:
•

Introduction to SQL Tuning
SQL tuning is the attempt to diagnose and repair SQL statements that fail to meet
a performance standard.

•

SQL Performance Methodology
This chapter describes the recommended methodology for SQL tuning.

1
Introduction to SQL Tuning
SQL tuning is the attempt to diagnose and repair SQL statements that fail to meet a
performance standard.
This chapter contains the following topics:
•

About SQL Tuning
SQL tuning is the iterative process of improving SQL statement performance to
meet specific, measurable, and achievable goals.

•

Purpose of SQL Tuning
A SQL statement becomes a problem when it fails to perform according to a
predetermined and measurable standard.

•

Prerequisites for SQL Tuning
SQL performance tuning requires a foundation of database knowledge.

•

Tasks and Tools for SQL Tuning
After you have identified the goal for a tuning session, for example, reducing user
response time from three minutes to less than a second, the problem becomes
how to accomplish this goal.

1.1 About SQL Tuning
SQL tuning is the iterative process of improving SQL statement performance to meet
specific, measurable, and achievable goals.
SQL tuning implies fixing problems in deployed applications. In contrast, application
design sets the security and performance goals before deploying an application.

See Also:
•

SQL Performance Methodology

•

"Guidelines for Designing Your Application" to learn how to design for
SQL performance

1.2 Purpose of SQL Tuning
A SQL statement becomes a problem when it fails to perform according to a
predetermined and measurable standard.
After you have identified the problem, a typical tuning session has one of the following
goals:
•

Reduce user response time, which means decreasing the time between when a
user issues a statement and receives a response

1-1

Chapter 1

Prerequisites for SQL Tuning

•

Improve throughput, which means using the least amount of resources necessary
to process all rows accessed by a statement

For a response time problem, consider an online book seller application that hangs for
three minutes after a customer updates the shopping cart. Contrast with a threeminute parallel query in a data warehouse that consumes all of the database host
CPU, preventing other queries from running. In each case, the user response time is
three minutes, but the cause of the problem is different, and so is the tuning goal.

1.3 Prerequisites for SQL Tuning
SQL performance tuning requires a foundation of database knowledge.
If you are tuning SQL performance, then this manual assumes that you have the
knowledge and skills shown in the following table.
Table 1-1

Required Knowledge

Required Knowledge

Description

To Learn More

Database architecture

Database architecture is not the
domain of administrators alone.
As a developer, you want to
develop applications in the least
amount of time against an Oracle
database, which requires
exploiting the database
architecture and features. For
example, not understanding
Oracle Database concurrency
controls and multiversioning read
consistency may make an
application corrupt the integrity of
the data, run slowly, and
decrease scalability.

Oracle Database Concepts
explains the basic relational data
structures, transaction
management, storage structures,
and instance architecture of
Oracle Database.

SQL and PL/SQL

Because of the existence of GUIbased tools, it is possible to
create applications and
administer a database without
knowing SQL. However, it is
impossible to tune applications or
a database without knowing SQL.

Oracle Database Concepts
includes an introduction to Oracle
SQL and PL/SQL. You must also
have a working knowledge of
Oracle Database SQL Language
Reference, Oracle Database
PL/SQL Packages and Types
Reference, and Oracle Database
PL/SQL Packages and Types
Reference.

SQL tuning tools

The database generates
performance statistics, and
provides SQL tuning tools that
interpret these statistics.

Oracle Database 2 Day +
Performance Tuning Guide
provides an introduction to the
principal SQL tuning tools.

1.4 Tasks and Tools for SQL Tuning
After you have identified the goal for a tuning session, for example, reducing user
response time from three minutes to less than a second, the problem becomes how to
accomplish this goal.

1-2

Chapter 1

Tasks and Tools for SQL Tuning

This section contains the following topics:
•

SQL Tuning Tasks
The specifics of a tuning session depend on many factors, including whether you
tune proactively or reactively.

•

SQL Tuning Tools
SQL tuning tools are either automated or manual.

•

User Interfaces to SQL Tuning Tools
Cloud Control is a system management tool that provides centralized
management of a database environment. Cloud Control provides access to most
tuning tools.

1.4.1 SQL Tuning Tasks
The specifics of a tuning session depend on many factors, including whether you tune
proactively or reactively.
In proactive SQL tuning, you regularly use SQL Tuning Advisor to determine whether
you can make SQL statements perform better. In reactive SQL tuning, you correct a
SQL-related problem that a user has experienced.
Whether you tune proactively or reactively, a typical SQL tuning session involves all or
most of the following tasks:
1.

Identifying high-load SQL statements
Review past execution history to find the statements responsible for a large share
of the application workload and system resources.

2.

Gathering performance-related data
The optimizer statistics are crucial to SQL tuning. If these statistics do not exist or
are no longer accurate, then the optimizer cannot generate the best plan. Other
data relevant to SQL performance include the structure of tables and views that
the statement accessed, and definitions of any indexes available to the statement.

3.

Determining the causes of the problem
Typically, causes of SQL performance problems include:
•

Inefficiently designed SQL statements
If a SQL statement is written so that it performs unnecessary work, then the
optimizer cannot do much to improve its performance. Examples of inefficient
design include

•

–

Neglecting to add a join condition, which leads to a Cartesian join

–

Using hints to specify a large table as the driving table in a join

–

Specifying UNION instead of UNION ALL

–

Making a subquery execute for every row in an outer query

Suboptimal execution plans
The query optimizer (also called the optimizer) is internal software that
determines which execution plan is most efficient. Sometimes the optimizer
chooses a plan with a suboptimal access path, which is the means by which
the database retrieves data from the database. For example, the plan for a

1-3

Chapter 1

Tasks and Tools for SQL Tuning

query predicate with low selectivity may use a full table scan on a large table
instead of an index.
You can compare the execution plan of an optimally performing SQL
statement to the plan of the statement when it performs suboptimally. This
comparison, along with information such as changes in data volumes, can
help identify causes of performance degradation.
•

Missing SQL access structures
Absence of SQL access structures, such as indexes and materialized views, is
a typical reason for suboptimal SQL performance. The optimal set of access
structures can improve SQL performance by orders of magnitude.

•

Stale optimizer statistics
Statistics gathered by DBMS_STATS can become stale when the statistics
maintenance operations, either automatic or manual, cannot keep up with the
changes to the table data caused by DML. Because stale statistics on a table
do not accurately reflect the table data, the optimizer can make decisions
based on faulty information and generate suboptimal execution plans.

•

Hardware problems
Suboptimal performance might be connected with memory, I/O, and CPU
problems.

4.

Defining the scope of the problem
The scope of the solution must match the scope of the problem. Consider a
problem at the database level and a problem at the statement level. For example,
the shared pool is too small, which causes cursors to age out quickly, which in turn
causes many hard parses. Using an initialization parameter to increase the shared
pool size fixes the problem at the database level and improves performance for all
sessions. However, if a single SQL statement is not using a helpful index, then
changing the optimizer initialization parameters for the entire database could harm
overall performance. If a single SQL statement has a problem, then an
appropriately scoped solution addresses just this problem with this statement.

5.

Implementing corrective actions for suboptimally performing SQL statements
These actions vary depending on circumstances. For example, you might rewrite a
SQL statement to be more efficient, avoiding unnecessary hard parsing by
rewriting the statement to use bind variables. You might also use equijoins,
remove functions from WHERE clauses, and break a complex SQL statement into
multiple simple statements.
In some cases, you improve SQL performance not by rewriting the statement, but
by restructuring schema objects. For example, you might index a new access
path, or reorder columns in a concatenated index. You might also partition a table,
introduce derived values, or even change the database design.

6.

Preventing SQL performance regressions
To ensure optimal SQL performance, verify that execution plans continue to
provide optimal performance, and choose better plans if they come available. You
can achieve these goals using optimizer statistics, SQL profiles, and SQL plan
baselines.

1-4

Chapter 1

Tasks and Tools for SQL Tuning

See Also:
•

"Shared Pool Check"

•

Oracle Database Concepts to learn more about the shared pool

1.4.2 SQL Tuning Tools
SQL tuning tools are either automated or manual.
In this context, a tool is automated if the database itself can provide diagnosis, advice,
or corrective actions. A manual tool requires you to perform all of these operations.
All tuning tools depend on the basic tools of the dynamic performance views, statistics,
and metrics that the database instance collects. The database itself contains the data
and metadata required to tune SQL statements.
This section contains the following topics:
•

Automated SQL Tuning Tools
Oracle Database provides several advisors relevant for SQL tuning.

•

Manual SQL Tuning Tools
In some situations, you may want to run manual tools in addition to the automated
tools. Alternatively, you may not have access to the automated tools.

1.4.2.1 Automated SQL Tuning Tools
Oracle Database provides several advisors relevant for SQL tuning.
Additionally, SQL plan management is a mechanism that can prevent performance
regressions and also help you to improve SQL performance.
All of the automated SQL tuning tools can use SQL tuning sets as input. A SQL tuning
set (STS) is a database object that includes one or more SQL statements along with
their execution statistics and execution context.
This section contains the following topics:
•

Automatic Database Diagnostic Monitor (ADDM)
ADDM is self-diagnostic software built into Oracle Database.

•

SQL Tuning Advisor
SQL Tuning Advisor is internal diagnostic software that identifies problematic
SQL statements and recommends how to improve statement performance.

•

SQL Access Advisor
SQL Access Advisor is internal diagnostic software that recommends which
materialized views, indexes, and materialized view logs to create, drop, or retain.

•

SQL Plan Management
SQL plan management is a preventative mechanism that enables the optimizer to
automatically manage execution plans, ensuring that the database uses only
known or verified plans.

1-5

Chapter 1

Tasks and Tools for SQL Tuning

•

SQL Performance Analyzer
SQL Performance Analyzer determines the effect of a change on a SQL workload
by identifying performance divergence for each SQL statement.

See Also:
•

"About SQL Tuning Sets"

•

Oracle Database 2 Day + Performance Tuning Guide to learn more
about managing SQL tuning sets

1.4.2.1.1 Automatic Database Diagnostic Monitor (ADDM)
ADDM is self-diagnostic software built into Oracle Database.
ADDM can automatically locate the root causes of performance problems, provide
recommendations for correction, and quantify the expected benefits. ADDM also
identifies areas where no action is necessary.
ADDM and other advisors use Automatic Workload Repository (AWR), which is an
infrastructure that provides services to database components to collect, maintain, and
use statistics. ADDM examines and analyzes statistics in AWR to determine possible
performance problems, including high-load SQL.
For example, you can configure ADDM to run nightly. In the morning, you can examine
the latest ADDM report to see what might have caused a problem and if there is a
recommended fix. The report might show that a particular SELECT statement consumed
a huge amount of CPU, and recommend that you run SQL Tuning Advisor.

See Also:
•

Oracle Database 2 Day + Performance Tuning Guide

•

Oracle Database Performance Tuning Guide

1.4.2.1.2 SQL Tuning Advisor
SQL Tuning Advisor is internal diagnostic software that identifies problematic SQL
statements and recommends how to improve statement performance.
When run during database maintenance windows as an automated maintenance task,
SQL Tuning Advisor is known as Automatic SQL Tuning Advisor.
SQL Tuning Advisor takes one or more SQL statements as an input and invokes the
Automatic Tuning Optimizer to perform SQL tuning on the statements. The advisor
performs the following types of analysis:
•

Checks for missing or stale statistics

•

Builds SQL profiles
A SQL profile is a set of auxiliary information specific to a SQL statement. A SQL
profile contains corrections for suboptimal optimizer estimates discovered during

1-6

Chapter 1

Tasks and Tools for SQL Tuning

Automatic SQL Tuning. This information can improve optimizer estimates for
cardinality, which is the number of rows that is estimated to be or actually is
returned by an operation in an execution plan, and selectivity. These improved
estimates lead the optimizer to select better plans.
•

Explores whether a different access path can significantly improve performance

•

Identifies SQL statements that lend themselves to suboptimal plans

The output is in the form of advice or recommendations, along with a rationale for each
recommendation and its expected benefit. The recommendation relates to a collection
of statistics on objects, creation of new indexes, restructuring of the SQL statement, or
creation of a SQL profile. You can choose to accept the recommendations to complete
the tuning of the SQL statements.

See Also:
•

"Analyzing SQL with SQL Tuning Advisor"

•

Oracle Database 2 Day + Performance Tuning Guide

1.4.2.1.3 SQL Access Advisor
SQL Access Advisor is internal diagnostic software that recommends which
materialized views, indexes, and materialized view logs to create, drop, or retain.
SQL Access Advisor takes an actual workload as input, or the advisor can derive a
hypothetical workload from the schema. SQL Access Advisor considers the trade-offs
between space usage and query performance, and recommends the most costeffective configuration of new and existing materialized views and indexes. The
advisor also makes recommendations about partitioning.

See Also:
•

"About SQL Access Advisor"

•

Oracle Database 2 Day + Performance Tuning Guide

1.4.2.1.4 SQL Plan Management
SQL plan management is a preventative mechanism that enables the optimizer to
automatically manage execution plans, ensuring that the database uses only known or
verified plans.
This mechanism can build a SQL plan baseline, which contains one or more accepted
plans for each SQL statement. By using baselines, SQL plan management can
prevent plan regressions from environmental changes, while permitting the optimizer
to discover and use better plans.

1-7

Chapter 1

Tasks and Tools for SQL Tuning

See Also:
•

"Overview of SQL Plan Management"

•

Oracle Database PL/SQL Packages and Types Reference
to learn about the DBMS_SPM package

1.4.2.1.5 SQL Performance Analyzer
SQL Performance Analyzer determines the effect of a change on a SQL workload by
identifying performance divergence for each SQL statement.
System changes such as upgrading a database or adding an index may cause
changes to execution plans, affecting SQL performance. By using SQL Performance
Analyzer, you can accurately forecast the effect of system changes on SQL
performance. Using this information, you can tune the database when SQL
performance regresses, or validate and measure the gain when SQL performance
improves.

See Also:
Oracle Database Testing Guide

1.4.2.2 Manual SQL Tuning Tools
In some situations, you may want to run manual tools in addition to the automated
tools. Alternatively, you may not have access to the automated tools.
This section contains the following topics:
•

Execution Plans
Execution plans are the principal diagnostic tool in manual SQL tuning. For
example, you can view plans to determine whether the optimizer selects the plan
you expect, or identify the effect of creating an index on a table.

•

Real-Time SQL Monitoring and Real-Time Database Operations
The Real-Time SQL Monitoring feature of Oracle Database enables you to monitor
the performance of SQL statements while they are executing. By default, SQL
monitoring starts automatically when a statement runs in parallel, or when it has
consumed at least 5 seconds of CPU or I/O time in a single execution.

•

Application Tracing
A SQL trace file provides performance information on individual SQL statements:
parse counts, physical and logical reads, misses on the library cache, and so on.

•

Optimizer Hints
A hint is an instruction passed to the optimizer through comments in a SQL
statement. Hints enable you to make decisions normally made automatically by
the optimizer.

1-8

Chapter 1

Tasks and Tools for SQL Tuning

1.4.2.2.1 Execution Plans
Execution plans are the principal diagnostic tool in manual SQL tuning. For example,
you can view plans to determine whether the optimizer selects the plan you expect, or
identify the effect of creating an index on a table.
You can display execution plans in multiple ways. The following tools are the most
commonly used:
•

EXPLAIN PLAN

This SQL statement enables you to view the execution plan that the optimizer
would use to execute a SQL statement without actually executing the statement.
See Oracle Database SQL Language Reference.
•

AUTOTRACE

The AUTOTRACE command in SQL*Plus generates the execution plan and statistics
about the performance of a query. This command provides statistics such as disk
reads and memory reads. See SQL*Plus User's Guide and Reference.
•

V$SQL_PLAN and related views

These views contain information about executed SQL statements, and their
execution plans, that are still in the shared pool. See Oracle Database Reference.
You can use the DBMS_XPLAN package methods to display the execution plan generated
by the EXPLAIN PLAN command and query of V$SQL_PLAN.

1.4.2.2.2 Real-Time SQL Monitoring and Real-Time Database Operations
The Real-Time SQL Monitoring feature of Oracle Database enables you to monitor the
performance of SQL statements while they are executing. By default, SQL monitoring
starts automatically when a statement runs in parallel, or when it has consumed at
least 5 seconds of CPU or I/O time in a single execution.
A database operation is a set of database tasks defined by end users or application
code, for example, a batch job or Extraction, Transformation, and Loading (ETL)
processing. You can define, monitor, and report on database operations. Real-Time
Database Operations provides the ability to monitor composite operations
automatically. The database automatically monitors parallel queries, DML, and DDL
statements as soon as execution begins.
Oracle Enterprise Manager Cloud Control (Cloud Control) provides easy-to-use SQL
monitoring pages. Alternatively, you can monitor SQL-related statistics using the
V$SQL_MONITOR and V$SQL_PLAN_MONITOR views. You can use these views with the
following views to get more information about executions that you are monitoring:
•

V$ACTIVE_SESSION_HISTORY

•

V$SESSION

•

V$SESSION_LONGOPS

•

V$SQL

•

V$SQL_PLAN

1-9

Chapter 1

Tasks and Tools for SQL Tuning

See Also:
•

"About Monitoring Database Operations"

•

Oracle Database Reference to learn about the V$ views

1.4.2.2.3 Application Tracing
A SQL trace file provides performance information on individual SQL statements:
parse counts, physical and logical reads, misses on the library cache, and so on.
Trace files are sometimes useful for diagnosing SQL performance problems. You can
enable and disable SQL tracing for a specific session using the DBMS_MONITOR or
DBMS_SESSION packages. Oracle Database implements tracing by generating a trace file
for each server process when you enable the tracing mechanism.
Oracle Database provides the following command-line tools for analyzing trace files:
•

TKPROF

This utility accepts as input a trace file produced by the SQL Trace facility, and
then produces a formatted output file.
•

trcsess

This utility consolidates trace output from multiple trace files based on criteria such
as session ID, client ID, and service ID. After trcsess merges the trace information
into a single output file, you can format the output file with TKPROF. trcsess is useful
for consolidating the tracing of a particular session for performance or debugging
purposes.
End-to-End Application Tracing simplifies the process of diagnosing performance
problems in multitier environments. In these environments, the middle tier routes a
request from an end client to different database sessions, making it difficult to track a
client across database sessions. End-to-End application tracing uses a client ID to
uniquely trace a specific end-client through all tiers to the database.

See Also:
Oracle Database PL/SQL Packages and Types Reference to learn more
about DBMS_MONITOR and DBMS_SESSION

1.4.2.2.4 Optimizer Hints
A hint is an instruction passed to the optimizer through comments in a SQL statement.
Hints enable you to make decisions normally made automatically by the optimizer.
In a test or development environment, hints are useful for testing the performance of a
specific access path. For example, you may know that a specific index is more
selective for certain queries. In this case, you may use hints to instruct the optimizer to
use a better execution plan, as in the following example:
SELECT /*+ INDEX (employees emp_department_ix) */
employee_id, department_id

1-10

Chapter 1

Tasks and Tools for SQL Tuning

FROM employees
WHERE department_id > 50;

See Also:
•

"Influencing the Optimizer with Hints"

•

Oracle Database SQL Language Reference to learn more about hints

1.4.3 User Interfaces to SQL Tuning Tools
Cloud Control is a system management tool that provides centralized management of
a database environment. Cloud Control provides access to most tuning tools.
By combining a graphical console, Oracle Management Servers, Oracle Intelligent
Agents, common services, and administrative tools, Cloud Control provides a
comprehensive system management platform.
You can also access all SQL tuning tools using a command-line interface. For
example, the DBMS_ADVISOR package is the command-line interface for SQL Tuning
Advisor.
Oracle recommends Cloud Control as the best interface for database administration
and tuning. In cases where the command-line interface better illustrates a particular
concept or task, this manual uses command-line examples. However, in these cases
the tuning tasks include a reference to the principal Cloud Control page associated
with the task.

1-11

2
SQL Performance Methodology
This chapter describes the recommended methodology for SQL tuning.

Note:
This book assumes that you have learned the Oracle Database performance
methodology described in Oracle Database 2 Day + Performance Tuning
Guide.

This chapter contains the following topics:
•

Guidelines for Designing Your Application
The key to obtaining good SQL performance is to design your application with
performance in mind.

•

Guidelines for Deploying Your Application
To achieve optimal performance, deploy your application with the same care that
you put into designing it.

2.1 Guidelines for Designing Your Application
The key to obtaining good SQL performance is to design your application with
performance in mind.
This section contains the following topics:
•

Guideline for Data Modeling
Data modeling is important to successful application design.

•

Guideline for Writing Efficient Applications
During the design and architecture phase of system development, ensure that the
application developers understand SQL execution efficiency.

2.1.1 Guideline for Data Modeling
Data modeling is important to successful application design.
You must perform data modeling in a way that represents the business practices.
Heated debates may occur about the correct data model. The important thing is to
apply greatest modeling efforts to those entities affected by the most frequent
business transactions.
In the modeling phase, there is a great temptation to spend too much time modeling
the non-core data elements, which results in increased development lead times. Use
of modeling tools can then rapidly generate schema definitions and can be useful
when a fast prototype is required.

2-1

Chapter 2

Guidelines for Designing Your Application

2.1.2 Guideline for Writing Efficient Applications
During the design and architecture phase of system development, ensure that the
application developers understand SQL execution efficiency.
To achieve this goal, the development environment must support the following
characteristics:
•

Good database connection management
Connecting to the database is an expensive operation that is not scalable.
Therefore, a best practice is to minimize the number of concurrent connections to
the database. A simple system, where a user connects at application initialization,
is ideal. However, in a web-based or multitiered application in which application
servers multiplex database connections to users, this approach can be difficult.
With these types of applications, design them to pool database connections, and
not reestablish connections for each user request.

•

Good cursor usage and management
Maintaining user connections is equally important to minimizing the parsing activity
on the system. Parsing is the process of interpreting a SQL statement and creating
an execution plan for it. This process has many phases, including syntax checking,
security checking, execution plan generation, and loading shared structures into
the shared pool. There are two types of parse operations:
–

Hard parsing
A SQL statement is submitted for the first time, and no match is found in the
shared pool. Hard parses are the most resource-intensive and unscalable,
because they perform all the operations involved in a parse.

–

Soft parsing
A SQL statement is submitted for the first time, and a match is found in the
shared pool. The match can be the result of previous execution by another
user. The SQL statement is shared, which is optimal for performance.
However, soft parses are not ideal, because they still require syntax and
security checking, which consume system resources.

Because parsing should be minimized as much as possible, application
developers should design their applications to parse SQL statements once and
execute them many times. This is done through cursors. Experienced SQL
programmers should be familiar with the concept of opening and re-executing
cursors.
•

Effective use of bind variables
Application developers must also ensure that SQL statements are shared within
the shared pool. To achieve this goal, use bind variables to represent the parts of
the query that change from execution to execution. If this is not done, then the
SQL statement is likely to be parsed once and never re-used by other users. To
ensure that SQL is shared, use bind variables and do not use string literals with
SQL statements. For example:
Statement with string literals:
SELECT *
FROM employees
WHERE last_name LIKE 'KING';

2-2

Chapter 2

Guidelines for Deploying Your Application

Statement with bind variables:
SELECT *
FROM employees
WHERE last_name LIKE :1;

The following example shows the results of some tests on a simple OLTP
application:
Test
#Users Supported
No Parsing all statements
270
Soft Parsing all statements
150
Hard Parsing all statements
60
Re-Connecting for each Transaction 30

These tests were performed on a four-CPU computer. The differences increase as
the number of CPUs on the system increase.

2.2 Guidelines for Deploying Your Application
To achieve optimal performance, deploy your application with the same care that you
put into designing it.
This section contains the following topics:
•

Guideline for Deploying in a Test Environment
The testing process mainly consists of functional and stability testing. At some
point in the process, you must perform performance testing.

•

Guidelines for Application Rollout
When new applications are rolled out, two strategies are commonly adopted: the
Big Bang approach, in which all users migrate to the new system at once, and the
trickle approach, in which users slowly migrate from existing systems to the new
one.

2.2.1 Guideline for Deploying in a Test Environment
The testing process mainly consists of functional and stability testing. At some point in
the process, you must perform performance testing.
The following list describes simple rules for performance testing an application. If
correctly documented, then this list provides important information for the production
application and the capacity planning process after the application has gone live.
•

Use the Automatic Database Diagnostic Monitor (ADDM) and SQL Tuning Advisor
for design validation.

•

Test with realistic data volumes and distributions.
All testing must be done with fully populated tables. The test database should
contain data representative of the production system in terms of data volume and
cardinality between tables. All the production indexes should be built and the
schema statistics should be populated correctly.

•

Use the correct optimizer mode.
Perform all testing with the optimizer mode that you plan to use in production.

•

Test a single user performance.

2-3

Chapter 2

Guidelines for Deploying Your Application

Test a single user on an idle or lightly-used database for acceptable performance.
If a single user cannot achieve acceptable performance under ideal conditions,
then multiple users cannot achieve acceptable performance under real conditions.
•

Obtain and document plans for all SQL statements.
Obtain an execution plan for each SQL statement. Use this process to verify that
the optimizer is obtaining an optimal execution plan, and that the relative cost of
the SQL statement is understood in terms of CPU time and physical I/Os. This
process assists in identifying the heavy use transactions that require the most
tuning and performance work in the future.

•

Attempt multiuser testing.
This process is difficult to perform accurately, because user workload and profiles
might not be fully quantified. However, transactions performing DML statements
should be tested to ensure that there are no locking conflicts or serialization
problems.

•

Test with the correct hardware configuration.
Test with a configuration as close to the production system as possible. Using a
realistic system is particularly important for network latencies, I/O subsystem
bandwidth, and processor type and speed. Failing to use this approach may result
in an incorrect analysis of potential performance problems.

•

Measure steady state performance.
When benchmarking, it is important to measure the performance under steady
state conditions. Each benchmark run should have a ramp-up phase, where users
are connected to the application and gradually start performing work on the
application. This process allows for frequently cached data to be initialized into the
cache and single execution operations—such as parsing—to be completed before
the steady state condition. Likewise, after a benchmark run, a ramp-down period is
useful so that the system frees resources, and users cease work and disconnect.

2.2.2 Guidelines for Application Rollout
When new applications are rolled out, two strategies are commonly adopted: the Big
Bang approach, in which all users migrate to the new system at once, and the trickle
approach, in which users slowly migrate from existing systems to the new one.
Both approaches have merits and disadvantages. The Big Bang approach relies on
reliable testing of the application at the required scale, but has the advantage of
minimal data conversion and synchronization with the old system, because it is simply
switched off. The Trickle approach allows debugging of scalability issues as the
workload increases, but might mean that data must be migrated to and from legacy
systems as the transition takes place.
It is difficult to recommend one approach over the other, because each technique has
associated risks that could lead to system outages as the transition takes place.
Certainly, the Trickle approach allows profiling of real users as they are introduced to
the new application, and allows the system to be reconfigured while only affecting the
migrated users. This approach affects the work of the early adopters, but limits the
load on support services. Thus, unscheduled outages only affect a small percentage of
the user population.
The decision on how to roll out a new application is specific to each business. Any
adopted approach has its own unique pressures and stresses. The more testing and

2-4

Chapter 2

Guidelines for Deploying Your Application

knowledge that you derive from the testing process, the more you realize what is best
for the rollout.

2-5

Part II
Query Optimizer Fundamentals
To tune Oracle SQL, you must understand the query optimizer. The optimizer is builtin software that determines the most efficient method for a statement to access data.
This part contains the following chapters:
•

SQL Processing
This chapter explains how database processes DDL statements to create objects,
DML to modify data, and queries to retrieve data.

•

Query Optimizer Concepts
This chapter describes the most important concepts relating to the query
optimizer, including its principal components.

•

Query Transformations
The optimizer employs many query transformation techniques. This chapter
describes some of the most important.

3
SQL Processing
This chapter explains how database processes DDL statements to create objects,
DML to modify data, and queries to retrieve data.
This chapter contains the following topics:
•

About SQL Processing
SQL processing is the parsing, optimization, row source generation, and
execution of a SQL statement.

•

How Oracle Database Processes DML
Most DML statements have a query component. In a query, execution of a cursor
places the results of the query into a set of rows called the result set.

•

How Oracle Database Processes DDL
Oracle Database processes DDL differently from DML.

3.1 About SQL Processing
SQL processing is the parsing, optimization, row source generation, and execution of
a SQL statement.
The following figure depicts the general stages of SQL processing. Depending on the
statement, the database may omit some of these stages.

3-1

Chapter 3

About SQL Processing

Figure 3-1

Stages of SQL Processing

SQL Statement

Parsing

Syntax
Check

Semantic
Check

Shared Pool
Check

Soft Parse

Hard Parse
Generation of
multiple
execution plans

Optimization

Generation of
query plan

Row Source
Generation

Execution

This section contains the following topics:
•

SQL Parsing
The first stage of SQL processing is parsing.

•

SQL Optimization
During optimization, Oracle Database must perform a hard parse at least once for
every unique DML statement and performs the optimization during this parse.

•

SQL Row Source Generation
The row source generator is software that receives the optimal execution plan
from the optimizer and produces an iterative execution plan that is usable by the
rest of the database.

•

SQL Execution
During execution, the SQL engine executes each row source in the tree produced
by the row source generator. This step is the only mandatory step in DML
processing.

3-2

Chapter 3

About SQL Processing

3.1.1 SQL Parsing
The first stage of SQL processing is parsing.
The parsing stage involves separating the pieces of a SQL statement into a data
structure that other routines can process. The database parses a statement when
instructed by the application, which means that only the application, and not the
database itself, can reduce the number of parses.
When an application issues a SQL statement, the application makes a parse call to the
database to prepare the statement for execution. The parse call opens or creates a
cursor, which is a handle for the session-specific private SQL area that holds a parsed
SQL statement and other processing information. The cursor and private SQL area are
in the program global area (PGA).
During the parse call, the database performs checks that identify the errors that can be
found before statement execution. Some errors cannot be caught by parsing. For
example, the database can encounter deadlocks or errors in data conversion only
during statement execution.
This section contains the following topics:
•

Syntax Check
Oracle Database must check each SQL statement for syntactic validity.

•

Semantic Check
The semantics of a statement are its meaning. A semantic check determines
whether a statement is meaningful, for example, whether the objects and columns
in the statement exist.

•

Shared Pool Check
During the parse, the database performs a shared pool check to determine
whether it can skip resource-intensive steps of statement processing.

See Also:
Oracle Database Concepts to learn about deadlocks

3.1.1.1 Syntax Check
Oracle Database must check each SQL statement for syntactic validity.
A statement that breaks a rule for well-formed SQL syntax fails the check. For
example, the following statement fails because the keyword FROM is misspelled as FORM:
SQL> SELECT * FORM employees;
SELECT * FORM employees
*
ERROR at line 1:
ORA-00923: FROM keyword not found where expected

3-3

Chapter 3

About SQL Processing

3.1.1.2 Semantic Check
The semantics of a statement are its meaning. A semantic check determines whether
a statement is meaningful, for example, whether the objects and columns in the
statement exist.
A syntactically correct statement can fail a semantic check, as shown in the following
example of a query of a nonexistent table:
SQL> SELECT * FROM nonexistent_table;
SELECT * FROM nonexistent_table
*
ERROR at line 1:
ORA-00942: table or view does not exist

3.1.1.3 Shared Pool Check
During the parse, the database performs a shared pool check to determine whether it
can skip resource-intensive steps of statement processing.
To this end, the database uses a hashing algorithm to generate a hash value for every
SQL statement. The statement hash value is the SQL ID shown in V$SQL.SQL_ID. This
hash value is deterministic within a version of Oracle Database, so the same
statement in a single instance or in different instances has the same SQL ID.
When a user submits a SQL statement, the database searches the shared SQL area
to see if an existing parsed statement has the same hash value. The hash value of a
SQL statement is distinct from the following values:
•

Memory address for the statement
Oracle Database uses the SQL ID to perform a keyed read in a lookup table. In
this way, the database obtains possible memory addresses of the statement.

•

Hash value of an execution plan for the statement
A SQL statement can have multiple plans in the shared pool. Typically, each plan
has a different hash value. If the same SQL ID has multiple plan hash values, then
the database knows that multiple plans exist for this SQL ID.

Parse operations fall into the following categories, depending on the type of statement
submitted and the result of the hash check:
•

Hard parse
If Oracle Database cannot reuse existing code, then it must build a new
executable version of the application code. This operation is known as a hard
parse, or a library cache miss.

Note:
The database always performs a hard parse of DDL.

During the hard parse, the database accesses the library cache and data
dictionary cache numerous times to check the data dictionary. When the database
accesses these areas, it uses a serialization device called a latch on required

3-4

Chapter 3

About SQL Processing

objects so that their definition does not change. Latch contention increases
statement execution time and decreases concurrency.
•

Soft parse
A soft parse is any parse that is not a hard parse. If the submitted statement is the
same as a reusable SQL statement in the shared pool, then Oracle Database
reuses the existing code. This reuse of code is also called a library cache hit.
Soft parses can vary in how much work they perform. For example, configuring the
session shared SQL area can sometimes reduce the amount of latching in the soft
parses, making them "softer."
In general, a soft parse is preferable to a hard parse because the database skips
the optimization and row source generation steps, proceeding straight to
execution.

The following graphic is a simplified representation of a shared pool check of an UPDATE
statement in a dedicated server architecture.
Figure 3-2

Shared Pool Check
User Global Area (SGA)
System
Shared Pool
Library Cache
Shared SQL Area
3667723989
3967354608
2190280494

Data
Dictionary
Cache

Server
Result
Cache

Private
SQL Area

Other

Reserved
Pool

Comparison of hash values
Update ...
PGA
Client
Process

Server
Process

SQL Work Areas
Session Memory

User

3967354608
Private SQL Area

If a check determines that a statement in the shared pool has the same hash value,
then the database performs semantic and environment checks to determine whether
the statements have the same meaning. Identical syntax is not sufficient. For example,
suppose two different users log in to the database and issue the following SQL
statements:
CREATE TABLE my_table ( some_col INTEGER );
SELECT * FROM my_table;

The SELECT statements for the two users are syntactically identical, but two separate
schema objects are named my_table. This semantic difference means that the second
statement cannot reuse the code for the first statement.

3-5

Chapter 3

About SQL Processing

Even if two statements are semantically identical, an environmental difference can
force a hard parse. In this context, the optimizer environment is the totality of session
settings that can affect execution plan generation, such as the work area size or
optimizer settings (for example, the optimizer mode). Consider the following series of
SQL statements executed by a single user:
ALTER SESSION SET OPTIMIZER_MODE=ALL_ROWS;
ALTER SYSTEM FLUSH SHARED_POOL;
SELECT * FROM sh.sales;

# optimizer environment 1

ALTER SESSION SET OPTIMIZER_MODE=FIRST_ROWS; # optimizer environment 2
SELECT * FROM sh.sales;
ALTER SESSION SET SQL_TRACE=true;
SELECT * FROM sh.sales;

# optimizer environment 3

In the preceding example, the same SELECT statement is executed in three different
optimizer environments. Consequently, the database creates three separate shared
SQL areas for these statements and forces a hard parse of each statement.

See Also:
•

Oracle Database Concepts to learn about private SQL areas and shared
SQL areas

•

Oracle Database Performance Tuning Guide to learn how to configure
the shared pool

•

Oracle Database Concepts to learn about latches

3.1.2 SQL Optimization
During optimization, Oracle Database must perform a hard parse at least once for
every unique DML statement and performs the optimization during this parse.
The database does not optimize DDL. The only exception is when the DDL includes a
DML component such as a subquery that requires optimization.
Related Topics
•

Query Optimizer Concepts
This chapter describes the most important concepts relating to the query
optimizer, including its principal components.

3.1.3 SQL Row Source Generation
The row source generator is software that receives the optimal execution plan from
the optimizer and produces an iterative execution plan that is usable by the rest of the
database.
The iterative plan is a binary program that, when executed by the SQL engine,
produces the result set. The plan takes the form of a combination of steps. Each step
returns a row set. The next step either uses the rows in this set, or the last step returns
the rows to the application issuing the SQL statement.

3-6

Chapter 3

About SQL Processing

A row source is a row set returned by a step in the execution plan along with a control
structure that can iteratively process the rows. The row source can be a table, view, or
result of a join or grouping operation.
The row source generator produces a row source tree, which is a collection of row
sources. The row source tree shows the following information:
•

An ordering of the tables referenced by the statement

•

An access method for each table mentioned in the statement

•

A join method for tables affected by join operations in the statement

•

Data operations such as filter, sort, or aggregation

Example 3-1

Execution Plan

This example shows the execution plan of a SELECT statement when AUTOTRACE is
enabled. The statement selects the last name, job title, and department name for all
employees whose last names begin with the letter A. The execution plan for this
statement is the output of the row source generator.
SELECT
FROM
WHERE
AND
AND

e.last_name, j.job_title, d.department_name
hr.employees e, hr.departments d, hr.jobs j
e.department_id = d.department_id
e.job_id = j.job_id
e.last_name LIKE 'A%';

Execution Plan
---------------------------------------------------------Plan hash value: 975837011
-------------------------------------------------------------------------------| Id| Operation
| Name
|Rows|Bytes|Cost(%CPU)|Time |
-------------------------------------------------------------------------------| 0 | SELECT STATEMENT
|
| 3 | 189 | 7(15)| 00:00:01 |
|*1 | HASH JOIN
|
| 3 | 189 | 7(15)| 00:00:01 |
|*2 | HASH JOIN
|
| 3 | 141 | 5(20)| 00:00:01 |
| 3 |
TABLE ACCESS BY INDEX ROWID| EMPLOYEES | 3 | 60 | 2 (0)| 00:00:01 |
|*4 |
INDEX RANGE SCAN
| EMP_NAME_IX | 3 |
| 1 (0)| 00:00:01 |
| 5 |
TABLE ACCESS FULL
| JOBS
| 19 | 513 | 2 (0)| 00:00:01 |
| 6 | TABLE ACCESS FULL
| DEPARTMENTS | 27 | 432 | 2 (0)| 00:00:01 |
-------------------------------------------------------------------------------Predicate Information (identified by operation id):
--------------------------------------------------1 - access("E"."DEPARTMENT_ID"="D"."DEPARTMENT_ID")
2 - access("E"."JOB_ID"="J"."JOB_ID")
4 - access("E"."LAST_NAME" LIKE 'A%')
filter("E"."LAST_NAME" LIKE 'A%')

3.1.4 SQL Execution
During execution, the SQL engine executes each row source in the tree produced by
the row source generator. This step is the only mandatory step in DML processing.
Figure 3-3 is an execution tree, also called a parse tree, that shows the flow of row
sources from one step to another in the plan in Example 3-1. In general, the order of
the steps in execution is the reverse of the order in the plan, so you read the plan from
the bottom up.

3-7

Chapter 3

About SQL Processing

Each step in an execution plan has an ID number. The numbers in Figure 3-3
correspond to the Id column in the plan shown in Example 3-1. Initial spaces in the
Operation column of the plan indicate hierarchical relationships. For example, if the
name of an operation is preceded by two spaces, then this operation is a child of an
operation preceded by one space. Operations preceded by one space are children of
the SELECT statement itself.

Figure 3-3

Row Source Tree

1
HASH JOIN

2

6

HASH JOIN

TABLE ACCESS
FULL
departments

3

5

TABLE ACCESS
BY INDEX ROWID
employees

TABLE ACCESS
FULL
jobs

4
INDEX RANGE
SCAN
emp_name_ix

In Figure 3-3, each node of the tree acts as a row source, which means that each step
of the execution plan in Example 3-1 either retrieves rows from the database or
accepts rows from one or more row sources as input. The SQL engine executes each
row source as follows:
•

Steps indicated by the black boxes physically retrieve data from an object in the
database. These steps are the access paths, or techniques for retrieving data from
the database.
–

Step 6 uses a full table scan to retrieve all rows from the departments table.

–

Step 5 uses a full table scan to retrieve all rows from the jobs table.

3-8

Chapter 3

How Oracle Database Processes DML

•

–

Step 4 scans the emp_name_ix index in order, looking for each key that begins
with the letter A and retrieving the corresponding rowid. For example, the rowid
corresponding to Atkinson is AAAPzRAAFAAAABSAAe.

–

Step 3 retrieves from the employees table the rows whose rowids were returned
by Step 4. For example, the database uses rowid AAAPzRAAFAAAABSAAe to
retrieve the row for Atkinson.

Steps indicated by the clear boxes operate on row sources.
–

Step 2 performs a hash join, accepting row sources from Steps 3 and 5,
joining each row from the Step 5 row source to its corresponding row in Step
3, and returning the resulting rows to Step 1.
For example, the row for employee Atkinson is associated with the job name
Stock Clerk.

–

Step 1 performs another hash join, accepting row sources from Steps 2 and 6,
joining each row from the Step 6 source to its corresponding row in Step 2,
and returning the result to the client.
For example, the row for employee Atkinson is associated with the department
named Shipping.

In some execution plans the steps are iterative and in others sequential. The hash join
shown in Example 3-1 is sequential. The database completes the steps in their entirety
based on the join order. The database starts with the index range scan of emp_name_ix.
Using the rowids that it retrieves from the index, the database reads the matching rows
in the employees table, and then scans the jobs table. After it retrieves the rows from
the jobs table, the database performs the hash join.
During execution, the database reads the data from disk into memory if the data is not
in memory. The database also takes out any locks and latches necessary to ensure
data integrity and logs any changes made during the SQL execution. The final stage of
processing a SQL statement is closing the cursor.

3.2 How Oracle Database Processes DML
Most DML statements have a query component. In a query, execution of a cursor
places the results of the query into a set of rows called the result set.
This section contains the following topics:
•

How Row Sets Are Fetched
Result set rows can be fetched either a row at a time or in groups.

•

Read Consistency
In general, a query retrieves data by using the Oracle Database read consistency
mechanism, which guarantees that all data blocks read by a query are consistent
to a single point in time.

•

Data Changes
DML statements that must change data use read consistency to retrieve only the
data that matched the search criteria when the modification began.

3.2.1 How Row Sets Are Fetched
Result set rows can be fetched either a row at a time or in groups.

3-9

Chapter 3

How Oracle Database Processes DDL

In the fetch stage, the database selects rows and, if requested by the query, orders the
rows. Each successive fetch retrieves another row of the result until the last row has
been fetched.
In general, the database cannot determine for certain the number of rows to be
retrieved by a query until the last row is fetched. Oracle Database retrieves the data in
response to fetch calls, so that the more rows the database reads, the more work it
performs. For some queries the database returns the first row as quickly as possible,
whereas for others it creates the entire result set before returning the first row.

3.2.2 Read Consistency
In general, a query retrieves data by using the Oracle Database read consistency
mechanism, which guarantees that all data blocks read by a query are consistent to a
single point in time.
Read consistency uses undo data to show past versions of data. For an example,
suppose a query must read 100 data blocks in a full table scan. The query processes
the first 10 blocks while DML in a different session modifies block 75. When the first
session reaches block 75, it realizes the change and uses undo data to retrieve the
old, unmodified version of the data and construct a noncurrent version of block 75 in
memory.

See Also:
Oracle Database Concepts to learn about multiversion read consistency

3.2.3 Data Changes
DML statements that must change data use read consistency to retrieve only the data
that matched the search criteria when the modification began.
Afterward, these statements retrieve the data blocks as they exist in their current state
and make the required modifications. The database must perform other actions related
to the modification of the data such as generating redo and undo data.

3.3 How Oracle Database Processes DDL
Oracle Database processes DDL differently from DML.
For example, when you create a table, the database does not optimize the CREATE
TABLE statement. Instead, Oracle Database parses the DDL statement and carries out
the command.
The database processes DDL differently because it is a means of defining an object in
the data dictionary. Typically, Oracle Database must parse and execute many
recursive SQL statements to execute a DDL statement. Suppose you create a table as
follows:
CREATE TABLE mytable (mycolumn INTEGER);

Typically, the database would run dozens of recursive statements to execute the
preceding statement. The recursive SQL would perform actions such as the following:

3-10

Chapter 3

How Oracle Database Processes DDL

•

Issue a COMMIT before executing the CREATE TABLE statement

•

Verify that user privileges are sufficient to create the table

•

Determine which tablespace the table should reside in

•

Ensure that the tablespace quota has not been exceeded

•

Ensure that no object in the schema has the same name

•

Insert rows that define the table into the data dictionary

•

Issue a COMMIT if the DDL statement succeeded or a ROLLBACK if it did not

See Also:
Oracle Database Development Guide to learn about processing DDL,
transaction control, and other types of statements

3-11

4
Query Optimizer Concepts
This chapter describes the most important concepts relating to the query optimizer,
including its principal components.
This chapter contains the following topics:
•

Introduction to the Query Optimizer
The query optimizer (called simply the optimizer) is built-in database software
that determines the most efficient method for a SQL statement to access
requested data.

•

About Optimizer Components
The optimizer contains three components: the transformer, estimator, and plan
generator.

•

About Automatic Tuning Optimizer
The optimizer performs different operations depending on how it is invoked.

•

About Adaptive Query Optimization
In Oracle Database, adaptive query optimization enables the optimizer to make
run-time adjustments to execution plans and discover additional information that
can lead to better statistics.

•

About Approximate Query Processing
Approximate query processing is a set of optimization techniques that speed
analytic queries by calculating results within an acceptable range of error.

•

About SQL Plan Management
SQL plan management enables the optimizer to automatically manage execution
plans, ensuring that the database uses only known or verified plans.

•

About the Expression Statistics Store (ESS)
The Expression Statistics Store (ESS) is a repository maintained by the
optimizer to store statistics about expression evaluation.

4.1 Introduction to the Query Optimizer
The query optimizer (called simply the optimizer) is built-in database software that
determines the most efficient method for a SQL statement to access requested data.
This section contains the following topics:
•

Purpose of the Query Optimizer
The optimizer attempts to generate the most optimal execution plan for a SQL
statement.

•

Cost-Based Optimization
Query optimization is the process of choosing the most efficient means of
executing a SQL statement.

•

Execution Plans
An execution plan describes a recommended method of execution for a SQL
statement.

4-1

Chapter 4

Introduction to the Query Optimizer

4.1.1 Purpose of the Query Optimizer
The optimizer attempts to generate the most optimal execution plan for a SQL
statement.
The optimizer choose the plan with the lowest cost among all considered candidate
plans. The optimizer uses available statistics to calculate cost. For a specific query in a
given environment, the cost computation accounts for factors of query execution such
as I/O, CPU, and communication.
For example, a query might request information about employees who are managers.
If the optimizer statistics indicate that 80% of employees are managers, then the
optimizer may decide that a full table scan is most efficient. However, if statistics
indicate that very few employees are managers, then reading an index followed by a
table access by rowid may be more efficient than a full table scan.
Because the database has many internal statistics and tools at its disposal, the
optimizer is usually in a better position than the user to determine the optimal method
of statement execution. For this reason, all SQL statements use the optimizer.

4.1.2 Cost-Based Optimization
Query optimization is the process of choosing the most efficient means of executing
a SQL statement.
SQL is a nonprocedural language, so the optimizer is free to merge, reorganize, and
process in any order. The database optimizes each SQL statement based on statistics
collected about the accessed data. The optimizer determines the optimal plan for a
SQL statement by examining multiple access methods, such as full table scan or index
scans, different join methods such as nested loops and hash joins, different join
orders, and possible transformations.
For a given query and environment, the optimizer assigns a relative numerical cost to
each step of a possible plan, and then factors these values together to generate an
overall cost estimate for the plan. After calculating the costs of alternative plans, the
optimizer chooses the plan with the lowest cost estimate. For this reason, the
optimizer is sometimes called the cost-based optimizer (CBO) to contrast it with the
legacy rule-based optimizer (RBO).

Note:
The optimizer may not make the same decisions from one version of Oracle
Database to the next. In recent versions, the optimizer might make different
decision because better information is available and more optimizer
transformations are possible.

Related Topics
•

Cost
The optimizer cost model accounts for the machine resources that a query is
predicted to use.

4-2

Chapter 4

Introduction to the Query Optimizer

4.1.3 Execution Plans
An execution plan describes a recommended method of execution for a SQL
statement.
The plan shows the combination of the steps Oracle Database uses to execute a SQL
statement. Each step either retrieves rows of data physically from the database or
prepares them for the user issuing the statement.
An execution plan displays the cost of the entire plan, indicated on line 0, and each
separate operation. The cost is an internal unit that the execution plan only displays to
allow for plan comparisons. Thus, you cannot tune or change the cost value.
In the following graphic, the optimizer generates two possible execution plans for an
input SQL statement, uses statistics to estimate their costs, compares their costs, and
then chooses the plan with the lowest cost.
Figure 4-1

Execution Plans

GB
NL

Plan
1
NL

GB

Plan
2

HJ

HJ

Generates Multiple
Plans and
Compares Them
Final Plan with
Lowest Cost

Parsed Representation
of SQL Statement
Input

Optimizer

Output

GB
HJ

1 0 1 1 0 0 1 0 0

Plan
2
HJ

Statistics

This section contains the following topics:
•

Query Blocks
The input to the optimizer is a parsed representation of a SQL statement.

•

Query Subplans
For each query block, the optimizer generates a query subplan.

•

Analogy for the Optimizer
One analogy for the optimizer is an online trip advisor.

4.1.3.1 Query Blocks
The input to the optimizer is a parsed representation of a SQL statement.

4-3

Chapter 4

Introduction to the Query Optimizer

Each SELECT block in the original SQL statement is represented internally by a query
block. A query block can be a top-level statement, subquery, or unmerged view.
Example 4-1

Query Blocks

The following SQL statement consists of two query blocks. The subquery in
parentheses is the inner query block. The outer query block, which is the rest of the
SQL statement, retrieves names of employees in the departments whose IDs were
supplied by the subquery. The query form determines how query blocks are
interrelated.
SELECT
FROM
WHERE
IN

first_name, last_name
hr.employees
department_id
(SELECT department_id
FROM hr.departments
WHERE location_id = 1800);

See Also:
•

"View Merging"

•

Oracle Database Concepts for an overview of SQL processing

4.1.3.2 Query Subplans
For each query block, the optimizer generates a query subplan.
The database optimizes query blocks separately from the bottom up. Thus, the
database optimizes the innermost query block first and generates a subplan for it, and
then generates the outer query block representing the entire query.
The number of possible plans for a query block is proportional to the number of objects
in the FROM clause. This number rises exponentially with the number of objects. For
example, the possible plans for a join of five tables are significantly higher than the
possible plans for a join of two tables.

4.1.3.3 Analogy for the Optimizer
One analogy for the optimizer is an online trip advisor.
A cyclist wants to know the most efficient bicycle route from point A to point B. A query
is like the directive "I need the most efficient route from point A to point B" or "I need
the most efficient route from point A to point B by way of point C." The trip advisor
uses an internal algorithm, which relies on factors such as speed and difficulty, to
determine the most efficient route. The cyclist can influence the trip advisor's decision
by using directives such as "I want to arrive as fast as possible" or "I want the easiest
ride possible."
In this analogy, an execution plan is a possible route generated by the trip advisor.
Internally, the advisor may divide the overall route into several subroutes (subplans),
and calculate the efficiency for each subroute separately. For example, the trip advisor
may estimate one subroute at 15 minutes with medium difficulty, an alternative
subroute at 22 minutes with minimal difficulty, and so on.

4-4

Chapter 4

About Optimizer Components

The advisor picks the most efficient (lowest cost) overall route based on user-specified
goals and the available statistics about roads and traffic conditions. The more accurate
the statistics, the better the advice. For example, if the advisor is not frequently notified
of traffic jams, road closures, and poor road conditions, then the recommended route
may turn out to be inefficient (high cost).

4.2 About Optimizer Components
The optimizer contains three components: the transformer, estimator, and plan
generator.
The following graphic illustrates the components.

Figure 4-2

Optimizer Components
Parsed Query
(from Parser)

Query
Transformer
Transformed query

statistics

Estimator

Data
Dictionary

Query + estimates

Plan
Generator

Query Plan
(to Row Source Generator)

A set of query blocks represents a parsed query, which is the input to the optimizer.
The following table describes the optimizer operations.
Table 4-1

Optimizer Operations

Phase

Operation

Description

To Learn More

1

Query
Transformer

The optimizer determines whether it is helpful "Query
to change the form of the query so that the
Transformer"
optimizer can generate a better execution
plan.

2

Estimator

The optimizer estimates the cost of each plan "Estimator"
based on statistics in the data dictionary.

4-5

Chapter 4

About Optimizer Components

Table 4-1

(Cont.) Optimizer Operations

Phase

Operation

Description

To Learn More

3

Plan Generator

The optimizer compares the costs of plans
"Plan Generator"
and chooses the lowest-cost plan, known as
the execution plan, to pass to the row source
generator.

This section contains the following topics:
•

Query Transformer
For some statements, the query transformer determines whether it is
advantageous to rewrite the original SQL statement into a semantically equivalent
SQL statement with a lower cost.

•

Estimator
The estimator is the component of the optimizer that determines the overall cost
of a given execution plan.

•

Plan Generator
The plan generator explores various plans for a query block by trying out different
access paths, join methods, and join orders.

4.2.1 Query Transformer
For some statements, the query transformer determines whether it is advantageous to
rewrite the original SQL statement into a semantically equivalent SQL statement with a
lower cost.
When a viable alternative exists, the database calculates the cost of the alternatives
separately and chooses the lowest-cost alternative. The following graphic shows the
query transformer rewriting an input query that uses OR into an output query that uses
UNION ALL.

4-6

Chapter 4

About Optimizer Components

Figure 4-3
SELECT
FROM
WHERE
OR

Query Transformer

*
sales
promo_id=33
prod_id=136;

Query Transformer

SELECT
FROM
WHERE
UNION
SELECT
FROM
WHERE
AND

*
sales
prod_id=136
ALL
*
sales
promo_id=33
LNNVL(prod_id=136);

Related Topics
•

Query Transformations
The optimizer employs many query transformation techniques. This chapter
describes some of the most important.

4.2.2 Estimator
The estimator is the component of the optimizer that determines the overall cost of a
given execution plan.
The estimator uses three different measures to determine cost:
•

Selectivity
The percentage of rows in the row set that the query selects, with 0 meaning no
rows and 1 meaning all rows. Selectivity is tied to a query predicate, such as WHERE
last_name LIKE 'A%', or a combination of predicates. A predicate becomes more
selective as the selectivity value approaches 0 and less selective (or more
unselective) as the value approaches 1.

Note:
Selectivity is an internal calculation that is not visible in the execution
plans.
•

Cardinality
The cardinality is the number of rows returned by each operation in an execution
plan. This input, which is crucial to obtaining an optimal plan, is common to all cost
functions. The estimator can derive cardinality from the table statistics collected by

4-7

Chapter 4

About Optimizer Components

DBMS_STATS, or derive it after accounting for effects from predicates (filter, join, and
so on), DISTINCT or GROUP BY operations, and so on. The Rows column in an

execution plan shows the estimated cardinality.
•

Cost
This measure represents units of work or resource used. The query optimizer uses
disk I/O, CPU usage, and memory usage as units of work.

As shown in the following graphic, if statistics are available, then the estimator uses
them to compute the measures. The statistics improve the degree of accuracy of the
measures.

Figure 4-4

Estimator
Cardinality
Selectivity

GB

Cost

Plan
Estimator

HJ

Total Cost

HJ

1 0 1 0 0
0 0 0 1 1
0 1 1 0 1

Statistics

For the query shown in Example 4-1, the estimator uses selectivity, estimated
cardinality (a total return of 10 rows), and cost measures to produce its total cost
estimate of 3:
-------------------------------------------------------------------------------| Id| Operation
|Name
|Rows|Bytes|Cost %CPU|Time|
-------------------------------------------------------------------------------| 0 | SELECT STATEMENT
|
| 10| 250| 3 (0)| 00:00:01|
| 1 | NESTED LOOPS
|
| |
|
|
|
| 2 | NESTED LOOPS
|
| 10| 250| 3 (0)| 00:00:01|
|*3 |
TABLE ACCESS FULL
|DEPARTMENTS
| 1| 7| 2 (0)| 00:00:01|
|*4 |
INDEX RANGE SCAN
|EMP_DEPARTMENT_IX| 10|
| 0 (0)| 00:00:01|
| 5 | TABLE ACCESS BY INDEX ROWID|EMPLOYEES
| 10| 180| 1 (0)| 00:00:01|
--------------------------------------------------------------------------------

This section contains the following topics:
•

Selectivity
The selectivity represents a fraction of rows from a row set.

•

Cardinality
The cardinality is the number of rows returned by each operation in an execution
plan.

•

Cost
The optimizer cost model accounts for the machine resources that a query is
predicted to use.

4-8

Chapter 4

About Optimizer Components

4.2.2.1 Selectivity
The selectivity represents a fraction of rows from a row set.
The row set can be a base table, a view, or the result of a join. The selectivity is tied to
a query predicate, such as last_name = 'Smith', or a combination of predicates, such
as last_name = 'Smith' AND job_id = 'SH_CLERK'.

Note:
Selectivity is an internal calculation that is not visible in execution plans.

A predicate filters a specific number of rows from a row set. Thus, the selectivity of a
predicate indicates how many rows pass the predicate test. Selectivity ranges from 0.0
to 1.0. A selectivity of 0.0 means that no rows are selected from a row set, whereas a
selectivity of 1.0 means that all rows are selected. A predicate becomes more
selective as the value approaches 0.0 and less selective (or more unselective) as the
value approaches 1.0.
The optimizer estimates selectivity depending on whether statistics are available:
•

Statistics not available
Depending on the value of the OPTIMIZER_DYNAMIC_SAMPLING initialization parameter,
the optimizer either uses dynamic statistics or an internal default value. The
database uses different internal defaults depending on the predicate type. For
example, the internal default for an equality predicate (last_name = 'Smith') is
lower than for a range predicate (last_name > 'Smith') because an equality
predicate is expected to return a smaller fraction of rows.

•

Statistics available
When statistics are available, the estimator uses them to estimate selectivity.
Assume there are 150 distinct employee last names. For an equality predicate
last_name = 'Smith', selectivity is the reciprocal of the number n of distinct values
of last_name, which in this example is .006 because the query selects rows that
contain 1 out of 150 distinct values.
If a histogram exists on the last_name column, then the estimator uses the
histogram instead of the number of distinct values. The histogram captures the
distribution of different values in a column, so it yields better selectivity estimates,
especially for columns that have data skew.

See Also:
•

"Histograms "

•

Oracle Database Reference to learn more about
OPTIMIZER_DYNAMIC_SAMPLING

4-9

Chapter 4

About Optimizer Components

4.2.2.2 Cardinality
The cardinality is the number of rows returned by each operation in an execution
plan.
For example, if the optimizer estimate for the number of rows returned by a full table
scan is 100, then the cardinality estimate for this operation is 100. The cardinality
estimate appears in the Rows column of the execution plan.
The optimizer determines the cardinality for each operation based on a complex set of
formulas that use both table and column level statistics, or dynamic statistics, as input.
The optimizer uses one of the simplest formulas when a single equality predicate
appears in a single-table query, with no histogram. In this case, the optimizer assumes
a uniform distribution and calculates the cardinality for the query by dividing the total
number of rows in the table by the number of distinct values in the column used in the
WHERE clause predicate.
For example, user hr queries the employees table as follows:
SELECT first_name, last_name
FROM employees
WHERE salary='10200';

The employees table contains 107 rows. The current database statistics indicate that
the number of distinct values in the salary column is 58. Therefore, the optimizer
estimates the cardinality of the result set as 2, using the formula 107/58=1.84.
Cardinality estimates must be as accurate as possible because they influence all
aspects of the execution plan. Cardinality is important when the optimizer determines
the cost of a join. For example, in a nested loops join of the employees and departments
tables, the number of rows in employees determines how often the database must
probe the departments table. Cardinality is also important for determining the cost of
sorts.

4.2.2.3 Cost
The optimizer cost model accounts for the machine resources that a query is
predicted to use.
The cost is an internal numeric measure that represents the estimated resource usage
for a plan. The cost is specific to a query in an optimizer environment. To estimate
cost, the optimizer considers factors such as the following:
•

System resources, which includes estimated I/O, CPU, and memory

•

Estimated number of rows returned (cardinality)

•

Size of the initial data sets

•

Distribution of the data

•

Access structures

4-10

Chapter 4

About Optimizer Components

Note:
The cost is an internal measure that the optimizer uses to compare different
plans for the same query. You cannot tune or change cost.

The execution time is a function of the cost, but cost does not equate directly to time.
For example, if the plan for query A has a lower cost than the plan for query B, then
the following outcomes are possible:
•

A executes faster than B.

•

A executes slower than B.

•

A executes in the same amount of time as B.

Therefore, you cannot compare the costs of different queries with one another. Also,
you cannot compare the costs of semantically equivalent queries that use different
optimizer modes.

4.2.3 Plan Generator
The plan generator explores various plans for a query block by trying out different
access paths, join methods, and join orders.
Many plans are possible because of the various combinations that the database can
use to produce the same result. The optimizer picks the plan with the lowest cost.
The following graphic shows the optimizer testing different plans for an input query.
Figure 4-5

Plan Generator
SELECT e.last_name, d.department_name
FROM
hr.employees e, hr.departments d
WHERE e.department_id = d.department_id;

Optimizer
Transformer
Join Method

Join Order

Hash, Nested
Loop, Sort Merge

departments 0 employees 1
employees 0 departments 1

Access Path
Index
Full Table Scan

Lowest Cost Plan
Hash Join
departments 0, employees 1

4-11

Chapter 4

About Optimizer Components

The following snippet from an optimizer trace file shows some computations that the
optimizer performs:
GENERAL PLANS
***************************************
Considering cardinality-based initial join order.
Permutations for Starting Table :0
Join order[1]: DEPARTMENTS[D]#0 EMPLOYEES[E]#1
***************
Now joining: EMPLOYEES[E]#1
***************
NL Join
Outer table: Card: 27.00 Cost: 2.01 Resp: 2.01 Degree: 1 Bytes: 16
Access path analysis for EMPLOYEES
. . .
Best NL cost: 13.17
. . .
SM Join
SM cost: 6.08
resc: 6.08 resc_io: 4.00 resc_cpu: 2501688
resp: 6.08 resp_io: 4.00 resp_cpu: 2501688
. . .
SM Join (with index on outer)
Access Path: index (FullScan)
. . .
HA Join
HA cost: 4.57
resc: 4.57 resc_io: 4.00 resc_cpu: 678154
resp: 4.57 resp_io: 4.00 resp_cpu: 678154
Best:: JoinMethod: Hash
Cost: 4.57 Degree: 1 Resp: 4.57 Card: 106.00 Bytes: 27
. . .
***********************
Join order[2]: EMPLOYEES[E]#1 DEPARTMENTS[D]#0
. . .
***************
Now joining: DEPARTMENTS[D]#0
***************
. . .
HA Join
HA cost: 4.58
resc: 4.58 resc_io: 4.00 resc_cpu: 690054
resp: 4.58 resp_io: 4.00 resp_cpu: 690054
Join order aborted: cost > best plan cost
***********************

The trace file shows the optimizer first trying the departments table as the outer table in
the join. The optimizer calculates the cost for three different join methods: nested
loops join (NL), sort merge (SM), and hash join (HA). The optimizer picks the hash join
as the most efficient method:
Best:: JoinMethod: Hash
Cost: 4.57 Degree: 1 Resp: 4.57 Card: 106.00 Bytes: 27

The optimizer then tries a different join order, using employees as the outer table. This
join order costs more than the previous join order, so it is abandoned.

4-12

Chapter 4

About Automatic Tuning Optimizer

The optimizer uses an internal cutoff to reduce the number of plans it tries when
finding the lowest-cost plan. The cutoff is based on the cost of the current best plan. If
the current best cost is large, then the optimizer explores alternative plans to find a
lower cost plan. If the current best cost is small, then the optimizer ends the search
swiftly because further cost improvement is not significant.

4.3 About Automatic Tuning Optimizer
The optimizer performs different operations depending on how it is invoked.
The database provides the following types of optimization:
•

Normal optimization
The optimizer compiles the SQL and generates an execution plan. The normal
mode generates a reasonable plan for most SQL statements. Under normal mode,
the optimizer operates with strict time constraints, usually a fraction of a second,
during which it must find an optimal plan.

•

SQL Tuning Advisor optimization
When SQL Tuning Advisor invokes the optimizer, the optimizer is known as
Automatic Tuning Optimizer. In this case, the optimizer performs additional
analysis to further improve the plan produced in normal mode. The optimizer
output is not an execution plan, but a series of actions, along with their rationale
and expected benefit for producing a significantly better plan.

See Also:
•

"Analyzing SQL with SQL Tuning Advisor"

•

Oracle Database 2 Day + Performance Tuning Guide to learn more
about SQL Tuning Advisor

4.4 About Adaptive Query Optimization
In Oracle Database, adaptive query optimization enables the optimizer to make runtime adjustments to execution plans and discover additional information that can lead
to better statistics.
Adaptive optimization is helpful when existing statistics are not sufficient to generate
an optimal plan. The following graphic shows the feature set for adaptive query
optimization.

4-13

Chapter 4

About Adaptive Query Optimization

Figure 4-6

Adaptive Query Optimization

Adaptive Query
Optimization

Adaptive
Plans

Join
Methods

Parallel
Distribution
Methods

Adaptive
Statistics

Bitmap
Index Pruning

Dynamic
Statistics

Automatic
Reoptimization

SQL Plan
Directives

This section contains the following topics:
•

Adaptive Query Plans
An adaptive query plan enables the optimizer to make a plan decision for a
statement during execution.

•

Adaptive Statistics
The optimizer can use adaptive statistics when query predicates are too complex
to rely on base table statistics alone. By default, adaptive statistics are disabled
(OPTIMIZER_ADAPTIVE_STATISTICS is false).

4.4.1 Adaptive Query Plans
An adaptive query plan enables the optimizer to make a plan decision for a
statement during execution.
Adaptive query plans enable the optimizer to fix some classes of problems at run time.
Adaptive plans are enabled by default.
This section contains the following topics:
•

About Adaptive Query Plans
An adaptive query plan contains multiple predetermined subplans, and an
optimizer statistics collector. Based on the statistics collected during execution, the
dynamic plan coordinator chooses the best plan at run time.

•

Purpose of Adaptive Query Plans
The ability of the optimizer to adapt a plan, based on statistics obtained during
execution, can greatly improve query performance.

•

How Adaptive Query Plans Work
For the first execution of a statement, the optimizer uses the default plan, and then
stores an adaptive plan. The database uses the adaptive plan for subsequent
executions unless specific conditions are met.

•

When Adaptive Query Plans Are Enabled
Adaptive query plans are enabled by default.

4-14

Chapter 4

About Adaptive Query Optimization

Related Topics
•

Introduction to Optimizer Statistics
The optimizer cost model relies on statistics collected about the objects involved
in a query, and the database and host where the query runs.

•

About SQL Tuning Advisor
SQL Tuning Advisor is SQL diagnostic software in the Oracle Database Tuning
Pack.

•

Overview of SQL Plan Management
SQL plan management is a preventative mechanism that enables the optimizer
to automatically manage execution plans, ensuring that the database uses only
known or verified plans.

4.4.1.1 About Adaptive Query Plans
An adaptive query plan contains multiple predetermined subplans, and an optimizer
statistics collector. Based on the statistics collected during execution, the dynamic plan
coordinator chooses the best plan at run time.
Dynamic Plans
To change plans at runtime, adaptive query plans use a dynamic plan, which is
represented as a set of subplan groups. A subplan group is a set of subplans. A
subplan is a portion of a plan that the optimizer can switch to as an alternative at run
time. For example, a nested loops join could switch to a hash join during execution.
The optimizer decides which subplan to use at run time. When notified of a new
statistic value relevant to a subplan group, the coordinator dispatches it to the handler
function for this subgroup.

Figure 4-7

Dynamic Plan Coordinator

Dynamic Plan
Subplan Group
Dynamic Plan
Coordinator

GB

Subplan Group
GB

Subplan
NL

Subplan
NL

NL

GB

NL

GB
Subplan

HJ

Subplan
HJ

HJ

HJ

Optimizer Statistics Collector
An optimizer statistics collector is a row source inserted into a plan at key points to
collect run-time statistics relating to cardinality and histograms. These statistics help
the optimizer make a final decision between multiple subplans. The collector also
supports optional buffering up to an internal threshold.

4-15

Chapter 4

About Adaptive Query Optimization

For parallel buffering statistics collectors, each parallel execution server collects the
statistics, which the parallel query coordinator aggregates and then sends to the
clients. In this context, a client is a consumer of the collected statistics, such as a
dynamic plan. Each client specifies a callback function to be executed on each parallel
server or on the query coordinator.

4.4.1.2 Purpose of Adaptive Query Plans
The ability of the optimizer to adapt a plan, based on statistics obtained during
execution, can greatly improve query performance.
Adaptive query plans are useful because the optimizer occasionally picks a suboptimal
default plan because of a cardinality misestimate. The ability of the optimizer to pick
the best plan at run time based on actual execution statistics results in a more optimal
final plan. After choosing the final plan, the optimizer uses it for subsequent
executions, thus ensuring that the suboptimal plan is not reused.

4.4.1.3 How Adaptive Query Plans Work
For the first execution of a statement, the optimizer uses the default plan, and then
stores an adaptive plan. The database uses the adaptive plan for subsequent
executions unless specific conditions are met.
During the first execution of a statement, the database performs the following steps:
1.

The database begins executing the statement using the default plan.

2.

The statistics collector gathers information about the in-progress execution, and
buffers some rows received by the subplan.
For parallel buffering statistics collectors, each slave process collects the statistics,
which the query coordinator aggregates before sending to the clients.

3.

Based on the statistics gathered by the collector, the optimizer chooses a subplan.
The dynamic plan coordinator decides which subplan to use at runtime for all such
subplan groups. When notified of a new statistic value relevant to a subplan group,
the coordinator dispatches it to the handler function for this subgroup.

4.

The collector stops collecting statistics and buffering rows, permitting rows to pass
through instead.

5.

The database stores the adaptive plan in the child cursor, so that the next
execution of the statement can use it.

On subsequent executions of the child cursor, the optimizer continues to use the same
adaptive plan unless one of the following conditions is true, in which case it picks a
new plan for the current execution:
•

The current plan ages out of the shared pool.

•

A different optimizer feature (for example, adaptive cursor sharing or statistics
feedback) invalidates the current plan.

This section contains the following topics:
•

Adaptive Query Plans: Join Method Example
This example shows how the optimizer can choose a different plan based on
information collected at runtime.

4-16

Chapter 4

About Adaptive Query Optimization

•

Adaptive Query Plans: Parallel Distribution Methods
Typically, parallel execution requires data redistribution to perform operations such
as parallel sorts, aggregations, and joins.

•

Adaptive Query Plans: Bitmap Index Pruning
Adaptive plans prune indexes that do not significantly reduce the number of
matched rows.

Related Topics
•

Reading Adaptive Query Plans
The adaptive optimizer is a feature of the optimizer that enables it to adapt plans
based on run-time statistics. All adaptive mechanisms can execute a final plan for
a statement that differs from the default plan.

•

Controlling Adaptive Optimization
In Oracle Database, adaptive query optimization is the process by which the
optimizer adapts an execution plan based on statistics collected at run time.

4.4.1.3.1 Adaptive Query Plans: Join Method Example
This example shows how the optimizer can choose a different plan based on
information collected at runtime.
The following query shows a join of the order_items and prod_info tables.
SELECT
FROM
WHERE
AND
AND

product_name
order_items o, prod_info p
o.unit_price = 15
quantity > 1
p.product_id = o.product_id

An adaptive query plan for this statement shows two possible plans, one with a nested
loops join and the other with a hash join:
SELECT * FROM TABLE(DBMS_XPLAN.display_cursor(FORMAT => 'ADAPTIVE'));
SQL_ID
7hj8dwwy6gm7p, child number 0
------------------------------------SELECT product_name FROM order_items o, prod_info p WHERE
o.unit_price = 15 AND
quantity > 1 AND
p.product_id = o.product_id
Plan hash value: 1553478007
----------------------------------------------------------------------------| Id | Operation
| Name
|Rows|Bytes|Cost (%CPU)|Time|
----------------------------------------------------------------------------| 0| SELECT STATEMENT
|
| |
|7(100)|
|
| * 1| HASH JOIN
|
|4| 128 | 7 (0)|00:00:01|
|- 2| NESTED LOOPS
|
|4| 128 | 7 (0)|00:00:01|
|- 3|
NESTED LOOPS
|
|4| 128 | 7 (0)|00:00:01|
|- 4|
STATISTICS COLLECTOR
|
| |
|
|
|
| * 5|
TABLE ACCESS FULL
| ORDER_ITEMS |4| 48 | 3 (0)|00:00:01|
|-* 6|
INDEX UNIQUE SCAN
| PROD_INFO_PK |1|
| 0 (0)|
|
|- 7|
TABLE ACCESS BY INDEX ROWID| PROD_INFO
|1| 20 | 1 (0)|00:00:01|
| 8| TABLE ACCESS FULL
| PROD_INFO
|1| 20 | 1 (0)|00:00:01|
----------------------------------------------------------------------------Predicate Information (identified by operation id):
---------------------------------------------------

4-17

Chapter 4

About Adaptive Query Optimization

1 - access("P"."PRODUCT_ID"="O"."PRODUCT_ID")
5 - filter(("O"."UNIT_PRICE"=15 AND "QUANTITY">1))
6 - access("P"."PRODUCT_ID"="O"."PRODUCT_ID")
Note
----- this is an adaptive plan (rows marked '-' are inactive)

A nested loops join is preferable if the database can avoid scanning a significant
portion of prod_info because its rows are filtered by the join predicate. If few rows are
filtered, however, then scanning the right table in a hash join is preferable.
The following graphic shows the adaptive process. For the query in the preceding
example, the adaptive portion of the default plan contains two subplans, each of which
uses a different join method. The optimizer automatically determines when each join
method is optimal, depending on the cardinality of the left side of the join.
The statistics collector buffers enough rows coming from the order_items table to
determine which join method to use. If the row count is below the threshold determined
by the optimizer, then the optimizer chooses the nested loops join; otherwise, the
optimizer chooses the hash join. In this case, the row count coming from the
order_items table is above the threshold, so the optimizer chooses a hash join for the
final plan, and disables buffering.

4-18

Chapter 4

About Adaptive Query Optimization

Figure 4-8

Adaptive Join Methods

Default plan is a nested loops join
Nested
Loops

Hash
Join

Statistics
Collector

Table scan
order_items

Index scan
prod_info_pk

Table scan
prod_info

The optimizer buffers rows coming from the order_items table
up to a point. If the row count is less than the threshold,
then use a nested loops join. Otherwise,
switch to a hash join.
Threshold exceeded,
so subplan switches

The optimizer disables the statistics collector after making the decision,
and lets the rows pass through.
Final plan is a hash join
Nested
Loops

Hash
Join

Statistics
Collector

Table scan
order_items

Index scan
prod_info_pk

Table scan
prod_info

The Note section of the execution plan indicates whether the plan is adaptive, and
which rows in the plan are inactive.

See Also:
•

"Controlling Adaptive Optimization"

•

"Reading Execution Plans: Advanced" for an extended example showing
an adaptive query plan

4-19

Chapter 4

About Adaptive Query Optimization

4.4.1.3.2 Adaptive Query Plans: Parallel Distribution Methods
Typically, parallel execution requires data redistribution to perform operations such as
parallel sorts, aggregations, and joins.
Oracle Database can use many different data distributions methods. The database
chooses the method based on the number of rows to be distributed and the number of
parallel server processes in the operation.
For example, consider the following alternative cases:
•

Many parallel server processes distribute few rows.
The database may choose the broadcast distribution method. In this case, each
parallel server process receives each row in the result set.

•

Few parallel server processes distribute many rows.
If a data skew is encountered during the data redistribution, then it could adversely
affect the performance of the statement. The database is more likely to pick a
hash distribution to ensure that each parallel server process receives an equal
number of rows.

The hybrid hash distribution technique is an adaptive parallel data distribution that
does not decide the final data distribution method until execution time. The optimizer
inserts statistic collectors in front of the parallel server processes on the producer side
of the operation. If the number of rows is less than a threshold, defined as twice the
degree of parallelism (DOP), then the data distribution method switches from hash to
broadcast. Otherwise, the distribution method is a hash.
Broadcast Distribution
The following graphic depicts a hybrid hash join between the departments and
employees tables, with a query coordinator directing 8 parallel server processes: P5-P8
are producers, whereas P1-P4 are consumers. Each producer has its own consumer.

4-20

Chapter 4

About Adaptive Query Optimization

Figure 4-9

Adaptive Query with DOP of 4

Query
Coordinator

Statistics collector
threshold is 2X
the DOP

The number of rows
returned is below
threshold, so optimizer
chooses broadcast
method.

P1

P2

P3

P4

1
departments

employees

P5
2
P6
3
P7
4
P8

The database inserts a statistics collector in front of each producer process scanning
the departments table. The query coordinator aggregates the collected statistics. The
distribution method is based on the run-time statistics. In Figure 4-9, the number of
rows is below the threshold (8), which is twice the DOP (4), so the optimizer chooses a
broadcast technique for the departments table.
Hybrid Hash Distribution
Consider an example that returns a greater number of rows. In the following plan, the
threshold is 8, or twice the specified DOP of 4. However, because the statistics
collector (Step 10) discovers that the number of rows (27) is greater than the threshold
(8), the optimizer chooses a hybrid hash distribution rather than a broadcast
distribution. (The time column should show 00:00:01, but shows 0:01 so the plan can fit
the page.)
EXPLAIN PLAN FOR
SELECT /*+ parallel(4) full(e) full(d) */ department_name, sum(salary)
FROM employees e, departments d
WHERE d.department_id=e.department_id
GROUP BY department_name;
Plan hash value: 2940813933
----------------------------------------------------------------------------------------------|Id|Operation
| Name
|Rows|Bytes|Cost |Time| TQ |IN-OUT|PQ Distrib|
----------------------------------------------------------------------------------------------| 0|SELECT STATEMENT
|DEPARTMENTS| 27|621 |6(34)|0:01|
|
|
|
| 1| PX COORDINATOR
|
| |
|
|
|
|
|
|
| 2| PX SEND QC (RANDOM)
| :TQ10003 | 27|621 |6(34)|0:01|Q1,03|P->S| QC (RAND) |

4-21

Chapter 4

About Adaptive Query Optimization

| 3| HASH GROUP BY
|
| 27|621 |6(34)|0:01|Q1,03|PCWP|
|
| 4|
PX RECEIVE
|
| 27|621 |6(34)|0:01|Q1,03|PCWP|
|
| 5|
PX SEND HASH
| :TQ10002 | 27|621 |6(34)|0:01|Q1,02|P->P| HASH
|
| 6|
HASH GROUP BY
|
| 27|621 |6(34)|0:01|Q1,02|PCWP|
|
|*7|
HASH JOIN
|
|106|2438|5(20)|0:01|Q1,02|PCWP|
|
| 8|
PX RECEIVE
|
| 27|432 |2 (0)|0:01|Q1,02|PCWP|
|
| 9|
PX SEND HYBRID HASH
| :TQ10000 | 27|432 |2 (0)|0:01|Q1,00|P->P|HYBRID HASH|
|10|
STATISTICS COLLECTOR
|
| |
|
|
|Q1,00|PCWC|
|
|11|
PX BLOCK ITERATOR
|
| 27|432 |2 (0)|0:01|Q1,00|PCWC|
|
|12|
TABLE ACCESS FULL
|DEPARTMENTS| 27|432 |2 (0)|0:01|Q1,00|PCWP|
|
|13|
PX RECEIVE
|
|107|749 |2 (0)|0:01|Q1,02|PCWP|
|
|14|
PX SEND HYBRID HASH (SKEW)| :TQ10001 |107|749 |2 (0)|0:01|Q1,01|P->P|HYBRID HASH|
|15|
PX BLOCK ITERATOR
|
|107|749 |2 (0)|0:01|Q1,01|PCWC|
|
|16|
TABLE ACCESS FULL
| EMPLOYEES |107|749 |2 (0)|0:01|Q1,01|PCWP|
|
----------------------------------------------------------------------------------------------Predicate Information (identified by operation id):
--------------------------------------------------7 - access("D"."DEPARTMENT_ID"="E"."DEPARTMENT_ID")
Note
----- Degree of Parallelism is 4 because of hint
32 rows selected.

See Also:
Oracle Database VLDB and Partitioning Guide to learn more about parallel
data redistribution techniques

4.4.1.3.3 Adaptive Query Plans: Bitmap Index Pruning
Adaptive plans prune indexes that do not significantly reduce the number of matched
rows.
When the optimizer generates a star transformation plan, it must choose the right
combination of bitmap indexes to reduce the relevant set of rowids as efficiently as
possible. If many indexes exist, some indexes might not reduce the rowid set
substantially, but nevertheless introduce significant processing cost during query
execution. Adaptive plans can solve this problem by not using indexes that degrade
performance.
Example 4-2

Bitmap Index Pruning

In this example, you issue the following star query, which joins the cars fact table with
multiple dimension tables (sample output included):
SELECT /*+ star_transformation(r) */ l.color_name, k.make_name,
h.filter_col, count(*)
FROM cars r, colors l, makes k, models d, hcc_tab h
WHERE r.make_id = k.make_id
AND
r.color_id = l.color_id
AND
r.model_id = d.model_id
AND
r.high_card_col = h.high_card_col
AND
d.model_name = 'RAV4'

4-22

Chapter 4

About Adaptive Query Optimization

AND
AND
AND
GROUP

k.make_name = 'Toyota'
l.color_name = 'Burgundy'
h.filter_col = 100
BY l.color_name, k.make_name, h.filter_col;

COLOR_NA MAKE_N FILTER_COL COUNT(*)
-------- ------ ---------- ---------Burgundy Toyota
100
15000

The following sample execution plan shows that the query generated no rows for the
bitmap node in Step 12 and Step 17. The adaptive optimizer determined that filtering
rows by using the CAR_MODEL_IDX and CAR_MAKE_IDX indexes was inefficient. The query
did not use the steps in the plan that begin with a dash (-).
----------------------------------------------------------| Id | Operation
| Name
|
----------------------------------------------------------| 0 | SELECT STATEMENT
|
|
| 1 | SORT GROUP BY NOSORT
|
|
| 2 | HASH JOIN
|
|
| 3 |
VIEW
| VW_ST_5497B905 |
| 4 |
NESTED LOOPS
|
|
| 5 |
BITMAP CONVERSION TO ROWIDS |
|
| 6 |
BITMAP AND
|
|
| 7 |
BITMAP MERGE
|
|
| 8 |
BITMAP KEY ITERATION
|
|
| 9 |
TABLE ACCESS FULL
| COLORS
|
| 10 |
BITMAP INDEX RANGE SCAN | CAR_COLOR_IDX |
|- 11 |
STATISTICS COLLECTOR
|
|
|- 12 |
BITMAP MERGE
|
|
|- 13 |
BITMAP KEY ITERATION
|
|
|- 14 |
TABLE ACCESS FULL
| MODELS
|
|- 15 |
BITMAP INDEX RANGE SCAN | CAR_MODEL_IDX |
|- 16 |
STATISTICS COLLECTOR
|
|
|- 17 |
BITMAP MERGE
|
|
|- 18 |
BITMAP KEY ITERATION
|
|
|- 19 |
TABLE ACCESS FULL
| MAKES
|
|- 20 |
BITMAP INDEX RANGE SCAN | CAR_MAKE_IDX |
| 21 |
TABLE ACCESS BY USER ROWID | CARS
|
| 22 |
MERGE JOIN CARTESIAN
|
|
| 23 |
MERGE JOIN CARTESIAN
|
|
| 24 |
MERGE JOIN CARTESIAN
|
|
| 25 |
TABLE ACCESS FULL
| MAKES
|
| 26 |
BUFFER SORT
|
|
| 27 |
TABLE ACCESS FULL
| MODELS
|
| 28 |
BUFFER SORT
|
|
| 29 |
TABLE ACCESS FULL
| COLORS
|
| 30 |
BUFFER SORT
|
|
| 31 |
TABLE ACCESS FULL
| HCC_TAB
|
----------------------------------------------------------Note
----- dynamic statistics used: dynamic sampling (level=2)
- star transformation used for this statement
- this is an adaptive plan (rows marked '-' are inactive)

4-23

Chapter 4

About Adaptive Query Optimization

4.4.1.4 When Adaptive Query Plans Are Enabled
Adaptive query plans are enabled by default.
Adaptive plans are enabled when the following initialization parameters are set:
•

OPTIMIZER_ADAPTIVE_PLANS is TRUE (default)

•

OPTIMIZER_FEATURES_ENABLE is 12.1.0.1 or later

•

OPTIMIZER_ADAPTIVE_REPORTING_ONLY is FALSE (default)

Adaptive plans control the following optimizations:
•

Nested loops and hash join selection

•

Star transformation bitmap pruning

•

Adaptive parallel distribution method

See Also:
•

"Controlling Adaptive Optimization"

•

Oracle Database Reference to learn more about
OPTIMIZER_ADAPTIVE_PLANS

4.4.2 Adaptive Statistics
The optimizer can use adaptive statistics when query predicates are too complex to
rely on base table statistics alone. By default, adaptive statistics are disabled
(OPTIMIZER_ADAPTIVE_STATISTICS is false).
The following topics describe types of adaptive statistics:
•

Dynamic Statistics
Dynamic statistics are an optimization technique in which the database executes
a recursive SQL statement to scan a small random sample of a table's blocks to
estimate predicate cardinalities.

•

Automatic Reoptimization
In automatic reoptimization, the optimizer changes a plan on subsequent
executions after the initial execution.

•

SQL Plan Directives
A SQL plan directive is additional information that the optimizer uses to generate
a more optimal plan.

•

When Adaptive Statistics Are Enabled
Adaptive statistics are disabled by default.

4.4.2.1 Dynamic Statistics
Dynamic statistics are an optimization technique in which the database executes a
recursive SQL statement to scan a small random sample of a table's blocks to
estimate predicate cardinalities.

4-24

Chapter 4

About Adaptive Query Optimization

During SQL compilation, the optimizer decides whether to use dynamic statistics by
considering whether available statistics are sufficient to generate an optimal plan. If
the available statistics are insufficient, then the optimizer uses dynamic statistics to
augment the statistics. To improve the quality of optimizer decisions, the optimizer can
use dynamic statistics for table scans, index access, joins, and GROUP BY operations.
Related Topics
•

Supplemental Dynamic Statistics
By default, when optimizer statistics are missing, stale, or insufficient, the
database automatically gathers dynamic statistics during a parse. The database
uses recursive SQL to scan a small random sample of table blocks.

4.4.2.2 Automatic Reoptimization
In automatic reoptimization, the optimizer changes a plan on subsequent executions
after the initial execution.
Adaptive query plans are not feasible for all kinds of plan changes. For example, a
query with an inefficient join order might perform suboptimally, but adaptive query
plans do not support adapting the join order during execution. At the end of the first
execution of a SQL statement, the optimizer uses the information gathered during
execution to determine whether automatic reoptimization has a cost benefit. If
execution information differs significantly from optimizer estimates, then the optimizer
looks for a replacement plan on the next execution.
The optimizer uses the information gathered during the previous execution to help
determine an alternative plan. The optimizer can reoptimize a query several times,
each time gathering additional data and further improving the plan.
Automatic reoptimization takes the following forms:
•

Reoptimization: Statistics Feedback
A form of reoptimization known as statistics feedback (formerly known as
cardinality feedback) automatically improves plans for repeated queries that have
cardinality misestimates.

•

Reoptimization: Performance Feedback
Another form of reoptimization is performance feedback. This reoptimization helps
improve the degree of parallelism automatically chosen for repeated SQL
statements when PARALLEL_DEGREE_POLICY is set to ADAPTIVE.

Related Topics
•

Controlling Adaptive Optimization
In Oracle Database, adaptive query optimization is the process by which the
optimizer adapts an execution plan based on statistics collected at run time.

4.4.2.2.1 Reoptimization: Statistics Feedback
A form of reoptimization known as statistics feedback (formerly known as cardinality
feedback) automatically improves plans for repeated queries that have cardinality
misestimates.
The optimizer can estimate cardinalities incorrectly for many reasons, such as missing
statistics, inaccurate statistics, or complex predicates. The basic process of
reoptimization using statistics feedback is as follows:

4-25

Chapter 4

About Adaptive Query Optimization

1.

During the first execution of a SQL statement, the optimizer generates an
execution plan.
The optimizer may enable monitoring for statistics feedback for the shared SQL
area in the following cases:

2.

•

Tables with no statistics

•

Multiple conjunctive or disjunctive filter predicates on a table

•

Predicates containing complex operators for which the optimizer cannot
accurately compute selectivity estimates

At the end of the first execution, the optimizer compares its initial cardinality
estimates to the actual number of rows returned by each operation in the plan
during execution.
If estimates differ significantly from actual cardinalities, then the optimizer stores
the correct estimates for subsequent use. The optimizer also creates a SQL plan
directive so that other SQL statements can benefit from the information obtained
during this initial execution.

3.

If the query executes again, then the optimizer uses the corrected cardinality
estimates instead of its usual estimates.

The OPTIMIZER_ADAPTIVE_STATISTICS initialization parameter does not control all features
of automatic reoptimization. Specifically, this parameter controls statistics feedback for
join cardinality only in the context of automatic reoptimization. For example, setting
OPTIMIZER_ADAPTIVE_STATISTICS to FALSE disables statistics feedback for join cardinality
misestimates, but it does not disable statistics feedback for single-table cardinality
misestimates.
Example 4-3

Statistics Feedback
This example shows how the database uses statistics feedback to adjust incorrect
estimates.
1.

The user oe runs the following query of the orders, order_items, and
product_information tables:
SELECT o.order_id, v.product_name
FROM orders o,
( SELECT order_id, product_name
FROM order_items o, product_information p
WHERE p.product_id = o.product_id
AND
list_price < 50
AND
min_price < 40 ) v
WHERE o.order_id = v.order_id

2.

Querying the plan in the cursor shows that the estimated rows (E-Rows) is far fewer
than the actual rows (A-Rows).

-------------------------------------------------------------------------------------------------| Id | Operation
| Name
|Starts|E-Rows|A-Rows|A-Time|Buffers|OMem|1Mem|O/1/M|
-------------------------------------------------------------------------------------------------| 0| SELECT STATEMENT
|
| 1|
| 269 |00:00:00.14|1338|
|
|
|
| 1| NESTED LOOPS
|
| 1| 1 | 269 |00:00:00.14|1338|
|
|
|
| 2| MERGE JOIN CARTESIAN|
| 1| 4 |9135 |00:00:00.05| 33|
|
|
|
|*3|
TABLE ACCESS FULL |PRODUCT_INFORMATION| 1| 1 | 87 |00:00:00.01| 32|
|
|
|
| 4|
BUFFER SORT
|
| 87| 105 |9135 |00:00:00.02| 1|4096|4096|1/0/0|
| 5|
INDEX FULL SCAN |ORDER_PK
| 1| 105 | 105 |00:00:00.01| 1|
|
|
|
|*6| INDEX UNIQUE SCAN |ORDER_ITEMS_UK
|9135| 1 | 269 |00:00:00.04|1305|
|
|
|
--------------------------------------------------------------------------------------------------

4-26

Chapter 4

About Adaptive Query Optimization

Predicate Information (identified by operation id):
--------------------------------------------------3 - filter(("MIN_PRICE"<40 AND "LIST_PRICE"<50))
6 - access("O"."ORDER_ID"="ORDER_ID" AND "P"."PRODUCT_ID"="O"."PRODUCT_ID")
3.

The user oe reruns the query in Step 1.

4.

Querying the plan in the cursor shows that the optimizer used statistics feedback
(shown in the Note) for the second execution, and also chose a different plan.

-------------------------------------------------------------------------------------------------|Id | Operation
| Name | Starts |E-Rows|A-Rows|A-Time|Buffers|Reads|OMem|1Mem|O/1/M|
-------------------------------------------------------------------------------------------------| 0| SELECT STATEMENT
|
| 1| | 269 |00:00:00.05|60|1|
|
|
|
| 1| NESTED LOOPS
|
| 1|269| 269 |00:00:00.05|60|1|
|
|
|
|*2| HASH JOIN
|
| 1|313| 269 |00:00:00.05|39|1|1398K|1398K|1/0/0|
|*3|
TABLE ACCESS FULL |PRODUCT_INFORMATION| 1| 87| 87 |00:00:00.01|15|0|
|
|
|
| 4|
INDEX FAST FULL SCAN|ORDER_ITEMS_UK
| 1|665| 665 |00:00:00.01|24|1|
|
|
|
|*5| INDEX UNIQUE SCAN
|ORDER_PK
|269| 1| 269 |00:00:00.01|21|0|
|
|
|
-------------------------------------------------------------------------------------------------Predicate Information (identified by operation id):
--------------------------------------------------2 - access("P"."PRODUCT_ID"="O"."PRODUCT_ID")
3 - filter(("MIN_PRICE"<40 AND "LIST_PRICE"<50))
5 - access("O"."ORDER_ID"="ORDER_ID")
Note
----- statistics feedback used for this statement

In the preceding output, the estimated number of rows (269) in Step 1 matches the
actual number of rows.

4.4.2.2.2 Reoptimization: Performance Feedback
Another form of reoptimization is performance feedback. This reoptimization helps
improve the degree of parallelism automatically chosen for repeated SQL statements
when PARALLEL_DEGREE_POLICY is set to ADAPTIVE.
The basic process of reoptimization using performance feedback is as follows:
1.

During the first execution of a SQL statement, when PARALLEL_DEGREE_POLICY is set
to ADAPTIVE, the optimizer determines whether to execute the statement in parallel,
and if so, which degree of parallelism to use.
The optimizer chooses the degree of parallelism based on the estimated
performance of the statement. Additional performance monitoring is enabled for all
statements.

2.

At the end of the initial execution, the optimizer compares the following:
•

The degree of parallelism chosen by the optimizer

•

The degree of parallelism computed based on the performance statistics (for
example, the CPU time) gathered during the actual execution of the statement

4-27

Chapter 4

About Adaptive Query Optimization

If the two values vary significantly, then the database marks the statement for
reparsing, and stores the initial execution statistics as feedback. This feedback
helps better compute the degree of parallelism for subsequent executions.
3.

If the query executes again, then the optimizer uses the performance statistics
gathered during the initial execution to better determine a degree of parallelism for
the statement.

Note:
Even if PARALLEL_DEGREE_POLICY is not set to ADAPTIVE, statistics feedback may
influence the degree of parallelism chosen for a statement.

4.4.2.3 SQL Plan Directives
A SQL plan directive is additional information that the optimizer uses to generate a
more optimal plan.
The directive is a “note to self” by the optimizer that it is misestimating cardinalities of
certain types of predicates, and also a reminder to DBMS_STATS to gather statistics
needed to correct the misestimates in the future.
For example, during query optimization, when deciding whether the table is a
candidate for dynamic statistics, the database queries the statistics repository for
directives on a table. If the query joins two tables that have a data skew in their join
columns, then a SQL plan directive can direct the optimizer to use dynamic statistics to
obtain an accurate cardinality estimate.
The optimizer collects SQL plan directives on query expressions rather than at the
statement level so that it can apply directives to multiple SQL statements. The
optimizer not only corrects itself, but also records information about the mistake, so
that the database can continue to correct its estimates even after a query—and any
similar query—is flushed from the shared pool.
The database automatically creates directives, and stores them in the SYSAUX
tablespace. You can alter, save to disk, and transport directives using the PL/SQL
package DBMS_SPD.

See Also:
•

"SQL Plan Directives"

•

"Managing SQL Plan Directives"

•

Oracle Database PL/SQL Packages and Types Reference to learn about
the DBMS_SPD package

4.4.2.4 When Adaptive Statistics Are Enabled
Adaptive statistics are disabled by default.
Adaptive statistics are enabled when the following initialization parameters are set:

4-28

Chapter 4

About Approximate Query Processing

•

OPTIMIZER_ADAPTIVE_STATISTICS is TRUE (the default is FALSE)

•

OPTIMIZER_FEATURES_ENABLE is 12.1.0.1 or later

Setting OPTIMIZER_ADAPTIVE_STATISTICS to TRUE enables the following features:
•

SQL plan directives

•

Statistics feedback for join cardinality

•

Performance feedback

•

Adaptive dynamic sampling

Note:
Setting OPTIMIZER_ADAPTIVE_STATISTICS to FALSE preserves statistics feedback
for single-table cardinality misestimates.

See Also:
•
•

"Controlling Adaptive Optimization"
Oracle Database Reference to learn more about
OPTIMIZER_ADAPTIVE_STATISTICS

4.5 About Approximate Query Processing
Approximate query processing is a set of optimization techniques that speed
analytic queries by calculating results within an acceptable range of error.
Business intelligence (BI) queries heavily rely on aggregate functions (SUM, RANK,
MEDIAN, and so on) that require sorting. For example, an application generates reports
showing how many distinct customers are logged on, or which products were most
popular last week. It is not uncommon for BI applications to have the following
requirements:
•

Queries must be able to process data sets that are orders of magnitude larger
than in traditional data warehouses.
For example, the daily volumes of web logs of a popular website can reach tens or
hundreds of terabytes a day.

•

Queries must provide near real-time response.
For example, a company requires quick detection and response to credit card
fraud.

•

Explorative queries of large data sets must be fast.
For example, a user might want to find out a list of departments whose sales have
approximately reached a specific threshold. A user would form targeted queries on
these departments to find more detailed information, such as the exact sales
number, the locations of these departments, and so on.

4-29

Chapter 4

About Approximate Query Processing

For large data sets, exact aggregation queries consume extensive memory, often
spilling to temp space, and can be unacceptably slow. Applications are often more
interested in a general pattern than exact results, so customers are willing to sacrifice
exactitude for speed. For example, if the goal is to show a bar chart depicting the most
popular products, then whether a product sold 1 million units or .999 million units is
statistically insignificant.
Oracle Database implements its solution through approximate query processing.
Typically, the accuracy of the approximate aggregation is over 97% (with 95%
confidence), but the processing time is orders of magnitude faster. The database uses
less CPU, and avoids writing to temp files.
You can implement approximate query processing without changing existing code by
using the APPROX_FOR_* initialization parameters. You can set these parameters at the
database or session level. The following table describes initialization parameters and
SQL functions relevant to approximation techniques.
Table 4-2

Approximate Query User Interface

User Interface

Description

See Also

APPROX_FOR_AGGREGATION initialization parameter

Enables approximate query
processing.

Oracle Database
Reference

Setting this parameter to false
disables all automatic conversion
from exact aggregate to
approximate aggregate,
regardless of the settings of the
APPROX_FOR_COUNT_DISTINCT and
APPROX_FOR_PERCENTILE
parameters.
APPROX_FOR_COUNT_DISTINCT initialization
parameter

Converts COUNT(DISTINCT) to
APPROX_COUNT_DISTINCT.

Oracle Database
Reference

APPROX_FOR_PERCENTILE initialization parameter

Converts eligible exact percentile
functions to their
APPROX_PERCENTILE_*
counterparts.

Oracle Database
Reference

APPROX_COUNT_DISTINCT function

Returns the approximate number Oracle Database SQL
of rows that contain distinct values Language Reference
of an expression.

APPROX_COUNT_DISTINCT_AGG function

Aggregates the precomputed
approximate count distinct
synopses to a higher level.

Oracle Database SQL
Language Reference

APPROX_COUNT_DISTINCT_DETAIL function

Returns the synopses of the
APPROX_COUNT_DISTINCT function
as a BLOB.

Oracle Database SQL
Language Reference

The database can persist the
returned result to disk for further
aggregation.

4-30

Chapter 4

About SQL Plan Management

Table 4-2

(Cont.) Approximate Query User Interface

User Interface

Description

See Also

APPROX_PERCENTILE function

Accepts a percentile value and a
sort specification, and returns an
approximate interpolated value
that falls into that percentile value
with respect to the sort
specification.

Oracle Database SQL
Language Reference

This function provides an
alternative to the PERCENTILE_CONT
function.
APPROX_MEDIAN function

Accepts a numeric or date-time
Oracle Database SQL
value, and returns an approximate Language Reference
middle or approximate interpolated
value that would be the middle
value when the values are sorted.
This function provides an
alternative to the MEDIAN function.

See Also:
"NDV Algorithms: Adaptive Sampling and HyperLogLog"

4.6 About SQL Plan Management
SQL plan management enables the optimizer to automatically manage execution
plans, ensuring that the database uses only known or verified plans.
SQL plan management can build a SQL plan baseline, which contains one or more
accepted plans for each SQL statement. The optimizer can access and manage the
plan history and SQL plan baselines of SQL statements. The main objectives are as
follows:
•

Identify repeatable SQL statements

•

Maintain plan history, and possibly SQL plan baselines, for a set of SQL
statements

•

Detect plans that are not in the plan history

•

Detect potentially better plans that are not in the SQL plan baseline

The optimizer uses the normal cost-based search method.

4-31

Chapter 4

About the Expression Statistics Store (ESS)

See Also:
•

"Managing SQL Plan Baselines"

•

Oracle Database PL/SQL Packages and Types Reference to learn about
the DBMS_SPM package

4.7 About the Expression Statistics Store (ESS)
The Expression Statistics Store (ESS) is a repository maintained by the optimizer to
store statistics about expression evaluation.
When an IM column store is enabled, the database leverages the ESS for its InMemory Expressions (IM expressions) feature. However, the ESS is independent of
the IM column store. The ESS is a permanent component of the database and cannot
be disabled.
The database uses the ESS to determine whether an expression is “hot” (frequently
accessed), and thus a candidate for an IM expression. During a hard parse of a query,
the ESS looks for active expressions in the SELECT list, WHERE clause, GROUP BY clause,
and so on.
For each segment, the ESS maintains expression statistics such as the following:
•

Frequency of execution

•

Cost of evaluation

•

Timestamp evaluation

The optimizer assigns each expression a weighted score based on cost and the
number of times it was evaluated. The values are approximate rather than exact. More
active expressions have higher scores. The ESS maintains an internal list of the most
frequently accessed expressions.
The ESS resides in the SGA and also persists on disk. The database saves the
statistics to disk every 15 minutes, or immediately using the
DBMS_STATS.FLUSH_DATABASE_MONITORING_INFO procedure. The statistics are visible in the
DBA_EXPRESSION_STATISTICS view.

See Also:
•

Oracle Database In-Memory Guide to learn more about the ESS

•

Oracle Database PL/SQL Packages and Types Reference to learn more
about DBMS_STATS.FLUSH_DATABASE_MONITORING_INFO

4-32

5
Query Transformations
The optimizer employs many query transformation techniques. This chapter describes
some of the most important.
This chapter contains the following topics:
•

OR Expansion
In OR expansion, the optimizer transforms a query block containing top-level
disjunctions into the form of a UNION ALL query that contains two or more branches.
The optimizer achieves this goal by splitting the disjunction into its components,
and then associating each component with a branch of a UNION ALL query.

•

View Merging
In view merging, the optimizer merges the query block representing a view into
the query block that contains it.

•

Predicate Pushing
In predicate pushing, the optimizer "pushes" the relevant predicates from the
containing query block into the view query block.

•

Subquery Unnesting
In subquery unnesting, the optimizer transforms a nested query into an
equivalent join statement, and then optimizes the join.

•

Query Rewrite with Materialized Views
A materialized view is a query result that the database materializes and stores in
a table.

•

Star Transformation
Star transformation is an optimizer transformation that avoids full table scans of
fact tables in a star schema.

•

In-Memory Aggregation (VECTOR GROUP BY)
The key optimization of in-memory aggregation is to aggregate while scanning.

•

Cursor-Duration Temporary Tables
To materialize the intermediate results of a query, Oracle Database may implicitly
create a cursor-duration temporary table in memory during query compilation.

•

Table Expansion
In table expansion, the optimizer generates a plan that uses indexes on the readmostly portion of a partitioned table, but not on the active portion of the table.

•

Join Factorization
In the cost-based transformation known as join factorization, the optimizer can
factorize common computations from branches of a UNION ALL query.

Related Topics
•

Query Transformer
For some statements, the query transformer determines whether it is
advantageous to rewrite the original SQL statement into a semantically equivalent
SQL statement with a lower cost.

5-1

Chapter 5

OR Expansion

5.1 OR Expansion
In OR expansion, the optimizer transforms a query block containing top-level
disjunctions into the form of a UNION ALL query that contains two or more branches.
The optimizer achieves this goal by splitting the disjunction into its components, and
then associating each component with a branch of a UNION ALL query.
The optimizer can choose OR expansion for various reasons. For example, it may
enable more efficient access paths or alternative join methods that avoid Cartesian
products. As always, the optimizer performs the expansion only if the cost of the
transformed statement is lower than the cost of the original statement.
In previous releases, the optimizer used the CONCATENATION operator to perform the OR
expansion. Starting in Oracle Database 12c Release 2 (12.2), the optimizer uses the
UNION-ALL operator instead. The framework provides the following enhancements:

Example 5-1

•

Enables interaction among various transformations

•

Avoids sharing query structures

•

Enables the exploration of various search strategies

•

Provides the reuse of cost annotation

•

Supports the standard SQL syntax

Transformed Query: UNION ALL Condition
To prepare for this example, log in to the database as an administrator, execute the
following statements to add a unique constraint to the hr.departments.department_name
column, and then add 100,000 rows to the hr.employees table:
ALTER TABLE hr.departments ADD CONSTRAINT department_name_uk UNIQUE
(department_name);
DELETE FROM hr.employees WHERE employee_id > 999;
DECLARE
v_counter NUMBER(7) := 1000;
BEGIN
FOR i IN 1..100000 LOOP
INSERT INTO hr.employees
VALUES (v_counter,null,'Doe','Doe' || v_counter || '@example.com',null,'07JUN-02','AC_ACCOUNT',null,null,null,50);
v_counter := v_counter + 1;
END LOOP;
END;
/
COMMIT;
EXEC DBMS_STATS.GATHER_TABLE_STATS ( ownname => 'hr', tabname => 'employees');

You then connect as the user hr, and execute the following query, which joins the
employees and departments tables:
SELECT
FROM
WHERE
AND

*
employees e, departments d
(e.email='SSTILES' OR d.department_name='Treasury')
e.department_id = d.department_id;

Without OR expansion, the optimizer treats e.email='SSTILES' OR
d.department_name='Treasury' as a single unit. Consequently, the optimizer cannot use

5-2

Chapter 5

View Merging

the index on either the e.email or d.department_name column, and so performs a full
table scan of employees and departments.
With OR expansion, the optimizer breaks the disjunctive predicate into two independent
predicates, as shown in the following example:
SELECT *
FROM employees e, departments d
WHERE e.email = 'SSTILES'
AND
e.department_id = d.department_id
UNION ALL
SELECT *
FROM employees e, departments d
WHERE d.department_name = 'Treasury'
AND
e.department_id = d.department_id;

This transformation enables the e.email and d.department_name columns to serve as
index keys. Performance improves because the database filters data using two unique
indexes instead of two full table scans, as shown in the following execution plan:
Plan hash value: 2512933241
----------------------------------------------------------------------------------------------| Id| Operation
| Name
|Rows|Bytes|Cost(%CPU)|Time |
----------------------------------------------------------------------------------------------| 0 | SELECT STATEMENT
|
|
|
|122 (100)|
|
| 1 | VIEW
|VW_ORE_19FF4E3E |9102|1679K|122 (5) |00:00:01|
| 2 | UNION-ALL
|
|
|
|
|
|
| 3 |
NESTED LOOPS
|
| 1 | 78 | 4 (0) |00:00:01|
| 4 |
TABLE ACCESS BY INDEX ROWID
| EMPLOYEES
| 1 | 57 | 3 (0) |00:00:01|
|*5 |
INDEX UNIQUE SCAN
| EMP_EMAIL_UK
| 1 |
| 2 (0) |00:00:01|
| 6 |
TABLE ACCESS BY INDEX ROWID
| DEPARTMENTS
| 1 | 21 | 1 (0) |00:00:01|
|*7 |
INDEX UNIQUE SCAN
| DEPT_ID_PK
| 1 |
| 0 (0) |
|
| 8 |
NESTED LOOPS
|
|9101| 693K|118 (5) |00:00:01|
| 9 |
TABLE ACCESS BY INDEX ROWID
| DEPARTMENTS
| 1 | 21 | 1 (0) |00:00:01|
|*10|
INDEX UNIQUE SCAN
|DEPARTMENT_NAME_UK| 1 |
| 0 (0) |
|
|*11|
TABLE ACCESS BY INDEX ROWID BATCHED| EMPLOYEES
|9101| 506K|117 (5) |00:00:01|
|*12|
INDEX RANGE SCAN
|EMP_DEPARTMENT_IX |9101|
| 35 (6) |00:00:01|
----------------------------------------------------------------------------------------------Predicate Information (identified by operation id):
--------------------------------------------------5
7
10
11
12

-

access("E"."EMAIL"='SSTILES')
access("E"."DEPARTMENT_ID"="D"."DEPARTMENT_ID")
access("D"."DEPARTMENT_NAME"='Treasury')
filter(LNNVL("E"."EMAIL"='SSTILES'))
access("E"."DEPARTMENT_ID"="D"."DEPARTMENT_ID")

35 rows selected.

5.2 View Merging
In view merging, the optimizer merges the query block representing a view into the
query block that contains it.
View merging can improve plans by enabling the optimizer to consider additional join
orders, access methods, and other transformations. For example, after a view has
been merged and several tables reside in one query block, a table inside a view may
permit the optimizer to use join elimination to remove a table outside the view.

5-3

Chapter 5

View Merging

For certain simple views in which merging always leads to a better plan, the optimizer
automatically merges the view without considering cost. Otherwise, the optimizer uses
cost to make the determination. The optimizer may choose not to merge a view for
many reasons, including cost or validity restrictions.
If OPTIMIZER_SECURE_VIEW_MERGING is true (default), then Oracle Database performs
checks to ensure that view merging and predicate pushing do not violate the security
intentions of the view creator. To disable these additional security checks for a specific
view, you can grant the MERGE VIEW privilege to a user for this view. To disable
additional security checks for all views for a specific user, you can grant the MERGE ANY
VIEW privilege to that user.

Note:
You can use hints to override view merging rejected because of cost or
heuristics, but not validity.

This section contains the following topics:
•

Query Blocks in View Merging
The optimizer represents each nested subquery or unmerged view by a separate
query block.

•

Simple View Merging
In simple view merging, the optimizer merges select-project-join views.

•

Complex View Merging
In view merging, the optimizer merges views containing GROUP BY and DISTINCT
views. Like simple view merging, complex merging enables the optimizer to
consider additional join orders and access paths.

See Also:
•

Oracle Database SQL Language Reference for more information about
the MERGE ANY VIEW and MERGE VIEW privileges

•

Oracle Database Reference for more information about the
OPTIMIZER_SECURE_VIEW_MERGING initialization parameter

5.2.1 Query Blocks in View Merging
The optimizer represents each nested subquery or unmerged view by a separate
query block.
The database optimizes query blocks separately from the bottom up. Thus, the
database optimizes the innermost query block first, generates the part of the plan for it,
and then generates the plan for the outer query block, representing the entire query.
The parser expands each view referenced in a query into a separate query block. The
block essentially represents the view definition, and thus the result of a view. One
option for the optimizer is to analyze the view query block separately, generate a view

5-4

Chapter 5

View Merging

subplan, and then process the rest of the query by using the view subplan to generate
an overall execution plan. However, this technique may lead to a suboptimal execution
plan because the view is optimized separately.
View merging can sometimes improve performance. As shown in "Example 5-2", view
merging merges the tables from the view into the outer query block, removing the
inner query block. Thus, separate optimization of the view is not necessary.

5.2.2 Simple View Merging
In simple view merging, the optimizer merges select-project-join views.
For example, a query of the employees table contains a subquery that joins the
departments and locations tables.
Simple view merging frequently results in a more optimal plan because of the
additional join orders and access paths available after the merge. A view may not be
valid for simple view merging because:
•

The view contains constructs not included in select-project-join views, including:
–

GROUP BY

–

DISTINCT

–

Outer join

–

MODEL

–

CONNECT BY

–

Set operators

–

Aggregation

•

The view appears on the right side of a semijoin or antijoin.

•

The view contains subqueries in the SELECT list.

•

The outer query block contains PL/SQL functions.

•

The view participates in an outer join, and does not meet one of the several
additional validity requirements that determine whether the view can be merged.

Example 5-2

Simple View Merging

The following query joins the hr.employees table with the dept_locs_v view, which
returns the street address for each department. dept_locs_v is a join of the departments
and locations tables.
SELECT e.first_name, e.last_name, dept_locs_v.street_address,
dept_locs_v.postal_code
FROM employees e,
( SELECT d.department_id, d.department_name,
l.street_address, l.postal_code
FROM departments d, locations l
WHERE d.location_id = l.location_id ) dept_locs_v
WHERE dept_locs_v.department_id = e.department_id
AND
e.last_name = 'Smith';

The database can execute the preceding query by joining departments and locations to
generate the rows of the view, and then joining this result to employees. Because the

5-5

Chapter 5

View Merging

query contains the view dept_locs_v, and this view contains two tables, the optimizer
must use one of the following join orders:
•

employees, dept_locs_v (departments, locations)

•

employees, dept_locs_v (locations, departments)

•

dept_locs_v (departments, locations), employees

•

dept_locs_v (locations, departments), employees

Join methods are also constrained. The index-based nested loops join is not feasible
for join orders that begin with employees because no index exists on the column from
this view. Without view merging, the optimizer generates the following execution plan:
----------------------------------------------------------------| Id | Operation
| Name
| Cost (%CPU)|
----------------------------------------------------------------| 0 | SELECT STATEMENT
|
|
7 (15)|
|* 1 | HASH JOIN
|
|
7 (15)|
| 2 | TABLE ACCESS BY INDEX ROWID| EMPLOYEES |
2 (0)|
|* 3 |
INDEX RANGE SCAN
| EMP_NAME_IX |
1 (0)|
| 4 | VIEW
|
|
5 (20)|
|* 5 |
HASH JOIN
|
|
5 (20)|
| 6 |
TABLE ACCESS FULL
| LOCATIONS |
2 (0)|
| 7 |
TABLE ACCESS FULL
| DEPARTMENTS |
2 (0)|
----------------------------------------------------------------Predicate Information (identified by operation id):
--------------------------------------------------1 - access("DEPT_LOCS_V"."DEPARTMENT_ID"="E"."DEPARTMENT_ID")
3 - access("E"."LAST_NAME"='Smith')
5 - access("D"."LOCATION_ID"="L"."LOCATION_ID")

View merging merges the tables from the view into the outer query block, removing the
inner query block. After view merging, the query is as follows:
SELECT
FROM
WHERE
AND
AND

e.first_name, e.last_name, l.street_address, l.postal_code
employees e, departments d, locations l
d.location_id = l.location_id
d.department_id = e.department_id
e.last_name = 'Smith';

Because all three tables appear in one query block, the optimizer can choose from the
following six join orders:
•

employees, departments, locations

•

employees, locations, departments

•

departments, employees, locations

•

departments, locations, employees

•

locations, employees, departments

•

locations, departments, employees

The joins to employees and departments can now be index-based. After view merging,
the optimizer chooses the following more efficient plan, which uses nested loops:
------------------------------------------------------------------| Id | Operation
| Name
| Cost (%CPU)|
-------------------------------------------------------------------

5-6

Chapter 5

View Merging

| 0 | SELECT STATEMENT
|
|
4 (0)|
| 1 | NESTED LOOPS
|
|
|
| 2 | NESTED LOOPS
|
|
4 (0)|
| 3 |
NESTED LOOPS
|
|
3 (0)|
| 4 |
TABLE ACCESS BY INDEX ROWID| EMPLOYEES |
2 (0)|
|* 5 |
INDEX RANGE SCAN
| EMP_NAME_IX |
1 (0)|
| 6 |
TABLE ACCESS BY INDEX ROWID| DEPARTMENTS |
1 (0)|
|* 7 |
INDEX UNIQUE SCAN
| DEPT_ID_PK |
0 (0)|
|* 8 |
INDEX UNIQUE SCAN
| LOC_ID_PK |
0 (0)|
| 9 | TABLE ACCESS BY INDEX ROWID | LOCATIONS |
1 (0)|
------------------------------------------------------------------Predicate Information (identified by operation id):
--------------------------------------------------5 - access("E"."LAST_NAME"='Smith')
7 - access("E"."DEPARTMENT_ID"="D"."DEPARTMENT_ID")
8 - access("D"."LOCATION_ID"="L"."LOCATION_ID")

See Also:
The Oracle Optimizer blog at https://blogs.oracle.com/optimizer/ to learn
about outer join view merging, which is a special case of simple view
merging

5.2.3 Complex View Merging
In view merging, the optimizer merges views containing GROUP BY and DISTINCT views.
Like simple view merging, complex merging enables the optimizer to consider
additional join orders and access paths.
The optimizer can delay evaluation of GROUP BY or DISTINCT operations until after it has
evaluated the joins. Delaying these operations can improve or worsen performance
depending on the data characteristics. If the joins use filters, then delaying the
operation until after joins can reduce the data set on which the operation is to be
performed. Evaluating the operation early can reduce the amount of data to be
processed by subsequent joins, or the joins could increase the amount of data to be
processed by the operation. The optimizer uses cost to evaluate view merging and
merges the view only when it is the lower cost option.
Aside from cost, the optimizer may be unable to perform complex view merging for the
following reasons:
•

The outer query tables do not have a rowid or unique column.

•

The view appears in a CONNECT BY query block.

•

The view contains GROUPING SETS, ROLLUP, or PIVOT clauses.

•

The view or outer query block contains the MODEL clause.

Example 5-3

Complex View Joins with GROUP BY

The following view uses a GROUP BY clause:
CREATE VIEW cust_prod_totals_v AS
SELECT SUM(s.quantity_sold) total, s.cust_id, s.prod_id
FROM sales s
GROUP BY s.cust_id, s.prod_id;

5-7

Chapter 5

View Merging

The following query finds all of the customers from the United States who have bought
at least 100 fur-trimmed sweaters:
SELECT
FROM
WHERE
AND
AND
AND
AND

c.cust_id, c.cust_first_name, c.cust_last_name, c.cust_email
customers c, products p, cust_prod_totals_v
c.country_id = 52790
c.cust_id = cust_prod_totals_v.cust_id
cust_prod_totals_v.total > 100
cust_prod_totals_v.prod_id = p.prod_id
p.prod_name = 'T3 Faux Fur-Trimmed Sweater';

The cust_prod_totals_v view is eligible for complex view merging. After merging, the
query is as follows:
SELECT c.cust_id, cust_first_name, cust_last_name, cust_email
FROM customers c, products p, sales s
WHERE c.country_id = 52790
AND
c.cust_id = s.cust_id
AND
s.prod_id = p.prod_id
AND
p.prod_name = 'T3 Faux Fur-Trimmed Sweater'
GROUP BY s.cust_id, s.prod_id, p.rowid, c.rowid, c.cust_email, c.cust_last_name,
c.cust_first_name, c.cust_id
HAVING SUM(s.quantity_sold) > 100;

The transformed query is cheaper than the untransformed query, so the optimizer
chooses to merge the view. In the untransformed query, the GROUP BY operator applies
to the entire sales table in the view. In the transformed query, the joins to products and
customers filter out a large portion of the rows from the sales table, so the GROUP BY
operation is lower cost. The join is more expensive because the sales table has not
been reduced, but it is not much more expensive because the GROUP BY operation does
not reduce the size of the row set very much in the original query. If any of the
preceding characteristics were to change, merging the view might no longer be lower
cost. The final plan, which does not include a view, is as follows:
-------------------------------------------------------| Id | Operation
| Name
| Cost (%CPU)|
-------------------------------------------------------| 0 | SELECT STATEMENT
|
| 2101 (18)|
|* 1 | FILTER
|
|
|
| 2 | HASH GROUP BY
|
| 2101 (18)|
|* 3 |
HASH JOIN
|
| 2099 (18)|
|* 4 |
HASH JOIN
|
| 1801 (19)|
|* 5 |
TABLE ACCESS FULL| PRODUCTS |
96 (5)|
| 6 |
TABLE ACCESS FULL| SALES
| 1620 (15)|
|* 7 |
TABLE ACCESS FULL | CUSTOMERS | 296 (11)|
-------------------------------------------------------Predicate Information (identified by operation id):
--------------------------------------------------1 - filter(SUM("QUANTITY_SOLD")>100)
3 - access("C"."CUST_ID"="CUST_ID")
4 - access("PROD_ID"="P"."PROD_ID")
5 - filter("P"."PROD_NAME"='T3 Faux Fur-Trimmed Sweater')
7 - filter("C"."COUNTRY_ID"='US')

Example 5-4

Complex View Joins with DISTINCT

The following query of the cust_prod_v view uses a DISTINCT operator:
SELECT c.cust_id, c.cust_first_name, c.cust_last_name, c.cust_email
FROM customers c, products p,
( SELECT DISTINCT s.cust_id, s.prod_id

5-8

Chapter 5

Predicate Pushing

WHERE
AND
AND
AND

FROM sales s) cust_prod_v
c.country_id = 52790
c.cust_id = cust_prod_v.cust_id
cust_prod_v.prod_id = p.prod_id
p.prod_name = 'T3 Faux Fur-Trimmed Sweater';

After determining that view merging produces a lower-cost plan, the optimizer rewrites
the query into this equivalent query:
SELECT nwvw.cust_id, nwvw.cust_first_name, nwvw.cust_last_name, nwvw.cust_email
FROM ( SELECT DISTINCT(c.rowid), p.rowid, s.prod_id, s.cust_id,
c.cust_first_name, c.cust_last_name, c.cust_email
FROM customers c, products p, sales s
WHERE c.country_id = 52790
AND
c.cust_id = s.cust_id
AND
s.prod_id = p.prod_id
AND
p.prod_name = 'T3 Faux Fur-Trimmed Sweater' ) nwvw;

The plan for the preceding query is as follows:
------------------------------------------| Id | Operation
| Name
|
------------------------------------------| 0 | SELECT STATEMENT
|
|
| 1 | VIEW
| VM_NWVW_1 |
| 2 | HASH UNIQUE
|
|
|* 3 |
HASH JOIN
|
|
|* 4 |
HASH JOIN
|
|
|* 5 |
TABLE ACCESS FULL| PRODUCTS |
| 6 |
TABLE ACCESS FULL| SALES
|
|* 7 |
TABLE ACCESS FULL | CUSTOMERS |
------------------------------------------Predicate Information (identified by operation id):
--------------------------------------------------3 - access("C"."CUST_ID"="S"."CUST_ID")
4 - access("S"."PROD_ID"="P"."PROD_ID")
5 - filter("P"."PROD_NAME"='T3 Faux Fur-Trimmed Sweater')
7 - filter("C"."COUNTRY_ID"='US')

The preceding plan contains a view named vm_nwvw_1, known as a projection view,
even after view merging has occurred. Projection views appear in queries in which a
DISTINCT view has been merged, or a GROUP BY view is merged into an outer query
block that also contains GROUP BY, HAVING, or aggregates. In the latter case, the
projection view contains the GROUP BY, HAVING, and aggregates from the original outer
query block.
In the preceding example of a projection view, when the optimizer merges the view, it
moves the DISTINCT operator to the outer query block, and then adds several additional
columns to maintain semantic equivalence with the original query. Afterward, the query
can select only the desired columns in the SELECT list of the outer query block. The
optimization retains all of the benefits of view merging: all tables are in one query
block, the optimizer can permute them as needed in the final join order, and the
DISTINCT operation has been delayed until after all of the joins complete.

5.3 Predicate Pushing
In predicate pushing, the optimizer "pushes" the relevant predicates from the
containing query block into the view query block.

5-9

Chapter 5

Subquery Unnesting

For views that are not merged, this technique improves the subplan of the unmerged
view. The database can use the pushed-in predicates to access indexes or to use as
filters.
For example, suppose you create a table hr.contract_workers as follows:
DROP TABLE contract_workers;
CREATE TABLE contract_workers AS (SELECT * FROM employees where 1=2);
INSERT INTO contract_workers VALUES (306, 'Bill', 'Jones', 'BJONES',
'555.555.2000', '07-JUN-02', 'AC_ACCOUNT', 8300, 0,205, 110);
INSERT INTO contract_workers VALUES (406, 'Jill', 'Ashworth', 'JASHWORTH',
'555.999.8181', '09-JUN-05', 'AC_ACCOUNT', 8300, 0,205, 50);
INSERT INTO contract_workers VALUES (506, 'Marcie', 'Lunsford', 'MLUNSFORD',
'555.888.2233', '22-JUL-01', 'AC_ACCOUNT', 8300, 0,205, 110);
COMMIT;
CREATE INDEX contract_workers_index ON contract_workers(department_id);

You create a view that references employees and contract_workers. The view is defined
with a query that uses the UNION set operator, as follows:
CREATE VIEW all_employees_vw AS
( SELECT employee_id, last_name, job_id, commission_pct, department_id
FROM employees )
UNION
( SELECT employee_id, last_name, job_id, commission_pct, department_id
FROM contract_workers );

You then query the view as follows:
SELECT last_name
FROM all_employees_vw
WHERE department_id = 50;

Because the view is a UNION set query, the optimizer cannot merge the view's query
into the accessing query block. Instead, the optimizer can transform the accessing
statement by pushing its predicate, the WHERE clause condition department_id=50, into
the view's UNION set query. The equivalent transformed query is as follows:
SELECT last_name
FROM ( SELECT employee_id, last_name, job_id, commission_pct, department_id
FROM employees
WHERE department_id=50
UNION
SELECT employee_id, last_name, job_id, commission_pct, department_id
FROM contract_workers
WHERE department_id=50 );

The transformed query can now consider index access in each of the query blocks.

5.4 Subquery Unnesting
In subquery unnesting, the optimizer transforms a nested query into an equivalent
join statement, and then optimizes the join.
This transformation enables the optimizer to consider the subquery tables during
access path, join method, and join order selection. The optimizer can perform this
transformation only if the resulting join statement is guaranteed to return the same
rows as the original statement, and if subqueries do not contain aggregate functions
such as AVG.

5-10

Chapter 5

Query Rewrite with Materialized Views

For example, suppose you connect as user sh and execute the following query:
SELECT *
FROM sales
WHERE cust_id IN ( SELECT cust_id
FROM customers );

Because the customers.cust_id column is a primary key, the optimizer can transform
the complex query into the following join statement that is guaranteed to return the
same data:
SELECT sales.*
FROM sales, customers
WHERE sales.cust_id = customers.cust_id;

If the optimizer cannot transform a complex statement into a join statement, it selects
execution plans for the parent statement and the subquery as though they were
separate statements. The optimizer then executes the subquery and uses the rows
returned to execute the parent query. To improve execution speed of the overall
execution plan, the optimizer orders the subplans efficiently.

5.5 Query Rewrite with Materialized Views
A materialized view is a query result that the database materializes and stores in a
table.
When the optimizer finds a user query compatible with the query associated with a
materialized view, then the database can rewrite the query in terms of the materialized
view. This technique improves query execution because the database has
precomputed most of the query result.
The optimizer looks for any materialized views that are compatible with the user query,
and then selects one or more materialized views to rewrite the user query. The use of
materialized views to rewrite a query is cost-based. Thus, the optimizer does not
rewrite the query when the plan generated unless the materialized views has a lower
cost than the plan generated with the materialized views.
Consider the following materialized view, cal_month_sales_mv, which aggregates the
dollar amount sold each month:
CREATE MATERIALIZED VIEW cal_month_sales_mv
ENABLE QUERY REWRITE
AS
SELECT t.calendar_month_desc, SUM(s.amount_sold) AS dollars
FROM sales s, times t
WHERE s.time_id = t.time_id
GROUP BY t.calendar_month_desc;

Assume that sales number is around one million in a typical month. The view has the
precomputed aggregates for the dollar amount sold for each month. Consider the
following query, which asks for the sum of the amount sold for each month:
SELECT t.calendar_month_desc, SUM(s.amount_sold)
FROM sales s, times t
WHERE s.time_id = t.time_id
GROUP BY t.calendar_month_desc;

Without query rewrite, the database must access sales directly and compute the sum
of the amount sold. This method involves reading many million rows from sales, which

5-11

Chapter 5

Star Transformation

invariably increases query response time. The join also further slows query response
because the database must compute the join on several million rows. With query
rewrite, the optimizer transparently rewrites the query as follows:
SELECT calendar_month, dollars
FROM cal_month_sales_mv;

See Also:
Oracle Database Data Warehousing Guide to learn more about query rewrite

5.6 Star Transformation
Star transformation is an optimizer transformation that avoids full table scans of fact
tables in a star schema.
This section contains the following topics:
•

About Star Schemas
A star schema divides data into facts and dimensions.

•

Purpose of Star Transformations
In joins of fact and dimension tables, a star transformation can avoid a full scan of
a fact table.

•

How Star Transformation Works
Star transformation adds subquery predicates, called bitmap semijoin
predicates, corresponding to the constraint dimensions.

•

Controls for Star Transformation
The STAR_TRANSFORMATION_ENABLED initialization parameter controls star
transformations.

•

Star Transformation: Scenario
This scenario demonstrates a star transformation of a star query.

•

Temporary Table Transformation: Scenario
In the preceding scenario, the optimizer does not join back the table channels to
the sales table because it is not referenced outside and the channel_id is unique.

5.6.1 About Star Schemas
A star schema divides data into facts and dimensions.
Facts are the measurements of an event such as a sale and are typically numbers.
Dimensions are the categories that identify facts, such as date, location, and product.
A fact table has a composite key made up of the primary keys of the dimension tables
of the schema. Dimension tables act as lookup or reference tables that enable you to
choose values that constrain your queries.
Diagrams typically show a central fact table with lines joining it to the dimension tables,
giving the appearance of a star. The following graphic shows sales as the fact table
and products, times, customers, and channels as the dimension tables.

5-12

Chapter 5

Star Transformation

Figure 5-1

Star Schema

products

times

sales
(amount_sold,
quantity_sold)
Fact Table
customers

channels

Dimension Table

Dimension Table

A snowflake schema is a star schema in which the dimension tables reference other
tables. A snowstorm schema is a combination of snowflake schemas.

See Also:
Oracle Database Data Warehousing Guide to learn more about star
schemas

5.6.2 Purpose of Star Transformations
In joins of fact and dimension tables, a star transformation can avoid a full scan of a
fact table.
The star transformation improves performance by fetching only relevant fact rows that
join to the constraint dimension rows. In some cases, queries have restrictive filters on
other columns of the dimension tables. The combination of filters can dramatically
reduce the data set that the database processes from the fact table.

5.6.3 How Star Transformation Works
Star transformation adds subquery predicates, called bitmap semijoin predicates,
corresponding to the constraint dimensions.
The optimizer performs the transformation when indexes exist on the fact join
columns. By driving bitmap AND and OR operations of key values supplied by the
subqueries, the database only needs to retrieve relevant rows from the fact table. If the
predicates on the dimension tables filter out significant data, then the transformation
can be more efficient than a full scan on the fact table.
After the database has retrieved the relevant rows from the fact table, the database
may need to join these rows back to the dimension tables using the original
predicates. The database can eliminate the join back of the dimension table when the
following conditions are met:
•

All the predicates on dimension tables are part of the semijoin subquery predicate.

•

The columns selected from the subquery are unique.

•

The dimension columns are not in the SELECT list, GROUP BY clause, and so on.

5-13

Chapter 5

Star Transformation

5.6.4 Controls for Star Transformation
The STAR_TRANSFORMATION_ENABLED initialization parameter controls star transformations.
This parameter takes the following values:
•

true

The optimizer performs the star transformation by identifying the fact and
constraint dimension tables automatically. The optimizer performs the star
transformation only if the cost of the transformed plan is lower than the
alternatives. Also, the optimizer attempts temporary table transformation
automatically whenever materialization improves performance (see "Temporary
Table Transformation: Scenario").
•

false (default)

The optimizer does not perform star transformations.
•

TEMP_DISABLE

This value is identical to true except that the optimizer does not attempt temporary
table transformation.

See Also:
Oracle Database Reference to learn about the STAR_TRANSFORMATION_ENABLED
initialization parameter

5.6.5 Star Transformation: Scenario
This scenario demonstrates a star transformation of a star query.
Example 5-5

Star Query

The following query finds the total Internet sales amount in all cities in California for
quarters Q1 and Q2 of year 1999:
SELECT c.cust_city,
t.calendar_quarter_desc,
SUM(s.amount_sold) sales_amount
FROM sales s,
times t,
customers c,
channels ch
WHERE s.time_id = t.time_id
AND
s.cust_id = c.cust_id
AND
s.channel_id = ch.channel_id
AND
c.cust_state_province = 'CA'
AND
ch.channel_desc = 'Internet'
AND
t.calendar_quarter_desc IN ('1999-01','1999-02')
GROUP BY c.cust_city, t.calendar_quarter_desc;

Sample output is as follows:
CUST_CITY
CALENDA SALES_AMOUNT
------------------------------ ------- ------------

5-14

Chapter 5

Star Transformation

Montara
Pala
Cloverdale
Cloverdale
. . .

1999-02
1999-01
1999-01
1999-02

1618.01
3263.93
52.64
266.28

In this example, sales is the fact table, and the other tables are dimension tables. The
sales table contains one row for every sale of a product, so it could conceivably
contain billions of sales records. However, only a few products are sold to customers
in California through the Internet for the specified quarters.
Example 5-6

Star Transformation

This example shows a star transformation of the query in Example 5-5. The
transformation avoids a full table scan of sales.
SELECT
FROM
WHERE
AND
AND
AND
AND

c.cust_city, t.calendar_quarter_desc, SUM(s.amount_sold) sales_amount
sales s, times t, customers c
s.time_id = t.time_id
s.cust_id = c.cust_id
c.cust_state_province = 'CA'
t.calendar_quarter_desc IN ('1999-01','1999-02')
s.time_id IN ( SELECT time_id
FROM times
WHERE calendar_quarter_desc IN('1999-01','1999-02') )
AND
s.cust_id IN ( SELECT cust_id
FROM customers
WHERE cust_state_province='CA' )
AND
s.channel_id IN ( SELECT channel_id
FROM channels
WHERE channel_desc = 'Internet' )
GROUP BY c.cust_city, t.calendar_quarter_desc;

Example 5-7

Partial Execution Plan for Star Transformation

This example shows an edited version of the execution plan for the star transformation
in Example 5-6.
Line 26 shows that the sales table has an index access path instead of a full table
scan. For each key value that results from the subqueries of channels (line 14), times
(line 19), and customers (line 24), the database retrieves a bitmap from the indexes on
the sales fact table (lines 15, 20, 25).
Each bit in the bitmap corresponds to a row in the fact table. The bit is set when the
key value from the subquery is same as the value in the row of the fact table. For
example, in the bitmap 101000... (the ellipses indicates that the values for the
remaining rows are 0), rows 1 and 3 of the fact table have matching key values from
the subquery.
The operations in lines 12, 17, and 22 iterate over the keys from the subqueries and
retrieve the corresponding bitmaps. In Example 5-6, the customers subquery seeks the
IDs of customers whose state or province is CA. Assume that the bitmap 101000...
corresponds to the customer ID key value 103515 from the customers table subquery.
Also assume that the customers subquery produces the key value 103516 with the
bitmap 010000..., which means that only row 2 in sales has a matching key value from
the subquery.
The database merges (using the OR operator) the bitmaps for each subquery (lines 11,
16, 21). In our customers example, the database produces a single bitmap 111000... for
the customers subquery after merging the two bitmaps:

5-15

Chapter 5

Star Transformation

101000...
010000...
--------111000...

# bitmap corresponding to key 103515
# bitmap corresponding to key 103516
# result of OR operation

In line 10, the database applies the AND operator to the merged bitmaps. Assume that
after the database has performed all OR operations, the resulting bitmap for channels is
100000... If the database performs an AND operation on this bitmap and the bitmap
from customers subquery, then the result is as follows:
100000...
111000...
--------100000...

# channels bitmap after all OR operations performed
# customers bitmap after all OR operations performed
# bitmap result of AND operation for channels and customers

In line 9, the database generates the corresponding rowids of the final bitmap. The
database retrieves rows from the sales fact table using the rowids (line 26). In our
example, the database generate only one rowid, which corresponds to the first row,
and thus fetches only a single row instead of scanning the entire sales table.
------------------------------------------------------------------------------| Id | Operation
| Name
------------------------------------------------------------------------------| 0 | SELECT STATEMENT
|
| 1 | HASH GROUP BY
|
|* 2 | HASH JOIN
|
|* 3 |
TABLE ACCESS FULL
| CUSTOMERS
|* 4 |
HASH JOIN
|
|* 5 |
TABLE ACCESS FULL
| TIMES
| 6 |
VIEW
| VW_ST_B1772830
| 7 |
NESTED LOOPS
|
| 8 |
PARTITION RANGE SUBQUERY
|
| 9 |
BITMAP CONVERSION TO ROWIDS|
| 10 |
BITMAP AND
|
| 11 |
BITMAP MERGE
|
| 12 |
BITMAP KEY ITERATION
|
| 13 |
BUFFER SORT
|
|* 14 |
TABLE ACCESS FULL
| CHANNELS
|* 15 |
BITMAP INDEX RANGE SCAN| SALES_CHANNEL_BIX
| 16 |
BITMAP MERGE
|
| 17 |
BITMAP KEY ITERATION
|
| 18 |
BUFFER SORT
|
|* 19 |
TABLE ACCESS FULL
| TIMES
|* 20 |
BITMAP INDEX RANGE SCAN| SALES_TIME_BIX
| 21 |
BITMAP MERGE
|
| 22 |
BITMAP KEY ITERATION
|
| 23 |
BUFFER SORT
|
|* 24 |
TABLE ACCESS FULL
| CUSTOMERS
|* 25 |
BITMAP INDEX RANGE SCAN| SALES_CUST_BIX
| 26 |
TABLE ACCESS BY USER ROWID | SALES
------------------------------------------------------------------------------Predicate Information (identified by operation id):
--------------------------------------------------2
3
4
5

-

access("ITEM_1"="C"."CUST_ID")
filter("C"."CUST_STATE_PROVINCE"='CA')
access("ITEM_2"="T"."TIME_ID")
filter(("T"."CALENDAR_QUARTER_DESC"='1999-01'
OR "T"."CALENDAR_QUARTER_DESC"='1999-02'))
14 - filter("CH"."CHANNEL_DESC"='Internet')

5-16

Chapter 5

Star Transformation

15 - access("S"."CHANNEL_ID"="CH"."CHANNEL_ID")
19 - filter(("T"."CALENDAR_QUARTER_DESC"='1999-01'
OR "T"."CALENDAR_QUARTER_DESC"='1999-02'))
20 - access("S"."TIME_ID"="T"."TIME_ID")
24 - filter("C"."CUST_STATE_PROVINCE"='CA')
25 - access("S"."CUST_ID"="C"."CUST_ID")
Note
----- star transformation used for this statement

5.6.6 Temporary Table Transformation: Scenario
In the preceding scenario, the optimizer does not join back the table channels to the
sales table because it is not referenced outside and the channel_id is unique.
If the optimizer cannot eliminate the join back, however, then the database stores the
subquery results in a temporary table to avoid rescanning the dimension table for
bitmap key generation and join back. Also, if the query runs in parallel, then the
database materializes the results so that each parallel execution server can select the
results from the temporary table instead of executing the subquery again.
Example 5-8

Star Transformation Using Temporary Table

In this example, the database materializes the results of the subquery on customers
into a temporary table:
SELECT t1.c1 cust_city, t.calendar_quarter_desc calendar_quarter_desc,
SUM(s.amount_sold) sales_amount
FROM sales s, sh.times t, sys_temp_0fd9d6621_e7e24 t1
WHERE s.time_id=t.time_id
AND
s.cust_id=t1.c0
AND
(t.calendar_quarter_desc='1999-q1' OR t.calendar_quarter_desc='1999-q2')
AND
s.cust_id IN
( SELECT t1.c0
FROM sys_temp_0fd9d6621_e7e24 t1 )
AND
s.channel_id IN ( SELECT ch.channel_id
FROM channels ch
WHERE ch.channel_desc='internet' )
AND
s.time_id IN
( SELECT t.time_id
FROM times t
WHERE t.calendar_quarter_desc='1999-q1'
OR
t.calendar_quarter_desc='1999-q2' )
GROUP BY t1.c1, t.calendar_quarter_desc

The optimizer replaces customers with the temporary table sys_temp_0fd9d6621_e7e24,
and replaces references to columns cust_id and cust_city with the corresponding
columns of the temporary table. The database creates the temporary table with two
columns: (c0 NUMBER, c1 VARCHAR2(30)). These columns correspond to cust_id and
cust_city of the customers table. The database populates the temporary table by
executing the following query at the beginning of the execution of the previous query:
SELECT c.cust_id, c.cust_city FROM customers WHERE c.cust_state_province = 'CA'

Example 5-9
Table

Partial Execution Plan for Star Transformation Using Temporary

The following example shows an edited version of the execution plan for the query in
Example 5-8:

5-17

Chapter 5

Star Transformation

------------------------------------------------------------------------------| Id | Operation
| Name
------------------------------------------------------------------------------| 0 | SELECT STATEMENT
|
| 1 | TEMP TABLE TRANSFORMATION
|
| 2 | LOAD AS SELECT
|
|* 3 |
TABLE ACCESS FULL
| CUSTOMERS
| 4 | HASH GROUP BY
|
|* 5 |
HASH JOIN
|
| 6 |
TABLE ACCESS FULL
| SYS_TEMP_0FD9D6613_C716F
|* 7 |
HASH JOIN
|
|* 8 |
TABLE ACCESS FULL
| TIMES
| 9 |
VIEW
| VW_ST_A3F94988
| 10 |
NESTED LOOPS
|
| 11 |
PARTITION RANGE SUBQUERY
|
| 12 |
BITMAP CONVERSION TO ROWIDS|
| 13 |
BITMAP AND
|
| 14 |
BITMAP MERGE
|
| 15 |
BITMAP KEY ITERATION
|
| 16 |
BUFFER SORT
|
|* 17 |
TABLE ACCESS FULL
| CHANNELS
|* 18 |
BITMAP INDEX RANGE SCAN| SALES_CHANNEL_BIX
| 19 |
BITMAP MERGE
|
| 20 |
BITMAP KEY ITERATION
|
| 21 |
BUFFER SORT
|
|* 22 |
TABLE ACCESS FULL
| TIMES
|* 23 |
BITMAP INDEX RANGE SCAN| SALES_TIME_BIX
| 24 |
BITMAP MERGE
|
| 25 |
BITMAP KEY ITERATION
|
| 26 |
BUFFER SORT
|
| 27 |
TABLE ACCESS FULL
| SYS_TEMP_0FD9D6613_C716F
|* 28 |
BITMAP INDEX RANGE SCAN| SALES_CUST_BIX
| 29 |
TABLE ACCESS BY USER ROWID | SALES
------------------------------------------------------------------------------Predicate Information (identified by operation id):
--------------------------------------------------3
5
7
8

-

17 18 22 23 28 -

filter("C"."CUST_STATE_PROVINCE"='CA')
access("ITEM_1"="C0")
access("ITEM_2"="T"."TIME_ID")
filter(("T"."CALENDAR_QUARTER_DESC"='1999-01' OR
"T"."CALENDAR_QUARTER_DESC"='1999-02'))
filter("CH"."CHANNEL_DESC"='Internet')
access("S"."CHANNEL_ID"="CH"."CHANNEL_ID")
filter(("T"."CALENDAR_QUARTER_DESC"='1999-01' OR
"T"."CALENDAR_QUARTER_DESC"='1999-02'))
access("S"."TIME_ID"="T"."TIME_ID")
access("S"."CUST_ID"="C0")

Lines 1, 2, and 3 of the plan materialize the customers subquery into the temporary
table. In line 6, the database scans the temporary table (instead of the subquery) to
build the bitmap from the fact table. Line 27 scans the temporary table for joining back
instead of scanning customers. The database does not need to apply the filter on
customers on the temporary table because the filter is applied while materializing the
temporary table.

5-18

Chapter 5

In-Memory Aggregation (VECTOR GROUP BY)

5.7 In-Memory Aggregation (VECTOR GROUP BY)
The key optimization of in-memory aggregation is to aggregate while scanning.
To optimize query blocks involving aggregation and joins from a single large table to
multiple small tables, such as in a typical star query, the transformation uses KEY
VECTOR and VECTOR GROUP BY operations. These operations use efficient in-memory
arrays for joins and aggregation, and are especially effective when the underlying
tables are in-memory columnar tables.

See Also:
Oracle Database In-Memory Guide to learn more about in-memory
aggregation

5.8 Cursor-Duration Temporary Tables
To materialize the intermediate results of a query, Oracle Database may implicitly
create a cursor-duration temporary table in memory during query compilation.
This section contains the following topics:
•

Purpose of Cursor-Duration Temporary Tables
Complex queries sometimes process the same query block multiple times, which
creates unnecessary performance overhead.

•

How Cursor-Duration Temporary Tables Work
The definition of the cursor-definition temporary table resides in memory. The table
definition is associated with the cursor, and is only visible to the session executing
the cursor.

•

Cursor-Duration Temporary Tables: Example
A WITH query that repeats the same subquery can sometimes benefit from a
cursor-duration temporary table.

5.8.1 Purpose of Cursor-Duration Temporary Tables
Complex queries sometimes process the same query block multiple times, which
creates unnecessary performance overhead.
To avoid this scenario, Oracle Database can automatically create temporary tables for
the query results and store them in memory for the duration of the cursor. For complex
operations such as WITH clause queries, star transformations, and grouping sets, this
optimization enhances the materialization of intermediate results from repetitively used
subqueries. In this way, cursor-duration temporary tables improve performance and
optimize I/O.

5-19

Chapter 5

Cursor-Duration Temporary Tables

5.8.2 How Cursor-Duration Temporary Tables Work
The definition of the cursor-definition temporary table resides in memory. The table
definition is associated with the cursor, and is only visible to the session executing the
cursor.
When using cursor-duration temporary tables, the database performs the following
steps:
1.

Chooses a plan that uses a cursor-duration temporary table

2.

Creates the temporary table using a unique name

3.

Rewrites the query to refer to the temporary table

4.

Loads data into memory until no memory remains, in which case it creates
temporary segments on disk

5.

Executes the query, returning data from the temporary table

6.

Truncates the table, releasing memory and any on-disk temporary segments

Note:
The metadata for the cursor-duration temporary table stays in memory as
long as the cursor is in memory. The metadata is not stored in the data
dictionary, which means it is not visible through data dictionary views. You
cannot drop the metadata explicitly.

The preceding scenario depends on the availability of memory. For serial queries, the
temporary tables use PGA memory.
The implementation of cursor-duration temporary tables is similar to sorts. If no more
memory is available, then the database writes data to temporary segments. For
cursor-duration temporary tables, the differences are as follows:
•

The database releases memory and temporary segments at the end of the query
rather than when the row source is no longer active.

•

Data in memory stays in memory, unlike in sorts where data can move between
memory and temporary segments.

When the database uses cursor-duration temporary tables, the keyword CURSOR
DURATION MEMORY appears in the execution plan.

5.8.3 Cursor-Duration Temporary Tables: Example
A WITH query that repeats the same subquery can sometimes benefit from a cursorduration temporary table.
The following query uses a WITH clause to create three subquery blocks:
WITH
q1 AS (SELECT department_id, SUM(salary) sum_sal FROM hr.employees GROUP BY
department_id),
q2 AS (SELECT * FROM q1),

5-20

Chapter 5

Table Expansion

q3 AS (SELECT department_id, sum_sal FROM q1)
SELECT * FROM q1
UNION ALL
SELECT * FROM q2
UNION ALL
SELECT * FROM q3;

The following sample plan shows the transformation:
SELECT * FROM TABLE(DBMS_XPLAN.DISPLAY_CURSOR(FORMAT=>'BASIC +ROWS +COST'));
PLAN_TABLE_OUTPUT
-------------------------------------------------------------------------------------------| Id | Operation
| Name
|Rows |Cost (%CPU)|
-------------------------------------------------------------------------------------------| 0 | SELECT STATEMENT
|
|
|6 (100)|
| 1 | TEMP TABLE TRANSFORMATION
|
|
|
|
| 2 | LOAD AS SELECT (CURSOR DURATION MEMORY) | SYS_TEMP_0FD9D6606_1AE004 |
|
|
| 3 |
HASH GROUP BY
|
| 11 | 3 (34)|
| 4 |
TABLE ACCESS FULL
| EMPLOYEES
| 107 | 2 (0) |
| 5 | UNION-ALL
|
|
|
|
| 6 |
VIEW
|
| 11 | 2 (0) |
| 7 |
TABLE ACCESS FULL
| SYS_TEMP_0FD9D6606_1AE004 | 11 | 2 (0) |
| 8 |
VIEW
|
| 11 | 2 (0) |
| 9 |
TABLE ACCESS FULL
| SYS_TEMP_0FD9D6606_1AE004 | 11 | 2 (0) |
| 10 |
VIEW
|
| 11 | 2 (0) |
| 11 |
TABLE ACCESS FULL
| SYS_TEMP_0FD9D6606_1AE004 | 11 | 2 (0) |
--------------------------------------------------------------------------------------------

In the preceding plan, TEMP TABLE TRANSFORMATION in Step 1 indicates that the database
used cursor-duration temporary tables to execute the query. The CURSOR DURATION
MEMORY keyword in Step 2 indicates that the database used memory, if available, to
store the results of SYS_TEMP_0FD9D6606_1AE004. If memory was unavailable, then the
database wrote the temporary data to disk.

5.9 Table Expansion
In table expansion, the optimizer generates a plan that uses indexes on the readmostly portion of a partitioned table, but not on the active portion of the table.
This section contains the following topics:
•

Purpose of Table Expansion
Index-based plans can improve performance, but index maintenance creates
overhead. In many databases, DML affects only a small portion of the data.

•

How Table Expansion Works
Table partitioning makes table expansion possible.

•

Table Expansion: Scenario
The optimizer keeps track of which partitions must be accessed from each table,
based on predicates that appear in the query. Partition pruning enables the
optimizer to use table expansion to generate more optimal plans.

•

Table Expansion and Star Transformation: Scenario
Star transformation enables specific types of queries to avoid accessing large
portions of big fact tables.

5-21

Chapter 5

Table Expansion

5.9.1 Purpose of Table Expansion
Index-based plans can improve performance, but index maintenance creates
overhead. In many databases, DML affects only a small portion of the data.
Table expansion uses index-based plans for high-update tables. You can create an
index only on the read-mostly data, eliminating index overhead on the active data. In
this way, table expansion improves performance while avoiding index maintenance.

5.9.2 How Table Expansion Works
Table partitioning makes table expansion possible.
If a local index exists on a partitioned table, then the optimizer can mark the index as
unusable for specific partitions. In effect, some partitions are not indexed.
In table expansion, the optimizer transforms the query into a UNION ALL statement, with
some subqueries accessing indexed partitions and other subqueries accessing
unindexed partitions. The optimizer can choose the most efficient access method
available for a partition, regardless of whether it exists for all of the partitions accessed
in the query.
The optimizer does not always choose table expansion:
•

Table expansion is cost-based.
While the database accesses each partition of the expanded table only once
across all branches of the UNION ALL, any tables that the database joins to it are
accessed in each branch.

•

Semantic issues may render expansion invalid.
For example, a table appearing on the right side of an outer join is not valid for
table expansion.

You can control table expansion with the hint EXPAND_TABLE hint. The hint overrides the
cost-based decision, but not the semantic checks.

See Also:
•

"Influencing the Optimizer with Hints"

•

Oracle Database SQL Language Reference to learn more about SQL
hints

5.9.3 Table Expansion: Scenario
The optimizer keeps track of which partitions must be accessed from each table,
based on predicates that appear in the query. Partition pruning enables the optimizer
to use table expansion to generate more optimal plans.
Assumptions
This scenario assumes the following:

5-22

Chapter 5

Table Expansion

•

You want to run a star query against the sh.sales table, which is range-partitioned
on the time_id column.

•

You want to disable indexes on specific partitions to see the benefits of table
expansion.

To use table expansion:
1.

Log in to the database as the sh user.

2.

Run the following query:
SELECT
FROM
WHERE
AND

3.

*
sales
time_id >= TO_DATE('2000-01-01 00:00:00', 'SYYYY-MM-DD HH24:MI:SS')
prod_id = 38;

Explain the plan by querying DBMS_XPLAN:
SET LINESIZE 150
SET PAGESIZE 0
SELECT * FROM TABLE(DBMS_XPLAN.DISPLAY_CURSOR(format => 'BASIC,PARTITION'));

As shown in the Pstart and Pstop columns in the following plan, the optimizer
determines from the filter that only 16 of the 28 partitions in the table must be
accessed:
Plan hash value: 3087065703
-------------------------------------------------------------------------|Id| Operation
| Name
|Pstart|Pstop|
-------------------------------------------------------------------------| 0| SELECT STATEMENT
|
|
|
|
| 1| PARTITION RANGE ITERATOR
|
| 13 | 28 |
| 2| TABLE ACCESS BY LOCAL INDEX ROWID BATCHED| SALES
| 13 | 28 |
| 3|
BITMAP CONVERSION TO ROWIDS
|
|
|
|
|*4|
BITMAP INDEX SINGLE VALUE
|SALES_PROD_BIX| 13 | 28 |
-------------------------------------------------------------------------Predicate Information (identified by operation id):
--------------------------------------------------4 - access("PROD_ID"=38)

After the optimizer has determined the partitions to be accessed, it considers any
index that is usable on all of those partitions. In the preceding plan, the optimizer
chose to use the sales_prod_bix bitmap index.
4.

Disable the index on the SALES_1995 partition of the sales table:
ALTER INDEX sales_prod_bix MODIFY PARTITION sales_1995 UNUSABLE;

The preceding DDL disables the index on partition 1, which contains all sales from
before 1996.

Note:
You can obtain the partition information by querying the
USER_IND_PARTITIONS view.

5-23

Chapter 5

Table Expansion

5.

Execute the query of sales again, and then query DBMS_XPLAN to obtain the plan.
The output shows that the plan did not change:
Plan hash value: 3087065703
--------------------------------------------------------------------------|Id| Operation
| Name
|Pstart|Pstop
--------------------------------------------------------------------------| 0| SELECT STATEMENT
|
|
|
|
| 1| PARTITION RANGE ITERATOR
|
| 13 | 28 |
| 2| TABLE ACCESS BY LOCAL INDEX ROWID BATCHED| SALES
| 13 | 28 |
| 3|
BITMAP CONVERSION TO ROWIDS
|
|
|
|
|*4|
BITMAP INDEX SINGLE VALUE
| SALES_PROD_BIX| 13 | 28 |
--------------------------------------------------------------------------Predicate Information (identified by operation id):
--------------------------------------------------4 - access("PROD_ID"=38)

The plan is the same because the disabled index partition is not relevant to the
query. If all partitions that the query accesses are indexed, then the database can
answer the query using the index. Because the query only accesses partitions 16
through 28, disabling the index on partition 1 does not affect the plan.
6.

Disable the indexes for partition 28 (SALES_Q4_2003), which is a partition that the
query needs to access:
ALTER INDEX sales_prod_bix MODIFY PARTITION sales_q4_2003 UNUSABLE;
ALTER INDEX sales_time_bix MODIFY PARTITION sales_q4_2003 UNUSABLE;

By disabling the indexes on a partition that the query does need to access, the
query can no longer use this index (without table expansion).
7.

Query the plan using DBMS_XPLAN.
As shown in the following plan, the optimizer does not use the index:
Plan hash value: 3087065703
--------------------------------------------------------------------------| Id| Operation
| Name
|Pstart|Pstop
--------------------------------------------------------------------------| 0 | SELECT STATEMENT
|
|
|
|
| 1 | PARTITION RANGE ITERATOR
|
| 13 | 28 |
|*2 | TABLE ACCESS FULL
| SALES
| 13 | 28 |
--------------------------------------------------------------------------Predicate Information (identified by operation id):
--------------------------------------------------2 - access("PROD_ID"=38)

In the preceding example, the query accesses 16 partitions. On 15 of these
partitions, an index is available, but no index is available for the final partition.
Because the optimizer has to choose one access path or the other, the optimizer
cannot use the index on any of the partitions.
8.

With table expansion, the optimizer rewrites the original query as follows:
SELECT *
FROM sales
WHERE time_id >= TO_DATE('2000-01-01 00:00:00', 'SYYYY-MM-DD HH24:MI:SS')

5-24

Chapter 5

Table Expansion

AND
time_id
AND
prod_id
UNION ALL
SELECT *
FROM sales
WHERE time_id
AND
time_id
AND
prod_id

< TO_DATE('2003-10-01 00:00:00', 'SYYYY-MM-DD HH24:MI:SS')
= 38

>= TO_DATE('2003-10-01 00:00:00', 'SYYYY-MM-DD HH24:MI:SS')
< TO_DATE('2004-01-01 00:00:00', 'SYYYY-MM-DD HH24:MI:SS')
= 38;

In the preceding query, the first query block in the UNION ALL accesses the
partitions that are indexed, while the second query block accesses the partition
that is not. The two subqueries enable the optimizer to choose to use the index in
the first query block, if it is more optimal than using a table scan of all of the
partitions that are accessed.
9.

Query the plan using DBMS_XPLAN.
The plan appears as follows:
Plan hash value: 2120767686
--------------------------------------------------------------------------|Id| Operation
| Name
|Pstart|Pstop|
--------------------------------------------------------------------------| 0| SELECT STATEMENT
|
| | |
| 1| VIEW
| VW_TE_2
| | |
| 2| UNION-ALL
|
| | |
| 3|
PARTITION RANGE ITERATOR
|
| 13| 27|
| 4|
TABLE ACCESS BY LOCAL INDEX ROWID BATCHED| SALES
| 13| 27|
| 5|
BITMAP CONVERSION TO ROWIDS
|
| | |
|*6|
BITMAP INDEX SINGLE VALUE
| SALES_PROD_BIX| 13| 27|
| 7|
PARTITION RANGE SINGLE
|
| 28| 28|
|*8|
TABLE ACCESS FULL
| SALES
| 28| 28|
--------------------------------------------------------------------------Predicate Information (identified by operation id):
--------------------------------------------------6 - access("PROD_ID"=38)
8 - filter("PROD_ID"=38)

As shown in the preceding plan, the optimizer uses a UNION ALL for two query
blocks (Step 2). The optimizer chooses an index to access partitions 13 to 27 in
the first query block (Step 6). Because no index is available for partition 28, the
optimizer chooses a full table scan in the second query block (Step 8).

5.9.4 Table Expansion and Star Transformation: Scenario
Star transformation enables specific types of queries to avoid accessing large portions
of big fact tables.
Star transformation requires defining several indexes, which in an actively updated
table can have overhead. With table expansion, you can define indexes on only the
inactive partitions so that the optimizer can consider star transformation on only the
indexed portions of the table.
Assumptions
This scenario assumes the following:

5-25

Chapter 5

Table Expansion

•

You query the same schema used in "Star Transformation: Scenario".

•

The last partition of sales is actively being updated, as is often the case with timepartitioned tables.

•

You want the optimizer to take advantage of table expansion.

To take advantage of table expansion in a star query:
1.

Disable the indexes on the last partition as follows:
ALTER INDEX sales_channel_bix MODIFY PARTITION sales_q4_2003 UNUSABLE;
ALTER INDEX sales_cust_bix MODIFY PARTITION sales_q4_2003 UNUSABLE;

2.

Execute the following star query:
SELECT t.calendar_quarter_desc, SUM(s.amount_sold) sales_amount
FROM sales s, times t, customers c, channels ch
WHERE s.time_id = t.time_id
AND
s.cust_id = c.cust_id
AND
s.channel_id = ch.channel_id
AND
c.cust_state_province = 'CA'
AND
ch.channel_desc = 'Internet'
AND
t.calendar_quarter_desc IN ('1999-01','1999-02')
GROUP BY t.calendar_quarter_desc;

3.

Query the cursor using DBMS_XPLAN, which shows the following plan:
--------------------------------------------------------------------------|Id| Operation
| Name
| Pstart| Pstop |
--------------------------------------------------------------------------| 0| SELECT STATEMENT
|
|
|
|
| 1| HASH GROUP BY
|
|
|
|
| 2| VIEW
|VW_TE_14
|
|
|
| 3|
UNION-ALL
|
|
|
|
| 4|
HASH JOIN
|
|
|
|
| 5|
TABLE ACCESS FULL
|TIMES
|
|
|
| 6|
VIEW
|VW_ST_1319B6D8 |
|
|
| 7|
NESTED LOOPS
|
|
|
|
| 8|
PARTITION RANGE SUBQUERY
|
|KEY(SQ)|KEY(SQ)|
| 9|
BITMAP CONVERSION TO ROWIDS|
|
|
|
|10|
BITMAP AND
|
|
|
|
|11|
BITMAP MERGE
|
|
|
|
|12|
BITMAP KEY ITERATION
|
|
|
|
|13|
BUFFER SORT
|
|
|
|
|14|
TABLE ACCESS FULL
|CHANNELS
|
|
|
|15|
BITMAP INDEX RANGE SCAN|SALES_CHANNEL_BIX|KEY(SQ)|KEY(SQ)|
|16|
BITMAP MERGE
|
|
|
|
|17|
BITMAP KEY ITERATION
|
|
|
|
|18|
BUFFER SORT
|
|
|
|
|19|
TABLE ACCESS FULL
|TIMES
|
|
|
|20|
BITMAP INDEX RANGE SCAN|SALES_TIME_BIX |KEY(SQ)|KEY(SQ)|
|21|
BITMAP MERGE
|
|
|
|
|22|
BITMAP KEY ITERATION
|
|
|
|
|23|
BUFFER SORT
|
|
|
|
|24|
TABLE ACCESS FULL
|CUSTOMERS
|
|
|
|25|
BITMAP INDEX RANGE SCAN|SALES_CUST_BIX |KEY(SQ)|KEY(SQ)|
|26|
TABLE ACCESS BY USER ROWID |SALES
| ROWID | ROWID |
|27|
NESTED LOOPS
|
|
|
|
|28|
NESTED LOOPS
|
|
|
|
|29|
NESTED LOOPS
|
|
|
|
|30|
NESTED LOOPS
|
|
|
|
|31|
PARTITION RANGE SINGLE
|
|
28 |
28 |

5-26

Chapter 5

Join Factorization

|32|
TABLE ACCESS FULL
|SALES
|
28 |
28 |
|33|
TABLE ACCESS BY INDEX ROWID|CHANNELS
|
|
|
|34|
INDEX UNIQUE SCAN
|CHANNELS_PK
|
|
|
|35|
TABLE ACCESS BY INDEX ROWID |CUSTOMERS
|
|
|
|36|
INDEX UNIQUE SCAN
|CUSTOMERS_PK
|
|
|
|37|
INDEX UNIQUE SCAN
|TIMES_PK
|
|
|
|38|
TABLE ACCESS BY INDEX ROWID |TIMES
|
|
|
---------------------------------------------------------------------------

The preceding plan uses table expansion. The UNION ALL branch that is accessing
every partition except the last partition uses star transformation. Because the
indexes on partition 28 are disabled, the database accesses the final partition
using a full table scan.
Related Topics
•

Star Transformation
Star transformation is an optimizer transformation that avoids full table scans of
fact tables in a star schema.

5.10 Join Factorization
In the cost-based transformation known as join factorization, the optimizer can
factorize common computations from branches of a UNION ALL query.
This section contains the following topics:
•

Purpose of Join Factorization
UNION ALL queries are common in database applications, especially in data

integration applications.
•

How Join Factorization Works
Join factorization can factorize multiple tables and from more than two UNION ALL
branches.

•

Factorization and Join Orders: Scenario
Join factorization can create more possibilities for join orders

•

Factorization of Outer Joins: Scenario
The database supports join factorization of outer joins, antijoins, and semijoins, but
only for the right tables in such joins.

5.10.1 Purpose of Join Factorization
UNION ALL queries are common in database applications, especially in data integration

applications.
Often, branches in a UNION ALL query refer to the same base tables. Without join
factorization, the optimizer evaluates each branch of a UNION ALL query independently,
which leads to repetitive processing, including data access and joins. Join factorization
transformation can share common computations across the UNION ALL branches.
Avoiding an extra scan of a large base table can lead to a huge performance
improvement.

5.10.2 How Join Factorization Works
Join factorization can factorize multiple tables and from more than two UNION ALL
branches.

5-27

Chapter 5

Join Factorization

Join factorization is best explained through examples.
Example 5-10

UNION ALL Query

The following query shows a query of four tables (t1, t2, t3, and t4) and two UNION ALL
branches:
SELECT t1.c1, t2.c2
FROM
t1, t2, t3
WHERE t1.c1 = t2.c1
AND
t1.c1 > 1
AND
t2.c2 = 2
AND
t2.c2 = t3.c2
UNION ALL
SELECT t1.c1, t2.c2
t1, t2, t4
FROM
WHERE t1.c1 = t2.c1
t1.c1 > 1
AND
AND
t2.c3 = t4.c3

In the preceding query, table t1 appears in both UNION ALL branches, as does the filter
predicate t1.c1 > 1 and the join predicate t1.c1 = t2.c1. Without any transformation,
the database must perform the scan and the filtering on table t1 twice, one time for
each branch.
Example 5-11

Factorized Query

Example 5-10
SELECT t1.c1, VW_JF_1.item_2
FROM
t1, (SELECT t2.c1 item_1, t2.c2 item_2
FROM t2, t3
WHERE t2.c2 = t3.c2
AND
t2.c2 = 2
UNION ALL
SELECT t2.c1 item_1, t2.c2 item_2
FROM t2, t4
WHERE t2.c3 = t4.c3) VW_JF_1
WHERE t1.c1 = VW_JF_1.item_1
AND
t1.c1 > 1

In this case, because table t1 is factorized, the database performs the table scan and
the filtering on t1 only one time. If t1 is large, then this factorization avoids the huge
performance cost of scanning and filtering t1 twice.

Note:
If the branches in a UNION ALL query have clauses that use the DISTINCT
function, then join factorization is not valid.

5.10.3 Factorization and Join Orders: Scenario
Join factorization can create more possibilities for join orders
Example 5-12

Query Involving Five Tables

In the following query, view V is same as the query as in Example 5-10:

5-28

Chapter 5

Join Factorization

SELECT *
FROM t5, (SELECT t1.c1, t2.c2
FROM t1, t2, t3
WHERE t1.c1 = t2.c1
AND
t1.c1 > 1
AND
t2.c2 = 2
AND
t2.c2 = t3.c2
UNION ALL
SELECT t1.c1, t2.c2
FROM t1, t2, t4
WHERE t1.c1 = t2.c1
AND
t1.c1 > 1
AND
t2.c3 = t4.c3) V
WHERE t5.c1 = V.c1
t1t2t3t5

Example 5-13

Factorization of t1 from View V

If join factorization factorizes t1 from view V, as shown in the following query, then the
database can join t1 with t5.:
SELECT *
t5, ( SELECT t1.c1, VW_JF_1.item_2
FROM
FROM
t1, (SELECT t2.c1 item_1, t2.c2 item_2
FROM t2, t3
WHERE t2.c2 = t3.c2
AND
t2.c2 = 2
UNION ALL
SELECT t2.c1 item_1, t2.c2 item_2
FROM t2, t4
WHERE t2.c3 = t4.c3) VW_JF_1
WHERE t1.c1 = VW_JF_1.item_1
AND
t1.c1 > 1 )
WHERE t5.c1 = V.c1

The preceding query transformation opens up new join orders. However, join
factorization imposes specific join orders. For example, in the preceding query, tables
t2 and t3 appear in the first branch of the UNION ALL query in view VW_JF_1. The
database must join t2 with t3 before it can join with t1, which is not defined within the
VW_JF_1 view. The imposed join order may not necessarily be the best join order. For
this reason, the optimizer performs join factorization using the cost-based
transformation framework. The optimizer calculates the cost of the plans with and
without join factorization, and then chooses the cheapest plan.
Example 5-14

Factorization of t1 from View V with View Definition Removed

The following query is the same query in Example 5-13, but with the view definition
removed so that the factorization is easier to see:
SELECT *
FROM
t5, (SELECT t1.c1, VW_JF_1.item_2
FROM
t1, VW_JF_1
WHERE t1.c1 = VW_JF_1.item_1
AND
t1.c1 > 1)
WHERE t5.c1 = V.c1

5-29

Chapter 5

Join Factorization

5.10.4 Factorization of Outer Joins: Scenario
The database supports join factorization of outer joins, antijoins, and semijoins, but
only for the right tables in such joins.
For example, join factorization can transform the following UNION ALL query by
factorizing t2:
SELECT t1.c2, t2.c2
FROM t1, t2
WHERE t1.c1 = t2.c1(+)
AND
t1.c1 = 1
UNION ALL
SELECT t1.c2, t2.c2
FROM t1, t2
WHERE t1.c1 = t2.c1(+)
AND
t1.c1 = 2

The following example shows the transformation. Table t2 now no longer appears in
the UNION ALL branches of the subquery.
SELECT VW_JF_1.item_2, t2.c2
FROM t2, (SELECT t1.c1 item_1, t1.c2 item_2
FROM t1
WHERE t1.c1 = 1
UNION ALL
SELECT t1.c1 item_1, t1.c2 item_2
FROM t1
WHERE t1.c1 = 2) VW_JF_1
WHERE VW_JF_1.item_1 = t2.c1(+)

5-30

Part III
Query Execution Plans
If a query has suboptimal performance, the execution plan is the key tool for
understanding the problem and supplying a solution.
This part contains the following chapters:
•

Generating and Displaying Execution Plans
A thorough understanding of execution plans is essential to SQL tuning.

•

Reading Execution Plans
Execution plans are represented as a tree of operations.

6
Generating and Displaying Execution Plans
A thorough understanding of execution plans is essential to SQL tuning.
This chapter contains the following topics:
•

Introduction to Execution Plans
The combination of the steps that Oracle Database uses to execute a statement is
an execution plan.

•

About Plan Generation and Display
The EXPLAIN PLAN statement displays execution plans that the optimizer chooses
for SELECT, UPDATE, INSERT, and DELETE statements.

•

Generating Execution Plans
The EXPLAIN PLAN statement enables you to examine the execution plan that the
optimizer chose for a SQL statement.

•

Displaying PLAN_TABLE Output
You can use scripts or a package to display the plan output.

6.1 Introduction to Execution Plans
The combination of the steps that Oracle Database uses to execute a statement is an
execution plan.
Each step either retrieves rows of data physically from the database or prepares them
for the user issuing the statement. An execution plan includes an access path for each
table that the statement accesses and an ordering of the tables (the join order) with
the appropriate join method.
Related Topics
•

Joins
Oracle Database provides several optimizations for joining row sets.

6.2 About Plan Generation and Display
The EXPLAIN PLAN statement displays execution plans that the optimizer chooses for
SELECT, UPDATE, INSERT, and DELETE statements.
This section contains the following topics:
•

About the Plan Explanation
A statement execution plan is the sequence of operations that the database
performs to run the statement.

•

Why Execution Plans Change
Execution plans can and do change as the underlying optimizer inputs change.

•

Guideline for Minimizing Throw-Away
Examining an explain plan enables you to look for rows that are thrown-away.

6-1

Chapter 6

About Plan Generation and Display

•

Guidelines for Evaluating Execution Plans Using EXPLAIN PLAN
The execution plan operation alone cannot differentiate between well-tuned
statements and those that perform suboptimally.

•

Guidelines for Evaluating Plans Using the V$SQL_PLAN Views
As an alternative to running the EXPLAIN PLAN command and displaying the plan,
you can display the plan by querying the V$SQL_PLAN view.

•

EXPLAIN PLAN Restrictions
Oracle Database does not support EXPLAIN PLAN for statements performing implicit
type conversion of date bind variables.

•

Guidelines for Creating PLAN_TABLE
The PLAN_TABLE is automatically created as a public synonym to a global temporary
table.

6.2.1 About the Plan Explanation
A statement execution plan is the sequence of operations that the database performs
to run the statement.
The row source tree is the core of the execution plan. The tree shows the following
information:
•

An ordering of the tables referenced by the statement

•

An access method for each table mentioned in the statement

•

A join method for tables affected by join operations in the statement

•

Data operations like filter, sort, or aggregation

In addition to the row source tree, the plan table contains information about the
following:
•

Optimization, such as the cost and cardinality of each operation

•

Partitioning, such as the set of accessed partitions

•

Parallel execution, such as the distribution method of join inputs

You can use the EXPLAIN PLAN results to determine whether the optimizer chose a
particular execution plan, such as a nested loops join. The results also help you to
understand the optimizer decisions, such as why the optimizer chose a nested loops
join instead of a hash join.

See Also:
•

"SQL Row Source Generation"

•

Oracle Database SQL Language Reference to learn about the EXPLAIN
PLAN statement

6.2.2 Why Execution Plans Change
Execution plans can and do change as the underlying optimizer inputs change.

6-2

Chapter 6

About Plan Generation and Display

EXPLAIN PLAN output shows how the database would run the SQL statement when the

statement was explained. This plan can differ from the actual execution plan a SQL
statement uses because of differences in the execution environment and explain plan
environment.

Note:
To avoid possible SQL performance regression that may result from
execution plan changes, consider using SQL plan management.

This section contains the following topics:
•

Different Schemas
Schemas can differ for various reasons.

•

Different Costs
Even if the schemas are the same, the optimizer can choose different execution
plans when the costs are different.

See Also:
•

"Overview of SQL Plan Management"

•

Oracle Database PL/SQL Packages and Types Reference to learn about
the DBMS_SPM package

6.2.2.1 Different Schemas
Schemas can differ for various reasons.
Principal reasons include the following:
•

The execution and explain plan occur on different databases.

•

The user explaining the statement is different from the user running the statement.
Two users might be pointing to different objects in the same database, resulting in
different execution plans.

•

Schema changes (often changes in indexes) between the two operations.

6.2.2.2 Different Costs
Even if the schemas are the same, the optimizer can choose different execution plans
when the costs are different.
Some factors that affect the costs include the following:
•

Data volume and statistics

•

Bind variable types and values

•

Initialization parameters set globally or at session level

6-3

Chapter 6

About Plan Generation and Display

6.2.3 Guideline for Minimizing Throw-Away
Examining an explain plan enables you to look for rows that are thrown-away.
The database often throws away rows in the following situations:
•

Full scans

•

Unselective range scans

•

Late predicate filters

•

Wrong join order

•

Late filter operations

In the plan shown in Example 6-1, the last step is a very unselective range scan that is
executed 76,563 times, accesses 11,432,983 rows, throws away 99% of them, and
retains 76,563 rows. Why access 11,432,983 rows to realize that only 76,563 rows are
needed?
Example 6-1

Looking for Thrown-Away Rows in an Explain Plan

Rows
Execution Plan
-------- ---------------------------------------------------12 SORT AGGREGATE
2 SORT GROUP BY
76563
NESTED LOOPS
76575
NESTED LOOPS
19
TABLE ACCESS FULL CN_PAYRUNS_ALL
76570
TABLE ACCESS BY INDEX ROWID CN_POSTING_DETAILS_ALL
76570
INDEX RANGE SCAN (object id 178321)
76563
TABLE ACCESS BY INDEX ROWID CN_PAYMENT_WORKSHEETS_ALL
11432983
INDEX RANGE SCAN (object id 186024)

6.2.4 Guidelines for Evaluating Execution Plans Using EXPLAIN PLAN
The execution plan operation alone cannot differentiate between well-tuned
statements and those that perform suboptimally.
For example, an EXPLAIN PLAN output that shows that a statement uses an index does
not necessarily mean that the statement runs efficiently. Sometimes indexes are
extremely inefficient. In this case, a good practice is to examine the following:
•

The columns of the index being used

•

Their selectivity (fraction of table being accessed)

It is best to use EXPLAIN PLAN to determine an access plan, and then later prove that it
is the optimal plan through testing. When evaluating a plan, examine the statement's
actual resource consumption.
This section contains the following topics:

6-4

Chapter 6

About Plan Generation and Display

6.2.5 Guidelines for Evaluating Plans Using the V$SQL_PLAN Views
As an alternative to running the EXPLAIN PLAN command and displaying the plan, you
can display the plan by querying the V$SQL_PLAN view.
V$SQL_PLAN contains the execution plan for every statement stored in the shared SQL
area. Its definition is similar to PLAN_TABLE.

The advantage of V$SQL_PLAN over EXPLAIN PLAN is that you do not need to know the
compilation environment that was used to execute a particular statement. For EXPLAIN
PLAN, you would need to set up an identical environment to get the same plan when
executing the statement.
The V$SQL_PLAN_STATISTICS view provides the actual execution statistics for every
operation in the plan, such as the number of output rows and elapsed time. All
statistics, except the number of output rows, are cumulative. For example, the
statistics for a join operation also includes the statistics for its two inputs. The statistics
in V$SQL_PLAN_STATISTICS are available for cursors that have been compiled with the
STATISTICS_LEVEL initialization parameter set to ALL.
The V$SQL_PLAN_STATISTICS_ALL view enables side by side comparisons of the
estimates that the optimizer provides for the number of rows and elapsed time. This
view combines both V$SQL_PLAN and V$SQL_PLAN_STATISTICS information for every
cursor.

See Also:
•
•

"PLAN_TABLE Columns"
"Monitoring Database Operations " for information about the
V$SQL_PLAN_MONITOR view

•

Oracle Database Reference for more information about V$SQL_PLAN views

•

Oracle Database Reference for information about the STATISTICS_LEVEL
initialization parameter

6.2.6 EXPLAIN PLAN Restrictions
Oracle Database does not support EXPLAIN PLAN for statements performing implicit type
conversion of date bind variables.
With bind variables in general, the EXPLAIN PLAN output might not represent the real
execution plan.
From the text of a SQL statement, TKPROF cannot determine the types of the bind
variables. It assumes that the type is VARCHAR, and gives an error message otherwise.
You can avoid this limitation by putting appropriate type conversions in the SQL
statement.

6-5

Chapter 6

Generating Execution Plans

See Also:
•

"Performing Application Tracing "

•

"Guideline for Avoiding the Argument Trap"

•

Oracle Database SQL Language Reference to learn more about SQL
data types

6.2.7 Guidelines for Creating PLAN_TABLE
The PLAN_TABLE is automatically created as a public synonym to a global temporary
table.
This temporary table holds the output of EXPLAIN PLAN statements for all users.
PLAN_TABLE is the default sample output table into which the EXPLAIN PLAN statement
inserts rows describing execution plans.
While a PLAN_TABLE table is automatically set up for each user, you can use the SQL
script catplan.sql to manually create the global temporary table and the PLAN_TABLE
synonym. The name and location of this script depends on your operating system. On
UNIX and Linux, the script is located in the $ORACLE_HOME/rdbms/admin directory.
For example, start a SQL*Plus session, connect with SYSDBA privileges, and run the
script as follows:
@$ORACLE_HOME/rdbms/admin/catplan.sql

Oracle recommends that you drop and rebuild your local PLAN_TABLE table after
upgrading the version of the database because the columns might change. This can
cause scripts to fail or cause TKPROF to fail, if you are specifying the table.
If you do not want to use the name PLAN_TABLE, create a new synonym after running the
catplan.sql script. For example:
CREATE OR REPLACE PUBLIC SYNONYM my_plan_table for plan_table$

See Also:
•

"PLAN_TABLE Columns" for a description of the columns in the table

•

Oracle Database SQL Language Reference to learn about CREATE
SYNONYM

6.3 Generating Execution Plans
The EXPLAIN PLAN statement enables you to examine the execution plan that the
optimizer chose for a SQL statement.
When the statement is issued, the optimizer chooses an execution plan and then
inserts data describing the plan into a database table. Issue the EXPLAIN PLAN
statement and then query the output table.

6-6

Chapter 6

Generating Execution Plans

This section contains the following topics:
•

Executing EXPLAIN PLAN for a Single Statement
Explain the plan using database-supplied scripts.

•

Executing EXPLAIN PLAN Using a Statement ID
With multiple statements, you can specify a statement identifier and use that to
identify your specific execution plan.

•

Directing EXPLAIN PLAN Output to a Nondefault Table
You can specify the INTO clause to specify a different table.

6.3.1 Executing EXPLAIN PLAN for a Single Statement
Explain the plan using database-supplied scripts.
The basics of using the EXPLAIN PLAN statement are as follows:
•

Use the SQL script catplan.sql to create a sample output table called PLAN_TABLE
in your schema.

•

Include the EXPLAIN PLAN FOR clause before the SQL statement.

•

After issuing the EXPLAIN PLAN statement, use a script or package provided by
Oracle Database to display the most recent plan table output.

•

The execution order in EXPLAIN PLAN output begins with the line that is the furthest
indented to the right. The next step is the parent of that line. If two lines are
indented equally, then the top line is normally executed first.

Note:
–

The EXPLAIN PLAN output tables in this chapter were displayed with
the utlxpls.sql script.

–

The steps in the EXPLAIN PLAN output in this chapter may be different
on your database. The optimizer may choose different execution
plans, depending on database configurations.

To explain a SQL statement, use the EXPLAIN PLAN FOR clause immediately before the
statement. For example:
EXPLAIN PLAN FOR
SELECT last_name FROM employees;

The preceding plan explains the plan and stores the output in the PLAN_TABLE table.
You can then select the execution plan from PLAN_TABLE.

6-7

Chapter 6

Displaying PLAN_TABLE Output

See Also:
•

"Guidelines for Creating PLAN_TABLE"

•

"Displaying PLAN_TABLE Output"

•

Oracle Database SQL Language Reference for the syntax and
semantics of EXPLAIN PLAN

6.3.2 Executing EXPLAIN PLAN Using a Statement ID
With multiple statements, you can specify a statement identifier and use that to identify
your specific execution plan.
Before using SET STATEMENT ID, remove any existing rows for that statement ID. In the
following example, st1 is specified as the statement identifier.
Example 6-2

Using EXPLAIN PLAN with the STATEMENT ID Clause

EXPLAIN PLAN
SET STATEMENT_ID = 'st1' FOR
SELECT last_name FROM employees;

6.3.3 Directing EXPLAIN PLAN Output to a Nondefault Table
You can specify the INTO clause to specify a different table.
The following statement directs output to my_plan_table:
EXPLAIN PLAN
INTO my_plan_table FOR
SELECT last_name FROM employees;

You can specify a statement ID when using the INTO clause, as in the following
statement:
EXPLAIN PLAN
SET STATEMENT_ID = 'st1'
INTO my_plan_table FOR
SELECT last_name FROM employees;

See Also:
Oracle Database SQL Language Reference for a complete description of
EXPLAIN PLAN syntax.

6.4 Displaying PLAN_TABLE Output
You can use scripts or a package to display the plan output.
After you have explained the plan, use the following SQL scripts or PL/SQL package
provided by Oracle Database to display the most recent plan table output:

6-8

Chapter 6

Displaying PLAN_TABLE Output

•

utlxpls.sql

This script displays the plan table output for serial processing. Example 6-4 is an
example of the plan table output when using the utlxpls.sql script.
•

utlxplp.sql

This script displays the plan table output including parallel execution columns.
•

DBMS_XPLAN.DISPLAY table function

This function accepts options for displaying the plan table output. You can specify:
–

A plan table name if you are using a table different than PLAN_TABLE

–

A statement ID if you have set a statement ID with the EXPLAIN PLAN

–

A format option that determines the level of detail: BASIC, SERIAL, TYPICAL, and
ALL

Examples of using DBMS_XPLAN to display PLAN_TABLE output are:
SELECT PLAN_TABLE_OUTPUT FROM TABLE(DBMS_XPLAN.DISPLAY());
SELECT PLAN_TABLE_OUTPUT
FROM TABLE(DBMS_XPLAN.DISPLAY('MY_PLAN_TABLE', 'st1','TYPICAL'));

This section contains the following topics:

See Also:
Oracle Database PL/SQL Packages and Types Reference for more
information about the DBMS_XPLAN package
•

Displaying an Execution Plan: Example
This example uses EXPLAIN PLAN to examine a SQL statement that selects the
employee_id, job_title, salary, and department_name for the employees whose IDs
are less than 103.

•

Customizing PLAN_TABLE Output
If you have specified a statement identifier, then you can write your own script to
query the PLAN_TABLE.

6.4.1 Displaying an Execution Plan: Example
This example uses EXPLAIN PLAN to examine a SQL statement that selects the
employee_id, job_title, salary, and department_name for the employees whose IDs are
less than 103.
Example 6-3

Using EXPLAIN PLAN

EXPLAIN PLAN FOR
SELECT e.employee_id, j.job_title, e.salary, d.department_name
FROM employees e, jobs j, departments d
WHERE e.employee_id < 103
AND
e.job_id = j.job_id
AND
e.department_id = d.department_id;

6-9

Chapter 6

Displaying PLAN_TABLE Output

Example 6-4

EXPLAIN PLAN Output
The following output table shows the execution plan that the optimizer chose to
execute the SQL statement in Example 6-3:

----------------------------------------------------------------------------------| Id | Operation
| Name
| Rows | Bytes | Cost (%CPU)|
----------------------------------------------------------------------------------| 0 | SELECT STATEMENT
|
|
3 | 189 |
10 (10)|
| 1 | NESTED LOOPS
|
|
3 | 189 |
10 (10)|
| 2 | NESTED LOOPS
|
|
3 | 141 |
7 (15)|
|* 3 |
TABLE ACCESS FULL
| EMPLOYEES
|
3 |
60 |
4 (25)|
| 4 |
TABLE ACCESS BY INDEX ROWID| JOBS
|
19 | 513 |
2 (50)|
|* 5 |
INDEX UNIQUE SCAN
| JOB_ID_PK
|
1 |
|
|
| 6 | TABLE ACCESS BY INDEX ROWID | DEPARTMENTS |
27 | 432 |
2 (50)|
|* 7 |
INDEX UNIQUE SCAN
| DEPT_ID_PK |
1 |
|
|
----------------------------------------------------------------------------------Predicate Information (identified by operation id):
--------------------------------------------------3 - filter("E"."EMPLOYEE_ID"<103)
5 - access("E"."JOB_ID"="J"."JOB_ID")
7 - access("E"."DEPARTMENT_ID"="D"."DEPARTMENT_ID"
-----------------------------------------------------------------------------------------------| Id | Operation
| Name
| Rows | Bytes | Cost (%CPU)| Time
|
-----------------------------------------------------------------------------------------------| 0 | SELECT STATEMENT
|
|
3 | 189 |
8 (13)| 00:00:01 |
| 1 | NESTED LOOPS
|
|
|
|
|
|
| 2 | NESTED LOOPS
|
|
3 | 189 |
8 (13)| 00:00:01 |
| 3 |
MERGE JOIN
|
|
3 | 141 |
5 (20)| 00:00:01 |
| 4 |
TABLE ACCESS BY INDEX ROWID | JOBS
|
19 | 513 |
2 (0)| 00:00:01 |
| 5 |
INDEX FULL SCAN
| JOB_ID_PK
|
19 |
|
1 (0)| 00:00:01 |
|* 6 |
SORT JOIN
|
|
3 |
60 |
3 (34)| 00:00:01 |
| 7 |
TABLE ACCESS BY INDEX ROWID| EMPLOYEES
|
3 |
60 |
2 (0)| 00:00:01 |
|* 8 |
INDEX RANGE SCAN
| EMP_EMP_ID_PK |
3 |
|
1 (0)| 00:00:01 |
|* 9 |
INDEX UNIQUE SCAN
| DEPT_ID_PK
|
1 |
|
0 (0)| 00:00:01 |
| 10 | TABLE ACCESS BY INDEX ROWID | DEPARTMENTS |
1 |
16 |
1 (0)| 00:00:01 |
-----------------------------------------------------------------------------------------------Predicate Information (identified by operation id):
--------------------------------------------------6 - access("E"."JOB_ID"="J"."JOB_ID")
filter("E"."JOB_ID"="J"."JOB_ID")
8 - access("E"."EMPLOYEE_ID"<103)
9 - access("E"."DEPARTMENT_ID"="D"."DEPARTMENT_ID")

6.4.2 Customizing PLAN_TABLE Output
If you have specified a statement identifier, then you can write your own script to query
the PLAN_TABLE.
For example:
•

Start with ID = 0 and given STATEMENT_ID.

•

Use the CONNECT BY clause to walk the tree from parent to child, the join keys being
STATEMENT_ID = PRIOR STATMENT_ID and PARENT_ID = PRIOR ID.

•

Use the pseudo-column LEVEL (associated with CONNECT BY) to indent the children.

6-10

Chapter 6

Displaying PLAN_TABLE Output

SELECT cardinality "Rows",
lpad(' ',level-1)||operation||' '||options||' '||object_name "Plan"
FROM PLAN_TABLE
CONNECT BY prior id = parent_id
AND prior statement_id = statement_id
START WITH id = 0
AND statement_id = 'st1'
ORDER BY id;
Rows Plan
------- ---------------------------------------SELECT STATEMENT
TABLE ACCESS FULL EMPLOYEES

The NULL in the Rows column indicates that the optimizer does not have any
statistics on the table. Analyzing the table shows the following:
Rows
------16957
16957

Plan
---------------------------------------SELECT STATEMENT
TABLE ACCESS FULL EMPLOYEES

You can also select the COST. This is useful for comparing execution plans or for
understanding why the optimizer chooses one execution plan over another.

Note:
These simplified examples are not valid for recursive SQL.

6-11

7
Reading Execution Plans
Execution plans are represented as a tree of operations.
This chapter contains the following topics:
•

Reading Execution Plans: Basic
This section uses EXPLAIN PLAN examples to illustrate execution plans.

•

Reading Execution Plans: Advanced
In some cases, execution plans can be complicated and challenging to read.

•

Execution Plan Reference
This section describes V$ views and PLAN_COLUMN columns.

7.1 Reading Execution Plans: Basic
This section uses EXPLAIN PLAN examples to illustrate execution plans.
The following query displays the execution plans:
SELECT PLAN_TABLE_OUTPUT
FROM TABLE(DBMS_XPLAN.DISPLAY(NULL, 'statement_id','BASIC'));

Examples of the output from this statement are shown in Example 7-4 and
Example 7-1.
Example 7-1

EXPLAIN PLAN for Statement ID ex_plan1

The following plan shows execution of a SELECT statement. The table employees is
accessed using a full table scan. Every row in the table employees is accessed, and the
WHERE clause criteria is evaluated for every row.
EXPLAIN PLAN
SET statement_id = 'ex_plan1' FOR
SELECT phone_number
FROM employees
WHERE phone_number LIKE '650%';
--------------------------------------| Id | Operation
| Name
|
--------------------------------------| 0 | SELECT STATEMENT |
|
| 1 | TABLE ACCESS FULL| EMPLOYEES |
---------------------------------------

Example 7-2

EXPLAIN PLAN for Statement ID ex_plan2

This following plan shows the execution of a SELECT statement. In this example, the
database range scans the EMP_NAME_IX index to evaluate the WHERE clause criteria.
EXPLAIN PLAN
SET statement_id = 'ex_plan2' FOR
SELECT last_name
FROM employees

7-1

Chapter 7

Reading Execution Plans: Advanced

WHERE last_name LIKE 'Pe%';
SELECT PLAN_TABLE_OUTPUT
FROM TABLE(DBMS_XPLAN.DISPLAY(NULL, 'ex_plan2','BASIC'));
---------------------------------------| Id | Operation
| Name
|
---------------------------------------| 0 | SELECT STATEMENT |
|
| 1 | INDEX RANGE SCAN| EMP_NAME_IX |
----------------------------------------

7.2 Reading Execution Plans: Advanced
In some cases, execution plans can be complicated and challenging to read.
This section contains the following topics:
•

Reading Adaptive Query Plans
The adaptive optimizer is a feature of the optimizer that enables it to adapt plans
based on run-time statistics. All adaptive mechanisms can execute a final plan for
a statement that differs from the default plan.

•

Viewing Parallel Execution with EXPLAIN PLAN
Plans for parallel queries differ in important ways from plans for serial queries.

•

Viewing Bitmap Indexes with EXPLAIN PLAN
Index row sources using bitmap indexes appear in the EXPLAIN PLAN output with the
word BITMAP indicating the type of the index.

•

Viewing Result Cache with EXPLAIN PLAN
When your query contains the result_cache hint, the ResultCache operator is
inserted into the execution plan.

•

Viewing Partitioned Objects with EXPLAIN PLAN
Use EXPLAIN PLAN to determine how Oracle Database accesses partitioned objects
for specific queries.

•

PLAN_TABLE Columns
The PLAN_TABLE used by the EXPLAIN PLAN statement contains the columns listed in
this topic.

7.2.1 Reading Adaptive Query Plans
The adaptive optimizer is a feature of the optimizer that enables it to adapt plans
based on run-time statistics. All adaptive mechanisms can execute a final plan for a
statement that differs from the default plan.
An adaptive query plan chooses among subplans during the current statement
execution. In contrast, automatic reoptimization changes a plan only on executions
that occur after the current statement execution.
You can determine whether the database used adaptive query optimization for a SQL
statement based on the comments in the Notes section of plan. The comments indicate
whether row sources are dynamic, or whether automatic reoptimization adapted a
plan.

7-2

Chapter 7

Reading Execution Plans: Advanced

Assumptions
This tutorial assumes the following:
•

The STATISTICS_LEVEL initialization parameter is set to ALL.

•

The database uses the default settings for adaptive execution.

•

As user oe, you want to issue the following separate queries:
SELECT o.order_id, v.product_name
FROM orders o,
( SELECT order_id, product_name
FROM order_items o, product_information p
WHERE p.product_id = o.product_id
AND
list_price < 50
AND
min_price < 40 ) v
WHERE o.order_id = v.order_id
SELECT
FROM
WHERE
AND
AND

product_name
order_items o, product_information p
o.unit_price = 15
quantity > 1
p.product_id = o.product_id

•

Before executing each query, you want to query DBMS_XPLAN.DISPLAY_PLAN to see
the default plan, that is, the plan that the optimizer chose before applying its
adaptive mechanism.

•

After executing each query, you want to query DBMS_XPLAN.DISPLAY_CURSOR to see
the final plan and adaptive query plan.

•

SYS has granted oe the following privileges:

–

GRANT SELECT ON V_$SESSION TO oe

–

GRANT SELECT ON V_$SQL TO oe

–

GRANT SELECT ON V_$SQL_PLAN TO oe

–

GRANT SELECT ON V_$SQL_PLAN_STATISTICS_ALL TO oe

To see the results of adaptive optimization:
1.

Start SQL*Plus, and then connect to the database as user oe.

2.

Query orders.
For example, use the following statement:
SELECT o.order_id, v.product_name
FROM orders o,
( SELECT order_id, product_name
FROM order_items o, product_information p
WHERE p.product_id = o.product_id
AND
list_price < 50
AND
min_price < 40 ) v
WHERE o.order_id = v.order_id;

3.

View the plan in the cursor.
For example, run the following commands:

7-3

Chapter 7

Reading Execution Plans: Advanced

SET LINESIZE 165
SET PAGESIZE 0
SELECT * FROM TABLE(DBMS_XPLAN.DISPLAY_CURSOR(FORMAT=>'+ALLSTATS'));

The following sample output has been reformatted to fit on the page. In this plan,
the optimizer chooses a nested loops join. The original optimizer estimates are
shown in the E-Rows column, whereas the actual statistics gathered during
execution are shown in the A-Rows column. In the MERGE JOIN operation, the
difference between the estimated and actual number of rows is significant.
-------------------------------------------------------------------------------------------|Id| Operation
| Name
|Start|E-Rows|A-Rows|A-Time|Buff|OMem|1Mem|O/1/M|
-------------------------------------------------------------------------------------------| 0| SELECT STATEMENT
|
| 1| | 269|00:00:00.09|1338|
|
|
|
| 1| NESTED LOOPS
|
| 1| 1| 269|00:00:00.09|1338|
|
|
|
| 2| MERGE JOIN CARTESIAN|
| 1| 4|9135|00:00:00.03| 33|
|
|
|
|*3|
TABLE ACCESS FULL |PRODUCT_INFORMAT| 1| 1| 87|00:00:00.01| 32|
|
|
|
| 4|
BUFFER SORT
|
| 87|105|9135|00:00:00.01| 1|4096|4096|1/0/0|
| 5|
INDEX FULL SCAN | ORDER_PK
| 1|105| 105|00:00:00.01| 1|
|
|
|
|*6| INDEX UNIQUE SCAN | ORDER_ITEMS_UK |9135| 1| 269|00:00:00.03|1305|
|
|
|
-------------------------------------------------------------------------------------------Predicate Information (identified by operation id):
--------------------------------------------------3 - filter(("MIN_PRICE"<40 AND "LIST_PRICE"<50))
6 - access("O"."ORDER_ID"="ORDER_ID" AND "P"."PRODUCT_ID"="O"."PRODUCT_ID")
4.

Run the same query of orders that you ran in Step 2.

5.

View the execution plan in the cursor by using the same SELECT statement that you
ran in Step 3.
The following example shows that the optimizer has chosen a different plan, using
a hash join. The Note section shows that the optimizer used statistics feedback to
adjust its cost estimates for the second execution of the query, thus illustrating
automatic reoptimization.

-------------------------------------------------------------------------------------------|Id| Operation
|Name
|Start|E-Rows|A-Rows|A-Time|Buff|Reads|OMem|1Mem|O/1/M|
-------------------------------------------------------------------------------------------| 0| SELECT STATEMENT
|
| 1 | |269|00:00:00.02|60|1|
|
|
|
| 1| NESTED LOOPS
|
| 1 |269|269|00:00:00.02|60|1|
|
|
|
|*2|
HASH JOIN
|
| 1 |313|269|00:00:00.02|39|1|1000K|1000K|1/0/0|
|*3|
TABLE ACCESS FULL |PRODUCT_INFORMA| 1 | 87| 87|00:00:00.01|15|0|
|
|
|
| 4|
INDEX FAST FULL SCAN|ORDER_ITEMS_UK | 1 |665|665|00:00:00.01|24|1|
|
|
|
|*5| INDEX UNIQUE SCAN
|ORDER_PK
|269| 1|269|00:00:00.01|21|0|
|
|
|
-------------------------------------------------------------------------------------------Predicate Information (identified by operation id):
--------------------------------------------------2 - access("P"."PRODUCT_ID"="O"."PRODUCT_ID")
3 - filter(("MIN_PRICE"<40 AND "LIST_PRICE"<50))
5 - access("O"."ORDER_ID"="ORDER_ID")
Note
----- statistics feedback used for this statement
6.

Query V$SQL to verify the performance improvement.

7-4

Chapter 7

Reading Execution Plans: Advanced

The following query shows the performance of the two statements (sample output
included).
SELECT CHILD_NUMBER, CPU_TIME, ELAPSED_TIME, BUFFER_GETS
FROM V$SQL
WHERE SQL_ID = 'gm2npz344xqn8';
CHILD_NUMBER CPU_TIME ELAPSED_TIME BUFFER_GETS
------------ ---------- ------------ ----------0
92006
131485
1831
1
12000
24156
60

The second statement executed, which is child number 1, used statistics feedback.
CPU time, elapsed time, and buffer gets are all significantly lower.
7.

Explain the plan for the query of order_items.
For example, use the following statement:
EXPLAIN PLAN FOR
SELECT product_name
FROM order_items o, product_information p
WHERE o.unit_price = 15
AND
quantity > 1
AND
p.product_id = o.product_id

8.

View the plan in the plan table.
For example, run the following statement:
SELECT * FROM TABLE(DBMS_XPLAN.DISPLAY);

Sample output appears below:
------------------------------------------------------------------------------|Id| Operation
| Name
|Rows|Bytes|Cost (%CPU)|Time|
------------------------------------------------------------------------------| 0| SELECT STATEMENT
|
|4|128|7 (0)|00:00:01|
| 1| NESTED LOOPS
|
| | |
|
|
| 2| NESTED LOOPS
|
|4|128|7 (0)|00:00:01|
|*3|
TABLE ACCESS FULL
|ORDER_ITEMS
|4|48 |3 (0)|00:00:01|
|*4|
INDEX UNIQUE SCAN
|PRODUCT_INFORMATION_PK|1| |0 (0)|00:00:01|
| 5| TABLE ACCESS BY INDEX ROWID|PRODUCT_INFORMATION |1|20 |1 (0)|00:00:01|
------------------------------------------------------------------------------Predicate Information (identified by operation id):
--------------------------------------------------3 - filter("O"."UNIT_PRICE"=15 AND "QUANTITY">1)
4 - access("P"."PRODUCT_ID"="O"."PRODUCT_ID")

In this plan, the optimizer chooses a nested loops join.
9.

Run the query that you previously explained.
For example, use the following statement:
SELECT
FROM
WHERE
AND
AND

product_name
order_items o, product_information p
o.unit_price = 15
quantity > 1
p.product_id = o.product_id

10. View the plan in the cursor.

7-5

Chapter 7

Reading Execution Plans: Advanced

For example, run the following commands:
SET LINESIZE 165
SET PAGESIZE 0
SELECT * FROM TABLE(DBMS_XPLAN.DISPLAY(FORMAT=>'+ADAPTIVE'));

Sample output appears below. Based on statistics collected at run time (Step 4),
the optimizer chose a hash join rather than the nested loops join. The dashes (-)
indicate the steps in the nested loops plan that the optimizer considered but do not
ultimately choose. The switch illustrates the adaptive query plan feature.
------------------------------------------------------------------------------|Id | Operation
| Name
|Rows|Bytes|Cost(%CPU)|Time
|
------------------------------------------------------------------------------| 0| SELECT STATEMENT
|
|4|128|7(0)|00:00:01|
| *1| HASH JOIN
|
|4|128|7(0)|00:00:01|
|- 2| NESTED LOOPS
|
| | |
|
|
|- 3|
NESTED LOOPS
|
| |128|7(0)|00:00:01|
|- 4|
STATISTICS COLLECTOR
|
| | |
|
|
| *5|
TABLE ACCESS FULL
| ORDER_ITEMS
|4| 48|3(0)|00:00:01|
|-*6|
INDEX UNIQUE SCAN
| PRODUCT_INFORMATI_PK|1| |0(0)|00:00:01|
|- 7|
TABLE ACCESS BY INDEX ROWID| PRODUCT_INFORMATION |1| 20|1(0)|00:00:01|
| 8| TABLE ACCESS FULL
| PRODUCT_INFORMATION |1| 20|1(0)|00:00:01|
------------------------------------------------------------------------------Predicate Information (identified by operation id):
--------------------------------------------------1 - access("P"."PRODUCT_ID"="O"."PRODUCT_ID")
5 - filter("O"."UNIT_PRICE"=15 AND "QUANTITY">1)
6 - access("P"."PRODUCT_ID"="O"."PRODUCT_ID")
Note
----- this is an adaptive plan (rows marked '-' are inactive)

See Also:
•

"Adaptive Query Plans"

•

"Table 7-8"

•

"Controlling Adaptive Optimization"

•

Oracle Database Reference to learn about the STATISTICS_LEVEL
initialization parameter

•

Oracle Database PL/SQL Packages and Types Reference to learn more
about DBMS_XPLAN

7.2.2 Viewing Parallel Execution with EXPLAIN PLAN
Plans for parallel queries differ in important ways from plans for serial queries.
This section contains the following topics:
•

About EXPLAIN PLAN and Parallel Queries
Tuning a parallel query begins much like a non-parallel query tuning exercise by
choosing the driving table. However, the rules governing the choice are different.

7-6

Chapter 7

Reading Execution Plans: Advanced

•

Viewing Parallel Queries with EXPLAIN PLAN: Example
When using EXPLAIN PLAN with parallel queries, the database compiles and
executes one parallel plan. This plan is derived from the serial plan by allocating
row sources specific to the parallel support in the QC plan.

7.2.2.1 About EXPLAIN PLAN and Parallel Queries
Tuning a parallel query begins much like a non-parallel query tuning exercise by
choosing the driving table. However, the rules governing the choice are different.
In the serial case, the best driving table produces the fewest numbers of rows after
applying limiting conditions. The database joins a small number of rows to larger
tables using non-unique indexes.
For example, consider a table hierarchy consisting of customer, account, and
transaction.

Figure 7-1

A Table Hierarchy
TRANSACTION
ACCOUNT

CUSTOMER

In this example, customer is the smallest table, whereas transaction is the largest
table. A typical OLTP query retrieves transaction information about a specific customer
account. The query drives from the customer table. The goal is to minimize logical I/O,
which typically minimizes other critical resources including physical I/O and CPU time.
For parallel queries, the driving table is usually the largest table. It would not be
efficient to use parallel query in this case because only a few rows from each table are
accessed. However, what if it were necessary to identify all customers who had
transactions of a certain type last month? It would be more efficient to drive from the
transaction table because no limiting conditions exist on the customer table. The
database would join rows from the transaction table to the account table, and then
finally join the result set to the customer table. In this case, the used on the account and
customer table are probably highly selective primary key or unique indexes rather than
the non-unique indexes used in the first query. Because the transaction table is large
and the column is not selective, it would be beneficial to use parallel query driving from
the transaction table.
Parallel operations include the following:
•

PARALLEL_TO_PARALLEL

•

PARALLEL_TO_SERIAL

A PARALLEL_TO_SERIAL operation is always the step that occurs when the query
coordinator consumes rows from a parallel operation. Another type of operation
that does not occur in this query is a SERIAL operation. If these types of operations
occur, then consider making them parallel operations to improve performance
because they too are potential bottlenecks.

7-7

Chapter 7

Reading Execution Plans: Advanced

•

PARALLEL_FROM_SERIAL

•

PARALLEL_TO_PARALLEL

If the workloads in each step are relatively equivalent, then the
PARALLEL_TO_PARALLEL operations generally produce the best performance.
•

PARALLEL_COMBINED_WITH_CHILD

•

PARALLEL_COMBINED_WITH_PARENT

A PARALLEL_COMBINED_WITH_PARENT operation occurs when the database performs
the step simultaneously with the parent step.
If a parallel step produces many rows, then the QC may not be able to consume the
rows as fast as they are produced. Little can be done to improve this situation.

See Also:
The OTHER_TAG column in "PLAN_TABLE Columns"

7.2.2.2 Viewing Parallel Queries with EXPLAIN PLAN: Example
When using EXPLAIN PLAN with parallel queries, the database compiles and executes
one parallel plan. This plan is derived from the serial plan by allocating row sources
specific to the parallel support in the QC plan.
The table queue row sources (PX Send and PX Receive), the granule iterator, and buffer
sorts, required by the two parallel execution server set PQ model, are directly inserted
into the parallel plan. This plan is the same plan for all parallel execution servers when
executed in parallel or for the QC when executed serially.
Example 7-3

Parallel Query Explain Plan
The following simple example illustrates an EXPLAIN PLAN for a parallel query:
CREATE TABLE emp2 AS SELECT * FROM employees;
ALTER TABLE emp2 PARALLEL 2;
EXPLAIN PLAN FOR
SELECT SUM(salary)
FROM emp2
GROUP BY department_id;
SELECT PLAN_TABLE_OUTPUT FROM TABLE(DBMS_XPLAN.DISPLAY());

-----------------------------------------------------------------------------------------------| Id | Operation
| Name
| Rows| Bytes |Cost %CPU|
TQ |IN-OUT| PQ Distrib |
-----------------------------------------------------------------------------------------------| 0 | SELECT STATEMENT
|
| 107 | 2782 | 3 (34) |
|
|
|
| 1 | PX COORDINATOR
|
|
|
|
|
|
|
|
| 2 | PX SEND QC (RANDOM)
| :TQ10001 | 107 | 2782 | 3 (34) | Q1,01 | P->S | QC (RAND) |
| 3 |
HASH GROUP BY
|
| 107 | 2782 | 3 (34) | Q1,01 | PCWP |
|
| 4 |
PX RECEIVE
|
| 107 | 2782 | 3 (34) | Q1,01 | PCWP |
|
| 5 |
PX SEND HASH
| :TQ10000 | 107 | 2782 | 3 (34) | Q1,00 | P->P | HASH
|
| 6 |
HASH GROUP BY
|
| 107 | 2782 | 3 (34) | Q1,00 | PCWP |
|
| 7 |
PX BLOCK ITERATOR |
| 107 | 2782 | 2 (0) | Q1,00 | PCWP |
|

7-8

Chapter 7

Reading Execution Plans: Advanced

| 8 |
TABLE ACCESS FULL| EMP2
| 107 | 2782 | 2 (0) | Q1,00 | PCWP |
|
------------------------------------------------------------------------------------------------

One set of parallel execution servers scans EMP2 in parallel, while the second set
performs the aggregation for the GROUP BY operation. The PX BLOCK ITERATOR row source
represents the splitting up of the table EMP2 into pieces to divide the scan workload
between the parallel execution servers. The PX SEND and PX RECEIVE row sources
represent the pipe that connects the two sets of parallel execution servers as rows
flow up from the parallel scan, get repartitioned through the HASH table queue, and then
read by and aggregated on the top set. The PX SEND QC row source represents the
aggregated values being sent to the QC in random (RAND) order. The PX COORDINATOR
row source represents the QC or Query Coordinator which controls and schedules the
parallel plan appearing below it in the plan tree.

7.2.3 Viewing Bitmap Indexes with EXPLAIN PLAN
Index row sources using bitmap indexes appear in the EXPLAIN PLAN output with the
word BITMAP indicating the type of the index.

Note:
Queries using bitmap join index indicate the bitmap join index access path.
The operation for bitmap join index is the same as bitmap index.

Example 7-4

EXPLAIN PLAN with Bitmap Indexes

In this example, the predicate c1=2 yields a bitmap from which a subtraction can take
place. From this bitmap, the bits in the bitmap for c2=6 are subtracted. Also, the bits in
the bitmap for c2 IS NULL are subtracted, explaining why there are two MINUS row
sources in the plan. The NULL subtraction is necessary for semantic correctness unless
the column has a NOT NULL constraint. The TO ROWIDS option generates the rowids
necessary for the table access.
EXPLAIN
FROM
WHERE
AND
OR

PLAN FOR SELECT *
t
c1 = 2
c2 <> 6
c3 BETWEEN 10 AND 20;

SELECT STATEMENT
TABLE ACCESS T BY INDEX ROWID
BITMAP CONVERSION TO ROWID
BITMAP OR
BITMAP MINUS
BITMAP MINUS
BITMAP INDEX C1_IND SINGLE VALUE
BITMAP INDEX C2_IND SINGLE VALUE
BITMAP INDEX C2_IND SINGLE VALUE
BITMAP MERGE
BITMAP INDEX C3_IND RANGE SCAN

7-9

Chapter 7

Reading Execution Plans: Advanced

7.2.4 Viewing Result Cache with EXPLAIN PLAN
When your query contains the result_cache hint, the ResultCache operator is inserted
into the execution plan.
For example, consider the following query:
SELECT /*+ result_cache */ deptno, avg(sal)
FROM emp
GROUP BY deptno;

To view the EXPLAIN PLAN for this query, use the following command:
EXPLAIN PLAN FOR
SELECT /*+ result_cache */ deptno, avg(sal)
FROM emp
GROUP BY deptno;
SELECT PLAN_TABLE_OUTPUT FROM TABLE (DBMS_XPLAN.DISPLAY());

The EXPLAIN PLAN output for this query should look similar to the following:
-------------------------------------------------------------------------------|Id| Operation
| Name
|Rows|Bytes|Cost(%CPU)|Time |
-------------------------------------------------------------------------------|0| SELECT STATEMENT
|
| 11 | 77 | 4 (25)| 00:00:01|
|1| RESULT CACHE
|b06ppfz9pxzstbttpbqyqnfbmy|
|
|
|
|
|2| HASH GROUP BY
|
| 11 | 77 | 4 (25)| 00:00:01|
|3|
TABLE ACCESS FULL| EMP
|107 | 749| 3 (0) | 00:00:01|
--------------------------------------------------------------------------------

In this EXPLAIN PLAN, the ResultCache operator is identified by its CacheId, which is
b06ppfz9pxzstbttpbqyqnfbmy. You can now run a query on the V$RESULT_CACHE_OBJECTS
view by using this CacheId.

7.2.5 Viewing Partitioned Objects with EXPLAIN PLAN
Use EXPLAIN PLAN to determine how Oracle Database accesses partitioned objects for
specific queries.
Partitions accessed after pruning are shown in the PARTITION START and PARTITION STOP
columns. The row source name for the range partition is PARTITION RANGE. For hash
partitions, the row source name is PARTITION HASH.
A join is implemented using partial partition-wise join if the DISTRIBUTION column of the
plan table of one of the joined tables contains PARTITION(KEY). Partial partition-wise join
is possible if one of the joined tables is partitioned on its join column and the table is
parallelized.
A join is implemented using full partition-wise join if the partition row source appears
before the join row source in the EXPLAIN PLAN output. Full partition-wise joins are
possible only if both joined tables are equipartitioned on their respective join columns.
Examples of execution plans for several types of partitioning follow.
This section contains the following topics:

7-10

Chapter 7

Reading Execution Plans: Advanced

•

Displaying Range and Hash Partitioning with EXPLAIN PLAN: Examples
This example illustrates pruning by using the emp_range table, which partitioned by
range on hire_date.

•

Pruning Information with Composite Partitioned Objects: Examples
To illustrate how Oracle Database displays pruning information for composite
partitioned objects, consider the table emp_comp. It is range-partitioned on hiredate
and subpartitioned by hash on deptno.

•

Examples of Partial Partition-Wise Joins
In these examples, the PQ_DISTRIBUTE hint explicitly forces a partial partition-wise
join because the query optimizer could have chosen a different plan based on cost
in this query.

•

Example of Full Partition-Wise Join
In this example, emp_comp and dept_hash are joined on their hash partitioning
columns, enabling use of a full partition-wise join.

•

Examples of INLIST ITERATOR and EXPLAIN PLAN
An INLIST ITERATOR operation appears in the EXPLAIN PLAN output if an index
implements an IN-list predicate.

•

Example of Domain Indexes and EXPLAIN PLAN
You can use EXPLAIN PLAN to derive user-defined CPU and I/O costs for domain
indexes.

7.2.5.1 Displaying Range and Hash Partitioning with EXPLAIN PLAN:
Examples
This example illustrates pruning by using the emp_range table, which partitioned by
range on hire_date.
Assume that the tables employees and departments from the Oracle Database sample
schema exist.
CREATE TABLE emp_range
PARTITION BY RANGE(hire_date)
(
PARTITION emp_p1 VALUES LESS
PARTITION emp_p2 VALUES LESS
PARTITION emp_p3 VALUES LESS
PARTITION emp_p4 VALUES LESS
PARTITION emp_p5 VALUES LESS
)
AS SELECT * FROM employees;

THAN
THAN
THAN
THAN
THAN

(TO_DATE('1-JAN-1992','DD-MON-YYYY')),
(TO_DATE('1-JAN-1994','DD-MON-YYYY')),
(TO_DATE('1-JAN-1996','DD-MON-YYYY')),
(TO_DATE('1-JAN-1998','DD-MON-YYYY')),
(TO_DATE('1-JAN-2001','DD-MON-YYYY'))

For the first example, consider the following statement:
EXPLAIN PLAN FOR
SELECT * FROM emp_range;

Oracle Database displays something similar to the following:
-------------------------------------------------------------------|Id| Operation
| Name
|Rows| Bytes|Cost|Pstart|Pstop|
-------------------------------------------------------------------| 0| SELECT STATEMENT
|
| 105| 13965 | 2 | |
|
| 1| PARTITION RANGE ALL|
| 105| 13965 | 2 | 1 |
5 |
| 2| TABLE ACCESS FULL | EMP_RANGE | 105| 13965 | 2 | 1 |
5 |
--------------------------------------------------------------------

7-11

Chapter 7

Reading Execution Plans: Advanced

The database creates a partition row source on top of the table access row source. It
iterates over the set of partitions to be accessed. In this example, the partition iterator
covers all partitions (option ALL), because a predicate was not used for pruning. The
PARTITION_START and PARTITION_STOP columns of the PLAN_TABLE show access to all
partitions from 1 to 5.
For the next example, consider the following statement:
EXPLAIN PLAN FOR
SELECT *
FROM emp_range
WHERE hire_date >= TO_DATE('1-JAN-1996','DD-MON-YYYY');
----------------------------------------------------------------------| Id | Operation
| Name |Rows|Bytes|Cost|Pstart|Pstop|
----------------------------------------------------------------------| 0 | SELECT STATEMENT
|
| 3 | 399 | 2 |
|
|
| 1 | PARTITION RANGE ITERATOR|
| 3 | 399 | 2 | 4 | 5 |
| *2 | TABLE ACCESS FULL
|EMP_RANGE| 3 | 399 | 2 | 4 | 5 |
-----------------------------------------------------------------------

In the previous example, the partition row source iterates from partition 4 to 5 because
the database prunes the other partitions using a predicate on hire_date.
Finally, consider the following statement:
EXPLAIN PLAN FOR
SELECT *
FROM emp_range
WHERE hire_date < TO_DATE('1-JAN-1992','DD-MON-YYYY');
----------------------------------------------------------------------| Id | Operation
| Name
|Rows|Bytes|Cost|Pstart|Pstop|
----------------------------------------------------------------------| 0 | SELECT STATEMENT
|
|
1 | 133 | 2 | | |
| 1 | PARTITION RANGE SINGLE|
|
1 | 133 | 2 | 1 | 1 |
|* 2 | TABLE ACCESS FULL
| EMP_RANGE |
1 | 133 | 2 | 1 | 1 |
-----------------------------------------------------------------------

In the previous example, only partition 1 is accessed and known at compile time; thus,
there is no need for a partition row source.

Note:
Oracle Database displays the same information for hash partitioned objects,
except the partition row source name is PARTITION HASH instead of PARTITION
RANGE. Also, with hash partitioning, pruning is only possible using equality or
IN-list predicates.

7-12

Chapter 7

Reading Execution Plans: Advanced

7.2.5.2 Pruning Information with Composite Partitioned Objects: Examples
To illustrate how Oracle Database displays pruning information for composite
partitioned objects, consider the table emp_comp. It is range-partitioned on hiredate and
subpartitioned by hash on deptno.
CREATE TABLE emp_comp PARTITION BY RANGE(hire_date)
SUBPARTITION BY HASH(department_id) SUBPARTITIONS 3
(
PARTITION emp_p1 VALUES LESS THAN (TO_DATE('1-JAN-1992','DD-MON-YYYY')),
PARTITION emp_p2 VALUES LESS THAN (TO_DATE('1-JAN-1994','DD-MON-YYYY')),
PARTITION emp_p3 VALUES LESS THAN (TO_DATE('1-JAN-1996','DD-MON-YYYY')),
PARTITION emp_p4 VALUES LESS THAN (TO_DATE('1-JAN-1998','DD-MON-YYYY')),
PARTITION emp_p5 VALUES LESS THAN (TO_DATE('1-JAN-2001','DD-MON-YYYY'))
)
AS SELECT * FROM employees;

For the first example, consider the following statement:
EXPLAIN PLAN FOR
SELECT * FROM emp_comp;
----------------------------------------------------------------------|Id| Operation
| Name
| Rows | Bytes |Cost|Pstart|Pstop|
----------------------------------------------------------------------| 0| SELECT STATEMENT
|
| 10120 | 1314K| 78 |
|
|
| 1| PARTITION RANGE ALL|
| 10120 | 1314K| 78 | 1 |
5 |
| 2| PARTITION HASH ALL|
| 10120 | 1314K| 78 | 1 |
3 |
| 3|
TABLE ACCESS FULL| EMP_COMP | 10120 | 1314K| 78 | 1 |
15 |
-----------------------------------------------------------------------

This example shows the plan when Oracle Database accesses all subpartitions of all
partitions of a composite object. The database uses two partition row sources for this
purpose: a range partition row source to iterate over the partitions, and a hash partition
row source to iterate over the subpartitions of each accessed partition.
In the following example, the range partition row source iterates from partition 1 to 5,
because the database performs no pruning. Within each partition, the hash partition
row source iterates over subpartitions 1 to 3 of the current partition. As a result, the
table access row source accesses subpartitions 1 to 15. In other words, the database
accesses all subpartitions of the composite object.
EXPLAIN PLAN FOR
SELECT *
FROM emp_comp
WHERE hire_date = TO_DATE('15-FEB-1998', 'DD-MON-YYYY');
----------------------------------------------------------------------| Id | Operation
| Name
|Rows|Bytes |Cost|Pstart|Pstop|
----------------------------------------------------------------------| 0 | SELECT STATEMENT
|
| 20 | 2660 | 17 |
|
|
| 1 | PARTITION RANGE SINGLE|
| 20 | 2660 | 17 | 5 | 5 |
| 2 | PARTITION HASH ALL |
| 20 | 2660 | 17 | 1 | 3 |
|* 3 |
TABLE ACCESS FULL | EMP_COMP | 20 | 2660 | 17 | 13 | 15 |
-----------------------------------------------------------------------

In the previous example, only the last partition, partition 5, is accessed. This partition is
known at compile time, so the database does not need to show it in the plan. The hash

7-13

Chapter 7

Reading Execution Plans: Advanced

partition row source shows accessing of all subpartitions within that partition; that is,
subpartitions 1 to 3, which translates into subpartitions 13 to 15 of the emp_comp table.
Now consider the following statement:
EXPLAIN PLAN FOR
SELECT *
FROM emp_comp
WHERE department_id = 20;
-----------------------------------------------------------------------| Id | Operation
|Name
|Rows | Bytes |Cost|Pstart|Pstop|
-----------------------------------------------------------------------| 0 | SELECT STATEMENT
|
| 101 | 13433 | 78 |
|
|
| 1 | PARTITION RANGE ALL |
| 101 | 13433 | 78 | 1 | 5 |
| 2 | PARTITION HASH SINGLE|
| 101 | 13433 | 78 | 3 | 3 |
|* 3 |
TABLE ACCESS FULL | EMP_COMP | 101 | 13433 | 78 |
|
|
------------------------------------------------------------------------

In the previous example, the predicate deptno=20 enables pruning on the hash
dimension within each partition. Therefore, Oracle Database only needs to access a
single subpartition. The number of this subpartition is known at compile time, so the
hash partition row source is not needed.
Finally, consider the following statement:
VARIABLE dno NUMBER;
EXPLAIN PLAN FOR
SELECT *
FROM emp_comp
WHERE department_id = :dno;
----------------------------------------------------------------------| Id| Operation
| Name
|Rows| Bytes |Cost|Pstart|Pstop|
----------------------------------------------------------------------| 0 | SELECT STATEMENT
|
| 101| 13433 | 78 |
|
|
| 1 | PARTITION RANGE ALL |
| 101| 13433 | 78 | 1 | 5 |
| 2 | PARTITION HASH SINGLE|
| 101| 13433 | 78 | KEY | KEY |
|*3 |
TABLE ACCESS FULL | EMP_COMP | 101| 13433 | 78 |
|
|
-----------------------------------------------------------------------

The last two examples are the same, except that department_id = :dno replaces
deptno=20. In this last case, the subpartition number is unknown at compile time, and a
hash partition row source is allocated. The option is SINGLE for this row source because
Oracle Database accesses only one subpartition within each partition. In Step 2, both
PARTITION_START and PARTITION_STOP are set to KEY. This value means that Oracle
Database determines the number of subpartitions at run time.

7.2.5.3 Examples of Partial Partition-Wise Joins
In these examples, the PQ_DISTRIBUTE hint explicitly forces a partial partition-wise join
because the query optimizer could have chosen a different plan based on cost in this
query.
Example 7-5

Partial Partition-Wise Join with Range Partition
emp_range_diddepartment_iddept2dept2

CREATE TABLE dept2 AS SELECT * FROM departments;
ALTER TABLE dept2 PARALLEL 2;

7-14

Chapter 7

Reading Execution Plans: Advanced

CREATE TABLE emp_range_did PARTITION BY RANGE(department_id)
(PARTITION emp_p1 VALUES LESS THAN (150),
PARTITION emp_p5 VALUES LESS THAN (MAXVALUE) )
AS SELECT * FROM employees;
ALTER TABLE emp_range_did PARALLEL 2;
EXPLAIN PLAN FOR
SELECT /*+ PQ_DISTRIBUTE(d NONE PARTITION) ORDERED */ e.last_name,
d.department_name
FROM emp_range_did e, dept2 d
WHERE e.department_id = d.department_id;
-----------------------------------------------------------------------------------------------|Id| Operation
|Name
|Row|Byte|Cost|Pstart|Pstop|TQ|IN-OUT|PQ Distrib|
-----------------------------------------------------------------------------------------------| 0| SELECT STATEMENT
|
|284 |16188|6 | | |
|
|
|
| 1| PX COORDINATOR
|
|
|
| | | |
|
|
|
| 2| PX SEND QC (RANDOM)
|:TQ10001
|284 |16188|6 | | | Q1,01 |P->S |QC (RAND) |
|*3|
HASH JOIN
|
|284 |16188|6 | | | Q1,01 |PCWP |
|
| 4|
PX PARTITION RANGE ALL |
|284 |7668 |2 | 1 | 2 | Q1,01 |PCWC |
|
| 5|
TABLE ACCESS FULL
|EMP_RANGE_DID|284 |7668 |2 | 1 | 2 | Q1,01 |PCWP |
|
| 6|
BUFFER SORT
|
|
|
| | | | Q1,01 |PCWC |
|
| 7|
PX RECEIVE
|
| 21 | 630 |2 | | | Q1,01 |PCWP |
|
| 8|
PX SEND PARTITION (KEY)|:TQ10000
| 21 | 630 |2 | | |
|S->P |PART (KEY)|
| 9|
TABLE ACCESS FULL
|DEPT2
| 21 | 630 |2 | | |
|
|
|
------------------------------------------------------------------------------------------------

The execution plan shows that the table dept2 is scanned serially and all rows with the
same partitioning column value of emp_range_did (department_id) are sent through a
PART (KEY), or partition key, table queue to the same parallel execution server doing
the partial partition-wise join.
Example 7-6

Partial Partition-Wise Join with Composite Partition
In the following example, emp_comp is joined on the partitioning column and is
parallelized, enabling use of a partial partition-wise join because dept2 is not
partitioned. The database dynamically partitions dept2 before the join.

ALTER TABLE emp_comp PARALLEL 2;
EXPLAIN PLAN FOR
SELECT /*+ PQ_DISTRIBUTE(d NONE PARTITION) ORDERED */ e.last_name,
d.department_name
FROM emp_comp e, dept2 d
WHERE e.department_id = d.department_id;
SELECT PLAN_TABLE_OUTPUT FROM TABLE(DBMS_XPLAN.DISPLAY());
-----------------------------------------------------------------------------------------------| Id| Operation
| Name |Rows |Bytes |Cost|Pstart|Pstop|TQ |IN-OUT|PQ Distrib|
-----------------------------------------------------------------------------------------------| 0 | SELECT STATEMENT
|
| 445 | 17800 | 5 | | |
|
|
|
| 1 | PX COORDINATOR
|
|
|
| | | |
|
|
|
| 2 | PX SEND QC (RANDOM)
|:TQ10001| 445 | 17800 | 5 | | | Q1,01 | P->S | QC (RAND)|
|*3 |
HASH JOIN
|
| 445 | 17800 | 5 | | | Q1,01 | PCWP |
|
| 4 |
PX PARTITION RANGE ALL |
| 107 | 1070 | 3 | 1 | 5 | Q1,01 | PCWC |
|
| 5 |
PX PARTITION HASH ALL |
| 107 | 1070 | 3 | 1 | 3 | Q1,01 | PCWC |
|
| 6 |
TABLE ACCESS FULL
|EMP_COMP| 107 | 1070 | 3 | 1 | 15| Q1,01 | PCWP |
|
| 7 |
PX RECEIVE
|
| 21 | 630 | 1 | | | Q1,01 | PCWP |
|

7-15

Chapter 7

Reading Execution Plans: Advanced

| 8 |
PX SEND PARTITION (KEY)|:TQ10000| 21 | 630 | 1 | | | Q1,00 | P->P |PART (KEY)|
| 9 |
PX BLOCK ITERATOR
|
| 21 | 630 | 1 | | | Q1,00 | PCWC |
|
|10 |
TABLE ACCESS FULL
|DEPT2 | 21 | 630 | 1 | | | Q1,00 | PCWP |
|
------------------------------------------------------------------------------------------------

The plan shows that the optimizer selects partial partition-wise join from one of two
columns. The PX SEND node type is PARTITION (KEY) and the PQ Distrib column
contains the text PART (KEY), or partition key. This implies that the table dept2 is repartitioned based on the join column department_id to be sent to the parallel execution
servers executing the scan of EMP_COMP and the join.

7.2.5.4 Example of Full Partition-Wise Join
In this example, emp_comp and dept_hash are joined on their hash partitioning columns,
enabling use of a full partition-wise join.
The PARTITION HASH row source appears on top of the join row source in the plan table
output.
CREATE TABLE dept_hash
PARTITION BY HASH(department_id)
PARTITIONS 3
PARALLEL 2
AS SELECT * FROM departments;
EXPLAIN PLAN FOR
SELECT /*+ PQ_DISTRIBUTE(e NONE NONE) ORDERED */ e.last_name,
d.department_name
FROM emp_comp e, dept_hash d
WHERE e.department_id = d.department_id;
----------------------------------------------------------------------------------------------|Id| Operation
| Name
|Rows|Bytes|Cost|Pstart|Pstop|TQ |IN-OUT|
PQ Distrib|
----------------------------------------------------------------------------------------------| 0| SELECT STATEMENT
|
| 106 | 2544 | 8 | |
|
|
|
|
| 1| PX COORDINATOR
|
|
|
| | |
|
|
|
|
| 2| PX SEND QC (RANDOM)
| :TQ10000 | 106 | 2544 | 8 | |
| Q1,00 | P->S
|QC (RAND)|
| 3|
PX PARTITION HASH ALL |
| 106 | 2544 | 8 | 1 | 3 | Q1,00 | PCWC
|
|
|*4|
HASH JOIN
|
| 106 | 2544 | 8 | |
| Q1,00 | PCWP
|
|
| 5|
PX PARTITION RANGE ALL|
| 107 | 1070 | 3 | 1 | 5 | Q1,00 | PCWC
|
|
| 6|
TABLE ACCESS FULL
| EMP_COMP | 107 | 1070 | 3 | 1 | 15 | Q1,00 | PCWP
|
|
| 7|
TABLE ACCESS FULL
| DEPT_HASH | 27 | 378 | 4 | 1 | 3 | Q1,00 | PCWP
|
|
-----------------------------------------------------------------------------------------------

The PX PARTITION HASH row source appears on top of the join row source in the plan
table output while the PX PARTITION RANGE row source appears over the scan of
emp_comp. Each parallel execution server performs the join of an entire hash partition of
emp_comp with an entire partition of dept_hash.

7-16

Chapter 7

Reading Execution Plans: Advanced

7.2.5.5 Examples of INLIST ITERATOR and EXPLAIN PLAN
An INLIST ITERATOR operation appears in the EXPLAIN PLAN output if an index
implements an IN-list predicate.
Consider the following statement:
SELECT * FROM emp WHERE empno IN (7876, 7900, 7902);

The EXPLAIN PLAN output appears as follows:
OPERATION
---------------SELECT STATEMENT
INLIST ITERATOR
TABLE ACCESS
INDEX

OPTIONS
---------------

OBJECT_NAME
--------------

BY ROWID
RANGE SCAN

EMP
EMP_EMPNO

The INLIST ITERATOR operation iterates over the next operation in the plan for each
value in the IN-list predicate. The following sections describe the three possible types
of IN-list columns for partitioned tables and indexes.
This section contains the following topics:
•

When the IN-List Column is an Index Column: Example
If the IN-list column empno is an index column but not a partition column, then the
IN-list operator appears before the table operation but after the partition operation
in the plan.

•

When the IN-List Column is an Index and a Partition Column: Example
If empno is an indexed and a partition column, then the plan contains an INLIST
ITERATOR operation before the partition operation.

•

When the IN-List Column is a Partition Column: Example
If empno is a partition column and no indexes exist, then no INLIST ITERATOR
operation is allocated.

7.2.5.5.1 When the IN-List Column is an Index Column: Example
If the IN-list column empno is an index column but not a partition column, then the IN-list
operator appears before the table operation but after the partition operation in the plan.
OPERATION
---------------SELECT STATEMENT
PARTITION RANGE
INLIST ITERATOR
TABLE ACCESS
INDEX

OPTIONS
------------

OBJECT_NAME PARTIT_START PARTITION_STOP
----------- ------------ --------------

ALL

KEY(INLIST)

KEY(INLIST)

BY LOCAL INDEX ROWID EMP
RANGE SCAN
EMP_EMPNO

KEY(INLIST)
KEY(INLIST)

KEY(INLIST)
KEY(INLIST)

The KEY(INLIST) designation for the partition start and stop keys specifies that an IN-list
predicate appears on the index start and stop keys.

7-17

Chapter 7

Reading Execution Plans: Advanced

7.2.5.5.2 When the IN-List Column is an Index and a Partition Column: Example
If empno is an indexed and a partition column, then the plan contains an INLIST
ITERATOR operation before the partition operation.
OPERATION
---------------SELECT STATEMENT
INLIST ITERATOR
PARTITION RANGE
TABLE ACCESS
INDEX

OPTIONS
------------

OBJECT_NAME PARTITION_START PARTITION_STOP
----------- --------------- --------------

ITERATOR
BY LOCAL INDEX ROWID EMP
RANGE SCAN
EMP_EMPNO

KEY(INLIST)
KEY(INLIST)
KEY(INLIST)

KEY(INLIST)
KEY(INLIST)
KEY(INLIST)

7.2.5.5.3 When the IN-List Column is a Partition Column: Example
If empno is a partition column and no indexes exist, then no INLIST ITERATOR operation is
allocated.
OPERATION
OPTIONS
---------------- -----------SELECT STATEMENT
PARTITION RANGE INLIST
TABLE ACCESS
FULL

OBJECT_NAME
-----------

PARTITION_START
---------------

PARTITION_STOP
--------------

EMP

KEY(INLIST)
KEY(INLIST)

KEY(INLIST)
KEY(INLIST)

If emp_empno is a bitmap index, then the plan is as follows:
OPERATION
---------------SELECT STATEMENT
INLIST ITERATOR
TABLE ACCESS
BITMAP CONVERSION
BITMAP INDEX

OPTIONS
---------------

OBJECT_NAME
--------------

BY INDEX ROWID
TO ROWIDS
SINGLE VALUE

EMP
EMP_EMPNO

7.2.5.6 Example of Domain Indexes and EXPLAIN PLAN
You can use EXPLAIN PLAN to derive user-defined CPU and I/O costs for domain
indexes.
EXPLAIN PLAN displays domain index statistics in the OTHER column of PLAN_TABLE. For
example, assume table emp has user-defined operator CONTAINS with a domain index
emp_resume on the resume column, and the index type of emp_resume supports the
operator CONTAINS. You explain the plan for the following query:
SELECT * FROM emp WHERE CONTAINS(resume, 'Oracle') = 1

The database could display the following plan:
OPERATION
----------------SELECT STATEMENT
TABLE ACCESS
DOMAIN INDEX

OPTIONS
OBJECT_NAME
----------- -----------BY ROWID

EMP
EMP_RESUME

OTHER
----------------

CPU: 300, I/O: 4

7.2.6 PLAN_TABLE Columns
The PLAN_TABLE used by the EXPLAIN PLAN statement contains the columns listed in this
topic.

7-18

Chapter 7

Reading Execution Plans: Advanced

Table 7-1

PLAN_TABLE Columns

Column

Type

Description

STATEMENT_ID

VARCHAR2(30)

Value of the optional STATEMENT_ID
parameter specified in the EXPLAIN PLAN
statement.

PLAN_ID

NUMBER

Unique identifier of a plan in the database.

TIMESTAMP

DATE

Date and time when the EXPLAIN PLAN
statement was generated.

REMARKS

VARCHAR2(80)

Any comment (of up to 80 bytes) you want to
associate with each step of the explained
plan. This column indicates whether the
database used an outline or SQL profile for
the query.
If you need to add or change a remark on
any row of the PLAN_TABLE, then use the
UPDATE statement to modify the rows of the
PLAN_TABLE.

OPERATION

VARCHAR2(30)

Name of the internal operation performed in
this step. In the first row generated for a
statement, the column contains one of the
following values:
•
DELETE STATEMENT
•
INSERT STATEMENT
•
SELECT STATEMENT
•
UPDATE STATEMENT
See Table 7-3 for more information about
values for this column.

OPTIONS

VARCHAR2(225)

A variation on the operation described in the
OPERATION column.
See Table 7-3 for more information about
values for this column.

OBJECT_NODE

VARCHAR2(128)

Name of the database link used to reference
the object (a table name or view name). For
local queries using parallel execution, this
column describes the order in which the
database consumes output from operations.

OBJECT_OWNER

VARCHAR2(30)

Name of the user who owns the schema
containing the table or index.

OBJECT_NAME

VARCHAR2(30)

Name of the table or index.

OBJECT_ALIAS

VARCHAR2(65)

Unique alias of a table or view in a SQL
statement. For indexes, it is the object alias
of the underlying table.

OBJECT_INSTANCE

NUMERIC

Number corresponding to the ordinal position
of the object as it appears in the original
statement. The numbering proceeds from left
to right, outer to inner for the original
statement text. View expansion results in
unpredictable numbers.

7-19

Chapter 7

Reading Execution Plans: Advanced

Table 7-1

(Cont.) PLAN_TABLE Columns

Column

Type

Description

OBJECT_TYPE

VARCHAR2(30)

Modifier that provides descriptive information
about the object; for example, NON-UNIQUE for
indexes.

OPTIMIZER

VARCHAR2(255)

Current mode of the optimizer.

SEARCH_COLUMNS

NUMBERIC

Not currently used.

ID

NUMERIC

A number assigned to each step in the
execution plan.

PARENT_ID

NUMERIC

The ID of the next execution step that
operates on the output of the ID step.

DEPTH

NUMERIC

Depth of the operation in the row source tree
that the plan represents. You can use the
value to indent the rows in a plan table
report.

POSITION

NUMERIC

For the first row of output, this indicates the
optimizer's estimated cost of executing the
statement. For the other rows, it indicates
the position relative to the other children of
the same parent.

COST

NUMERIC

Cost of the operation as estimated by the
optimizer's query approach. Cost is not
determined for table access operations. The
value of this column does not have any
particular unit of measurement; it is a
weighted value used to compare costs of
execution plans. The value of this column is
a function of the CPU_COST and IO_COST
columns.

CARDINALITY

NUMERIC

Estimate by the query optimization approach
of the number of rows that the operation
accessed.

BYTES

NUMERIC

Estimate by the query optimization approach
of the number of bytes that the operation
accessed.

7-20

Chapter 7

Reading Execution Plans: Advanced

Table 7-1

(Cont.) PLAN_TABLE Columns

Column

Type

Description

OTHER_TAG

VARCHAR2(255)

Describes the contents of the OTHER column.
Values are:
•

•
•

•

•

•

•

PARTITION_START

VARCHAR2(255)

SERIAL (blank): Serial execution.
Currently, SQL is not loaded in the
OTHER column for this case.
SERIAL_FROM_REMOTE (S -> R): Serial
execution at a remote site.
PARALLEL_FROM_SERIAL (S -> P):
Serial execution. Output of step is
partitioned or broadcast to parallel
execution servers.
PARALLEL_TO_SERIAL (P -> S): Parallel
execution. Output of step is returned to
serial QC process.
PARALLEL_TO_PARALLEL (P -> P):
Parallel execution. Output of step is
repartitioned to second set of parallel
execution servers.
PARALLEL_COMBINED_WITH_PARENT
(PWP): Parallel execution; Output of step
goes to next step in same parallel
process. No interprocess
communication to parent.
PARALLEL_COMBINED_WITH_CHILD (PWC):
Parallel execution. Input of step comes
from prior step in same parallel process.
No interprocess communication from
child.

Start partition of a range of accessed
partitions. It can take one of the following
values:
n indicates that the start partition has been
identified by the SQL compiler, and its
partition number is given by n.
KEY indicates that the start partition is
identified at run time from partitioning key
values.
ROW REMOVE_LOCATION indicates that the
database computes the start partition (same
as the stop partition) at run time from the
location of each retrieved record. The record
location is obtained by a user or from a
global index.
INVALID indicates that the range of accessed
partitions is empty.

7-21

Chapter 7

Reading Execution Plans: Advanced

Table 7-1

(Cont.) PLAN_TABLE Columns

Column

Type

Description

PARTITION_STOP

VARCHAR2(255)

Stop partition of a range of accessed
partitions. It can take one of the following
values:
n indicates that the stop partition has been
identified by the SQL compiler, and its
partition number is given by n.
KEY indicates that the stop partition is
identified at run time from partitioning key
values.
ROW REMOVE_LOCATION indicates that the
database computes the stop partition (same
as the start partition) at run time from the
location of each retrieved record. The record
location is obtained by a user or from a
global index.
INVALID indicates that the range of accessed
partitions is empty.

PARTITION_ID

NUMERIC

Step that has computed the pair of values of
the PARTITION_START and PARTITION_STOP
columns.

OTHER

LONG

Other information that is specific to the
execution step that a user might find useful.
See the OTHER_TAG column.

DISTRIBUTION

VARCHAR2(30)

Method used to distribute rows from
producer query servers to consumer query
servers.
See Table 7-2 for more information about the
possible values for this column. For more
information about consumer and producer
query servers, see Oracle Database VLDB
and Partitioning Guide.

CPU_COST

NUMERIC

CPU cost of the operation as estimated by
the query optimizer's approach. The value of
this column is proportional to the number of
machine cycles required for the operation.
For statements that use the rule-based
approach, this column is null.

IO_COST

NUMERIC

I/O cost of the operation as estimated by the
query optimizer's approach. The value of this
column is proportional to the number of data
blocks read by the operation. For statements
that use the rule-based approach, this
column is null.

TEMP_SPACE

NUMERIC

Temporary space, in bytes, that the
operation uses as estimated by the query
optimizer's approach. For statements that
use the rule-based approach, or for
operations that do not use any temporary
space, this column is null.

7-22

Chapter 7

Reading Execution Plans: Advanced

Table 7-1

(Cont.) PLAN_TABLE Columns

Column

Type

Description

ACCESS_PREDICATES

VARCHAR2(4000)

Predicates used to locate rows in an access
structure. For example, start or stop
predicates for an index range scan.

FILTER_PREDICATES

VARCHAR2(4000)

Predicates used to filter rows before
producing them.

PROJECTION

VARCHAR2(4000)

Expressions produced by the operation.

TIME

NUMBER(20,2)

Elapsed time in seconds of the operation as
estimated by query optimization. For
statements that use the rule-based
approach, this column is null. The
DBMS_XPLAN.DISPLAY_PLAN out, the time is in
the HH:MM:SS format.

QBLOCK_NAME

VARCHAR2(30)

Name of the query block, either systemgenerated or defined by the user with the
QB_NAME hint.

Table 7-2 describes the values that can appear in the DISTRIBUTION column:
Table 7-2

Values of DISTRIBUTION Column of the PLAN_TABLE

DISTRIBUTION Text Interpretation
PARTITION (ROWID)

Maps rows to query servers based on the partitioning of a table or index using the rowid of
the row to UPDATE or DELETE.

PARTITION (KEY)

Maps rows to query servers based on the partitioning of a table or index using a set of
columns. Used for partial partition-wise join, PARALLEL INSERT, CREATE TABLE AS SELECT of
a partitioned table, and CREATE PARTITIONED GLOBAL INDEX.

HASH

Maps rows to query servers using a hash function on the join key. Used for PARALLEL JOIN
or PARALLEL GROUP BY.

RANGE

Maps rows to query servers using ranges of the sort key. Used when the statement
contains an ORDER BY clause.

ROUND-ROBIN

Randomly maps rows to query servers.

BROADCAST

Broadcasts the rows of the entire table to each query server. Used for a parallel join when
one table is very small compared to the other.

QC (ORDER)

The QC consumes the input in order, from the first to the last query server. Used when the
statement contains an ORDER BY clause.

QC (RANDOM)

The QC consumes the input randomly. Used when the statement does not have an ORDER
BY clause.

Table 7-3 lists each combination of OPERATION and OPTIONS produced by the EXPLAIN
PLAN statement and its meaning within an execution plan.

7-23

Chapter 7

Reading Execution Plans: Advanced

Table 7-3

OPERATION and OPTIONS Values Produced by EXPLAIN PLAN

Operation

Option

Operation accepting multiple sets of rowids, returning
the intersection of the sets, eliminating duplicates.
Used for the single-column indexes access path.

AND-EQUAL

BITMAP

Description

CONVERSION

TO ROWIDS converts bitmap representations to actual
rowids that you can use to access the table.
FROM ROWIDS converts the rowids to a bitmap
representation.
COUNT returns the number of rowids if the actual values
are not needed.

BITMAP

INDEX

SINGLE VALUE looks up the bitmap for a single key
value in the index.
RANGE SCAN retrieves bitmaps for a key value range.
FULL SCAN performs a full scan of a bitmap index if
there is no start or stop key.

BITMAP

MERGE

Merges several bitmaps resulting from a range scan
into one bitmap.

BITMAP

MINUS

Subtracts bits of one bitmap from another. Row source
is used for negated predicates. Use this option only if
there are nonnegated predicates yielding a bitmap from
which the subtraction can take place. An example
appears in "Viewing Bitmap Indexes with EXPLAIN
PLAN".

BITMAP

OR

Computes the bitwise OR of two bitmaps.

BITMAP

AND

Computes the bitwise AND of two bitmaps.

BITMAP

KEY ITERATION

Takes each row from a table row source and finds the
corresponding bitmap from a bitmap index. This set of
bitmaps are then merged into one bitmap in a following
BITMAP MERGE operation.

CONNECT BY

Retrieves rows in hierarchical order for a query
containing a CONNECT BY clause.

CONCATENATION

Operation accepting multiple sets of rows returning the
union-all of the sets.

COUNT

Operation counting the number of rows selected from a
table.

COUNT

STOPKEY

Count operation where the number of rows returned is
limited by the ROWNUM expression in the WHERE clause.
Uses inner joins for all cube access.

CUBE SCAN
CUBE SCAN

PARTIAL OUTER

Uses an outer join for at least one dimension, and inner
joins for the other dimensions.

CUBE SCAN

OUTER

Uses outer joins for all cube access.

DOMAIN INDEX

Retrieval of one or more rowids from a domain index.
The options column contain information supplied by a
user-defined domain index cost function, if any.

FILTER

Operation accepting a set of rows, eliminates some of
them, and returns the rest.

FIRST ROW

Retrieval of only the first row selected by a query.

7-24

Chapter 7

Reading Execution Plans: Advanced

Table 7-3
PLAN

(Cont.) OPERATION and OPTIONS Values Produced by EXPLAIN

Operation

Option

Description
Operation retrieving and locking the rows selected by a
query containing a FOR UPDATE clause.

FOR UPDATE
HASH

GROUP BY

Operation hashing a set of rows into groups for a query
with a GROUP BY clause.

HASH

GROUP BY PIVOT

Operation hashing a set of rows into groups for a query
with a GROUP BY clause. The PIVOT option indicates a
pivot-specific optimization for the HASH GROUP BY
operator.
Operation joining two sets of rows and returning the
result. This join method is useful for joining large data
sets of data (DSS, Batch). The join condition is an
efficient way of accessing the second table.

HASH JOIN
(These are join
operations.)

Query optimizer uses the smaller of the two tables/data
sources to build a hash table on the join key in
memory. Then it scans the larger table, probing the
hash table to find the joined rows.
HASH JOIN

ANTI

Hash (left) antijoin

HASH JOIN

SEMI

Hash (left) semijoin

HASH JOIN

RIGHT ANTI

Hash right antijoin

HASH JOIN

RIGHT SEMI

Hash right semijoin

HASH JOIN

OUTER

Hash (left) outer join

HASH JOIN

RIGHT OUTER

Hash right outer join

INDEX

UNIQUE SCAN

Retrieval of a single rowid from an index.

INDEX

RANGE SCAN

Retrieval of one or more rowids from an index. Indexed
values are scanned in ascending order.

INDEX

RANGE SCAN
DESCENDING

Retrieval of one or more rowids from an index. Indexed
values are scanned in descending order.

INDEX

FULL SCAN

Retrieval of all rowids from an index when there is no
start or stop key. Indexed values are scanned in
ascending order.

INDEX

FULL SCAN
DESCENDING

Retrieval of all rowids from an index when there is no
start or stop key. Indexed values are scanned in
descending order.

INDEX

FAST FULL SCAN

Retrieval of all rowids (and column values) using
multiblock reads. No sorting order can be defined.
Compares to a full table scan on only the indexed
columns. Only available with the cost based optimizer.

INDEX

SKIP SCAN

Retrieval of rowids from a concatenated index without
using the leading column(s) in the index. Only available
with the cost based optimizer.

(These are access
methods.)

INLIST ITERATOR

Iterates over the next operation in the plan for each
value in the IN-list predicate.

7-25

Chapter 7

Reading Execution Plans: Advanced

Table 7-3
PLAN

(Cont.) OPERATION and OPTIONS Values Produced by EXPLAIN

Operation

Option

Description

INTERSECTION

Operation accepting two sets of rows and returning the
intersection of the sets, eliminating duplicates.

MERGE JOIN

Operation accepting two sets of rows, each sorted by a
value, combining each row from one set with the
matching rows from the other, and returning the result.

(These are join
operations.)
MERGE JOIN

OUTER

Merge join operation to perform an outer join
statement.

MERGE JOIN

ANTI

Merge antijoin.

MERGE JOIN

SEMI

Merge semijoin.

MERGE JOIN

CARTESIAN

Can result from 1 or more of the tables not having any
join conditions to any other tables in the statement.
Can occur even with a join and it may not be flagged as
CARTESIAN in the plan.
Retrieval of rows in hierarchical order for a query
containing a CONNECT BY clause.

CONNECT BY
FULL

Retrieval of all rows from a materialized view.

MAT_VIEW REWITE
ACCESS

SAMPLE

Retrieval of sampled rows from a materialized view.

MAT_VIEW REWITE
ACCESS

CLUSTER

Retrieval of rows from a materialized view based on a
value of an indexed cluster key.

MAT_VIEW REWITE
ACCESS

HASH

Retrieval of rows from materialized view based on hash
cluster key value.

MAT_VIEW REWITE
ACCESS

BY ROWID RANGE

Retrieval of rows from a materialized view based on a
rowid range.

MAT_VIEW REWITE
ACCESS

SAMPLE BY ROWID Retrieval of sampled rows from a materialized view
based on a rowid range.
RANGE

MAT_VIEW REWITE
ACCESS

BY USER ROWID

If the materialized view rows are located using usersupplied rowids.

MAT_VIEW REWITE
ACCESS

BY INDEX ROWID

If the materialized view is nonpartitioned and rows are
located using index(es).

MAT_VIEW REWITE
ACCESS

BY GLOBAL INDEX If the materialized view is partitioned and rows are
located using only global indexes.
ROWID

MAT_VIEW REWITE
ACCESS
(These are access
methods.)

7-26

Chapter 7

Reading Execution Plans: Advanced

Table 7-3
PLAN

(Cont.) OPERATION and OPTIONS Values Produced by EXPLAIN

Operation

Option

Description

MAT_VIEW REWITE
ACCESS

BY LOCAL INDEX
ROWID

If the materialized view is partitioned and rows are
located using one or more local indexes and possibly
some global indexes.
Partition Boundaries:
The partition boundaries might have been computed
by:
A previous PARTITION step, in which case the
PARTITION_START and PARTITION_STOP column values
replicate the values present in the PARTITION step, and
the PARTITION_ID contains the ID of the PARTITION
step. Possible values for PARTITION_START and
PARTITION_STOP are NUMBER(n), KEY, INVALID.
The MAT_VIEW REWRITE ACCESS or INDEX step itself, in
which case the PARTITION_ID contains the ID of the
step. Possible values for PARTITION_START and
PARTITION_STOP are NUMBER(n), KEY, ROW
REMOVE_LOCATION (MAT_VIEW REWRITE ACCESS only),
and INVALID.

MINUS

Operation accepting two sets of rows and returning
rows appearing in the first set but not in the second,
eliminating duplicates.

NESTED LOOPS

Operation accepting two sets of rows, an outer set and
an inner set. Oracle Database compares each row of
the outer set with each row of the inner set, returning
rows that satisfy a condition. This join method is useful
for joining small subsets of data (OLTP). The join
condition is an efficient way of accessing the second
table.

(These are join
operations.)

NESTED LOOPS

OUTER

Nested loops operation to perform an outer join
statement.
Iterates over the next operation in the plan for each
partition in the range given by the PARTITION_START
and PARTITION_STOP columns. PARTITION describes
partition boundaries applicable to a single partitioned
object (table or index) or to a set of equipartitioned
objects (a partitioned table and its local indexes). The
partition boundaries are provided by the values of
PARTITION_START and PARTITION_STOP of the
PARTITION. Refer to Table 7-1 for valid values of
partition start and stop.

PARTITION

PARTITION

SINGLE

Access one partition.

PARTITION

ITERATOR

Access many partitions (a subset).

PARTITION

ALL

Access all partitions.

PARTITION

INLIST

Similar to iterator, but based on an IN-list predicate.

PARTITION

INVALID

Indicates that the partition set to be accessed is empty.

PX ITERATOR

BLOCK, CHUNK

Implements the division of an object into block or chunk
ranges among a set of parallel execution servers.

7-27

Chapter 7

Reading Execution Plans: Advanced

Table 7-3
PLAN

(Cont.) OPERATION and OPTIONS Values Produced by EXPLAIN

Operation

Option

Description

PX COORDINATOR

Implements the query coordinator that controls,
schedules, and executes the parallel plan below it
using parallel execution servers. It also represents a
serialization point, as the end of the part of the plan
executed in parallel and always has a PX SEND QC
operation below it.

PX PARTITION

Same semantics as the regular PARTITION operation
except that it appears in a parallel plan.

PX RECEIVE

Shows the consumer/receiver parallel execution node
reading repartitioned data from a send/producer (QC or
parallel execution server) executing on a PX SEND
node. This information was formerly displayed into the
DISTRIBUTION column. See Table 7-2.

PX SEND

QC (RANDOM),
HASH, RANGE

Implements the distribution method taking place
between two parallel execution servers. Shows the
boundary between two sets and how data is
repartitioned on the send/producer side. This
information was formerly displayed into the
DISTRIBUTION column. See Table 7-2.

REMOTE

Retrieval of data from a remote database.

SEQUENCE

Operation involving accessing values of a sequence.

SORT

AGGREGATE

Retrieval of a single row that is the result of applying a
group function to a group of selected rows.

SORT

UNIQUE

Operation sorting a set of rows to eliminate duplicates.

SORT

GROUP BY

Operation sorting a set of rows into groups for a query
with a GROUP BY clause.

SORT

GROUP BY PIVOT

Operation sorting a set of rows into groups for a query
with a GROUP BY clause. The PIVOT option indicates a
pivot-specific optimization for the SORT GROUP BY
operator.

SORT

JOIN

Operation sorting a set of rows before a merge-join.

SORT

ORDER BY

Operation sorting a set of rows for a query with an
ORDER BY clause.

TABLE ACCESS

FULL

Retrieval of all rows from a table.

TABLE ACCESS

SAMPLE

Retrieval of sampled rows from a table.

TABLE ACCESS

CLUSTER

Retrieval of rows from a table based on a value of an
indexed cluster key.

TABLE ACCESS

HASH

Retrieval of rows from table based on hash cluster key
value.

TABLE ACCESS

BY ROWID RANGE

Retrieval of rows from a table based on a rowid range.

TABLE ACCESS

SAMPLE BY ROWID Retrieval of sampled rows from a table based on a
rowid range.
RANGE

(These are access
methods.)

7-28

Chapter 7

Execution Plan Reference

Table 7-3
PLAN

(Cont.) OPERATION and OPTIONS Values Produced by EXPLAIN

Operation

Option

Description

TABLE ACCESS

BY USER ROWID

If the table rows are located using user-supplied
rowids.

TABLE ACCESS

BY INDEX ROWID

If the table is nonpartitioned and rows are located using
indexes.

TABLE ACCESS

BY GLOBAL INDEX If the table is partitioned and rows are located using
only global indexes.
ROWID

TABLE ACCESS

BY LOCAL INDEX
ROWID

If the table is partitioned and rows are located using
one or more local indexes and possibly some global
indexes.
Partition Boundaries:
The partition boundaries might have been computed
by:
A previous PARTITION step, in which case the
PARTITION_START and PARTITION_STOP column values
replicate the values present in the PARTITION step, and
the PARTITION_ID contains the ID of the PARTITION
step. Possible values for PARTITION_START and
PARTITION_STOP are NUMBER(n), KEY, INVALID.
The TABLE ACCESS or INDEX step itself, in which case
the PARTITION_ID contains the ID of the step. Possible
values for PARTITION_START and PARTITION_STOP are
NUMBER(n), KEY, ROW REMOVE_LOCATION (TABLE ACCESS
only), and INVALID.

TRANSPOSE

Operation evaluating a PIVOT operation by transposing
the results of GROUP BY to produce the final pivoted
data.

UNION

Operation accepting two sets of rows and returns the
union of the sets, eliminating duplicates.

UNPIVOT

Operation that rotates data from columns into rows.

VIEW

Operation performing a view's query and then returning
the resulting rows to another operation.

See Also:
Oracle Database Reference for more information about PLAN_TABLE

7.3 Execution Plan Reference
This section describes V$ views and PLAN_COLUMN columns.
This section contains the following topics:
•

Execution Plan Views
The following dynamic performance and data dictionary views provide information
on execution plans.

7-29

Chapter 7

Execution Plan Reference

•

PLAN_TABLE Columns
The PLAN_TABLE is used by the EXPLAIN PLAN statement.

•

DBMS_XPLAN Program Units
The functions DISPLAY_PLAN and DISPLAY_CURSOR in DBMS_XPLAN are relevant for
accessing adapted plans.

7.3.1 Execution Plan Views
The following dynamic performance and data dictionary views provide information on
execution plans.
Table 7-4

Execution Plan Views

View

Description

V$SQL_SHARED_CURSOR

Explains why a particular child cursor is not shared with
existing child cursors. Each column identifies a specific
reason why the cursor cannot be shared.
The USE_FEEDBACK_STATS column shows whether a
child cursor fails to match because of reoptimization.

V$SQL_PLAN

Includes a superset of all rows appearing in all final
plans. PLAN_LINE_ID is consecutively numbered, but for
a single final plan, the IDs may not be consecutive.

V$SQL_PLAN_STATISTICS_ALL

Contains memory usage statistics for row sources that
use SQL memory (sort or hash join). This view
concatenates information in V$SQL_PLAN with execution
statistics from V$SQL_PLAN_STATISTICS and
V$SQL_WORKAREA.

7.3.2 PLAN_TABLE Columns
The PLAN_TABLE is used by the EXPLAIN PLAN statement.
PLAN_TABLE contains the columns listed in Table 7-5.

Table 7-5

PLAN_TABLE Columns

Column

Type

Description

STATEMENT_ID

VARCHAR2(30)

Value of the optional STATEMENT_ID parameter specified
in the EXPLAIN PLAN statement.

PLAN_ID

NUMBER

Unique identifier of a plan in the database.

TIMESTAMP

DATE

Date and time when the EXPLAIN PLAN statement was
generated.

REMARKS

VARCHAR2(80)

Any comment (of up to 80 bytes) you want to associate
with each step of the explained plan. This column
indicates whether the database used an outline or SQL
profile for the query.
If you need to add or change a remark on any row of the
PLAN_TABLE, then use the UPDATE statement to modify the
rows of the PLAN_TABLE.

7-30

Chapter 7

Execution Plan Reference

Table 7-5

(Cont.) PLAN_TABLE Columns

Column

Type

Description

OPERATION

VARCHAR2(30)

Name of the internal operation performed in this step. In
the first row generated for a statement, the column
contains one of the following values:
•
DELETE STATEMENT
•
INSERT STATEMENT
•
SELECT STATEMENT
•
UPDATE STATEMENT
See Table 7-6 for more information about values for this
column.

OPTIONS

VARCHAR2(225)

A variation on the operation that the OPERATION column
describes.
See Table 7-6 for more information about values for this
column.

OBJECT_NODE

VARCHAR2(128)

Name of the database link used to reference the object
(a table name or view name). For local queries using
parallel execution, this column describes the order in
which the database consumes output from operations.

OBJECT_OWNER

VARCHAR2(30)

Name of the user who owns the schema containing the
table or index.

OBJECT_NAME

VARCHAR2(30)

Name of the table or index.

OBJECT_ALIAS

VARCHAR2(65)

Unique alias of a table or view in a SQL statement. For
indexes, it is the object alias of the underlying table.

OBJECT_INSTANCE

NUMERIC

Number corresponding to the ordinal position of the
object as it appears in the original statement. The
numbering proceeds from left to right, outer to inner for
the original statement text. View expansion results in
unpredictable numbers.

OBJECT_TYPE

VARCHAR2(30)

Modifier that provides descriptive information about the
object; for example, NONUNIQUE for indexes.

OPTIMIZER

VARCHAR2(255)

Current mode of the optimizer.

SEARCH_COLUMNS

NUMBERIC

Not currently used.

ID

NUMERIC

A number assigned to each step in the execution plan.

PARENT_ID

NUMERIC

The ID of the next execution step that operates on the
output of the ID step.

DEPTH

NUMERIC

Depth of the operation in the row source tree that the
plan represents. You can use this value to indent the
rows in a plan table report.

POSITION

NUMERIC

For the first row of output, this indicates the optimizer's
estimated cost of executing the statement. For the other
rows, it indicates the position relative to the other
children of the same parent.

7-31

Chapter 7

Execution Plan Reference

Table 7-5

(Cont.) PLAN_TABLE Columns

Column

Type

Description

COST

NUMERIC

Cost of the operation as estimated by the optimizer's
query approach. Cost is not determined for table access
operations. The value of this column does not have any
particular unit of measurement; it is a weighted value
used to compare costs of execution plans. The value of
this column is a function of the CPU_COST and IO_COST
columns.

CARDINALITY

NUMERIC

Estimate by the query optimization approach of the
number of rows that the operation accessed.

BYTES

NUMERIC

Estimate by the query optimization approach of the
number of bytes that the operation accessed.

OTHER_TAG

VARCHAR2(255)

Describes the contents of the OTHER column. Values are:
•
•
•

•
•

•

•

PARTITION_START

VARCHAR2(255)

SERIAL (blank): Serial execution. Currently, SQL is
not loaded in the OTHER column for this case.
SERIAL_FROM_REMOTE (S -> R): Serial execution at
a remote site.
PARALLEL_FROM_SERIAL (S -> P): Serial execution.
Output of step is partitioned or broadcast to parallel
execution servers.
PARALLEL_TO_SERIAL (P -> S): Parallel execution.
Output of step is returned to serial QC process.
PARALLEL_TO_PARALLEL (P -> P): Parallel
execution. Output of step is repartitioned to second
set of parallel execution servers.
PARALLEL_COMBINED_WITH_PARENT (PWP): Parallel
execution; Output of step goes to next step in same
parallel process. No interprocess communication to
parent.
PARALLEL_COMBINED_WITH_CHILD (PWC): Parallel
execution. Input of step comes from prior step in
same parallel process. No interprocess
communication from child.

Start partition of a range of accessed partitions. It can
take one of the following values:
n indicates that the start partition has been identified by
the SQL compiler, and its partition number is given by n.
KEY indicates that the start partition is identified at run
time from partitioning key values.
ROW REMOVE_LOCATION indicates that the database
computes the start partition (same as the stop partition)
at run time from the location of each retrieved record.
The record location is obtained by a user or from a global
index.
INVALID indicates that the range of accessed partitions is
empty.

7-32

Chapter 7

Execution Plan Reference

Table 7-5

(Cont.) PLAN_TABLE Columns

Column

Type

Description

PARTITION_STOP

VARCHAR2(255)

Stop partition of a range of accessed partitions. It can
take one of the following values:
n indicates that the stop partition has been identified by
the SQL compiler, and its partition number is given by n.
KEY indicates that the stop partition is identified at run
time from partitioning key values.
ROW REMOVE_LOCATION indicates that the database
computes the stop partition (same as the start partition)
at run time from the location of each retrieved record.
The record location is obtained by a user or from a global
index.
INVALID indicates that the range of accessed partitions is
empty.

PARTITION_ID

NUMERIC

Step that has computed the pair of values of the
PARTITION_START and PARTITION_STOP columns.

OTHER

LONG

Other information that is specific to the execution step
that a user might find useful. See the OTHER_TAG column.

DISTRIBUTION

VARCHAR2(30)

Method used to distribute rows from producer query
servers to consumer query servers.
See "Table 7-6" for more information about the possible
values for this column. For more information about
consumer and producer query servers, see Oracle
Database VLDB and Partitioning Guide.

CPU_COST

NUMERIC

CPU cost of the operation as estimated by the query
optimizer's approach. The value of this column is
proportional to the number of machine cycles required for
the operation. For statements that use the rule-based
approach, this column is null.

IO_COST

NUMERIC

I/O cost of the operation as estimated by the query
optimizer's approach. The value of this column is
proportional to the number of data blocks read by the
operation. For statements that use the rule-based
approach, this column is null.

TEMP_SPACE

NUMERIC

Temporary space, in bytes, used by the operation as
estimated by the query optimizer's approach. For
statements that use the rule-based approach, or for
operations that do not use any temporary space, this
column is null.

ACCESS_PREDICATES

VARCHAR2(4000)

Predicates used to locate rows in an access structure.
For example, start or stop predicates for an index range
scan.

FILTER_PREDICATES

VARCHAR2(4000)

Predicates used to filter rows before producing them.

PROJECTION

VARCHAR2(4000)

Expressions produced by the operation.

TIME

NUMBER(20,2)

Elapsed time in seconds of the operation as estimated by
query optimization. For statements that use the rulebased approach, this column is null.

QBLOCK_NAME

VARCHAR2(30)

Name of the query block, either system-generated or
defined by the user with the QB_NAME hint.

7-33

Chapter 7

Execution Plan Reference

Table 7-6

Values of DISTRIBUTION Column of the PLAN_TABLE

DISTRIBUTION Text

Interpretation

PARTITION (ROWID)

Maps rows to query servers based on the partitioning of a table or index using
the rowid of the row to UPDATE/DELETE.

PARTITION (KEY)

Maps rows to query servers based on the partitioning of a table or index using a
set of columns. Used for partial partition-wise join, PARALLEL INSERT, CREATE
TABLE AS SELECT of a partitioned table, and CREATE PARTITIONED GLOBAL
INDEX.

HASH

Maps rows to query servers using a hash function on the join key. Used for
PARALLEL JOIN or PARALLEL GROUP BY.

RANGE

Maps rows to query servers using ranges of the sort key. Used when the
statement contains an ORDER BY clause.

ROUND-ROBIN

Randomly maps rows to query servers.

BROADCAST

Broadcasts the rows of the entire table to each query server. Used for a parallel
join when one table is very small compared to the other.

QC (ORDER)

The QC consumes the input in order, from the first to the last query server.
Used when the statement contains an ORDER BY clause.

QC (RANDOM)

The QC consumes the input randomly. Used when the statement does not have
an ORDER BY clause.

Table 7-7 lists each combination of OPERATION and OPTIONS produced by the EXPLAIN
PLAN statement and its meaning within an execution plan.
Table 7-7
Operation

OPERATION and OPTIONS Values Produced by EXPLAIN PLAN
Option

Operation accepting multiple sets of rowids, returning the
intersection of the sets, eliminating duplicates. Used for the singlecolumn indexes access path.

AND-EQUAL

BITMAP

Description

CONVERSION

TO ROWIDS converts bitmap representations to actual rowids that you
can use to access the table.
FROM ROWIDS converts the rowids to a bitmap representation.
COUNT returns the number of rowids if the actual values are not
needed.

BITMAP

INDEX

SINGLE VALUE looks up the bitmap for a single key value in the
index.
RANGE SCAN retrieves bitmaps for a key value range.
FULL SCAN performs a full scan of a bitmap index if there is no start
or stop key.

BITMAP

MERGE

Merges several bitmaps resulting from a range scan into one
bitmap.

BITMAP

MINUS

Subtracts bits of one bitmap from another. Row source is used for
negated predicates. This option is usable only if there are nonnegated predicates yielding a bitmap from which the subtraction can
take place.

BITMAP

OR

Computes the bitwise OR of two bitmaps.

BITMAP

AND

Computes the bitwise AND of two bitmaps.

7-34

Chapter 7

Execution Plan Reference

Table 7-7

(Cont.) OPERATION and OPTIONS Values Produced by EXPLAIN PLAN

Operation

Option

Description

BITMAP

KEY ITERATION

Takes each row from a table row source and finds the
corresponding bitmap from a bitmap index. This set of bitmaps are
then merged into one bitmap in a following BITMAP MERGE operation.

CONNECT BY

Retrieves rows in hierarchical order for a query containing a CONNECT
BY clause.

CONCATENATION

Operation accepting multiple sets of rows returning the union-all of
the sets.

COUNT

Operation counting the number of rows selected from a table.

COUNT

STOPKEY

Count operation where the number of rows returned is limited by the
ROWNUM expression in the WHERE clause.
Joins a table or view on the left and a cube on the right.

CUBE JOIN

See Oracle Database SQL Language Reference to learn about the
NO_USE_CUBE and USE_CUBE hints.
CUBE JOIN

ANTI

Uses an antijoin for a table or view on the left and a cube on the
right.

CUBE JOIN

ANTI SNA

Uses an antijoin (single-sided null aware) for a table or view on the
left and a cube on the right. The join column on the right (cube side)
is NOT NULL.

CUBE JOIN

OUTER

Uses an outer join for a table or view on the left and a cube on the
right.

CUBE JOIN

RIGHT SEMI

Uses a right semijoin for a table or view on the left and a cube on
the right.
Uses inner joins for all cube access.

CUBE SCAN
CUBE SCAN

PARTIAL OUTER

Uses an outer join for at least one dimension, and inner joins for the
other dimensions.

CUBE SCAN

OUTER

Uses outer joins for all cube access.

DOMAIN INDEX

Retrieval of one or more rowids from a domain index. The options
column contain information supplied by a user-defined domain index
cost function, if any.

FILTER

Operation accepting a set of rows, eliminates some of them, and
returns the rest.

FIRST ROW

Retrieval of only the first row selected by a query.

FOR UPDATE

Operation retrieving and locking the rows selected by a query
containing a FOR UPDATE clause.

HASH

GROUP BY

Operation hashing a set of rows into groups for a query with a GROUP
BY clause.

HASH

GROUP BY PIVOT

Operation hashing a set of rows into groups for a query with a GROUP
BY clause. The PIVOT option indicates a pivot-specific optimization
for the HASH GROUP BY operator.

HASH JOIN
(These are join
operations.)

Operation joining two sets of rows and returning the result. This join
method is useful for joining large data sets of data (DSS, Batch).
The join condition is an efficient way of accessing the second table.
Query optimizer uses the smaller of the two tables/data sources to
build a hash table on the join key in memory. Then it scans the
larger table, probing the hash table to find the joined rows.

7-35

Chapter 7

Execution Plan Reference

Table 7-7

(Cont.) OPERATION and OPTIONS Values Produced by EXPLAIN PLAN

Operation

Option

Description

HASH JOIN

ANTI

Hash (left) antijoin

HASH JOIN

SEMI

Hash (left) semijoin

HASH JOIN

RIGHT ANTI

Hash right antijoin

HASH JOIN

RIGHT SEMI

Hash right semijoin

HASH JOIN

OUTER

Hash (left) outer join

HASH JOIN

RIGHT OUTER

Hash right outer join

INDEX

UNIQUE SCAN

Retrieval of a single rowid from an index.

INDEX

RANGE SCAN

Retrieval of one or more rowids from an index. Indexed values are
scanned in ascending order.

INDEX

RANGE SCAN
DESCENDING

Retrieval of one or more rowids from an index. Indexed values are
scanned in descending order.

INDEX

FULL SCAN

Retrieval of all rowids from an index when there is no start or stop
key. Indexed values are scanned in ascending order.

INDEX

FULL SCAN
DESCENDING

Retrieval of all rowids from an index when there is no start or stop
key. Indexed values are scanned in descending order.

INDEX

FAST FULL SCAN

Retrieval of all rowids (and column values) using multiblock reads.
No sorting order can be defined. Compares to a full table scan on
only the indexed columns. Only available with the cost based
optimizer.

INDEX

SKIP SCAN

Retrieval of rowids from a concatenated index without using the
leading column(s) in the index. Only available with the cost based
optimizer.

(These are access
methods.)

INLIST ITERATOR

Iterates over the next operation in the plan for each value in the INlist predicate.

INTERSECTION

Operation accepting two sets of rows and returning the intersection
of the sets, eliminating duplicates.

MERGE JOIN

Operation accepting two sets of rows, each sorted by a value,
combining each row from one set with the matching rows from the
other, and returning the result.

(These are join
operations.)
MERGE JOIN

OUTER

Merge join operation to perform an outer join statement.

MERGE JOIN

ANTI

Merge antijoin.

MERGE JOIN

SEMI

Merge semijoin.

MERGE JOIN

CARTESIAN

Can result from 1 or more of the tables not having any join
conditions to any other tables in the statement. Can occur even with
a join and it may not be flagged as CARTESIAN in the plan.
Retrieval of rows in hierarchical order for a query containing a
CONNECT BY clause.

CONNECT BY
MAT_VIEW REWITE
ACCESS

FULL

Retrieval of all rows from a materialized view.

(These are access
methods.)

7-36

Chapter 7

Execution Plan Reference

Table 7-7

(Cont.) OPERATION and OPTIONS Values Produced by EXPLAIN PLAN

Operation

Option

Description

MAT_VIEW REWITE
ACCESS

SAMPLE

Retrieval of sampled rows from a materialized view.

MAT_VIEW REWITE
ACCESS

CLUSTER

Retrieval of rows from a materialized view based on a value of an
indexed cluster key.

MAT_VIEW REWITE
ACCESS

HASH

Retrieval of rows from materialized view based on hash cluster key
value.

MAT_VIEW REWITE
ACCESS

BY ROWID RANGE

Retrieval of rows from a materialized view based on a rowid range.

MAT_VIEW REWITE
ACCESS

SAMPLE BY ROWID
RANGE

Retrieval of sampled rows from a materialized view based on a
rowid range.

MAT_VIEW REWITE
ACCESS

BY USER ROWID

If the materialized view rows are located using user-supplied rowids.

MAT_VIEW REWITE
ACCESS

BY INDEX ROWID

If the materialized view is nonpartitioned and rows are located using
indexes.

MAT_VIEW REWITE
ACCESS

BY GLOBAL INDEX
ROWID

If the materialized view is partitioned and rows are located using
only global indexes.

MAT_VIEW REWITE
ACCESS

BY LOCAL INDEX
ROWID

If the materialized view is partitioned and rows are located using one
or more local indexes and possibly some global indexes.
Partition Boundaries:
The partition boundaries might have been computed by:
A previous PARTITION step, in which case the PARTITION_START and
PARTITION_STOP column values replicate the values present in the
PARTITION step, and the PARTITION_ID contains the ID of the
PARTITION step. Possible values for PARTITION_START and
PARTITION_STOP are NUMBER(n), KEY, and INVALID.
The MAT_VIEW REWRITE ACCESS or INDEX step itself, in which case
the PARTITION_ID contains the ID of the step. Possible values for
PARTITION_START and PARTITION_STOP are NUMBER(n), KEY, ROW
REMOVE_LOCATION (MAT_VIEW REWRITE ACCESS only), and INVALID.

MINUS

Operation accepting two sets of rows and returning rows appearing
in the first set but not in the second, eliminating duplicates.

NESTED LOOPS

Operation accepting two sets of rows, an outer set and an inner set.
Oracle Database compares each row of the outer set with each row
of the inner set, returning rows that satisfy a condition. This join
method is useful for joining small subsets of data (OLTP). The join
condition is an efficient way of accessing the second table.

(These are join
operations.)

NESTED LOOPS

OUTER

Nested loops operation to perform an outer join statement.
Iterates over the next operation in the plan for each partition in the
range given by the PARTITION_START and PARTITION_STOP columns.
PARTITION describes partition boundaries applicable to a single
partitioned object (table or index) or to a set of equipartitioned
objects (a partitioned table and its local indexes). The partition
boundaries are provided by the values of PARTITION_START and
PARTITION_STOP of the PARTITION. Refer to Table 7-4 for valid
values of partition start and stop.

PARTITION

PARTITION

SINGLE

Access one partition.

PARTITION

ITERATOR

Access many partitions (a subset).

7-37

Chapter 7

Execution Plan Reference

Table 7-7

(Cont.) OPERATION and OPTIONS Values Produced by EXPLAIN PLAN

Operation

Option

Description

PARTITION

ALL

Access all partitions.

PARTITION

INLIST

Similar to iterator, but based on an IN-list predicate.

PARTITION

INVALID

Indicates that the partition set to be accessed is empty.

PX ITERATOR

BLOCK, CHUNK

Implements the division of an object into block or chunk ranges
among a set of parallel execution servers.

PX COORDINATOR

Implements the Query Coordinator which controls, schedules, and
executes the parallel plan below it using parallel execution servers.
It also represents a serialization point, as the end of the part of the
plan executed in parallel and always has a PX SEND QC operation
below it.

PX PARTITION

Same semantics as the regular PARTITION operation except that it
appears in a parallel plan.

PX RECEIVE

Shows the consumer/receiver parallel execution node reading
repartitioned data from a send/producer (QC or parallel execution
server) executing on a PX SEND node. This information was
formerly displayed into the DISTRIBUTION column. See Table 7-5.

PX SEND

QC (RANDOM), HASH,
RANGE

Implements the distribution method taking place between two sets of
parallel execution servers. Shows the boundary between two sets
and how data is repartitioned on the send/producer side (QC or side.
This information was formerly displayed into the DISTRIBUTION
column. See Table 7-5.

REMOTE

Retrieval of data from a remote database.

SEQUENCE

Operation involving accessing values of a sequence.

SORT

AGGREGATE

Retrieval of a single row that is the result of applying a group
function to a group of selected rows.

SORT

UNIQUE

Operation sorting a set of rows to eliminate duplicates.

SORT

GROUP BY

Operation sorting a set of rows into groups for a query with a GROUP
BY clause.

SORT

GROUP BY PIVOT

Operation sorting a set of rows into groups for a query with a GROUP
BY clause. The PIVOT option indicates a pivot-specific optimization
for the SORT GROUP BY operator.

SORT

JOIN

Operation sorting a set of rows before a merge-join.

SORT

ORDER BY

Operation sorting a set of rows for a query with an ORDER BY clause.

TABLE ACCESS

FULL

Retrieval of all rows from a table.

TABLE ACCESS

SAMPLE

Retrieval of sampled rows from a table.

TABLE ACCESS

CLUSTER

Retrieval of rows from a table based on a value of an indexed
cluster key.

TABLE ACCESS

HASH

Retrieval of rows from table based on hash cluster key value.

TABLE ACCESS

BY ROWID RANGE

Retrieval of rows from a table based on a rowid range.

TABLE ACCESS

SAMPLE BY ROWID
RANGE

Retrieval of sampled rows from a table based on a rowid range.

TABLE ACCESS

BY USER ROWID

If the table rows are located using user-supplied rowids.

(These are access
methods.)

7-38

Chapter 7

Execution Plan Reference

Table 7-7

(Cont.) OPERATION and OPTIONS Values Produced by EXPLAIN PLAN

Operation

Option

Description

TABLE ACCESS

BY INDEX ROWID

If the table is nonpartitioned and rows are located using index(es).

TABLE ACCESS

BY GLOBAL INDEX
ROWID

If the table is partitioned and rows are located using only global
indexes.

TABLE ACCESS

BY LOCAL INDEX
ROWID

If the table is partitioned and rows are located using one or more
local indexes and possibly some global indexes.
Partition Boundaries:
The partition boundaries might have been computed by:
A previous PARTITION step, in which case the PARTITION_START and
PARTITION_STOP column values replicate the values present in the
PARTITION step, and the PARTITION_ID contains the ID of the
PARTITION step. Possible values for PARTITION_START and
PARTITION_STOP are NUMBER(n), KEY, and INVALID.
The TABLE ACCESS or INDEX step itself, in which case the
PARTITION_ID contains the ID of the step. Possible values for
PARTITION_START and PARTITION_STOP are NUMBER(n), KEY, ROW
REMOVE_LOCATION (TABLE ACCESS only), and INVALID.

TRANSPOSE

Operation evaluating a PIVOT operation by transposing the results of
GROUP BY to produce the final pivoted data.

UNION

Operation accepting two sets of rows and returns the union of the
sets, eliminating duplicates.

UNPIVOT

Operation that rotates data from columns into rows.

VIEW

Operation performing a view's query and then returning the resulting
rows to another operation.

See Also:
Oracle Database Reference for more information about PLAN_TABLE

7.3.3 DBMS_XPLAN Program Units
The functions DISPLAY_PLAN and DISPLAY_CURSOR in DBMS_XPLAN are relevant for accessing
adapted plans.

7-39

Chapter 7

Execution Plan Reference

Table 7-8
Queries

DBMS_XPLAN Functions and Parameters Relevant for Adaptive

Functions

Notes

DISPLAY_PLAN

The FORMAT argument supports the modifier ADAPTIVE.
When you specify ADAPTIVE, the output includes the default plan. For
each dynamic subplan, the plan shows a list of the row sources from
the original that may be replaced, and the row sources that would
replace them.
If the format argument specifies the outline display, then the function
displays the hints for each option in the dynamic subplan. If the plan is
not an adaptive query plan, then the function displays the default plan.
When you do not specify ADAPTIVE, the plan is shown as-is, but with
additional comments in the Note section that show any row sources
that are dynamic.

DISPLAY_CURSOR

The FORMAT argument supports the modifier ADAPTIVE.
When you specify ADAPTIVE, the output includes:
•

•
•
•

The final plan. If the execution has not completed, then the output
shows the current plan. This section also includes notes about
run-time optimizations that affect the plan.
Recommended plan. In reporting mode, the output includes the
plan that would be chosen based on execution statistics.
Dynamic plan. The output summarizes the portions of the plan
that differ from the default plan chosen by the optimizer.
Reoptimization. The output displays the plan that would be chosen
on a subsequent execution because of reoptimization.

See Also:
Oracle Database PL/SQL Packages and Types Reference to learn more
about DBMS_XPLAN

7-40

Part IV
SQL Operators: Access Paths and Joins
A row source is a set of rows returned by a step in the execution plan. A SQL
operator acts on a row source.
A unary operator acts on one input, as with access paths. A binary operator acts on
two outputs, as with joins.
This part contains the following chapters:
•

Optimizer Access Paths
An access path is a technique used by a query to retrieve rows from a row
source.

•

Joins
Oracle Database provides several optimizations for joining row sets.

8
Optimizer Access Paths
An access path is a technique used by a query to retrieve rows from a row source.
This chapter contains the following topics:
•

Introduction to Access Paths
A row source is a set of rows returned by a step in an execution plan. A row
source can be a table, view, or result of a join or grouping operation.

•

Table Access Paths
A table is the basic unit of data organization in an Oracle database.

•

B-Tree Index Access Paths
An index is an optional structure, associated with a table or table cluster, that can
sometimes speed data access.

•

Bitmap Index Access Paths
Bitmap indexes combine the indexed data with a rowid range.

•

Table Cluster Access Paths
A table cluster is a group of tables that share common columns and store related
data in the same blocks. When tables are clustered, a single data block can
contain rows from multiple tables.

8.1 Introduction to Access Paths
A row source is a set of rows returned by a step in an execution plan. A row source
can be a table, view, or result of a join or grouping operation.
A unary operation such as an access path, which is a technique used by a query to
retrieve rows from a row source, accepts a single row source as input. For example, a
full table scan is the retrieval of rows of a single row source. In contrast, a join
operation is binary and receives inputs from two row sources
The database uses different access paths for different relational data structures. The
following table summarizes common access paths for the major data structures.
Table 8-1

Data Structures and Access Paths

Access Path

Heap-Organized
Tables

Full Table Scans

x

Table Access by Rowid

x

Sample Table Scans

x

B-Tree Indexes
and IOTs

Index Unique Scans

x

Index Range Scans

x

Index Full Scans

x

Index Fast Full Scans

x

Bitmap Indexes

Table Clusters

8-1

Chapter 8

Table Access Paths

Table 8-1

(Cont.) Data Structures and Access Paths

Access Path

Heap-Organized
Tables

B-Tree Indexes
and IOTs

Index Skip Scans

x

Index Join Scans

x

Bitmap Indexes

Bitmap Index Single Value

x

Bitmap Index Range Scans

x

Bitmap Merge

x

Bitmap Index Range Scans

x

Table Clusters

Cluster Scans

x

Hash Scans

x

The optimizer considers different possible execution plans, and then assigns each
plan a cost. The optimizer chooses the plan with the lowest cost. In general, index
access paths are more efficient for statements that retrieve a small subset of table
rows, whereas full table scans are more efficient when accessing a large portion of a
table.

See Also:
•

"Joins"

•

"Cost-Based Optimization"

•

Oracle Database Concepts for an overview of these structures

8.2 Table Access Paths
A table is the basic unit of data organization in an Oracle database.
Relational tables are the most common table type. Relational tables have with the
following organizational characteristics:
•

A heap-organized table does not store rows in any particular order.

•

An index-organized table orders rows according to the primary key values.

•

An external table is a read-only table whose metadata is stored in the database
but whose data is stored outside the database.

This section explains optimizer access paths for heap-organized tables, and contains
the following topics:
•

About Heap-Organized Table Access
By default, a table is organized as a heap, which means that the database places
rows where they fit best rather than in a user-specified order.

8-2

Chapter 8

Table Access Paths

•

Full Table Scans
A full table scan reads all rows from a table, and then filters out those rows that
do not meet the selection criteria.

•

Table Access by Rowid
A rowid is an internal representation of the storage location of data.

•

Sample Table Scans
A sample table scan retrieves a random sample of data from a simple table or a
complex SELECT statement, such as a statement involving joins and views.

•

In-Memory Table Scans
An In-Memory scan retrieves rows from the In-Memory Column Store (IM column
store).

See Also:
•

Oracle Database Concepts for an overview of tables

•

Oracle Database Administrator’s Guide to learn how to manage tables

8.2.1 About Heap-Organized Table Access
By default, a table is organized as a heap, which means that the database places rows
where they fit best rather than in a user-specified order.
As users add rows, the database places the rows in the first available free space in the
data segment. Rows are not guaranteed to be retrieved in the order in which they were
inserted.
This section contains the following topics:
•

Row Storage in Data Blocks and Segments: A Primer
The database stores rows in data blocks. In tables, the database can write a row
anywhere in the bottom part of the block. Oracle Database uses the block
overhead, which contains the row directory and table directory, to manage the
block itself.

•

Importance of Rowids for Row Access
Every row in a heap-organized table has a rowid unique to this table that
corresponds to the physical address of a row piece. A rowid is a 10-byte physical
address of a row.

•

Direct Path Reads
In a direct path read, the database reads buffers from disk directly into the PGA,
bypassing the SGA entirely.

8.2.1.1 Row Storage in Data Blocks and Segments: A Primer
The database stores rows in data blocks. In tables, the database can write a row
anywhere in the bottom part of the block. Oracle Database uses the block overhead,
which contains the row directory and table directory, to manage the block itself.
An extent is made up of logically contiguous data blocks. The blocks may not be
physically contiguous on disk. A segment is a set of extents that contains all the data
for a logical storage structure within a tablespace. For example, Oracle Database

8-3

Chapter 8

Table Access Paths

allocates one or more extents to form the data segment for a table. The database also
allocates one or more extents to form the index segment for a table.
By default, the database uses automatic segment space management (ASSM) for
permanent, locally managed tablespaces. When a session first inserts data into a
table, the database formats a bitmap block. The bitmap tracks the blocks in the
segment. The database uses the bitmap to find free blocks and then formats each
block before writing to it. ASSM spread out inserts among blocks to avoid concurrency
issues.
The high water mark (HWM) is the point in a segment beyond which data blocks are
unformatted and have never been used. Below the HWM, a block may be formatted
and written to, formatted and empty, or unformatted. The low high water mark (low
HWM) marks the point below which all blocks are known to be formatted because they
either contain data or formerly contained data.
During a full table scan, the database reads all blocks up to the low HWM, which are
known to be formatted, and then reads the segment bitmap to determine which blocks
between the HWM and low HWM are formatted and safe to read. The database knows
not to read past the HWM because these blocks are unformatted.

See Also:
Oracle Database Concepts to learn about data block storage

8.2.1.2 Importance of Rowids for Row Access
Every row in a heap-organized table has a rowid unique to this table that corresponds
to the physical address of a row piece. A rowid is a 10-byte physical address of a row.
The rowid points to a specific file, block, and row number. For example, in the rowid
AAAPecAAFAAAABSAAA, the final AAA represents the row number. The row number is an

index into a row directory entry. The row directory entry contains a pointer to the
location of the row on the block.
The database can sometimes move a row in the bottom part of the block. For
example, if row movement is enabled, then the row can move because of partition key
updates, Flashback Table operations, shrink table operations, and so on. If the
database moves a row within a block, then the database updates the row directory
entry to modify the pointer. The rowid stays constant.
Oracle Database uses rowids internally for the construction of indexes. For example,
each key in a B-tree index is associated with a rowid that points to the address of the
associated row. Physical rowids provide the fastest possible access to a table row,
enabling the database to retrieve a row in as little as a single I/O.

See Also:
Oracle Database Concepts to learn about rowids

8-4

Chapter 8

Table Access Paths

8.2.1.3 Direct Path Reads
In a direct path read, the database reads buffers from disk directly into the PGA,
bypassing the SGA entirely.
The following figure shows the difference between scattered and sequential reads,
which store buffers in the SGA, and direct path reads.

Figure 8-1

Direct Path Reads

Database Buffer
Cache

Database Buffer
Cache

SGA Buffer Cache

SGA Buffer Cache

Process PGA
Sort Area

Hash Area

Session
Memory

Bitmap Merge
Area

Persistent
Area

Runtime
Area

Direct path
read

DB File
Sequential Read

DB File
Scattered Read

Direct Path
Read

Situations in which Oracle Database may perform direct path reads include:
•

Execution of a CREATE TABLE AS SELECT statement

•

Execution of an ALTER REBUILD or ALTER MOVE statement

•

Reads from a temporary tablespace

•

Parallel queries

•

Reads from a LOB segment

See Also:
Oracle Database Performance Tuning Guide to learn about wait events for
direct path reads

8-5

Chapter 8

Table Access Paths

8.2.2 Full Table Scans
A full table scan reads all rows from a table, and then filters out those rows that do
not meet the selection criteria.
This section contains the following topics:
•

When the Optimizer Considers a Full Table Scan
In general, the optimizer chooses a full table scan when it cannot use a different
access path, or another usable access path is higher cost.

•

How a Full Table Scan Works
In a full table scan, the database sequentially reads every formatted block under
the high water mark. The database reads each block only once.

•

Full Table Scan: Example
This example scans the hr.employees table.

8.2.2.1 When the Optimizer Considers a Full Table Scan
In general, the optimizer chooses a full table scan when it cannot use a different
access path, or another usable access path is higher cost.
The following table shows typical reasons for choosing a full table scan.
Table 8-2

Typical Reasons for a Full Table Scan

Reason

Explanation

To Learn More

No index exists.

If no index exists, then the optimizer
uses a full table scan.

Oracle Database
Concepts

The query predicate
applies a function to the
indexed column.

Unless the index is a function-based "Guidelines for Using
index, the database indexes the
Function-Based Indexes
values of the column, not the values
for Performance"
of the column with the function
applied. A typical application-level
mistake is to index a character
column, such as char_col, and then
query the column using syntax such
as WHERE char_col=1. The database
implicitly applies a TO_NUMBER function
to the constant number 1, which
prevents use of the index.

A SELECT COUNT(*) query
is issued, and an index
exists, but the indexed
column contains nulls.

The optimizer cannot use the index to "B-Tree Indexes and
count the number of table rows
Nulls"
because the index cannot contain null
entries.

The query predicate does For example, an index might exist on "Index Skip Scans"
not use the leading edge of employees(first_name,last_name).
a B-tree index.
If a user issues a query with the
predicate WHERE last_name='KING',
then the optimizer may not choose an
index because column first_name is
not in the predicate. However, in this
situation the optimizer may choose to
use an index skip scan.

8-6

Chapter 8

Table Access Paths

Table 8-2

(Cont.) Typical Reasons for a Full Table Scan

Reason

Explanation

To Learn More

The query is unselective.

If the optimizer determines that the
query requires most of the blocks in
the table, then it uses a full table
scan, even though indexes are
available. Full table scans can use
larger I/O calls. Making fewer large
I/O calls is cheaper than making
many smaller calls.

"Selectivity"

The table statistics are
stale.

For example, a table was small, but
"Introduction to Optimizer
now has grown large. If the table
Statistics"
statistics are stale and do not reflect
the current size of the table, then the
optimizer does not know that an index
is now most efficient than a full table
scan.

The table is small.

If a table contains fewer than n blocks Oracle Database
under the high water mark, where n
Reference
equals the setting for the
DB_FILE_MULTIBLOCK_READ_COUNT
initialization parameter, then a full
table scan may be cheaper than an
index range scan. The scan may be
less expensive regardless of the
fraction of tables being accessed or
indexes present.

The table has a high
degree of parallelism.

A high degree of parallelism for a
table skews the optimizer toward full
table scans over range scans. Query
the value in the ALL_TABLES.DEGREE
column to determine the degree of
parallelism.

The query uses a full table
scan hint.

The hint FULL(table alias) instructs Oracle Database SQL
the optimizer to use a full table scan. Language Reference

Oracle Database
Reference

8.2.2.2 How a Full Table Scan Works
In a full table scan, the database sequentially reads every formatted block under the
high water mark. The database reads each block only once.
The following graphic depicts a scan of a table segment, showing how the scan skips
unformatted blocks below the high water mark.

8-7

Chapter 8

Table Access Paths

Figure 8-2

High Water Mark
Low HWM

Sequential
Read

HWM

Used

Never Used,
Unformatted

Because the blocks are adjacent, the database can speed up the scan by making I/O
calls larger than a single block, known as a multiblock read. The size of a read call
ranges from one block to the number of blocks specified by the
DB_FILE_MULTIBLOCK_READ_COUNT initialization parameter. For example, setting this
parameter to 4 instructs the database to read up to 4 blocks in a single call.
The algorithms for caching blocks during full table scans are complex. For example,
the database caches blocks differently depending on whether tables are small or large.

See Also:
•

"Table 19-2"

•

Oracle Database Concepts for an overview of the default caching mode

•

Oracle Database Reference to learn about the
DB_FILE_MULTIBLOCK_READ_COUNT initialization parameter

8.2.2.3 Full Table Scan: Example
This example scans the hr.employees table.
The following statement queries monthly salaries over $4000:
SELECT salary
FROM hr.employees
WHERE salary > 4000;

Example 8-1

Full Table Scan

The following plan was retrieved using the DBMS_XPLAN.DISPLAY_CURSOR function.
Because no index exists on the salary column, the optimizer cannot use an index
range scan, and so uses a full table scan.

8-8

Chapter 8

Table Access Paths

SQL_ID 54c20f3udfnws, child number 0
------------------------------------select salary from hr.employees where salary > 4000
Plan hash value: 3476115102
------------------------------------------------------------------------------| Id | Operation
| Name
| Rows | Bytes | Cost (%CPU)| Time
|
------------------------------------------------------------------------------| 0 | SELECT STATEMENT |
|
|
|
3 (100)|
|
|* 1 | TABLE ACCESS FULL| EMPLOYEES |
98 | 6762 |
3 (0)| 00:00:01 |
------------------------------------------------------------------------------Predicate Information (identified by operation id):
--------------------------------------------------1 - filter("SALARY">4000)

8.2.3 Table Access by Rowid
A rowid is an internal representation of the storage location of data.
The rowid of a row specifies the data file and data block containing the row and the
location of the row in that block. Locating a row by specifying its rowid is the fastest
way to retrieve a single row because it specifies the exact location of the row in the
database.

Note:
Rowids can change between versions. Accessing data based on position is
not recommended because rows can move.

This section contains the following topics:
•

When the Optimizer Chooses Table Access by Rowid
In most cases, the database accesses a table by rowid after a scan of one or more
indexes.

•

How Table Access by Rowid Works
To access a table by rowid, the database performs multiple steps.

•

Table Access by Rowid: Example
This example demonstrates rowid access of the hr.employees table.

See Also:
Oracle Database Development Guide to learn more about rowids

8.2.3.1 When the Optimizer Chooses Table Access by Rowid
In most cases, the database accesses a table by rowid after a scan of one or more
indexes.

8-9

Chapter 8

Table Access Paths

However, table access by rowid need not follow every index scan. If the index contains
all needed columns, then access by rowid might not occur.
Related Topics
•

Index Fast Full Scans
An index fast full scan reads the index blocks in unsorted order, as they exist on
disk. This scan does not use the index to probe the table, but reads the index
instead of the table, essentially using the index itself as a table.

8.2.3.2 How Table Access by Rowid Works
To access a table by rowid, the database performs multiple steps.
The database does the following:
1.

Obtains the rowids of the selected rows, either from the statement WHERE clause or
through an index scan of one or more indexes
Table access may be needed for columns in the statement not present in the
index.

2.

Locates each selected row in the table based on its rowid

8.2.3.3 Table Access by Rowid: Example
This example demonstrates rowid access of the hr.employees table.
Assume that you run the following query:
SELECT *
FROM employees
WHERE employee_id > 190;

Step 2 of the following plan shows a range scan of the emp_emp_id_pk index on the
hr.employees table. The database uses the rowids obtained from the index to find the
corresponding rows from the employees table, and then retrieve them. The BATCHED
access shown in Step 1 means that the database retrieves a few rowids from the
index, and then attempts to access rows in block order to improve the clustering and
reduce the number of times that the database must access a block.
-------------------------------------------------------------------------------|Id| Operation
| Name
|Rows|Bytes|Cost(%CPU)|Time|
-------------------------------------------------------------------------------| 0| SELECT STATEMENT
|
| |
|2(100)|
|
| 1| TABLE ACCESS BY INDEX ROWID BATCHED|EMPLOYEES
|16|1104|2 (0)|00:00:01|
|*2| INDEX RANGE SCAN
|EMP_EMP_ID_PK|16|
|1 (0)|00:00:01|
-------------------------------------------------------------------------------Predicate Information (identified by operation id):
--------------------------------------------------2 - access("EMPLOYEE_ID">190)

8.2.4 Sample Table Scans
A sample table scan retrieves a random sample of data from a simple table or a
complex SELECT statement, such as a statement involving joins and views.

8-10

Chapter 8

Table Access Paths

This section contains the following topics:
•

When the Optimizer Chooses a Sample Table Scan
The database uses a sample table scan when a statement FROM clause includes
the SAMPLE keyword.

•

Sample Table Scans: Example
This example uses a sample table scan to access 1% of the employees table,
sampling by blocks instead of rows.

8.2.4.1 When the Optimizer Chooses a Sample Table Scan
The database uses a sample table scan when a statement FROM clause includes the
SAMPLE keyword.

The SAMPLE clause has the following forms:
•

SAMPLE (sample_percent)

The database reads a specified percentage of rows in the table to perform a
sample table scan.
•

SAMPLE BLOCK (sample_percent)

The database reads a specified percentage of table blocks to perform a sample
table scan.
The sample_percent specifies the percentage of the total row or block count to include
in the sample. The value must be in the range .000001 up to, but not including, 100.
This percentage indicates the probability of each row, or each cluster of rows in block
sampling, being selected for the sample. It does not mean that the database retrieves
exactly sample_percent of the rows.

Note:
Block sampling is possible only during full table scans or index fast full
scans. If a more efficient execution path exists, then the database does not
sample blocks. To guarantee block sampling for a specific table or index, use
the FULL or INDEX_FFS hint.

See Also:
•

"Influencing the Optimizer with Hints"

•

Oracle Database SQL Language Reference to learn about the SAMPLE
clause

8.2.4.2 Sample Table Scans: Example
This example uses a sample table scan to access 1% of the employees table, sampling
by blocks instead of rows.

8-11

Chapter 8

Table Access Paths

Example 8-2

Sample Table Scan

SELECT * FROM hr.employees SAMPLE BLOCK (1);

The EXPLAIN PLAN output for this statement might look as follows:
------------------------------------------------------------------------| Id | Operation
| Name
| Rows | Bytes | Cost (%CPU)|
------------------------------------------------------------------------| 0 | SELECT STATEMENT
|
|
1 |
68 |
3 (34)|
| 1 | TABLE ACCESS SAMPLE | EMPLOYEES |
1 |
68 |
3 (34)|
-------------------------------------------------------------------------

8.2.5 In-Memory Table Scans
An In-Memory scan retrieves rows from the In-Memory Column Store (IM column
store).
The IM column store is an optional SGA area that stores copies of tables and
partitions in a special columnar format optimized for rapid scans.
This section contains the following topics:
•

When the Optimizer Chooses an In-Memory Table Scan
The optimizer cost model is aware of the content of the IM column store.

•

In-Memory Query Controls
You can control In-Memory queries using initialization parameters.

•

In-Memory Table Scans: Example
This example shows an execution plan that includes the TABLE ACCESS INMEMORY
operation.

See Also:
Oracle Database In-Memory Guide for an introduction to the IM column store

8.2.5.1 When the Optimizer Chooses an In-Memory Table Scan
The optimizer cost model is aware of the content of the IM column store.
When a user executes a query that references a table in the IM column store, the
optimizer calculates the cost of all possible access methods—including the In-Memory
table scan—and selects the access method with the lowest cost.

8.2.5.2 In-Memory Query Controls
You can control In-Memory queries using initialization parameters.
The following database initialization parameters affect the In-Memory features:
•

INMEMORY_QUERY

This parameter enables or disables In-Memory queries for the database at the
session or system level. This parameter is helpful when you want to test workloads
with and without the use of the IM column store.

8-12

Chapter 8

Table Access Paths

•

OPTIMIZER_INMEMORY_AWARE

This parameter enables (TRUE) or disables (FALSE) all of the In-Memory
enhancements made to the optimizer cost model, table expansion, bloom filters,
and so on. Setting the parameter to FALSE causes the optimizer to ignore the InMemory property of tables during the optimization of SQL statements.
•

OPTIMIZER_FEATURES_ENABLE

When set to values lower than 12.1.0.2, this parameter has the same effect as
setting OPTIMIZER_INMEMORY_AWARE to FALSE.
To enable or disable In-Memory queries, you can specify the INMEMORY or NO_INMEMORY
hints, which are the per-query equivalent of the INMEMORY_QUERY initialization parameter.
If a SQL statement uses the INMEMORY hint, but the object it references is not already
loaded in the IM column store, then the database does not wait for the object to be
populated in the IM column store before executing the statement. However, initial
access of the object triggers the object population in the IM column store.

See Also:
•

Oracle Database Reference to learn more about the INMEMORY_QUERY,
OPTIMIZER_INMEMORY_AWARE, and OPTIMIZER_FEATURES_ENABLE initialization
parameters

•

Oracle Database SQL Language Reference to learn more about the
INMEMORY hints

8.2.5.3 In-Memory Table Scans: Example
This example shows an execution plan that includes the TABLE ACCESS INMEMORY
operation.
The following example shows a query of the oe.product_information table, which has
been altered with the INMEMORY HIGH option.
Example 8-3

In-Memory Table Scan
SELECT *
FROM oe.product_information
WHERE list_price > 10
ORDER BY product_id

The plan for this statement might look as follows, with the INMEMORY keyword in Step 2
indicating that some or all of the object was accessed from the IM column store:
SQL> SELECT * FROM TABLE(DBMS_XPLAN.DISPLAY_CURSOR);
SQL_ID 2mb4h57x8pabw, child number 0
------------------------------------select * from oe.product_information where list_price > 10 order byproduct_id
Plan hash value: 2256295385
-------------------------------------------------------------------------------------------|Id| Operation
| Name
|Rows|Bytes |TempSpc|Cost(%CPU)|Time|
-------------------------------------------------------------------------------------------| 0| SELECT STATEMENT
|
|
|
|
|21 (100)|
|

8-13

Chapter 8

B-Tree Index Access Paths

| 1| SORT ORDER BY
|
| 285| 62415|82000|21 (5)|00:00:01|
|*2|
TABLE ACCESS INMEMORY FULL| PRODUCT_INFORMATION | 285| 62415|
| 5 (0)|00:00:01|
-------------------------------------------------------------------------------------------Predicate Information (identified by operation id):
--------------------------------------------------2 - inmemory("LIST_PRICE">10)
filter("LIST_PRICE">10)

8.3 B-Tree Index Access Paths
An index is an optional structure, associated with a table or table cluster, that can
sometimes speed data access.
By creating an index on one or more columns of a table, you gain the ability in some
cases to retrieve a small set of randomly distributed rows from the table. Indexes are
one of many means of reducing disk I/O.
This section contains the following topics:
•

About B-Tree Index Access
B-trees, short for balanced trees, are the most common type of database index.

•

Index Unique Scans
An index unique scan returns at most 1 rowid.

•

Index Range Scans
An index range scan is an ordered scan of values.

•

Index Full Scans
An index full scan reads the entire index in order. An index full scan can eliminate
a separate sorting operation because the data in the index is ordered by index
key.

•

Index Fast Full Scans
An index fast full scan reads the index blocks in unsorted order, as they exist on
disk. This scan does not use the index to probe the table, but reads the index
instead of the table, essentially using the index itself as a table.

•

Index Skip Scans
An index skip scan occurs when the initial column of a composite index is
"skipped" or not specified in the query.

•

Index Join Scans
An index join scan is a hash join of multiple indexes that together return all
columns requested by a query. The database does not need to access the table
because all data is retrieved from the indexes.

See Also:
•

Oracle Database Concepts for an overview of indexes

•

Oracle Database Administrator’s Guide to learn how to manage indexes

8-14

Chapter 8

B-Tree Index Access Paths

8.3.1 About B-Tree Index Access
B-trees, short for balanced trees, are the most common type of database index.
A B-tree index is an ordered list of values divided into ranges. By associating a key
with a row or range of rows, B-trees provide excellent retrieval performance for a wide
range of queries, including exact match and range searches.
This section contains the following topics:
•

B-Tree Index Structure
A B-tree index has two types of blocks: branch blocks for searching and leaf
blocks that store values.

•

How Index Storage Affects Index Scans
Bitmap index blocks can appear anywhere in the index segment.

•

Unique and Nonunique Indexes
In a nonunique index, the database stores the rowid by appending it to the key as
an extra column. The entry adds a length byte to make the key unique.

•

B-Tree Indexes and Nulls
B-tree indexes never store completely null keys, which is important for how the
optimizer chooses access paths. A consequence of this rule is that single-column
B-tree indexes never store nulls.

8.3.1.1 B-Tree Index Structure
A B-tree index has two types of blocks: branch blocks for searching and leaf blocks
that store values.
The following graphic illustrates the logical structure of a B-tree index. Branch blocks
store the minimum key prefix needed to make a branching decision between two keys.
The leaf blocks contain every indexed data value and a corresponding rowid used to
locate the actual row. Each index entry is sorted by (key, rowid). The leaf blocks are
doubly linked.

8-15

Chapter 8

B-Tree Index Access Paths

Figure 8-3

B-Tree Index Structure

Branch Blocks
0..40
41..80
81..120
....
200..250

0..10
11..19
20..25
....
32..40

41..48
49..53
54..65
....
78..80

...

200..209
210..220
221..228
....
246..250

Leaf Blocks

0,rowid
0,rowid
....
10,rowid

11,rowid
11,rowid
12,rowid
....
19,rowid

...

221,rowid
222,rowid
223,rowid
....
228,rowid

...

246,rowid
248,rowid
248,rowid
....
250,rowid

8.3.1.2 How Index Storage Affects Index Scans
Bitmap index blocks can appear anywhere in the index segment.
Figure 8-3 shows the leaf blocks as adjacent to each other. For example, the 1-10
block is next to and before the 11-19 block. This sequencing illustrates the linked lists
that connect the index entries. However, index blocks need not be stored in order
within an index segment. For example, the 246-250 block could appear anywhere in the
segment, including directly before the 1-10 block. For this reason, ordered index scans
must perform single-block I/O. The database must read an index block to determine
which index block it must read next.
The index block body stores the index entries in a heap, just like table rows. For
example, if the value 10 is inserted first into a table, then the index entry with key 10
might be inserted at the bottom of the index block. If 0 is inserted next into the table,
then the index entry for key 0 might be inserted on top of the entry for 10. Thus, the
index entries in the block body are not stored in key order. However, within the index
block, the row header stores records in key order. For example, the first record in the
header points to the index entry with key 0, and so on sequentially up to the record that
points to the index entry with key 10. Thus, index scans can read the row header to
determine where to begin and end range scans, avoiding the necessity of reading
every entry in the block.

8-16

Chapter 8

B-Tree Index Access Paths

See Also:
Oracle Database Concepts to learn about index blocks

8.3.1.3 Unique and Nonunique Indexes
In a nonunique index, the database stores the rowid by appending it to the key as an
extra column. The entry adds a length byte to make the key unique.
For example, the first index key in the nonunique index shown in Figure 8-3 is the pair
0,rowid and not simply 0. The database sorts the data by index key values and then by

rowid ascending. For example, the entries are sorted as follows:
0,AAAPvCAAFAAAAFaAAa
0,AAAPvCAAFAAAAFaAAg
0,AAAPvCAAFAAAAFaAAl
2,AAAPvCAAFAAAAFaAAm

In a unique index, the index key does not include the rowid. The database sorts the
data only by the index key values, such as 0, 1, 2, and so on.

See Also:
Oracle Database Concepts for an overview of unique and nonunique indexes

8.3.1.4 B-Tree Indexes and Nulls
B-tree indexes never store completely null keys, which is important for how the
optimizer chooses access paths. A consequence of this rule is that single-column Btree indexes never store nulls.
An example helps illustrate. The hr.employees table has a primary key index on
employee_id, and a unique index on department_id. The department_id column can
contain nulls, making it a nullable column, but the employee_id column cannot.
SQL> SELECT COUNT(*) FROM employees WHERE department_id IS NULL;
COUNT(*)
---------1
SQL> SELECT COUNT(*) FROM employees WHERE employee_id IS NULL;
COUNT(*)
---------0

The following example shows that the optimizer chooses a full table scan for a query
of all department IDs in hr.employees. The optimizer cannot use the index on
employees.department_id because the index is not guaranteed to include entries for
every row in the table.

8-17

Chapter 8

B-Tree Index Access Paths

SQL> EXPLAIN PLAN FOR SELECT department_id FROM employees;
Explained.
SQL> SELECT PLAN_TABLE_OUTPUT FROM TABLE(DBMS_XPLAN.DISPLAY());
PLAN_TABLE_OUTPUT
-------------------------------------------------------------------------------Plan hash value: 3476115102
------------------------------------------------------------------------------| Id | Operation
| Name
| Rows | Bytes | Cost (%CPU)| Time
|
------------------------------------------------------------------------------| 0 | SELECT STATEMENT |
| 107 | 321 |
2 (0)| 00:00:01 |
| 1 | TABLE ACCESS FULL| EMPLOYEES | 107 | 321 |
2 (0)| 00:00:01 |
------------------------------------------------------------------------------8 rows selected.

The following example shows the optimizer can use the index on department_id for a
query of a specific department ID because all non-null rows are indexed.
SQL> EXPLAIN PLAN FOR SELECT department_id FROM employees WHERE department_id=10;
Explained.
SQL> SELECT PLAN_TABLE_OUTPUT FROM TABLE(DBMS_XPLAN.DISPLAY());
PLAN_TABLE_OUTPUT
-------------------------------------------------------------------------------Plan hash value: 67425611
-------------------------------------------------------------------------------| Id| Operation
| Name
| Rows |Bytes| Cost (%CPU)| Time
|
-------------------------------------------------------------------------------PLAN_TABLE_OUTPUT
-------------------------------------------------------------------------------| 0 | SELECT STATEMENT |
| 1 | 3 |
1 (0)| 00:0 0:01|
|*1 | INDEX RANGE SCAN| EMP_DEPARTMENT_IX | 1 | 3 |
1 (0)| 00:0 0:01|
-------------------------------------------------------------------------------Predicate Information (identified by operation id):
PLAN_TABLE_OUTPUT
-------------------------------------------------------------------------------1 - access("DEPARTMENT_ID"=10)

The following example shows that the optimizer chooses an index scan when the
predicate excludes null values:
SQL> EXPLAIN PLAN FOR SELECT department_id FROM employees
WHERE department_id IS NOT NULL;
Explained.
SQL> SELECT PLAN_TABLE_OUTPUT FROM TABLE(DBMS_XPLAN.DISPLAY());
PLAN_TABLE_OUTPUT
-------------------------------------------------------------------------------Plan hash value: 1590637672

8-18

Chapter 8

B-Tree Index Access Paths

-------------------------------------------------------------------------------| Id| Operation
| Name
| Rows|Bytes| Cost (%CPU)| Time
|
-------------------------------------------------------------------------------| 0 | SELECT STATEMENT |
| 106 | 318 |
1 (0)| 00:0 0:01 |
|*1 | INDEX FULL SCAN | EMP_DEPARTMENT_IX | 106 | 318 |
1 (0)| 00:0 0:01 |
-------------------------------------------------------------------------------Predicate Information (identified by operation id):
PLAN_TABLE_OUTPUT
-------------------------------------------------------------------------------1 - filter("DEPARTMENT_ID" IS NOT NULL)

8.3.2 Index Unique Scans
An index unique scan returns at most 1 rowid.
This section contains the following topics:
•

When the Optimizer Considers Index Unique Scans
An index unique scan requires an equality predicate.

•

How Index Unique Scans Work
The scan searches the index in order for the specified key. An index unique scan
stops processing as soon as it finds the first record because no second record is
possible. The database obtains the rowid from the index entry, and then retrieves
the row specified by the rowid.

•

Index Unique Scans: Example
This example uses a unique scan to retrieve a row from the products table.

8.3.2.1 When the Optimizer Considers Index Unique Scans
An index unique scan requires an equality predicate.
Specifically, the database performs a unique scan only when a query predicate
references all columns in a unique index key using an equality operator, such as WHERE
prod_id=10.
A unique or primary key constraint is insufficient by itself to produce an index unique
scan because a non-unique index on the column may already exist. Consider the
following example, which creates the t_table table and then creates a non-unique
index on numcol:
SQL> CREATE TABLE t_table(numcol INT);
SQL> CREATE INDEX t_table_idx ON t_table(numcol);
SQL> SELECT UNIQUENESS FROM USER_INDEXES WHERE INDEX_NAME = 'T_TABLE_IDX';
UNIQUENES
--------NONUNIQUE

The following code creates a primary key constraint on a column with a non-unique
index, resulting in an index range scan rather than an index unique scan:
SQL> ALTER TABLE t_table ADD CONSTRAINT t_table_pk PRIMARY KEY(numcol);
SQL> SET AUTOTRACE TRACEONLY EXPLAIN
SQL> SELECT * FROM t_table WHERE numcol = 1;

8-19

Chapter 8

B-Tree Index Access Paths

Execution Plan
---------------------------------------------------------Plan hash value: 868081059
-------------------------------------------------------------------------------| Id | Operation
| Name
| Rows | Bytes | Cost (%CPU)| Time
|
-------------------------------------------------------------------------------| 0 | SELECT STATEMENT |
|
1 |
13 |
1 (0)| 00:00:01 |
|* 1 | INDEX RANGE SCAN| T_TABLE_IDX |
1 |
13 |
1 (0)| 00:00:01 |
-------------------------------------------------------------------------------Predicate Information (identified by operation id):
--------------------------------------------------1 - access("NUMCOL"=1)

You can use the INDEX(alias index_name) hint to specify the index to use, but not a
specific type of index access path.

See Also:
•
•

Oracle Database Concepts for more details on index structures and for
detailed information on how a B-tree is searched
Oracle Database SQL Language Reference to learn more about the
INDEX hint

8.3.2.2 How Index Unique Scans Work
The scan searches the index in order for the specified key. An index unique scan
stops processing as soon as it finds the first record because no second record is
possible. The database obtains the rowid from the index entry, and then retrieves the
row specified by the rowid.
The following figure illustrates an index unique scan. The statement requests the
record for product ID 19 in the prod_id column, which has a primary key index.

8-20

Chapter 8

B-Tree Index Access Paths

Figure 8-4

Index Unique Scan

Branch Blocks
0..40
41..80
81..120
....
200..250

0..10
11..19
20..25
....
32..40

41..48
49..53
54..65
....
78..80

...

200..209
210..220
221..228
....
246..250

Leaf Blocks

0,rowid
1,rowid
....
10,rowid

11,rowid
12,rowid
13,rowid
....
19,rowid

...

221,rowid
222,rowid
223,rowid
....
228,rowid

...

246,rowid
247,rowid
248,rowid
....
250,rowid

8.3.2.3 Index Unique Scans: Example
This example uses a unique scan to retrieve a row from the products table.
The following statement queries the record for product 19 in the sh.products table:
SELECT *
FROM sh.products
WHERE prod_id = 19;

Because a primary key index exists on the products.prod_id column, and the WHERE
clause references all of the columns using an equality operator, the optimizer chooses
a unique scan:
SQL_ID 3ptq5tsd5vb3d, child number 0
------------------------------------select * from sh.products where prod_id = 19
Plan hash value: 4047888317
-------------------------------------------------------------------------------| Id| Operation
| Name
|Rows|Bytes|Cost (%CPU)|Time |
-------------------------------------------------------------------------------| 0| SELECT STATEMENT
|
|
|
| 1 (100)|
|
| 1| TABLE ACCESS BY INDEX ROWID| PRODUCTS
| 1 | 173 | 1 (0)| 00:00:01|
|* 2|
INDEX UNIQUE SCAN
| PRODUCTS_PK | 1 |
| 0 (0)|
|

8-21

Chapter 8

B-Tree Index Access Paths

-------------------------------------------------------------------------------Predicate Information (identified by operation id):
--------------------------------------------------2 - access("PROD_ID"=19)

8.3.3 Index Range Scans
An index range scan is an ordered scan of values.
The range in the scan can be bounded on both sides, or unbounded on one or both
sides. The optimizer typically chooses a range scan for queries with high selectivity.
By default, the database stores indexes in ascending order, and scans them in the
same order. For example, a query with the predicate department_id >= 20 uses a range
scan to return rows sorted by index keys 20, 30, 40, and so on. If multiple index entries
have identical keys, then the database returns them in ascending order by rowid, so
that 0,AAAPvCAAFAAAAFaAAa is followed by 0,AAAPvCAAFAAAAFaAAg, and so on.
An index range scan descending is identical to an index range scan except that the
database returns rows in descending order. Usually, the database uses a descending
scan when ordering data in a descending order, or when seeking a value less than a
specified value.
This section contains the following topics:
•

When the Optimizer Considers Index Range Scans
For an index range scan, multiple values must be possible for the index key.

•

How Index Range Scans Work
During an index range scan, Oracle Database proceeds from root to branch.

•

Index Range Scan: Example
This example retrieves a set of values from the employees table using an index
range scan.

•

Index Range Scan Descending: Example
This example uses an index to retrieve rows from the employees table in sorted
order.

8.3.3.1 When the Optimizer Considers Index Range Scans
For an index range scan, multiple values must be possible for the index key.
Specifically, the optimizer considers index range scans in the following circumstances:
•

One or more leading columns of an index are specified in conditions.
A condition specifies a combination of one or more expressions and logical
(Boolean) operators and returns a value of TRUE, FALSE, or UNKNOWN. Examples of
conditions include:
–

department_id = :id

–

department_id < :id

–

department_id > :id

–

AND combination of the preceding conditions for leading columns in the index,
such as department_id > :low AND department_id < :hi.

8-22

Chapter 8

B-Tree Index Access Paths

Note:
For the optimizer to consider a range scan, wild-card searches of the
form col1 LIKE '%ASD' must not be in a leading position.
•

0, 1, or more values are possible for an index key.

Tip:
If you require sorted data, then use the ORDER BY clause, and do not rely on an
index. If an index can satisfy an ORDER BY clause, then the optimizer uses this
option and thereby avoids a sort.

The optimizer considers an index range scan descending when an index can satisfy
an ORDER BY DESCENDING clause.
If the optimizer chooses a full table scan or another index, then a hint may be required
to force this access path. The INDEX(tbl_alias ix_name) and INDEX_DESC(tbl_alias
ix_name) hints instruct the optimizer to use a specific index.

See Also:
Oracle Database SQL Language Reference to learn more about the INDEX
and INDEX_DESC hints

8.3.3.2 How Index Range Scans Work
During an index range scan, Oracle Database proceeds from root to branch.
In general, the scan algorithm is as follows:
1.

Read the root block.

2.

Read the branch block.

3.

Alternate the following steps until all data is retrieved:
a.

Read a leaf block to obtain a rowid.

b.

Read a table block to retrieve a row.

Note:
In some cases, an index scan reads a set of index blocks, sorts the
rowids, and then reads a set of table blocks.
Thus, to scan the index, the database moves backward or forward through the leaf
blocks. For example, a scan for IDs between 20 and 40 locates the first leaf block that

8-23

Chapter 8

B-Tree Index Access Paths

has the lowest key value that is 20 or greater. The scan proceeds horizontally through
the linked list of leaf nodes until it finds a value greater than 40, and then stops.
The following figure illustrates an index range scan using ascending order. A
statement requests the employees records with the value 20 in the department_id
column, which has a nonunique index. In this example, 2 index entries for department
20 exist.
Figure 8-5

Index Range Scan

Branch Blocks
0..40
41..80
81..120
....
200..250

0..10
11..2

41..48
49..53
54..65
....
78..80

....
32..40

...

200..209
210..220
221..228
....
246..250

Leaf Blocks

0,rowid
0,rowid
....
10,rowid

11,rowid
11,rowid
12,rowid
....
20,rowid
20, rowid

...

221,rowid
222,rowid
223,rowid
....
228,rowid

...

246,rowid
248,rowid
248,rowid
....
250,rowid

8.3.3.3 Index Range Scan: Example
This example retrieves a set of values from the employees table using an index range
scan.
The following statement queries the records for employees in department 20 with
salaries greater than 1000:
SELECT
FROM
WHERE
AND

*
employees
department_id = 20
salary > 1000;

The preceding query has low cardinality (returns few rows), so the query uses the
index on the department_id column. The database scans the index, fetches the records
from the employees table, and then applies the salary > 1000 filter to these fetched
records to generate the result.

8-24

Chapter 8

B-Tree Index Access Paths

SQL_ID brt5abvbxw9tq, child number 0
------------------------------------SELECT * FROM employees WHERE department_id = 20 AND

salary > 1000

Plan hash value: 2799965532
------------------------------------------------------------------------------------------|Id | Operation
| Name
|Rows|Bytes|Cost(%CPU)| Time |
------------------------------------------------------------------------------------------| 0 | SELECT STATEMENT
|
|
|
| 2 (100)|
|
|*1 | TABLE ACCESS BY INDEX ROWID BATCHED| EMPLOYEES
| 2 | 138 | 2 (0)|00:00:01|
|*2 |
INDEX RANGE SCAN
| EMP_DEPARTMENT_IX| 2 |
| 1 (0)|00:00:01|
------------------------------------------------------------------------------------------Predicate Information (identified by operation id):
--------------------------------------------------1 - filter("SALARY">1000)
2 - access("DEPARTMENT_ID"=20)

8.3.3.4 Index Range Scan Descending: Example
This example uses an index to retrieve rows from the employees table in sorted order.
The following statement queries the records for employees in department 20 in
descending order:
SELECT *
FROM employees
WHERE department_id < 20
ORDER BY department_id DESC;

This preceding query has low cardinality, so the query uses the index on the
department_id column.
SQL_ID 8182ndfj1ttj6, child number 0
------------------------------------SELECT * FROM employees WHERE department_id < 20 ORDER BY department_id DESC
Plan hash value: 1681890450
-------------------------------------------------------------------------------|Id| Operation
| Name
|Rows|Bytes|Cost(%CPU)|Time |
-------------------------------------------------------------------------------| 0| SELECT STATEMENT
|
| | | 2 (100)|
|
| 1| TABLE ACCESS BY INDEX ROWID | EMPLOYEES
| 2|138| 2 (0)|00:00:01|
|*2|
INDEX RANGE SCAN DESCENDING| EMP_DEPARTMENT_IX | 2| | 1 (0)|00:00:01|
-------------------------------------------------------------------------------Predicate Information (identified by operation id):
--------------------------------------------------2 - access("DEPARTMENT_ID"<20)

The database locates the first index leaf block that contains the highest key value that
is 20 or less. The scan then proceeds horizontally to the left through the linked list of
leaf nodes. The database obtains the rowid from each index entry, and then retrieves
the row specified by the rowid.

8-25

Chapter 8

B-Tree Index Access Paths

8.3.4 Index Full Scans
An index full scan reads the entire index in order. An index full scan can eliminate a
separate sorting operation because the data in the index is ordered by index key.
This section contains the following topics:
•

When the Optimizer Considers Index Full Scans
The optimizer considers an index full scan in a variety of situations.

•

How Index Full Scans Work
The database reads the root block, and then navigates down the left hand side of
the index (or right if doing a descending full scan) until it reaches a leaf block.

•

Index Full Scans: Example
This example uses an index full scan to satisfy a query with an ORDER BY clause.

8.3.4.1 When the Optimizer Considers Index Full Scans
The optimizer considers an index full scan in a variety of situations.
The situations include the following:
•

A predicate references a column in the index. This column need not be the leading
column.

•

No predicate is specified, but all of the following conditions are met:

•

–

All columns in the table and in the query are in the index.

–

At least one indexed column is not null.

A query includes an ORDER BY on indexed non-nullable columns.

8.3.4.2 How Index Full Scans Work
The database reads the root block, and then navigates down the left hand side of the
index (or right if doing a descending full scan) until it reaches a leaf block.
Then the database reaches a leaf block, the scan proceeds across the bottom of the
index, one block at a time, in sorted order. The database uses single-block I/O rather
than multiblock I/O.
The following graphic illustrates an index full scan. A statement requests the
departments records ordered by department_id.

8-26

Chapter 8

B-Tree Index Access Paths

Figure 8-6

Index Full Scan

Branch Blocks
0..40
41..80
81..120
....
200..250

0..10
11..19
20..25
....
32..40

41..48
49..53
54..65
....
78..80

...

200..209
210..220
221..228
....
246..250

Leaf Blocks

0,rowid
1,rowid
....
10,rowid

11,rowid
12,rowid
13,rowid
....
19,rowid

...

221,rowid
222,rowid
223,rowid
....
228,rowid

...

246,rowid
247,rowid
248,rowid
....
250,rowid

8.3.4.3 Index Full Scans: Example
This example uses an index full scan to satisfy a query with an ORDER BY clause.
The following statement queries the ID and name for departments in order of
department ID:
SELECT department_id, department_name
FROM departments
ORDER BY department_id;

The following plan shows that the optimizer chose an index full scan:
SQL_ID 94t4a20h8what, child number 0
------------------------------------select department_id, department_name from departments order by department_id
Plan hash value: 4179022242
-------------------------------------------------------------------------------|Id | Operation
| Name
|Rows|Bytes|Cost(%CPU)|Time
|
-------------------------------------------------------------------------------| 0 | SELECT STATEMENT
|
|
|
| 2 (100)|
|
| 1 | TABLE ACCESS BY INDEX ROWID| DEPARTMENTS | 27 | 432 | 2 (0)| 00:00:01 |
| 2 |
INDEX FULL SCAN
| DEPT_ID_PK | 27 |
| 1 (0)| 00:00:01 |
--------------------------------------------------------------------------------

8-27

Chapter 8

B-Tree Index Access Paths

The database locates the first index leaf block, and then proceeds horizontally to the
right through the linked list of leaf nodes. For each index entry, the database obtains
the rowid from the entry, and then retrieves the table row specified by the rowid.
Because the index is sorted on department_id, the database avoids a separate
operation to sort the retrieved rows.

8.3.5 Index Fast Full Scans
An index fast full scan reads the index blocks in unsorted order, as they exist on
disk. This scan does not use the index to probe the table, but reads the index instead
of the table, essentially using the index itself as a table.
This section contains the following topics:
•

When the Optimizer Considers Index Fast Full Scans
The optimizer considers this scan when a query only accesses attributes in the
index.

•

How Index Fast Full Scans Work
The database uses multiblock I/O to read the root block and all of the leaf and
branch blocks. The databases ignores the branch and root blocks and reads the
index entries on the leaf blocks.

•

Index Fast Full Scans: Example
This examples uses a fast full index scan as a result of an optimizer hint.

8.3.5.1 When the Optimizer Considers Index Fast Full Scans
The optimizer considers this scan when a query only accesses attributes in the index.

Note:
Unlike a full scan, a fast full scan cannot eliminate a sort operation because it
does not read the index in order.

The INDEX_FFS(table_name index_name) hint forces a fast full index scan.

See Also:
Oracle Database SQL Language Reference to learn more about the INDEX
hint

8.3.5.2 How Index Fast Full Scans Work
The database uses multiblock I/O to read the root block and all of the leaf and branch
blocks. The databases ignores the branch and root blocks and reads the index entries
on the leaf blocks.

8-28

Chapter 8

B-Tree Index Access Paths

8.3.5.3 Index Fast Full Scans: Example
This examples uses a fast full index scan as a result of an optimizer hint.
The following statement queries the ID and name for departments in order of
department ID:
SELECT /*+ INDEX_FFS(departments dept_id_pk) */ COUNT(*)
FROM departments;

The following plan shows that the optimizer chose a fast full index scan:
SQL_ID fu0k5nvx7sftm, child number 0
------------------------------------select /*+ index_ffs(departments dept_id_pk) */ count(*) from departments
Plan hash value: 3940160378
-------------------------------------------------------------------------| Id | Operation
| Name
| Rows |Cost (%CPU)| Time
|
-------------------------------------------------------------------------| 0 | SELECT STATEMENT
|
|
|
2 (100)|
|
| 1 | SORT AGGREGATE
|
|
1 |
|
|
| 2 |
INDEX FAST FULL SCAN| DEPT_ID_PK |
27 |
2 (0)| 00:00:01 |
--------------------------------------------------------------------------

8.3.6 Index Skip Scans
An index skip scan occurs when the initial column of a composite index is "skipped"
or not specified in the query.
This section contains the following topics:
•

When the Optimizer Considers Index Skips Scans
Often, skip scanning index blocks is faster than scanning table blocks, and faster
than performing full index scans.

•

How Index Skip Scans Work
An index skip scan logically splits a composite index into smaller subindexes.

•

Index Skip Scans: Example
This example uses an index skip scan to satisfy a query of the customers table.

See Also:
Oracle Database Concepts

8.3.6.1 When the Optimizer Considers Index Skips Scans
Often, skip scanning index blocks is faster than scanning table blocks, and faster than
performing full index scans.
The optimizer considers a skip scan when the following criteria are met:
•

The leading column of a composite index is not specified in the query predicate.

8-29

Chapter 8

B-Tree Index Access Paths

For example, the query predicate does not reference the cust_gender column, and
the composite index key is (cust_gender,cust_email).
•

Few distinct values exist in the leading column of the composite index, but many
distinct values exist in the nonleading key of the index.
For example, if the composite index key is (cust_gender,cust_email), then the
cust_gender column has only two distinct values, but cust_email has thousands.

8.3.6.2 How Index Skip Scans Work
An index skip scan logically splits a composite index into smaller subindexes.
The number of distinct values in the leading columns of the index determines the
number of logical subindexes. The lower the number, the fewer logical subindexes the
optimizer must create, and the more efficient the scan becomes. The scan reads each
logical index separately, and "skips" index blocks that do not meet the filter condition
on the non-leading column.

8.3.6.3 Index Skip Scans: Example
This example uses an index skip scan to satisfy a query of the customers table.
The customers table contains a column cust_gender whose values are either M or F. You
create a composite index on the columns (cust_gender, cust_email) as follows:
CREATE INDEX cust_gender_email_ix
ON sh.customers (cust_gender, cust_email);

Conceptually, a portion of the index might look as follows, with the gender value of F or
M as the leading edge of the index.
F,Wolf@company.example.com,rowid
F,Wolsey@company.example.com,rowid
F,Wood@company.example.com,rowid
F,Woodman@company.example.com,rowid
F,Yang@company.example.com,rowid
F,Zimmerman@company.example.com,rowid
M,Abbassi@company.example.com,rowid
M,Abbey@company.example.com,rowid

You run the following query for a customer in the sh.customers table:
SELECT *
FROM sh.customers
WHERE cust_email = 'Abbey@company.example.com';

The database can use a skip scan of the customers_gender_email index even though
cust_gender is not specified in the WHERE clause. In the sample index, the leading
column cust_gender has two possible values: F and M. The database logically splits the
index into two. One subindex has the key F, with entries in the following form:
F,Wolf@company.example.com,rowid
F,Wolsey@company.example.com,rowid
F,Wood@company.example.com,rowid
F,Woodman@company.example.com,rowid
F,Yang@company.example.com,rowid
F,Zimmerman@company.example.com,rowid

The second subindex has the key M, with entries in the following form:

8-30

Chapter 8

B-Tree Index Access Paths

M,Abbassi@company.example.com,rowid
M,Abbey@company.example.com,rowid

When searching for the record for the customer whose email is Abbey@company.com, the
database searches the subindex with the leading value F first, and then searches the
subindex with the leading value M. Conceptually, the database processes the query as
follows:
( SELECT *
FROM sh.customers
WHERE cust_gender = 'F'
AND
cust_email = 'Abbey@company.com' )
UNION ALL
( SELECT *
FROM sh.customers
WHERE cust_gender = 'M'
AND
cust_email = 'Abbey@company.com' )

The plan for the query is as follows:
SQL_ID d7a6xurcnx2dj, child number 0
------------------------------------SELECT * FROM sh.customers WHERE cust_email = 'Abbey@company.example.com'
Plan hash value: 797907791
----------------------------------------------------------------------------------------|Id| Operation
| Name
|Rows|Bytes|Cost(%CPU)|Time|
----------------------------------------------------------------------------------------| 0|SELECT STATEMENT
|
| |
|10(100)|
|
| 1| TABLE ACCESS BY INDEX ROWID BATCHED| CUSTOMERS
|33|6237| 10(0)|00:00:01|
|*2| INDEX SKIP SCAN
| CUST_GENDER_EMAIL_IX |33|
| 4(0)|00:00:01|
----------------------------------------------------------------------------------------Predicate Information (identified by operation id):
--------------------------------------------------2 - access("CUST_EMAIL"='Abbey@company.example.com')
filter("CUST_EMAIL"='Abbey@company.example.com')

See Also:
Oracle Database Concepts to learn more about skip scans

8.3.7 Index Join Scans
An index join scan is a hash join of multiple indexes that together return all columns
requested by a query. The database does not need to access the table because all
data is retrieved from the indexes.
This section contains the following topics:
•

When the Optimizer Considers Index Join Scans
In some cases, avoiding table access is the most cost efficient option.

8-31

Chapter 8

B-Tree Index Access Paths

•

How Index Join Scans Work
An index join involves scanning multiple indexes, and then using a hash join on
the rowids obtained from these scans to return the rows.

•

Index Join Scans: Example
This example queries the last name and email for employees whose last name
begins with A, specifying an index join.

8.3.7.1 When the Optimizer Considers Index Join Scans
In some cases, avoiding table access is the most cost efficient option.
The optimizer considers an index join in the following circumstances:
•

A hash join of multiple indexes retrieves all data requested by the query, without
requiring table access.

•

The cost of retrieving rows from the table is higher than reading the indexes
without retrieving rows from the table. An index join is often expensive. For
example, when scanning two indexes and joining them, it is often less costly to
choose the most selective index, and then probe the table.

You can specify an index join with the INDEX_JOIN(table_name) hint.

See Also:
Oracle Database SQL Language Reference

8.3.7.2 How Index Join Scans Work
An index join involves scanning multiple indexes, and then using a hash join on the
rowids obtained from these scans to return the rows.
In an index join scan, table access is always avoided. For example, the process for
joining two indexes on a single table is as follows:
1.

Scan the first index to retrieve rowids.

2.

Scan the second index to retrieve rowids.

3.

Perform a hash join by rowid to obtain the rows.

8.3.7.3 Index Join Scans: Example
This example queries the last name and email for employees whose last name begins
with A, specifying an index join.
SELECT /*+ INDEX_JOIN(employees) */ last_name, email
FROM employees
WHERE last_name like 'A%';

Separate indexes exist on the (last_name,first_name) and email columns. Part of the
emp_name_ix index might look as follows:
Banda,Amit,AAAVgdAALAAAABSABD
Bates,Elizabeth,AAAVgdAALAAAABSABI
Bell,Sarah,AAAVgdAALAAAABSABc

8-32

Chapter 8

Bitmap Index Access Paths

Bernstein,David,AAAVgdAALAAAABSAAz
Bissot,Laura,AAAVgdAALAAAABSAAd
Bloom,Harrison,AAAVgdAALAAAABSABF
Bull,Alexis,AAAVgdAALAAAABSABV

The first part of the emp_email_uk index might look as follows:
ABANDA,AAAVgdAALAAAABSABD
ABULL,AAAVgdAALAAAABSABV
ACABRIO,AAAVgdAALAAAABSABX
AERRAZUR,AAAVgdAALAAAABSAAv
AFRIPP,AAAVgdAALAAAABSAAV
AHUNOLD,AAAVgdAALAAAABSAAD
AHUTTON,AAAVgdAALAAAABSABL

The following example retrieves the plan using the DBMS_XPLAN.DISPLAY_CURSOR function.
The database retrieves all rowids in the emp_email_uk index, and then retrieves rowids
in emp_name_ix for last names that begin with A. The database uses a hash join to
search both sets of rowids for matches. For example, rowid AAAVgdAALAAAABSABD occurs
in both sets of rowids, so the database probes the employees table for the record
corresponding to this rowid.
Example 8-4

Index Join Scan

SQL_ID d2djchyc9hmrz, child number 0
------------------------------------SELECT /*+ INDEX_JOIN(employees) */ last_name, email FROM
WHERE last_name like 'A%'

employees

Plan hash value: 3719800892
------------------------------------------------------------------------------------------| Id | Operation
| Name
| Rows | Bytes | Cost (%CPU)| Time
|
------------------------------------------------------------------------------------------| 0 | SELECT STATEMENT
|
|
|
|
3 (100)|
|
|* 1 | VIEW
| index$_join$_001 |
3 |
48 |
3 (34)| 00:00:01 |
|* 2 |
HASH JOIN
|
|
|
|
|
|
|* 3 |
INDEX RANGE SCAN
| EMP_NAME_IX
|
3 |
48 |
1 (0)| 00:00:01 |
| 4 |
INDEX FAST FULL SCAN| EMP_EMAIL_UK
|
3 |
48 |
1 (0)| 00:00:01 |
------------------------------------------------------------------------------------------Predicate Information (identified by operation id):
--------------------------------------------------1 - filter("LAST_NAME" LIKE 'A%')
2 - access(ROWID=ROWID)
3 - access("LAST_NAME" LIKE 'A%')

8.4 Bitmap Index Access Paths
Bitmap indexes combine the indexed data with a rowid range.
This section explains how bitmap indexes, and describes some of the more common
bitmap index access paths:
•

About Bitmap Index Access
In a conventional B-tree index, one index entry points to a single row. In a bitmap
index, the key is the combination of the indexed data and the rowid range.

8-33

Chapter 8

Bitmap Index Access Paths

•

Bitmap Conversion to Rowid
A bitmap conversion translates between an entry in the bitmap and a row in a
table. The conversion can go from entry to row (TO ROWID), or from row to entry
(FROM ROWID).

•

Bitmap Index Single Value
This type of access path uses a bitmap index to look up a single key value.

•

Bitmap Index Range Scans
This type of access path uses a bitmap index to look up a range of values.

•

Bitmap Merge
This access path merges multiple bitmaps, and returns a single bitmap as a result.

8.4.1 About Bitmap Index Access
In a conventional B-tree index, one index entry points to a single row. In a bitmap
index, the key is the combination of the indexed data and the rowid range.
The database stores at least one bitmap for each index key. Each value in the bitmap,
which is a series of 1 and 0 values, points to a row within a rowid range. Thus, in a
bitmap index, one index entry points to a set of rows rather than a single row.
This section contains the following topics:
•

Differences Between Bitmap and B-Tree Indexes
A bitmap index uses a different key from a B-tree index, but is stored in a B-tree
structure.

•

Purpose of Bitmap Indexes
Bitmap indexes are typically suitable for infrequently modified data with a low or
medium number of distinct values (NDV).

•

Bitmaps and Rowids
For a particular value in a bitmap, the value is 1 if the row values match the bitmap
condition, and 0 if it does not. Based on these values, the database uses an
internal algorithm to map bitmaps onto rowids.

•

Bitmap Join Indexes
A bitmap join index is a bitmap index for the join of two or more tables.

•

Bitmap Storage
A bitmap index resides in a B-tree structure, using branch blocks and leaf blocks
just as in a B-tree.

8.4.1.1 Differences Between Bitmap and B-Tree Indexes
A bitmap index uses a different key from a B-tree index, but is stored in a B-tree
structure.
The following table shows the differences among types of index entries.

8-34

Chapter 8

Bitmap Index Access Paths

Table 8-3

Index Entries for B-Trees and Bitmaps

Index Entry

Key

Data

Example

Unique B-tree

Indexed data only

Rowid

In an entry of the index on the employees.employee_id
column, employee ID 101 is the key, and the rowid
AAAPvCAAFAAAAFaAAa is the data:
101,AAAPvCAAFAAAAFaAAa

Nonunique B-tree Indexed data
combined with rowid

None

In an entry of the index on the employees.last_name
column, the name and rowid combination
Smith,AAAPvCAAFAAAAFaAAa is the key, and there is no
data:
Smith,AAAPvCAAFAAAAFaAAa

Bitmap

Indexed data
combined with rowid
range

Bitmap

In an entry of the index on the customers.cust_gender
column, the M,low-rowid,high-rowid part is the key, and
the series of 1 and 0 values is the data:
M,low-rowid,high-rowid,1000101010101010

The database stores a bitmap index in a B-tree structure. The database can search
the B-tree quickly on the first part of the key, which is the set of attributes on which the
index is defined, and then obtain the corresponding rowid range and bitmap.

See Also:
•

"Bitmap Storage"

•

Oracle Database Concepts for an overview of bitmap indexes

•

Oracle Database Data Warehousing Guide for more information about
bitmap indexes

8.4.1.2 Purpose of Bitmap Indexes
Bitmap indexes are typically suitable for infrequently modified data with a low or
medium number of distinct values (NDV).
In general, B-tree indexes are suitable for columns with high NDV and frequent DML
activity. For example, the optimizer might choose a B-tree index for a query of a
sales.amount column that returns few rows. In contrast, the customers.state and
customers.county columns are candidates for bitmap indexes because they have few
distinct values, are infrequently updated, and can benefit from efficient AND and OR
operations.
Bitmap indexes are a useful way to speed ad hoc queries in a data warehouse. They
are fundamental to star transformations. Specifically, bitmap indexes are useful in
queries that contain the following:
•

Multiple conditions in the WHERE clause

8-35

Chapter 8

Bitmap Index Access Paths

Before the table itself is accessed, the database filters out rows that satisfy some,
but not all, conditions.
•

AND, OR, and NOT operations on columns with low or medium NDV

Combining bitmap indexes makes these operations more efficient. The database
can merge bitmaps from bitmap indexes very quickly. For example, if bitmap
indexes exist on the customers.state and customers.county columns, then these
indexes can enormously improve the performance of the following query:
SELECT
FROM
WHERE
AND

*
customers
state = 'CA'
county = 'San Mateo'

The database can convert 1 values in the merged bitmap into rowids efficiently.
•

The COUNT function
The database can scan the bitmap index without needing to scan the table.

•

Predicates that select for null values
Unlike B-tree indexes, bitmap indexes can contain nulls. Queries that count the
number of nulls in a column can use the bitmap index without scanning the table.

•

Columns that do not experience heavy DML
The reason is that one index key points to many rows. If a session modifies the
indexed data, then the database cannot lock a single bit in the bitmap: rather, the
database locks the entire index entry, which in practice locks the rows pointed to
by the bitmap. For example, if the county of residence for a specific customer
changes from San Mateo to Alameda, then the database must get exclusive access
to the San Mateo index entry and Alameda index entry in the bitmap. Rows
containing these two values cannot be modified until COMMIT.

See Also:
•

"Star Transformation"

•

Oracle Database SQL Language Reference to learn about the COUNT
function

8.4.1.3 Bitmaps and Rowids
For a particular value in a bitmap, the value is 1 if the row values match the bitmap
condition, and 0 if it does not. Based on these values, the database uses an internal
algorithm to map bitmaps onto rowids.
The bitmap entry contains the indexed value, the rowid range (start and end rowids),
and a bitmap. Each 0 or 1 value in the bitmap is an offset into the rowid range, and
maps to a potential row in the table, even if the row does not exist. Because the
number of possible rows in a block is predetermined, the database can use the range
endpoints to determine the rowid of an arbitrary row in the range.

8-36

Chapter 8

Bitmap Index Access Paths

Note:
The Hakan factor is an optimization used by the bitmap index algorithms to
limit the number of rows that Oracle Database assumes can be stored in a
single block. By artificially limiting the number of rows, the database reduces
the size of the bitmaps.

Table 8-4 shows part of a sample bitmap for the sh.customers.cust_marital_status
column, which is nullable. The actual index has 12 distinct values. Only 3 are shown in
the sample: null, married, and single.
Table 8-4

Bitmap Index Entries

Column
Start
Value for
Rowid in
cust_marital Range
_status

End
Rowid in
Range

1st
2nd
3rd
4th
5th
6th
Row in Row in Row in Row in Row in Row in
Range Range Range Range Range Range

(null)

AAA ...

CCC ...

0

0

0

0

0

1

married

AAA ...

CCC ...

1

0

1

1

1

0

single

AAA ...

CCC ...

0

1

0

0

0

0

single

DDD ...

EEE ...

1

0

1

0

1

1

As shown in Table 8-4, bitmap indexes can include keys that consist entirely of null
values, unlike B-tree indexes. In Table 8-4, the null has a value of 1 for the 6th row in
the range, which means that the cust_marital_status value is null for the 6th row in the
range. Indexing nulls can be useful for some SQL statements, such as queries with the
aggregate function COUNT.

See Also:
Oracle Database Concepts to learn about rowid formats

8.4.1.4 Bitmap Join Indexes
A bitmap join index is a bitmap index for the join of two or more tables.
The optimizer can use a bitmap join index to reduce or eliminate the volume of data
that must be joined during plan execution. Bitmap join indexes can be much more
efficient in storage than materialized join views.
The following example creates a bitmap index on the sh.sales and sh.customers tables:
CREATE BITMAP INDEX cust_sales_bji ON sales(c.cust_city)
FROM sales s, customers c
WHERE c.cust_id = s.cust_id LOCAL;

The FROM and WHERE clause in the preceding CREATE statement represent the join
condition between the tables. The customers.cust_city column is the index key.

8-37

Chapter 8

Bitmap Index Access Paths

Each key value in the index represents a possible city in the customers table.
Conceptually, key values for the index might look as follows, with one bitmap
associated with each key value:
San Francisco
San Mateo
Smithville
.
.
.

0 0 0 1 0 1 0 0 0 1 0 0 0 0 0 . . .
0 1 0 0 0 0 0 0 1 0 0 0 0 0 1 . . .
1 0 0 0 1 0 0 1 0 0 1 0 1 0 0 . . .

Each bit in a bitmap corresponds to one row in the sales table. In the Smithville key,
the value 1 means that the first row in the sales table corresponds to a product sold to
a Smithville customer, whereas the value 0 means that the second row corresponds to
a product not sold to a Smithville customer.
Consider the following query of the number of separate sales to Smithville customers:
SELECT
FROM
WHERE
AND

COUNT (*)
sales s, customers c
c.cust_id = s.cust_id
c.cust_city = 'Smithville';

The following plan shows that the database reads the Smithville bitmap to derive the
number of Smithville sales (Step 4), thereby avoiding a join of the customers and sales
tables.
SQL_ID 57s100mh142wy, child number 0
------------------------------------SELECT COUNT (*) FROM sales s, customers c WHERE c.cust_id = s.cust_id
AND c.cust_city = 'Smithville'
Plan hash value: 3663491772
-----------------------------------------------------------------------------------|Id| Operation
| Name |Rows|Bytes|Cost (%CPU)| Time|Pstart|Pstop|
-----------------------------------------------------------------------------------| 0| SELECT STATEMENT
|
|
|
|29 (100)|
| | |
| 1| SORT AGGREGATE
|
| 1 | 5|
|
| | |
| 2| PARTITION RANGE ALL
|
| 1708|8540|29 (0)|00:00:01|1|28|
| 3|
BITMAP CONVERSION COUNT |
| 1708|8540|29 (0)|00:00:01| | |
|*4|
BITMAP INDEX SINGLE VALUE|CUST_SALES_BJI|
|
|
|
|1|28|
-----------------------------------------------------------------------------------Predicate Information (identified by operation id):
--------------------------------------------------4 - access("S"."SYS_NC00008$"='Smithville')

See Also:
Oracle Database Concepts to learn about the CREATE INDEX statement

8.4.1.5 Bitmap Storage
A bitmap index resides in a B-tree structure, using branch blocks and leaf blocks just
as in a B-tree.

8-38

Chapter 8

Bitmap Index Access Paths

For example, if the customers.cust_marital_status column has 12 distinct values, then
one branch block might point to the keys married,rowid-range and single,rowid-range,
another branch block might point to the widowed,rowid-range key, and so on.
Alternatively, a single branch block could point to a leaf block containing all 12 distinct
keys.
Each indexed column value may have one or more bitmap pieces, each with its own
rowid range occupying a contiguous set of rows in one or more extents. The database
can use a bitmap piece to break up an index entry that is large relative to the size of a
block. For example, the database could break a single index entry into three pieces,
with the first two pieces in separate blocks in the same extent, and the last piece in a
separate block in a different extent.
To conserve space, Oracle Database can compression consecutive ranges of 0
values.

8.4.2 Bitmap Conversion to Rowid
A bitmap conversion translates between an entry in the bitmap and a row in a table.
The conversion can go from entry to row (TO ROWID), or from row to entry (FROM ROWID).
This section contains the following topics:
•

When the Optimizer Chooses Bitmap Conversion to Rowid
The optimizer uses a conversion whenever it retrieves a row from a table using a
bitmap index entry.

•

How Bitmap Conversion to Rowid Works
Conceptually, a bitmap can be represented as table.

•

Bitmap Conversion to Rowid: Example
In this example, the optimizer chooses a bitmap conversion operation to satisfy a
query using a range predicate.

8.4.2.1 When the Optimizer Chooses Bitmap Conversion to Rowid
The optimizer uses a conversion whenever it retrieves a row from a table using a
bitmap index entry.

8.4.2.2 How Bitmap Conversion to Rowid Works
Conceptually, a bitmap can be represented as table.
For example, Table 8-4 represents the bitmap as a table with customers row numbers
as columns and cust_marital_status values as rows. Each field in Table 8-4 has the
value 1 or 0, and represents a column value in a row. Conceptually, the bitmap
conversion uses an internal algorithm that says, "Field F in the bitmap corresponds to
the Nth row of the Mth block of the table," or "The Nth row of the Mth block in the table
corresponds to field F in the bitmap."

8.4.2.3 Bitmap Conversion to Rowid: Example
In this example, the optimizer chooses a bitmap conversion operation to satisfy a
query using a range predicate.
A query of the sh.customers table selects the names of all customers born before 1918:

8-39

Chapter 8

Bitmap Index Access Paths

SELECT cust_last_name, cust_first_name
FROM customers
WHERE cust_year_of_birth < 1918;

The following plan shows that the database uses a range scan to find all key values
less than 1918 (Step 3), converts the 1 values in the bitmap to rowids (Step 2), and
then uses the rowids to obtain the rows from the customers table (Step 1):
--------------------------------------------------------------------------------------------|Id| Operation
| Name
|Rows|Bytes|Cost(%CPU)| Time
|
--------------------------------------------------------------------------------------------| 0| SELECT STATEMENT
|
|
|
|421 (100)|
|
| 1| TABLE ACCESS BY INDEX ROWID BATCHED| CUSTOMERS
|3604|68476|421 (1)| 00:00:01 |
| 2|
BITMAP CONVERSION TO ROWIDS
|
|
|
|
|
|
|*3|
BITMAP INDEX RANGE SCAN
| CUSTOMERS_YOB_BIX|
|
|
|
|
--------------------------------------------------------------------------------------------Predicate Information (identified by operation id):
--------------------------------------------------3 - access("CUST_YEAR_OF_BIRTH"<1918)
filter("CUST_YEAR_OF_BIRTH"<1918)

8.4.3 Bitmap Index Single Value
This type of access path uses a bitmap index to look up a single key value.
This section contains the following topics:
•

When the Optimizer Considers Bitmap Index Single Value
The optimizer considers this access path when the predicate contains an equality
operator.

•

How Bitmap Index Single Value Works
The query scans a single bitmap for positions containing a 1 value. The database
converts the 1 values into rowids, and then uses the rowids to find the rows.

•

Bitmap Index Single Value: Example
In this example, the optimizer chooses a bitmap index single value operation to
satisfy a query that uses an equality predicate.

8.4.3.1 When the Optimizer Considers Bitmap Index Single Value
The optimizer considers this access path when the predicate contains an equality
operator.

8.4.3.2 How Bitmap Index Single Value Works
The query scans a single bitmap for positions containing a 1 value. The database
converts the 1 values into rowids, and then uses the rowids to find the rows.
The database only needs to process a single bitmap. For example, the following table
represents the bitmap index (in two bitmap pieces) for the value widowed in the
sh.customers.cust_marital_status column. To satisfy a query of customers with the
status widowed, the database can search for each 1 value in the widowed bitmap and find
the rowid of the corresponding row.

8-40

Chapter 8

Bitmap Index Access Paths

Table 8-5

Bitmap Index Entries

Column
Value

Start
Rowid in
Range

End
Rowid in
Range

1st
2nd
Row in Row in
Range Range

3rd
4th
Row in Row in
Range Range

5th
6th
Row in Row in
Range Range

widowed

AAA ...

CCC ...

0

1

0

0

0

0

widowed

DDD ...

EEE ...

1

0

1

0

1

1

8.4.3.3 Bitmap Index Single Value: Example
In this example, the optimizer chooses a bitmap index single value operation to satisfy
a query that uses an equality predicate.
A query of the sh.customers table selects all widowed customers:
SELECT *
FROM customers
WHERE cust_marital_status = 'Widowed';

The following plan shows that the database reads the entry with the Widowed key in the
customers bitmap index (Step 3), converts the 1 values in the bitmap to rowids (Step 2),
and then uses the rowids to obtain the rows from the customers table (Step 1):
SQL_ID ff5an2xsn086h, child number 0
------------------------------------SELECT * FROM customers WHERE cust_marital_status = 'Widowed'
Plan hash value: 2579015045
--------------------------------------------------------------------------------------------|Id| Operation
| Name
|Rows|Bytes|Cost (%CPU)| Time|
--------------------------------------------------------------------------------------------| 0| SELECT STATEMENT
|
|
|
|412 (100)|
|
| 1| TABLE ACCESS BY INDEX ROWID BATCHED|CUSTOMERS
|3461|638K|412 (2)|00:00:01|
| 2| BITMAP CONVERSION TO ROWIDS
|
|
|
|
|
|
|*3|
BITMAP INDEX SINGLE VALUE
|CUSTOMERS_MARITAL_BIX|
|
|
|
|
--------------------------------------------------------------------------------------------Predicate Information (identified by operation id):
--------------------------------------------------3 - access("CUST_MARITAL_STATUS"='Widowed')

8.4.4 Bitmap Index Range Scans
This type of access path uses a bitmap index to look up a range of values.
This section contains the following topics:
•

When the Optimizer Considers Bitmap Index Range Scans
The optimizer considers this access path when the predicate selects a range of
values.

•

How Bitmap Index Range Scans Work
This scan works similarly to a B-tree range scan.

•

Bitmap Index Range Scans: Example
This example uses a range scan to select customers born before a single year.

8-41

Chapter 8

Bitmap Index Access Paths

8.4.4.1 When the Optimizer Considers Bitmap Index Range Scans
The optimizer considers this access path when the predicate selects a range of
values.
The range in the scan can be bounded on both sides, or unbounded on one or both
sides. The optimizer typically chooses a range scan for selective queries.

See Also:
"Index Range Scans"

8.4.4.2 How Bitmap Index Range Scans Work
This scan works similarly to a B-tree range scan.
For example, the following table represents three values in the bitmap index for the
sh.customers.cust_year_of_birth column. If a query requests all customers born before
1917, then the database can scan this index for values lower than 1917, and then
obtain the rowids for rows that have a 1.
Table 8-6

Bitmap Index Entries

Column
Value

Start
Rowid in
Range

End
Rowid in
Range

1st
2nd
Row in Row in
Range Range

3rd
4th
Row in Row in
Range Range

5th
6th
Row in Row in
Range Range

1913

AAA ...

CCC ...

0

0

0

0

0

1

1917

AAA ...

CCC ...

1

0

1

1

1

0

1918

AAA ...

CCC ...

0

1

0

0

0

0

1918

DDD ...

EEE ...

1

0

1

0

1

1

See Also:
"Index Range Scans"

8.4.4.3 Bitmap Index Range Scans: Example
This example uses a range scan to select customers born before a single year.
A query of the sh.customers table selects the names of customers born before 1918:
SELECT cust_last_name, cust_first_name
FROM customers
WHERE cust_year_of_birth < 1918

8-42

Chapter 8

Bitmap Index Access Paths

The following plan shows that the database obtains all bitmaps for cust_year_of_birth
keys lower than 1918 (Step 3), converts the bitmaps to rowids (Step 2), and then
fetches the rows (Step 1):
SQL_ID 672z2h9rawyjg, child number 0
------------------------------------SELECT cust_last_name, cust_first_name FROM
cust_year_of_birth < 1918

customers WHERE

Plan hash value: 4198466611
--------------------------------------------------------------------------------------------|Id| Operation
| Name
|Rows|Bytes|Cost(%CPU)|Time
|
--------------------------------------------------------------------------------------------| 0| SELECT STATEMENT
|
|
|
|421 (100)|
|
| 1| TABLE ACCESS BY INDEX ROWID BATCHED| CUSTOMERS
|3604|68476|421 (1)|00:00:01 |
| 2| BITMAP CONVERSION TO ROWIDS
|
|
|
|
|
|
BITMAP INDEX RANGE SCAN
|*3|
| CUSTOMERS_YOB_BIX |
|
|
|
|
--------------------------------------------------------------------------------------------Predicate Information (identified by operation id):
--------------------------------------------------3 - access("CUST_YEAR_OF_BIRTH"<1918)
filter("CUST_YEAR_OF_BIRTH"<1918)

8.4.5 Bitmap Merge
This access path merges multiple bitmaps, and returns a single bitmap as a result.
A bitmap merge is indicated by the BITMAP MERGE operation in an execution plan.
This section contains the following topics:
•

When the Optimizer Considers Bitmap Merge
The optimizer typically uses a bitmap merge to combine bitmaps generated from a
bitmap index range scan.

•

How Bitmap Merge Works
A merge uses a Boolean OR operation between two bitmaps. The resulting bitmap
selects all rows from the first bitmap, plus all rows from every subsequent bitmap.

•

Bitmap Merge: Example
This example shows how the database merges bitmaps to optimize a query using
a range predicate.

8.4.5.1 When the Optimizer Considers Bitmap Merge
The optimizer typically uses a bitmap merge to combine bitmaps generated from a
bitmap index range scan.

8.4.5.2 How Bitmap Merge Works
A merge uses a Boolean OR operation between two bitmaps. The resulting bitmap
selects all rows from the first bitmap, plus all rows from every subsequent bitmap.
A query might select all customers born before 1918. The following example shows
sample bitmaps for three customers.cust_year_of_birth keys: 1917, 1916, and 1915. If
any position in any bitmap has a 1, then the merged bitmap has a 1 in the same
position. Otherwise, the merged bitmap has a 0.

8-43

Chapter 8

Table Cluster Access Paths

1917
1 0 1 0 0 0 0 0 0 0 0 0 0 1
1916
0 1 0 0 0 0 0 0 0 0 0 0 0 0
1915
0 0 0 0 0 0 0 0 1 0 0 0 0 0
-----------------------------------merged: 1 1 1 0 0 0 0 0 1 0 0 0 0 1

The 1 values in resulting bitmap correspond to rows that contain the values 1915, 1916,
or 1917.

8.4.5.3 Bitmap Merge: Example
This example shows how the database merges bitmaps to optimize a query using a
range predicate.
A query of the sh.customers table selects the names of female customers born before
1918:
SELECT
FROM
WHERE
AND

cust_last_name, cust_first_name
customers
cust_gender = 'F'
cust_year_of_birth < 1918

The following plan shows that the database obtains all bitmaps for cust_year_of_birth
keys lower than 1918 (Step 6), and then merges these bitmaps using OR logic to create
a single bitmap (Step 5). The database obtains a single bitmap for the cust_gender key
of F (Step 4), and then performs an AND operation on these two bitmaps. The result is a
single bitmap that contains 1 values for the requested rows (Step 3).
SQL_ID 1xf59h179zdg2, child number 0
------------------------------------select cust_last_name, cust_first_name from customers where cust_gender
= 'F' and cust_year_of_birth < 1918
Plan hash value: 49820847
--------------------------------------------------------------------------------------------|Id| Operation
| Name
|Rows|Bytes|Cost(%CPU)| Time |
--------------------------------------------------------------------------------------------| 0| SELECT STATEMENT
|
|
|
|288 (100)|
|
| 1| TABLE ACCESS BY INDEX ROWID BATCHED|CUSTOMERS
|1802|37842|288 (1)|00:00:01|
| 2| BITMAP CONVERSION TO ROWIDS
|
|
|
|
|
|
| 3|
BITMAP AND
|
|
|
|
|
|
|*4|
BITMAP INDEX SINGLE VALUE
|CUSTOMERS_GENDER_BIX|
|
|
|
|
| 5|
BITMAP MERGE
|
|
|
|
|
|
|*6|
BITMAP INDEX RANGE SCAN
|CUSTOMERS_YOB_BIX |
|
|
|
|
--------------------------------------------------------------------------------------------Predicate Information (identified by operation id):
--------------------------------------------------4 - access("CUST_GENDER"='F')
6 - access("CUST_YEAR_OF_BIRTH"<1918)
filter("CUST_YEAR_OF_BIRTH"<1918)

8.5 Table Cluster Access Paths
A table cluster is a group of tables that share common columns and store related
data in the same blocks. When tables are clustered, a single data block can contain
rows from multiple tables.
This section contains the following topics:

8-44

Chapter 8

Table Cluster Access Paths

•

Cluster Scans
An index cluster is a table cluster that uses an index to locate data.

•

Hash Scans
A hash cluster is like an indexed cluster, except the index key is replaced with a
hash function. No separate cluster index exists.

See Also:
Oracle Database Concepts for an overview of table clusters

8.5.1 Cluster Scans
An index cluster is a table cluster that uses an index to locate data.
The cluster index is a B-tree index on the cluster key. A cluster scan retrieves all rows
that have the same cluster key value from a table stored in an indexed cluster.
This section contains the following topics:
•

When the Optimizer Considers Cluster Scans
The database considers a cluster scan when a query accesses a table in an
indexed cluster.

•

How a Cluster Scan Works
In an indexed cluster, the database stores all rows with the same cluster key value
in the same data block.

•

Cluster Scans: Example
This example clusters the employees and departments tables on the department_id
column, and then queries the cluster for a single department.

8.5.1.1 When the Optimizer Considers Cluster Scans
The database considers a cluster scan when a query accesses a table in an indexed
cluster.

8.5.1.2 How a Cluster Scan Works
In an indexed cluster, the database stores all rows with the same cluster key value in
the same data block.
For example, if the hr.employees2 and hr.departments2 tables are clustered in
emp_dept_cluster, and if the cluster key is department_id, then the database stores all
employees in department 10 in the same block, all employees in department 20 in the
same block, and so on.
The B-tree cluster index associates the cluster key value with the database block
address (DBA) of the block containing the data. For example, the index entry for key 30
shows the address of the block that contains rows for employees in department 30:
30,AADAAAA9d

8-45

Chapter 8

Table Cluster Access Paths

When a user requests rows in the cluster, the database scans the index to obtain the
DBAs of the blocks containing the rows. Oracle Database then locates the rows based
on these DBAs.

8.5.1.3 Cluster Scans: Example
This example clusters the employees and departments tables on the department_id
column, and then queries the cluster for a single department.
As user hr, you create a table cluster, cluster index, and tables in the cluster as
follows:
CREATE CLUSTER employees_departments_cluster
(department_id NUMBER(4)) SIZE 512;
CREATE INDEX idx_emp_dept_cluster
ON CLUSTER employees_departments_cluster;
CREATE TABLE employees2
CLUSTER employees_departments_cluster (department_id)
AS SELECT * FROM employees;
CREATE TABLE departments2
CLUSTER employees_departments_cluster (department_id)
AS SELECT * FROM departments;

You query the employees in department 30 as follows:
SELECT *
FROM employees2
WHERE department_id = 30;

To perform the scan, Oracle Database first obtains the rowid of the row describing
department 30 by scanning the cluster index (Step 2). Oracle Database then locates
the rows in employees2 using this rowid (Step 1).
SQL_ID b7xk1jzuwdc6t, child number 0
------------------------------------SELECT * FROM employees2 WHERE department_id = 30
Plan hash value: 49826199
-------------------------------------------------------------------------------|Id| Operation
| Name
|Rows|Bytes|Cost(%CPU)| Time
|
-------------------------------------------------------------------------------| 0| SELECT STATEMENT
|
|
|
| 2 (100)|
|
| 1| TABLE ACCESS CLUSTER| EMPLOYEES2
| 6 | 798 | 2 (0)| 00:00:01|
|*2| INDEX UNIQUE SCAN |IDX_EMP_DEPT_CLUSTER| 1 |
| 1 (0)| 00:00:01|
-------------------------------------------------------------------------------Predicate Information (identified by operation id):
--------------------------------------------------2 - access("DEPARTMENT_ID"=30)

See Also:
Oracle Database Concepts to learn about indexed clusters

8-46

Chapter 8

Table Cluster Access Paths

8.5.2 Hash Scans
A hash cluster is like an indexed cluster, except the index key is replaced with a hash
function. No separate cluster index exists.
In a hash cluster, the data is the index. The database uses a hash scan to locate rows
in a hash cluster based on a hash value.
This section contains the following topics:
•

When the Optimizer Considers a Hash Scan
The database considers a hash scan when a query accesses a table in a hash
cluster.

•

How a Hash Scan Works
In a hash cluster, all rows with the same hash value are stored in the same data
block.

•

Hash Scans: Example
This example hashes the employees and departments tables on the department_id
column, and then queries the cluster for a single department.

8.5.2.1 When the Optimizer Considers a Hash Scan
The database considers a hash scan when a query accesses a table in a hash cluster.

8.5.2.2 How a Hash Scan Works
In a hash cluster, all rows with the same hash value are stored in the same data block.
To perform a hash scan of the cluster, Oracle Database first obtains the hash value by
applying a hash function to a cluster key value specified by the statement. Oracle
Database then scans the data blocks containing rows with this hash value.

8.5.2.3 Hash Scans: Example
This example hashes the employees and departments tables on the department_id
column, and then queries the cluster for a single department.
You create a hash cluster and tables in the cluster as follows:
CREATE CLUSTER employees_departments_cluster
(department_id NUMBER(4)) SIZE 8192 HASHKEYS 100;
CREATE TABLE employees2
CLUSTER employees_departments_cluster (department_id)
AS SELECT * FROM employees;
CREATE TABLE departments2
CLUSTER employees_departments_cluster (department_id)
AS SELECT * FROM departments;

You query the employees in department 30 as follows:
SELECT *
FROM employees2
WHERE department_id = 30

8-47

Chapter 8

Table Cluster Access Paths

To perform a hash scan, Oracle Database first obtains the hash value by applying a
hash function to the key value 30, and then uses this hash value to scan the data
blocks and retrieve the rows (Step 1).
SQL_ID 919x7hyyxr6p4, child number 0
------------------------------------SELECT * FROM employees2 WHERE department_id = 30
Plan hash value: 2399378016
---------------------------------------------------------------| Id | Operation
| Name
| Rows | Bytes | Cost |
---------------------------------------------------------------| 0 | SELECT STATEMENT |
|
|
|
1 |
|* 1 | TABLE ACCESS HASH| EMPLOYEES2 |
10 | 1330 |
|
---------------------------------------------------------------Predicate Information (identified by operation id):
--------------------------------------------------1 - access("DEPARTMENT_ID"=30)

See Also:
Oracle Database Concepts to learn about hash clusters

8-48

9
Joins
Oracle Database provides several optimizations for joining row sets.
This chapter contains the following topics:
•

About Joins
A join combines the output from exactly two row sources, such as tables or views,
and returns one row source. The returned row source is the data set.

•

Join Methods
A join method is the mechanism for joining two row sources.

•

Join Types
A join type is determined by the type of join condition.

•

Join Optimizations
Join optimizations enable joins to be more efficient.

9.1 About Joins
A join combines the output from exactly two row sources, such as tables or views, and
returns one row source. The returned row source is the data set.
A join is characterized by multiple tables in the WHERE (non-ANSI) or FROM ... JOIN
(ANSI) clause of a SQL statement. Whenever multiple tables exist in the FROM clause,
Oracle Database performs a join.
A join condition compares two row sources using an expression. The join condition
defines the relationship between the tables. If the statement does not specify a join
condition, then the database performs a Cartesian join, matching every row in one
table with every row in the other table.
This section contains the following topics:
•

Join Trees
Typically, a join tree is represented as an upside-down tree structure.

•

How the Optimizer Executes Join Statements
The database joins pairs of row sources. When multiple tables exist in the FROM
clause, the optimizer must determine which join operation is most efficient for each
pair.

•

How the Optimizer Chooses Execution Plans for Joins
When determining the join order and method, the optimizer goal is to reduce the
number of rows early so it performs less work throughout the execution of the SQL
statement.

9-1

Chapter 9

About Joins

See Also:
•

"Cartesian Joins"

•

Oracle Database SQL Language Reference for a concise discussion of
joins in Oracle SQL

9.1.1 Join Trees
Typically, a join tree is represented as an upside-down tree structure.
As shown in the following graphic, table1 is the left table, and table2 is the right table.
The optimizer processes the join from left to right. For example, if this graphic depicted
a nested loops join, then table1 is the outer loop, and table2 is the inner loop.

Figure 9-1

Join Tree

result set

table1

table2

The input of a join can be the result set from a previous join. If the right child of every
internal node of a join tree is a table, then the tree is a left deep join tree, as shown in
the following example. Most join trees are left deep joins.

Figure 9-2

Left Deep Join Tree
result set

table4

table3

table1

table2

9-2

Chapter 9

About Joins

If the left child of every internal node of a join tree is a table, then the tree is called a
right deep join tree, as shown in the following diagram.

Figure 9-3

Right Deep Join Tree

result set

table1

table2

table3

table4

If the left or the right child of an internal node of a join tree can be a join node, then the
tree is called a bushy join tree. In the following example, table4 is a right child of a join
node, table1 is the left child of a join node, and table2 is the left child of a join node.

Figure 9-4

Bushy Join Tree
result set

table4

table1

table2

table3

In yet another variation, both inputs of a join are the results of a previous join.

9.1.2 How the Optimizer Executes Join Statements
The database joins pairs of row sources. When multiple tables exist in the FROM clause,
the optimizer must determine which join operation is most efficient for each pair.
The optimizer must make the interrelated decisions shown in the following table.

9-3

Chapter 9

About Joins

Table 9-1

Join Operations

Operation

Explanation

To Learn More

Access paths

As for simple statements, the optimizer
must choose an access path to retrieve
data from each table in the join statement.
For example, the optimizer might choose
between a full table scan or an index scan..

"Optimizer Access Paths"

Join methods

To join each pair of row sources, Oracle
Database must decide how to do it. The
"how" is the join method. The possible join
methods are nested loop, sort merge, and
hash joins. A Cartesian join requires one of
the preceding join methods. Each join
method has specific situations in which it is
more suitable than the others.

"Join Methods"

Join types

The join condition determines the join type.
For example, an inner join retrieves only
rows that match the join condition. An outer
join retrieves rows that do not match the
join condition.

"Join Types"

Join order

To execute a statement that joins more than N/A
two tables, Oracle Database joins two
tables and then joins the resulting row
source to the next table. This process
continues until all tables are joined into the
result. For example, the database joins two
tables, and then joins the result to a third
table, and then joins this result to a fourth
table, and so on.

9.1.3 How the Optimizer Chooses Execution Plans for Joins
When determining the join order and method, the optimizer goal is to reduce the
number of rows early so it performs less work throughout the execution of the SQL
statement.
The optimizer generates a set of execution plans, according to possible join orders,
join methods, and available access paths. The optimizer then estimates the cost of
each plan and chooses the one with the lowest cost. When choosing an execution
plan, the optimizer considers the following factors:
•

The optimizer first determines whether joining two or more tables results in a row
source containing at most one row.
The optimizer recognizes such situations based on UNIQUE and PRIMARY KEY
constraints on the tables. If such a situation exists, then the optimizer places these
tables first in the join order. The optimizer then optimizes the join of the remaining
set of tables.

•

For join statements with outer join conditions, the table with the outer join operator
typically comes after the other table in the condition in the join order.
In general, the optimizer does not consider join orders that violate this guideline,
although the optimizer overrides this ordering condition in certain circumstances.
Similarly, when a subquery has been converted into an antijoin or semijoin, the

9-4

Chapter 9

Join Methods

tables from the subquery must come after those tables in the outer query block to
which they were connected or correlated. However, hash antijoins and semijoins
are able to override this ordering condition in certain circumstances.
The optimizer estimates cost of a query plan by computing the estimated I/Os and
CPU. These I/Os have specific costs associated with them: one cost for a single block
I/O, and another cost for multiblock I/Os. Also, different functions and expressions
have CPU costs associated with them. The optimizer determines the total cost of a
query plan using these metrics. These metrics may be influenced by many initialization
parameter and session settings at compile time, such as the
DB_FILE_MULTI_BLOCK_READ_COUNT setting, system statistics, and so on.
For example, the optimizer estimates costs in the following ways:
•

The cost of a nested loops join depends on the cost of reading each selected row
of the outer table and each of its matching rows of the inner table into memory.
The optimizer estimates these costs using statistics in the data dictionary.

•

The cost of a sort merge join depends largely on the cost of reading all the
sources into memory and sorting them.

•

The cost of a hash join largely depends on the cost of building a hash table on one
of the input sides to the join and using the rows from the other side of the join to
probe it.

Example 9-1

Estimating Costs for Join Order and Method

Conceptually, the optimizer constructs a matrix of join orders and methods and the
cost associated with each. For example, the optimizer must determine how best to join
the date_dim and lineorder tables in a query. The following table shows the possible
variations of methods and orders, and the cost for each. In this example, a nested
loops join in the order date_dim, lineorder has the lowest cost.
Table 9-2

Sample Costs for Join of date_dim and lineorder Tables

Join Method

Cost of date_dim, lineorder

Cost of lineorder, date_dim

Nested Loops

39,480

6,187,540

Hash Join

187,528

194,909

Sort Merge

217,129

217,129

See Also:
•

"Introduction to Optimizer Statistics"

•

"Influencing the Optimizer " for more information about optimizer hints

•

Oracle Database Reference to learn about
DB_FILE_MULTIBLOCK_READ_COUNT

9.2 Join Methods
A join method is the mechanism for joining two row sources.

9-5

Chapter 9

Join Methods

Depending on the statistics, the optimizer chooses the method with the lowest
estimated cost. As shown in Figure 9-5, each join method has two children: the driving
(also called outer) row source and the driven-to (also called inner) row source.
Figure 9-5

Join Method
Join Method
(Nested Loops, Hash
Join, or Sort Merge)

Driving Row Source,
Outer row Source

Driven-To Row Source,
Inner Row Source

This section contains the following topics:
•

Nested Loops Joins
Nested loops join an outer data set to an inner data set.

•

Hash Joins
The database uses a hash join to join larger data sets.

•

Sort Merge Joins
A sort merge join is a variation on a nested loops join.

•

Cartesian Joins
The database uses a Cartesian join when one or more of the tables does not
have any join conditions to any other tables in the statement.

9.2.1 Nested Loops Joins
Nested loops join an outer data set to an inner data set.
For each row in the outer data set that matches the single-table predicates, the
database retrieves all rows in the inner data set that satisfy the join predicate. If an
index is available, then the database can use it to access the inner data set by rowid.
This section contains the following topics:
•

When the Optimizer Considers Nested Loops Joins
Nested loops joins are useful when the database joins small subsets of data, the
database joins large sets of data with the optimizer mode set to FIRST_ROWS, or the
join condition is an efficient method of accessing the inner table.

•

How Nested Loops Joins Work
Conceptually, nested loops are equivalent to two nested for loops.

•

Nested Nested Loops
The outer loop of a nested loop can itself be a row source generated by a different
nested loop.

•

Current Implementation for Nested Loops Joins
Oracle Database 11g introduced a new implementation for nested loops that
reduces overall latency for physical I/O.

9-6

Chapter 9

Join Methods

•

Original Implementation for Nested Loops Joins
In the current release, both the new and original implementation of nested loops
are possible.

•

Nested Loops Controls
For some SQL statements, the data is small enough for the optimizer to prefer full
table scans and hash joins. However, you can add the USE_NL hint to instruct the
optimizer to join each specified table to another row source with a nested loops
join, using the specified table as the inner table.

9.2.1.1 When the Optimizer Considers Nested Loops Joins
Nested loops joins are useful when the database joins small subsets of data, the
database joins large sets of data with the optimizer mode set to FIRST_ROWS, or the join
condition is an efficient method of accessing the inner table.

Note:
The number of rows expected from the join is what drives the optimizer
decision, not the size of the underlying tables. For example, a query might
join two tables of a billion rows each, but because of the filters the optimizer
expects data sets of 5 rows each.

In general, nested loops joins work best on small tables with indexes on the join
conditions. If a row source has only one row, as with an equality lookup on a primary
key value (for example, WHERE employee_id=101), then the join is a simple lookup. The
optimizer always tries to put the smallest row source first, making it the driving table.
Various factors enter into the optimizer decision to use nested loops. For example, the
database may read several rows from the outer row source in a batch. Based on the
number of rows retrieved, the optimizer may choose either a nested loop or a hash join
to the inner row source. For example, if a query joins departments to driving table
employees, and if the predicate specifies a value in employees.last_name, then the
database might read enough entries in the index on last_name to determine whether an
internal threshold is passed. If the threshold is not passed, then the optimizer picks a
nested loop join to departments, and if the threshold is passed, then the database
performs a hash join, which means reading the rest of employees, hashing it into
memory, and then joining to departments.
If the access path for the inner loop is not dependent on the outer loop, then the result
can be a Cartesian product: for every iteration of the outer loop, the inner loop
produces the same set of rows. To avoid this problem, use other join methods to join
two independent row sources.

See Also:
•

"Table 19-2"

•

"Adaptive Query Plans"

9-7

Chapter 9

Join Methods

9.2.1.2 How Nested Loops Joins Work
Conceptually, nested loops are equivalent to two nested for loops.
For example, if a query joins employees and departments, then a nested loop in
pseudocode might be:
FOR erow IN (select * from employees where X=Y) LOOP
FOR drow IN (select * from departments where erow is matched) LOOP
output values from erow and drow
END LOOP
END LOOP

The inner loop is executed for every row of the outer loop. The employees table is the
"outer" data set because it is in the exterior for loop. The outer table is sometimes
called a driving table. The departments table is the "inner" data set because it is in the
interior for loop.
A nested loops join involves the following basic steps:
1.

The optimizer determines the driving row source and designates it as the outer
loop.
The outer loop produces a set of rows for driving the join condition. The row
source can be a table accessed using an index scan, a full table scan, or any other
operation that generates rows.
The number of iterations of the inner loop depends on the number of rows
retrieved in the outer loop. For example, if 10 rows are retrieved from the outer
table, then the database must perform 10 lookups in the inner table. If 10,000,000
rows are retrieved from the outer table, then the database must perform
10,000,000 lookups in the inner table.

2.

The optimizer designates the other row source as the inner loop.
The outer loop appears before the inner loop in the execution plan, as follows:
NESTED LOOPS
outer_loop
inner_loop

3.

For every fetch request from the client, the basic process is as follows:
a.

Fetch a row from the outer row source

b.

Probe the inner row source to find rows that match the predicate criteria

c.

Repeat the preceding steps until all rows are obtained by the fetch request

Sometimes the database sorts rowids to obtain a more efficient buffer access
pattern.

9.2.1.3 Nested Nested Loops
The outer loop of a nested loop can itself be a row source generated by a different
nested loop.
The database can nest two or more outer loops to join as many tables as needed.
Each loop is a data access method. The following template shows how the database
iterates through three nested loops:

9-8

Chapter 9

Join Methods

SELECT STATEMENT
NESTED LOOPS 3
NESTED LOOPS 2
NESTED LOOPS 1
OUTER LOOP 1.1
INNER LOOP 1.2
INNER LOOP 2.2
INNER LOOP 3.2

- Row source becomes OUTER LOOP 3.1
- Row source becomes OUTER LOOP 2.1

The database orders the loops as follows:
1.

The database iterates through NESTED LOOPS 1:
NESTED LOOPS 1
OUTER LOOP 1.1
INNER LOOP 1.2

The output of NESTED LOOP 1 is a row source.
2.

The database iterates through NESTED LOOPS 2, using the row source generated by
NESTED LOOPS 1 as its outer loop:
NESTED LOOPS 2
OUTER LOOP 2.1
INNER LOOP 2.2

- Row source generated by NESTED LOOPS 1

The output of NESTED LOOPS 2 is another row source.
3.

The database iterates through NESTED LOOPS 3, using the row source generated by
NESTED LOOPS 2 as its outer loop:
NESTED LOOPS 3
OUTER LOOP 3.1
INNER LOOP 3.2

Example 9-2

- Row source generated by NESTED LOOPS 2

Nested Nested Loops Join
Suppose you join the employees and departments tables as follows:
SELECT
FROM
WHERE
AND

/*+ ORDERED USE_NL(d) */ e.last_name, e.first_name, d.department_name
employees e, departments d
e.department_id=d.department_id
e.last_name like 'A%';

The plan reveals that the optimizer chose two nested loops (Step 1 and Step 2) to
access the data:
SQL_ID ahuavfcv4tnz4, child number 0
------------------------------------SELECT /*+ ORDERED USE_NL(d) */ e.last_name, d.department_name FROM
employees e, departments d WHERE e.department_id=d.department_id AND
e.last_name like 'A%'
Plan hash value: 1667998133
---------------------------------------------------------------------------------|Id| Operation
|Name
|Rows|Bytes|Cost(%CPU)|Time|
---------------------------------------------------------------------------------| 0| SELECT STATEMENT
|
| | |5 (100)|
|
| 1| NESTED LOOPS
|
| | |
|
|
| 2| NESTED LOOPS
|
| 3|102|5 (0)|00:00:01|
| 3|
TABLE ACCESS BY INDEX ROWID BATCHED| EMPLOYEES | 3| 54|2 (0)|00:00:01|
|*4|
INDEX RANGE SCAN
| EMP_NAME_IX | 3| |1 (0)|00:00:01|

9-9

Chapter 9

Join Methods

|*5|
INDEX UNIQUE SCAN
| DEPT_ID_PK | 1| |0 (0)|
|
| 6| TABLE ACCESS BY INDEX ROWID
| DEPARTMENTS | 1| 16|1 (0)|00:00:01|
---------------------------------------------------------------------------------Predicate Information (identified by operation id):
--------------------------------------------------4 - access("E"."LAST_NAME" LIKE 'A%')
filter("E"."LAST_NAME" LIKE 'A%')
5 - access("E"."DEPARTMENT_ID"="D"."DEPARTMENT_ID")

In this example, the basic process is as follows:
1.

The database begins iterating through the inner nested loop (Step 2) as follows:
a.

The database searches the emp_name_ix for the rowids for all last names that
begins with A (Step 4).
For example:
Abel,employees_rowid
Ande,employees_rowid
Atkinson,employees_rowid
Austin,employees_rowid

b.

Using the rowids from the previous step, the database retrieves a batch of
rows from the employees table (Step 3). For example:
Abel,Ellen,80
Abel,John,50

These rows become the outer row source for the innermost nested loop.
The batch step is typically part of adaptive execution plans. To determine
whether a nested loop is better than a hash join, the optimizer needs to
determine many rows come back from the row source. If too many rows are
returned, then the optimizer switches to a different join method.
c.

For each row in the outer row source, the database scans the dept_id_pk index
to obtain the rowid in departments of the matching department ID (Step 5), and
joins it to the employees rows. For example:
Abel,Ellen,80,departments_rowid
Ande,Sundar,80,departments_rowid
Atkinson,Mozhe,50,departments_rowid
Austin,David,60,departments_rowid

These rows become the outer row source for the outer nested loop (Step 1).
2.

The database iterates through the outer nested loop as follows:
a.

The database reads the first row in outer row source.
For example:
Abel,Ellen,80,departments_rowid

b.

The database uses the departments rowid to retrieve the corresponding row
from departments (Step 6), and then joins the result to obtain the requested
values (Step 1).
For example:
Abel,Ellen,80,Sales

9-10

Chapter 9

Join Methods

c.

The database reads the next row in the outer row source, uses the departments
rowid to retrieve the corresponding row from departments (Step 6), and iterates
through the loop until all rows are retrieved.
The result set has the following form:
Abel,Ellen,80,Sales
Ande,Sundar,80,Sales
Atkinson,Mozhe,50,Shipping
Austin,David,60,IT

9.2.1.4 Current Implementation for Nested Loops Joins
Oracle Database 11g introduced a new implementation for nested loops that reduces
overall latency for physical I/O.
When an index or a table block is not in the buffer cache and is needed to process the
join, a physical I/O is required. The database can batch multiple physical I/O requests
and process them using a vector I/O (array) instead of one at a time. The database
sends an array of rowids to the operating system, which performs the read.
As part of the new implementation, two NESTED LOOPS join row sources might appear in
the execution plan where only one would have appeared in prior releases. In such
cases, Oracle Database allocates one NESTED LOOPS join row source to join the values
from the table on the outer side of the join with the index on the inner side. A second
row source is allocated to join the result of the first join, which includes the rowids
stored in the index, with the table on the inner side of the join.
Consider the query in "Original Implementation for Nested Loops Joins". In the current
implementation, the execution plan for this query might be as follows:
------------------------------------------------------------------------------------| Id | Operation
| Name
|Rows|Bytes|Cost%CPU| Time |
------------------------------------------------------------------------------------| 0 | SELECT STATEMENT
|
| 19 | 722 | 3 (0)|00:00:01|
| 1 | NESTED LOOPS
|
|
|
|
|
|
| 2 | NESTED LOOPS
|
| 19 | 722 | 3 (0)|00:00:01|
|* 3 |
TABLE ACCESS FULL
| DEPARTMENTS
| 2 | 32 | 2 (0)|00:00:01|
|* 4 |
INDEX RANGE SCAN
| EMP_DEPARTMENT_IX | 10 |
| 0 (0)|00:00:01|
| 5 | TABLE ACCESS BY INDEX ROWID| EMPLOYEES
| 10 | 220 | 1 (0)|00:00:01|
------------------------------------------------------------------------------------Predicate Information (identified by operation id):
--------------------------------------------------3 - filter("D"."DEPARTMENT_NAME"='Marketing' OR "D"."DEPARTMENT_NAME"='Sales')
4 - access("E"."DEPARTMENT_ID"="D"."DEPARTMENT_ID")

In this case, rows from the hr.departments table form the outer row source (Step 3) of
the inner nested loop (Step 2). The index emp_department_ix is the inner row source
(Step 4) of the inner nested loop. The results of the inner nested loop form the outer
row source (Row 2) of the outer nested loop (Row 1). The hr.employees table is the
outer row source (Row 5) of the outer nested loop.
For each fetch request, the basic process is as follows:
1.

The database iterates through the inner nested loop (Step 2) to obtain the rows
requested in the fetch:
a.

The database reads the first row of departments to obtain the department IDs
for departments named Marketing or Sales (Step 3). For example:

9-11

Chapter 9

Join Methods

Marketing,20

This row set is the outer loop. The database caches the data in the PGA.
b.

The database scans emp_department_ix, which is an index on the employees
table, to find employees rowids that correspond to this department ID (Step 4),
and then joins the result (Step 2).
The result set has the following form:
Marketing,20,employees_rowid
Marketing,20,employees_rowid
Marketing,20,employees_rowid

c.

The database reads the next row of departments, scans emp_department_ix to
find employees rowids that correspond to this department ID, and then iterates
through the loop until the client request is satisfied.
In this example, the database only iterates through the outer loop twice
because only two rows from departments satisfy the predicate filter.
Conceptually, the result set has the following form:
Marketing,20,employees_rowid
Marketing,20,employees_rowid
Marketing,20,employees_rowid
.
.
.
Sales,80,employees_rowid
Sales,80,employees_rowid
Sales,80,employees_rowid
.
.
.

These rows become the outer row source for the outer nested loop (Step 1).
This row set is cached in the PGA.
2.

The database organizes the rowids obtained in the previous step so that it can
more efficiently access them in the cache.

3.

The database begins iterating through the outer nested loop as follows:
a.

The database retrieves the first row from the row set obtained in the previous
step, as in the following example:
Marketing,20,employees_rowid

b.

Using the rowid, the database retrieves a row from employees to obtain the
requested values (Step 1), as in the following example:
Michael,Hartstein,13000,Marketing

c.

The database retrieves the next row from the row set, uses the rowid to probe
employees for the matching row, and iterates through the loop until all rows are
retrieved.
The result set has the following form:
Michael,Hartstein,13000,Marketing
Pat,Fay,6000,Marketing
John,Russell,14000,Sales
Karen,Partners,13500,Sales
Alberto,Errazuriz,12000,Sales
.

9-12

Chapter 9

Join Methods

.
.

In some cases, a second join row source is not allocated, and the execution plan looks
the same as it did before Oracle Database 11g. The following list describes such
cases:
•

All of the columns needed from the inner side of the join are present in the index,
and there is no table access required. In this case, Oracle Database allocates only
one join row source.

•

The order of the rows returned might be different from the order returned in
releases earlier than Oracle Database 12c. Thus, when Oracle Database tries to
preserve a specific ordering of the rows, for example to eliminate the need for an
ORDER BY sort, Oracle Database might use the original implementation for nested
loops joins.

•

The OPTIMIZER_FEATURES_ENABLE initialization parameter is set to a release before
Oracle Database 11g. In this case, Oracle Database uses the original
implementation for nested loops joins.

9.2.1.5 Original Implementation for Nested Loops Joins
In the current release, both the new and original implementation of nested loops are
possible.
For an example of the original implementation, consider the following join of the
hr.employees and hr.departments tables:
SELECT
FROM
WHERE
AND

e.first_name, e.last_name, e.salary, d.department_name
hr.employees e, hr.departments d
d.department_name IN ('Marketing', 'Sales')
e.department_id = d.department_id;

In releases before Oracle Database 11g, the execution plan for this query might
appear as follows:
-----------------------------------------------------------------------------------------------| Id | Operation
| Name
| Rows | Bytes | Cost (%CPU)| Time
|
-----------------------------------------------------------------------------------------------| 0 | SELECT STATEMENT
|
|
19 | 722 |
3 (0)| 00:00:01 |
| 1 | TABLE ACCESS BY INDEX ROWID| EMPLOYEES
|
10 | 220 |
1 (0)| 00:00:01 |
| 2 | NESTED LOOPS
|
|
19 | 722 |
3 (0)| 00:00:01 |
|* 3 |
TABLE ACCESS FULL
| DEPARTMENTS
|
2 |
32 |
2 (0)| 00:00:01 |
|* 4 |
INDEX RANGE SCAN
| EMP_DEPARTMENT_IX |
10 |
|
0 (0)| 00:00:01 |
-----------------------------------------------------------------------------------------------Predicate Information (identified by operation id):
--------------------------------------------------3 - filter("D"."DEPARTMENT_NAME"='Marketing' OR "D"."DEPARTMENT_NAME"='Sales')
4 - access("E"."DEPARTMENT_ID"="D"."DEPARTMENT_ID")

For each fetch request, the basic process is as follows:
1.

The database iterates through the loop to obtain the rows requested in the fetch:
a.

The database reads the first row of departments to obtain the department IDs
for departments named Marketing or Sales (Step 3). For example:
Marketing,20

9-13

Chapter 9

Join Methods

This row set is the outer loop. The database caches the row in the PGA.
b.

The database scans emp_department_ix, which is an index on the
employees.department_id column, to find employees rowids that correspond to
this department ID (Step 4), and then joins the result (Step 2).
Conceptually, the result set has the following form:
Marketing,20,employees_rowid
Marketing,20,employees_rowid
Marketing,20,employees_rowid

c.

The database reads the next row of departments, scans emp_department_ix to
find employees rowids that correspond to this department ID, and iterates
through the loop until the client request is satisfied.
In this example, the database only iterates through the outer loop twice
because only two rows from departments satisfy the predicate filter.
Conceptually, the result set has the following form:
Marketing,20,employees_rowid
Marketing,20,employees_rowid
Marketing,20,employees_rowid
.
.
.
Sales,80,employees_rowid
Sales,80,employees_rowid
Sales,80,employees_rowid
.
.
.

2.

Depending on the circumstances, the database may organize the cached rowids
obtained in the previous step so that it can more efficiently access them.

3.

For each employees rowid in the result set generated by the nested loop, the
database retrieves a row from employees to obtain the requested values (Step 1).
Thus, the basic process is to read a rowid and retrieve the matching employees
row, read the next rowid and retrieve the matching employees row, and so on.
Conceptually, the result set has the following form:
Michael,Hartstein,13000,Marketing
Pat,Fay,6000,Marketing
John,Russell,14000,Sales
Karen,Partners,13500,Sales
Alberto,Errazuriz,12000,Sales
.
.
.

9.2.1.6 Nested Loops Controls
For some SQL statements, the data is small enough for the optimizer to prefer full
table scans and hash joins. However, you can add the USE_NL hint to instruct the
optimizer to join each specified table to another row source with a nested loops join,
using the specified table as the inner table.
The related hint USE_NL_WITH_INDEX(table index) hint instructs the optimizer to join the
specified table to another row source with a nested loops join using the specified table

9-14

Chapter 9

Join Methods

as the inner table. The index is optional. If no index is specified, then the nested loops
join uses an index with at least one join predicate as the index key.
Example 9-3

Nested Loops Hint

Assume that the optimizer chooses a hash join for the following query:
SELECT e.last_name, d.department_name
FROM employees e, departments d
WHERE e.department_id=d.department_id;

The plan looks as follows:
-----------------------------------------------------------------------------|Id | Operation
| Name
| Rows | Bytes |Cost(%CPU)| Time
|
-----------------------------------------------------------------------------| 0 | SELECT STATEMENT |
|
|
| 5 (100)|
|
|*1 | HASH JOIN
|
| 106 | 2862 | 5 (20)| 00:00:01 |
| 2 | TABLE ACCESS FULL| DEPARTMENTS |
27 | 432 | 2 (0)| 00:00:01 |
| 3 | TABLE ACCESS FULL| EMPLOYEES | 107 | 1177 | 2 (0)| 00:00:01 |
------------------------------------------------------------------------------

To force a nested loops join using departments as the inner table, add the USE_NL hint
as in the following query:
SELECT /*+ ORDERED USE_NL(d) */ e.last_name, d.department_name
FROM employees e, departments d
WHERE e.department_id=d.department_id;

The plan looks as follows:
-------------------------------------------------------------------------------| Id | Operation
| Name
| Rows | Bytes |Cost (%CPU)| Time
|
-------------------------------------------------------------------------------| 0 | SELECT STATEMENT |
|
|
| 34 (100)|
|
| 1 | NESTED LOOPS
|
| 106 | 2862 | 34 (3)| 00:00:01 |
| 2 | TABLE ACCESS FULL| EMPLOYEES | 107 | 1177 |
2 (0)| 00:00:01 |
|* 3 | TABLE ACCESS FULL| DEPARTMENTS |
1 |
16 |
0 (0)|
|
-------------------------------------------------------------------------------Predicate Information (identified by operation id):
--------------------------------------------------3 - filter("E"."DEPARTMENT_ID"="D"."DEPARTMENT_ID")

The database obtains the result set as follows:
1.

In the nested loop, the database reads employees to obtain the last name and
department ID for an employee (Step 2). For example:
De Haan,90

2.

For the row obtained in the previous step, the database scans departments to find
the department name that matches the employees department ID (Step 3), and
joins the result (Step 1). For example:
De Haan,Executive

3.

The database retrieves the next row in employees, retrieves the matching row from
departments, and then repeats this process until all rows are retrieved.
The result set has the following form:

9-15

Chapter 9

Join Methods

De Haan,Executive
Kochnar,Executive
Baer,Public Relations
King,Executive
.
.
.

See Also:
•

"Guidelines for Join Order Hints" to learn more about the USE_NL hint

•

Oracle Database SQL Language Reference to learn about the USE_NL
hint

9.2.2 Hash Joins
The database uses a hash join to join larger data sets.
The optimizer uses the smaller of two data sets to build a hash table on the join key in
memory, using a deterministic hash function to specify the location in the hash table in
which to store each row. The database then scans the larger data set, probing the
hash table to find the rows that meet the join condition.
This section contains the following topics:
•

When the Optimizer Considers Hash Joins
In general, the optimizer considers a hash join when a relatively large amount of
data must be joined (or a large percentage of a small table must be joined), and
the join is an equijoin.

•

How Hash Joins Work
A hashing algorithm takes a set of inputs and applies a deterministic hash function
to generate a hash value between 1 and n, where n is the size of the hash table.

•

How Hash Joins Work When the Hash Table Does Not Fit in the PGA
The database must use a different technique when the hash table does not fit
entirely in the PGA. In this case, the database uses a temporary space to hold
portions (called partitions) of the hash table, and sometimes portions of the larger
table that probes the hash table.

•

Hash Join Controls
The USE_HASH hint instructs the optimizer to use a hash join when joining two tables
together.

9.2.2.1 When the Optimizer Considers Hash Joins
In general, the optimizer considers a hash join when a relatively large amount of data
must be joined (or a large percentage of a small table must be joined), and the join is
an equijoin.
A hash join is most cost effective when the smaller data set fits in memory. In this
case, the cost is limited to a single read pass over the two data sets.

9-16

Chapter 9

Join Methods

Because the hash table is in the PGA, Oracle Database can access rows without
latching them. This technique reduces logical I/O by avoiding the necessity of
repeatedly latching and reading blocks in the database buffer cache.
If the data sets do not fit in memory, then the database partitions the row sources, and
the join proceeds partition by partition. This can use a lot of sort area memory, and I/O
to the temporary tablespace. This method can still be the most cost effective,
especially when the database uses parallel query servers.

9.2.2.2 How Hash Joins Work
A hashing algorithm takes a set of inputs and applies a deterministic hash function to
generate a hash value between 1 and n, where n is the size of the hash table.
In a hash join, the input values are the join keys. The output values are indexes (slots)
in an array, which is the hash table.
This section contains the following topics:
•

Hash Tables
To illustrate a hash table, assume that the database hashes hr.departments in a
join of departments and employees. The join key column is department_id.

•

Hash Join: Basic Steps
The optimizer uses the smaller data source to build a hash table on the join key in
memory, and then scans the larger table to find the joined rows.

9.2.2.2.1 Hash Tables
To illustrate a hash table, assume that the database hashes hr.departments in a join of
departments and employees. The join key column is department_id.
The first 5 rows of departments are as follows:
SQL> select * from departments where rownum < 6;
DEPARTMENT_ID
------------10
20
30
40
50

DEPARTMENT_NAME
MANAGER_ID LOCATION_ID
------------------------------ ---------- ----------Administration
200
1700
Marketing
201
1800
Purchasing
114
1700
Human Resources
203
2400
Shipping
121
1500

The database applies the hash function to each department_id in the table, generating
a hash value for each. For this illustration, the hash table has 5 slots (it could have
more or less). Because n is 5, the possible hash values range from 1 to 5. The hash
functions might generate the following values for the department IDs:
f(10)
f(20)
f(30)
f(40)
f(50)

=
=
=
=
=

4
1
4
2
5

Note that the hash function happens to generate the same hash value of 4 for
departments 10 and 30. This is known as a hash collision. In this case, the database
puts the records for departments 10 and 30 in the same slot, using a linked list.
Conceptually, the hash table looks as follows:

9-17

Chapter 9

Join Methods

1
2
3
4
5

20,Marketing,201,1800
40,Human Resources,203,2400
10,Administration,200,1700 -> 30,Purchasing,114,1700
50,Shipping,121,1500

9.2.2.2.2 Hash Join: Basic Steps
The optimizer uses the smaller data source to build a hash table on the join key in
memory, and then scans the larger table to find the joined rows.
The basic steps are as follows:
1.

The database performs a full scan of the smaller data set, called the build table,
and then applies a hash function to the join key in each row to build a hash table in
the PGA.
In pseudocode, the algorithm might look as follows:
FOR small_table_row IN (SELECT * FROM small_table)
LOOP
slot_number := HASH(small_table_row.join_key);
INSERT_HASH_TABLE(slot_number,small_table_row);
END LOOP;

2.

The database probes the second data set, called the probe table, using
whichever access mechanism has the lowest cost.
Typically, the database performs a full scan of both the smaller and larger data
set. The algorithm in pseudocode might look as follows:
FOR large_table_row IN (SELECT * FROM large_table)
LOOP
slot_number := HASH(large_table_row.join_key);
small_table_row = LOOKUP_HASH_TABLE(slot_number,large_table_row.join_key);
IF small_table_row FOUND
THEN
output small_table_row + large_table_row;
END IF;
END LOOP;

For each row retrieved from the larger data set, the database does the following:
a.

Applies the same hash function to the join column or columns to calculate the
number of the relevant slot in the hash table.
For example, to probe the hash table for department ID 30, the database
applies the hash function to 30, which generates the hash value 4.

b.

Probes the hash table to determine whether rows exists in the slot.
If no rows exist, then the database processes the next row in the larger data
set. If rows exist, then the database proceeds to the next step.

c.

Checks the join column or columns for a match. If a match occurs, then the
database either reports the rows or passes them to the next step in the plan,
and then processes the next row in the larger data set.
If multiple rows exist in the hash table slot, the database walks through the
linked list of rows, checking each one. For example, if department 30 hashes
to slot 4, then the database checks each row until it finds 30.

9-18

Chapter 9

Join Methods

Example 9-4

Hash Joins

An application queries the oe.orders and oe.order_items tables, joining on the order_id
column.
SELECT o.customer_id, l.unit_price * l.quantity
FROM orders o, order_items l
WHERE l.order_id = o.order_id;

The execution plan is as follows:
-------------------------------------------------------------------------| Id | Operation
| Name
| Rows | Bytes | Cost (%CPU)|
-------------------------------------------------------------------------| 0 | SELECT STATEMENT
|
| 665 | 13300 |
8 (25)|
|* 1 | HASH JOIN
|
| 665 | 13300 |
8 (25)|
| 2 | TABLE ACCESS FULL | ORDERS
| 105 | 840 |
4 (25)|
| 3 | TABLE ACCESS FULL | ORDER_ITEMS | 665 | 7980 |
4 (25)|
-------------------------------------------------------------------------Predicate Information (identified by operation id):
--------------------------------------------------1 - access("L"."ORDER_ID"="O"."ORDER_ID")

Because the orders table is small relative to the order_items table, which is 6 times
larger, the database hashes orders. In a hash join, the data set for the build table
always appears first in the list of operations (Step 2). In Step 3, the database performs
a full scan of the larger order_items later, probing the hash table for each row.

9.2.2.3 How Hash Joins Work When the Hash Table Does Not Fit in the PGA
The database must use a different technique when the hash table does not fit entirely
in the PGA. In this case, the database uses a temporary space to hold portions (called
partitions) of the hash table, and sometimes portions of the larger table that probes the
hash table.
The basic process is as follows:
1.

The database performs a full scan of the smaller data set, and then builds an array
of hash buckets in both the PGA and on disk.
When the PGA hash area fills up, the database finds the largest partition within the
hash table and writes it to temporary space on disk. The database stores any new
row that belongs to this on-disk partition on disk, and all other rows in the PGA.
Thus, part of the hash table is in memory and part of it on disk.

2.

The database takes a first pass at reading the other data set.
For each row, the database does the following:
a.

Applies the same hash function to the join column or columns to calculate the
number of the relevant hash bucket.

b.

Probes the hash table to determine whether rows exist in the bucket in
memory.
If the hashed value points to a row in memory, then the database completes
the join and returns the row. If the value points to a hash partition on disk,
however, then the database stores this row in the temporary tablespace, using
the same partitioning scheme used for the original data set.

3.

The database reads each on-disk temporary partition one by one

9-19

Chapter 9

Join Methods

4.

The database joins each partition row to the row in the corresponding on-disk
temporary partition.

9.2.2.4 Hash Join Controls
The USE_HASH hint instructs the optimizer to use a hash join when joining two tables
together.

See Also:
•

"Guidelines for Join Order Hints"

•

Oracle Database SQL Language Reference to learn about USE_HASH

9.2.3 Sort Merge Joins
A sort merge join is a variation on a nested loops join.
If the two data sets in the join are not already sorted, then the database sorts them.
These are the SORT JOIN operations. For each row in the first data set, the database
probes the second data set for matching rows and joins them, basing its start position
on the match made in the previous iteration. This is the MERGE JOIN operation.
Figure 9-6

Sort Merge Join

MERGE JOIN

SORT JOIN

SORT JOIN

First Row
Source

Second Row
Source

This section contains the following topics:
•

When the Optimizer Considers Sort Merge Joins
A hash join requires one hash table and one probe of this table, whereas a sort
merge join requires two sorts.

•

How Sort Merge Joins Work
As in a nested loops join, a sort merge join reads two data sets, but sorts them
when they are not already sorted. For each row in the first data set, the database
finds a starting row in the second data set, and then reads the second data set
until it finds a nonmatching row.

9-20

Chapter 9

Join Methods

•

Sort Merge Join Controls
The USE_MERGE hint instructs the optimizer to use a sort merge join.

9.2.3.1 When the Optimizer Considers Sort Merge Joins
A hash join requires one hash table and one probe of this table, whereas a sort merge
join requires two sorts.
The optimizer may choose a sort merge join over a hash join for joining large amounts
of data when any of the following conditions is true:
•

The join condition between two tables is not an equijoin, that is, uses an inequality
condition such as <, <=, >, or >=.
In contrast to sort merges, hash joins require an equality condition.

•

Because of sorts required by other operations, the optimizer finds it cheaper to use
a sort merge.
If an index exists, then the database can avoid sorting the first data set. However,
the database always sorts the second data set, regardless of indexes.

A sort merge has the same advantage over a nested loops join as the hash join: the
database accesses rows in the PGA rather than the SGA, reducing logical I/O by
avoiding the necessity of repeatedly latching and reading blocks in the database buffer
cache. In general, hash joins perform better than sort merge joins because sorting is
expensive. However, sort merge joins offer the following advantages over a hash join:
•

After the initial sort, the merge phase is optimized, resulting in faster generation of
output rows.

•

A sort merge can be more cost-effective than a hash join when the hash table
does not fit completely in memory.
A hash join with insufficient memory requires both the hash table and the other
data set to be copied to disk. In this case, the database may have to read from
disk multiple times. In a sort merge, if memory cannot hold the two data sets, then
the database writes them both to disk, but reads each data set no more than once.

9.2.3.2 How Sort Merge Joins Work
As in a nested loops join, a sort merge join reads two data sets, but sorts them when
they are not already sorted. For each row in the first data set, the database finds a
starting row in the second data set, and then reads the second data set until it finds a
nonmatching row.
In pseudocode, the high-level algorithm for sort merge might look as follows:
READ data_set_1 SORT BY JOIN KEY TO temp_ds1
READ data_set_2 SORT BY JOIN KEY TO temp_ds2
READ ds1_row FROM temp_ds1
READ ds2_row FROM temp_ds2
WHILE NOT eof ON temp_ds1,temp_ds2
LOOP
IF ( temp_ds1.key = temp_ds2.key ) OUTPUT JOIN ds1_row,ds2_row
ELSIF ( temp_ds1.key <= temp_ds2.key ) READ ds1_row FROM temp_ds1
ELSIF ( temp_ds1.key => temp_ds2.key ) READ ds2_row FROM temp_ds2
END LOOP

9-21

Chapter 9

Join Methods

For example, the following table shows sorted values in two data sets: temp_ds1 and
temp_ds2.

Table 9-3

Sorted Data Sets

temp_ds1

temp_ds2

10

20

20

20

30

40

40

40

50

40

60

40

70

40

.

60

.

70

.

70

As shown in the following table, the database begins by reading 10 in temp_ds1, and
then reads the first value in temp_ds2. Because 20 in temp_ds2 is higher than 10 in
temp_ds1, the database stops reading temp_ds2.
Table 9-4

Start at 10 in temp_ds1

temp_ds1

temp_ds2

Action

10 [start here]

20 [start here] [stop
here]

20 in temp_ds2 is higher than 10 in temp_ds1. Stop.
Start again with next row in temp_ds1.

20

20

30

40

40

40

50

40

60

40

70

40

.

60

.

70

.

70

The database proceeds to the next value in temp_ds1, which is 20. The database
proceeds through temp_ds2 as shown in the following table.
Table 9-5

Start at 20 in temp_ds1

temp_ds1

temp_ds2

Action

10

20 [start here]

Match. Proceed to next value in temp_ds2.

20 [start here]

20

Match. Proceed to next value in temp_ds2.

30

40 [stop here]

40 in temp_ds2 is higher than 20 in temp_ds1. Stop.
Start again with next row in temp_ds1.

40

40

50

40

60

40

9-22

Chapter 9

Join Methods

Table 9-5

(Cont.) Start at 20 in temp_ds1

temp_ds1

temp_ds2

70

40

.

60

.

70

.

70

Action

The database proceeds to the next row in temp_ds1, which is 30. The database starts at
the number of its last match, which was 20, and then proceeds through temp_ds2
looking for a match, as shown in the following table.
Table 9-6

Start at 30 in temp_ds1

temp_ds1

temp_ds2

Action

10

20

20

20 [start at last match] 20 in temp_ds2 is lower than 30 in temp_ds1.
Proceed to next value in temp_ds2.

30 [start here]

40 [stop here]

40

40

50

40

60

40

70

40

.

60

.

70

.

70

40 in temp_ds2 is higher than 30 in temp_ds1. Stop.
Start again with next row in temp_ds1.

The database proceeds to the next row in temp_ds1, which is 40. As shown in the
following table, the database starts at the number of its last match in temp_ds2, which
was 20, and then proceeds through temp_ds2 looking for a match.
Table 9-7

Start at 40 in temp_ds1

temp_ds1

temp_ds2

Action

10

20

20

20 [start at last match] 20 in temp_ds2 is lower than 40 in temp_ds1.
Proceed to next value in temp_ds2.

30

40

Match. Proceed to next value in temp_ds2.

40 [start here]

40

Match. Proceed to next value in temp_ds2.

50

40

Match. Proceed to next value in temp_ds2.

60

40

Match. Proceed to next value in temp_ds2.

70

40

Match. Proceed to next value in temp_ds2.

.

60 [stop here]

60 in temp_ds2 is higher than 40 in temp_ds1. Stop.
Start again with next row in temp_ds1.

.

70

.

70

9-23

Chapter 9

Join Methods

The database continues in this way until it has matched the final 70 in temp_ds2. This
scenario demonstrates that the database, as it reads through temp_ds1, does not need
to read every row in temp_ds2. This is an advantage over a nested loops join.
Example 9-5

Sort Merge Join Using Index

The following query joins the employees and departments tables on the department_id
column, ordering the rows on department_id as follows:
SELECT e.employee_id, e.last_name, e.first_name, e.department_id,
d.department_name
FROM employees e, departments d
WHERE e.department_id = d.department_id
ORDER BY department_id;

A query of DBMS_XPLAN.DISPLAY_CURSOR shows that the plan uses a sort merge join:
-------------------------------------------------------------------------------|Id| Operation
| Name
|Rows|Bytes |Cost (%CPU)| Time |
-------------------------------------------------------------------------------| 0| SELECT STATEMENT
|
|
|
| 5(100)|
|
| 1| MERGE JOIN
|
|106 | 4028 | 5 (20)| 00:00:01 |
| 2| TABLE ACCESS BY INDEX ROWID| DEPARTMENTS | 27 | 432 | 2 (0)| 00:00:01 |
| 3|
INDEX FULL SCAN
| DEPT_ID_PK | 27 |
| 1 (0)| 00:00:01 |
|*4| SORT JOIN
|
|107 | 2354 | 3 (34)| 00:00:01 |
| 5|
TABLE ACCESS FULL
| EMPLOYEES |107 | 2354 | 2 (0)| 00:00:01 |
-------------------------------------------------------------------------------Predicate Information (identified by operation id):
--------------------------------------------------4 - access("E"."DEPARTMENT_ID"="D"."DEPARTMENT_ID")
filter("E"."DEPARTMENT_ID"="D"."DEPARTMENT_ID")

The two data sets are the departments table and the employees table. Because an index
orders the departments table by department_id, the database can read this index and
avoid a sort (Step 3). The database only needs to sort the employees table (Step 4),
which is the most CPU-intensive operation.
Example 9-6

Sort Merge Join Without an Index

You join the employees and departments tables on the department_id column, ordering
the rows on department_id as follows. In this example, you specify NO_INDEX and
USE_MERGE to force the optimizer to choose a sort merge:
SELECT /*+ USE_MERGE(d e) NO_INDEX(d) */ e.employee_id, e.last_name, e.first_name,
e.department_id, d.department_name
FROM employees e, departments d
WHERE e.department_id = d.department_id
ORDER BY department_id;

A query of DBMS_XPLAN.DISPLAY_CURSOR shows that the plan uses a sort merge join:
-------------------------------------------------------------------------------| Id| Operation
| Name
| Rows | Bytes | Cost (%CPU)| Time
|
-------------------------------------------------------------------------------| 0 | SELECT STATEMENT
|
|
|
|
6 (100)|
|
| 1 | MERGE JOIN
|
| 106 | 9540 |
6 (34)| 00:00:01|
| 2 | SORT JOIN
|
|
27 | 567 |
3 (34)| 00:00:01|
| 3 |
TABLE ACCESS FULL| DEPARTMENTS |
27 | 567 |
2 (0)| 00:00:01|
|*4 | SORT JOIN
|
| 107 | 7383 |
3 (34)| 00:00:01|

9-24

Chapter 9

Join Methods

| 5 |
TABLE ACCESS FULL| EMPLOYEES | 107 | 7383 |
2 (0)| 00:00:01|
-------------------------------------------------------------------------------Predicate Information (identified by operation id):
--------------------------------------------------4 - access("E"."DEPARTMENT_ID"="D"."DEPARTMENT_ID")
filter("E"."DEPARTMENT_ID"="D"."DEPARTMENT_ID")

Because the departments.department_id index is ignored, the optimizer performs a sort,
which increases the combined cost of Step 2 and Step 3 by 67% (from 3 to 5).

9.2.3.3 Sort Merge Join Controls
The USE_MERGE hint instructs the optimizer to use a sort merge join.
In some situations it may make sense to override the optimizer with the USE_MERGE hint.
For example, the optimizer can choose a full scan on a table and avoid a sort
operation in a query. However, there is an increased cost because a large table is
accessed through an index and single block reads, as opposed to faster access
through a full table scan.

See Also:
Oracle Database SQL Language Reference to learn about the USE_MERGE hint

9.2.4 Cartesian Joins
The database uses a Cartesian join when one or more of the tables does not have
any join conditions to any other tables in the statement.
The optimizer joins every row from one data source with every row from the other data
source, creating the Cartesian product of the two sets. Therefore, the total number of
rows resulting from the join is calculated using the following formula, where rs1 is the
number of rows in first row set and rs2 is the number of rows in the second row set:
rs1 X rs2 = total rows in result set

This section contains the following topics:
•

When the Optimizer Considers Cartesian Joins
The optimizer uses a Cartesian join for two row sources only in specific
circumstances.

•

How Cartesian Joins Work
A Cartesian join uses nested FOR loops.

•

Cartesian Join Controls
The ORDERED hint instructs the optimizer to join tables in the order in which they
appear in the FROM clause. By forcing a join between two row sources that have no
direct connection, the optimizer must perform a Cartesian join.

9.2.4.1 When the Optimizer Considers Cartesian Joins
The optimizer uses a Cartesian join for two row sources only in specific circumstances.

9-25

Chapter 9

Join Methods

Typically, the situation is one of the following:
•

No join condition exists.
In some cases, the optimizer could pick up a common filter condition between the
two tables as a possible join condition.

Note:
If a Cartesian join appears in a query plan, it could be caused by an
inadvertently omitted join condition. In general, if a query joins n tables,
then n-1 join conditions are required to avoid a Cartesian join.
•

A Cartesian join is an efficient method.
For example, the optimizer may decide to generate a Cartesian product of two
very small tables that are both joined to the same large table.

•

The ORDERED hint specifies a table before its join table is specified.

9.2.4.2 How Cartesian Joins Work
A Cartesian join uses nested FOR loops.
At a high level, the algorithm for a Cartesian join looks as follows, where ds1 is typically
the smaller data set, and ds2 is the larger data set:
FOR ds1_row IN ds1 LOOP
FOR ds2_row IN ds2 LOOP
output ds1_row and ds2_row
END LOOP
END LOOP

Example 9-7

Cartesian Join

In this example, a user intends to perform an inner join of the employees and
departments tables, but accidentally leaves off the join condition:
SELECT e.last_name, d.department_name
FROM employees e, departments d

The execution plan is as follows:
-------------------------------------------------------------------------------| Id| Operation
| Name
| Rows | Bytes |Cost (%CPU)| Time |
-------------------------------------------------------------------------------| 0 | SELECT STATEMENT
|
|
|
|11 (100)|
|
| 1 | MERGE JOIN CARTESIAN |
| 2889 | 57780 |11 (0)| 00:00:01 |
| 2 | TABLE ACCESS FULL
| DEPARTMENTS |
27 | 324 | 2 (0)| 00:00:01 |
| 3 | BUFFER SORT
|
| 107 | 856 | 9 (0)| 00:00:01 |
| 4 |
INDEX FAST FULL SCAN| EMP_NAME_IX | 107 | 856 | 0 (0)|
|
--------------------------------------------------------------------------------

In Step 1 of the preceding plan, the CARTESIAN keyword indicates the presence of a
Cartesian join. The number of rows (2889) is the product of 27 and 107.
In Step 3, the BUFFER SORT operation indicates that the database is copying the data
blocks obtained by the scan of emp_name_ix from the SGA to the PGA. This strategy

9-26

Chapter 9

Join Types

avoids multiple scans of the same blocks in the database buffer cache, which would
generate many logical reads and permit resource contention.

9.2.4.3 Cartesian Join Controls
The ORDERED hint instructs the optimizer to join tables in the order in which they appear
in the FROM clause. By forcing a join between two row sources that have no direct
connection, the optimizer must perform a Cartesian join.
Example 9-8

ORDERED Hint

In the following example, the ORDERED hint instructs the optimizer to join employees and
locations, but no join condition connects these two row sources:
SELECT
FROM
WHERE
AND

/*+ORDERED*/ e.last_name, d.department_name, l.country_id, l.state_province
employees e, locations l, departments d
e.department_id = d.department_id
d.location_id = l.location_id

The following execution plan shows a Cartesian product (Step 3) between locations
(Step 6) and employees (Step 4), which is then joined to the departments table (Step 2):
-------------------------------------------------------------------------------| Id| Operation
| Name
| Rows | Bytes |Cost (%CPU)|Time
|
-------------------------------------------------------------------------------| 0 | SELECT STATEMENT
|
|
|
| 37 (100)|
|
|*1 | HASH JOIN
|
| 106 | 4664 | 37 (6)| 00:00:01 |
| 2 | TABLE ACCESS FULL | DEPARTMENTS |
27 | 513 | 2 (0)| 00:00:01 |
| 3 |
MERGE JOIN CARTESIAN|
| 2461 | 61525 | 34 (3)| 00:00:01 |
| 4 |
TABLE ACCESS FULL | EMPLOYEES | 107 | 1177 | 2 (0)| 00:00:01 |
| 5 |
BUFFER SORT
|
|
23 | 322 | 32 (4)| 00:00:01 |
| 6 |
TABLE ACCESS FULL | LOCATIONS |
23 | 322 | 0 (0)|
|
--------------------------------------------------------------------------------

See Also:
Oracle Database SQL Language Reference to learn about the ORDERED hint

9.3 Join Types
A join type is determined by the type of join condition.
This section contains the following topics:
•

Inner Joins
An inner join (sometimes called a simple join) is a join that returns only rows that
satisfy the join condition. Inner joins are either equijoins or nonequijoins.

•

Outer Joins
An outer join returns all rows that satisfy the join condition and also rows from
one table for which no rows from the other table satisfy the condition. Thus, the
result set of an outer join is the superset of an inner join.

•

Semijoins
A semijoin is a join between two data sets that returns a row from the first set
when a matching row exists in the subquery data set.

9-27

Chapter 9

Join Types

•

Antijoins
An antijoin is a join between two data sets that returns a row from the first set
when a matching row does not exist in the subquery data set.

9.3.1 Inner Joins
An inner join (sometimes called a simple join) is a join that returns only rows that
satisfy the join condition. Inner joins are either equijoins or nonequijoins.
This section contains the following topics:
•

Equijoins
An equijoin is an inner join whose join condition contains an equality operator.

•

Nonequijoins
A nonequijoin is an inner join whose join condition contains an operator that is
not an equality operator.

•

Band Joins
A band join is a special type of nonequijoin in which key values in one data set
must fall within the specified range (“band”) of the second data set. The same
table can serve as both the first and second data sets.

9.3.1.1 Equijoins
An equijoin is an inner join whose join condition contains an equality operator.
The following example is an equijoin because the join condition contains only an
equality operator:
SELECT e.employee_id, e.last_name, d.department_name
FROM employees e, departments d
WHERE e.department_id=d.department_id;

In the preceding query, the join condition is e.department_id=d.department_id. If a row
in the employees table has a department ID that matches the value in a row in the
departments table, then the database returns the joined result; otherwise, the database
does not return a result.

9.3.1.2 Nonequijoins
A nonequijoin is an inner join whose join condition contains an operator that is not an
equality operator.
The following query lists all employees whose hire date occurred when employee 176
(who is listed in job_history because he changed jobs in 2007) was working at the
company:
SELECT
FROM
WHERE
AND

e.employee_id, e.first_name, e.last_name, e.hire_date
employees e, job_history h
h.employee_id = 176
e.hire_date BETWEEN h.start_date AND h.end_date;

In the preceding example, the condition joining employees and job_history does not
contain an equality operator, so it is a nonequijoin. Nonequijoins are relatively rare.
Note that a hash join requires at least a partial equijoin. The following SQL script
contains an equality join condition (e1.empno = e2.empno) and a nonequality condition:

9-28

Chapter 9

Join Types

SET AUTOTRACE TRACEONLY EXPLAIN
SELECT *
FROM scott.emp e1 JOIN scott.emp e2
ON
( e1.empno = e2.empno
AND
e1.hiredate BETWEEN e2.hiredate-1 AND e2.hiredate+1 )

The optimizer chooses a hash join for the preceding query, as shown in the following
plan:
Execution Plan
---------------------------------------------------------Plan hash value: 3638257876
--------------------------------------------------------------------------| Id | Operation
| Name | Rows | Bytes | Cost (%CPU)| Time
|
--------------------------------------------------------------------------| 0 | SELECT STATEMENT |
|
1 | 174 |
5 (20)| 00:00:01 |
|* 1 | HASH JOIN
|
|
1 | 174 |
5 (20)| 00:00:01 |
| 2 | TABLE ACCESS FULL| EMP |
14 | 1218 |
2 (0)| 00:00:01 |
| 3 | TABLE ACCESS FULL| EMP |
14 | 1218 |
2 (0)| 00:00:01 |
--------------------------------------------------------------------------Predicate Information (identified by operation id):
--------------------------------------------------1 - access("E1"."EMPNO"="E2"."EMPNO")
filter("E1"."HIREDATE">=INTERNAL_FUNCTION("E2"."HIREDATE")-1 AND
"E1"."HIREDATE"<=INTERNAL_FUNCTION("E2"."HIREDATE")+1)

9.3.1.3 Band Joins
A band join is a special type of nonequijoin in which key values in one data set must
fall within the specified range (“band”) of the second data set. The same table can
serve as both the first and second data sets.
Starting in Oracle Database 12c Release 2 (12.2), the database evaluates band joins
more efficiently. The optimization avoids the unnecessary scanning of rows that fall
outside the defined bands.
The optimizer uses a cost estimate to choose the join method (hash, nested loops, or
sort merge) and the parallel data distribution method. In most cases, optimized
performance is comparable to an equijoin.
This following examples query employees whose salaries are between $100 less
and $100 more than the salary of each employee. Thus, the band has a width of $200.
The examples assume that it is permissible to compare the salary of every employee
with itself. The following query includes partial sample output:
SELECT e1.last_name ||
' has salary between 100 less and 100 more than ' ||
e2.last_name AS "SALARY COMPARISON"
FROM
employees e1,
employees e2
WHERE e1.salary
BETWEEN e2.salary - 100
AND
e2.salary + 100;
SALARY COMPARISON
------------------------------------------------------------King has salary between 100 less and 100 more than King

9-29

Chapter 9

Join Types

Kochhar has salary between 100 less and 100 more than Kochhar
Kochhar has salary between 100 less and 100 more than De Haan
De Haan has salary between 100 less and 100 more than Kochhar
De Haan has salary between 100 less and 100 more than De Haan
Russell has salary between 100 less and 100 more than Russell
Partners has salary between 100 less and 100 more than Partners
...

Example 9-9

Query Without Band Join Optimization

Without the band join optimization, the database uses the following query plan:
-----------------------------------------PLAN_TABLE_OUTPUT
----------------------------------------------------------------------------------| Id | Operation
| Name
|
-----------------------------------------| 0 | SELECT STATEMENT
|
|
| 1 | MERGE JOIN
|
|
| 2 | SORT JOIN
|
|
| 3 |
TABLE ACCESS FULL | EMPLOYEES |
|* 4 | FILTER
|
|
|* 5 |
SORT JOIN
|
|
| 6 |
TABLE ACCESS FULL| EMPLOYEES |
-----------------------------------------Predicate Information (identified by operation id):
--------------------------------------------------4 - filter("E1"."SAL"<="E2"."SAL"+100)
5 - access(INTERNAL_FUNCTION("E1"."SAL")>="E2"."SAL"-100)
filter(INTERNAL_FUNCTION("E1"."SAL")>="E2"."SAL"-100)

In this plan, Step 2 sorts the e1 row source, and Step 5 sorts the e2 row source. The
sorted row sources are illustrated in the following table.
Table 9-8

Sorted row Sources

e1 Sorted (Step 2 of Plan)

e2 Sorted (Step 5 of Plan)

24000 (King)

24000 (King)

17000 (Kochhar)

17000 (Kochhar)

17000 (De Haan)

17000 (De Haan)

14000 (Russell)

14000 (Russell)

13500 (Partners)

13500 (Partners)

The join begins by iterating through the sorted input (e1), which is the left branch of the
join, corresponding to Step 2 of the plan. The original query contains two predicates:
•

e1.sal >= e2.sal–100, which is the Step 5 filter

•

e1.sal >= e2.sal+100, which is the Step 4 filter

For each iteration of the sorted row source e1, the database iterates through row
source e2, checking every row against Step 5 filter e1.sal >= e2.sal–100. If the row
passes the Step 5 filter, then the database sends it to the Step 4 filter, and then
proceeds to test the next row in e2 against the Step 5 filter. However, if a row fails the

9-30

Chapter 9

Join Types

Step 5 filter, then the scan of e2 stops, and the database proceeds through the next
iteration of e1.
The following table shows the first iteration of e1, which begins with 24000 (King) in
data set e1. The database determines that the first row in e2, which is 24000 (King),
passes the Step 5 filter. The database then sends the row to the Step 4 filter, e1.sal
<= w2.sal+100, which also passes. The database sends this row to the MERGE row
source. Next, the database checks 17000 (Kochhar) against the Step 5 filter, which also
passes. However, the row fails the Step 4 filter, and is discarded. The database
proceeds to test 17000 (De Haan) against the Step 5 filter.
Table 9-9

First Iteration of e1: Separate SORT JOIN and FILTER

Scan e2

Step 5 Filter (e1.sal >= e2.sal–100) Step 4 Filter (e1.sal <= e2.sal+100)

24000 (King)

Pass because 24000 >= 23900.
Send to Step 4 filter.

Pass because 24000 <= 24100.
Return row for merging.

17000
(Kochhar)

Pass because 24000 >= 16900.
Send to Step 4 filter.

Fail because 24000 <=17100 is false.
Discard row. Scan next row in e2.

17000 (De
Haan)

Pass because 24000 >= 16900.
Send to Step 4 filter.

Fail because 24000 <=17100 is false.
Discard row. Scan next row in e2.

14000 (Russell) Pass because 24000 >= 13900.
Send to Step 4 filter.

Fail because 24000 <=14100 is false.
Discard row. Scan next row in e2.

13500
(Partners)

Fail because 24000 <=13600 is false.
Discard row. Scan next row in e2.

Pass because 24000 >= 13400.
Send to Step 4 filter.

As shown in the preceding table, every e2 row necessarily passes the Step 5 filter
because the e2 salaries are sorted in descending order. Thus, the Step 5 filter always
sends the row to the Step 4 filter. Because the e2 salaries are sorted in descending
order, the Step 4 filter necessarily fails every row starting with 17000 (Kochhar). The
inefficiency occurs because the database tests every subsequent row in e2 against the
Step 5 filter, which necessarily passes, and then against the Step 4 filter, which
necessarily fails.
Example 9-10

Query With Band Join Optimization

Starting in Oracle Database 12c Release 2 (12.2), the database optimizes the band
join by using the following plan, which does not have a separate FILTER operation:
-----------------------------------------PLAN_TABLE_OUTPUT
-----------------------------------------| Id | Operation
| Name
|
-----------------------------------------| 0 | SELECT STATEMENT
|
|
| 1 | MERGE JOIN
|
|
| 2 | SORT JOIN
|
|
| 3 |
TABLE ACCESS FULL | EMPLOYEES |
|* 4 | SORT JOIN
|
|
| 5 |
TABLE ACCESS FULL | EMPLOYEES |
-----------------------------------------Predicate Information (identified by operation id):
--------------------------------------------------4 - access(INTERNAL_FUNCTION("E1"."SALARY")>="E2"."SALARY"-100)
filter(("E1"."SALARY"<="E2"."SALARY"+100 AND
INTERNAL_FUNCTION("E1"."SALARY")>="E2"."SALARY"-100))

9-31

Chapter 9

Join Types

The difference is that Step 4 uses Boolean AND logic for the two predicates to create a
single filter. Instead of checking a row against one filter, and then sending it to a
different row source for checking against a second filter, the database performs one
check against one filter. If the check fails, then processing stops.
In this example, the query begins the first iteration of e1, which begins with 24000
(King). The following figure represents the range. e2 values below 23900 and above
24100 fall outside the range.
Figure 9-7

Band Join

The following table shows that the database tests the first row of e2, which is 24000
(King), against the Step 4 filter. The row passes the test, so the database sends the
row to be merged. The next row in e2 is 17000 (Kochhar). This row falls outside of the
range (band) and thus does not satisfy the filter predicate, so the database stops
testing e2 rows in this iteration. The database stops testing because the descending
sort of e2 ensures that all subsequent rows in e2 fail the filter test. Thus, the database
can proceed to the second iteration of e1.
Table 9-10

First Iteration of e1: Single SORT JOIN

Scan e2

Filter 4 (e1.sal >= e2.sal – 100) AND (e1.sal <= e2.sal + 100)

24000 (King)

Passes test because it is true that (24000 >= 23900) AND (24000 <=
24100).
Send row to MERGE. Test next row.

17000 (Kochhar)

Fails test because it is false that (24000 >= 16900) AND (24000 <=
17100).
Stop scanning e2. Begin next iteration of e1.

17000 (De Haan)

n/a

14000 (Russell)

n/a

13500 (Partners)

n/a

In this way, the band join optimization eliminates unnecessary processing. Instead of
scanning every row in e2 as in the unoptimized case, the database scans only the
minimum two rows.

9.3.2 Outer Joins
An outer join returns all rows that satisfy the join condition and also rows from one
table for which no rows from the other table satisfy the condition. Thus, the result set
of an outer join is the superset of an inner join.
In ANSI syntax, the OUTER JOIN clause specifies an outer join. In the FROM clause, the
left table appears to the left of the OUTER JOIN keywords, and the right table appears to

9-32

Chapter 9

Join Types

the right of these keywords. The left table is also called the outer table, and the right
table is also called the inner table. For example, in the following statement the
employees table is the left or outer table:
SELECT employee_id, last_name, first_name
FROM employees LEFT OUTER JOIN departments
ON
(employees.department_id=departments.departments_id);

Outer joins require the outer-joined table to be the driving table. In the preceding
example, employees is the driving table, and departments is the driven-to table.
This section contains the following topics:
•

Nested Loops Outer Joins
The database uses this operation to loop through an outer join between two
tables. The outer join returns the outer (preserved) table rows, even when no
corresponding rows are in the inner (optional) table.

•

Hash Join Outer Joins
The optimizer uses hash joins for processing an outer join when either the data
volume is large enough to make a hash join efficient, or it is impossible to drive
from the outer table to the inner table.

•

Sort Merge Outer Joins
When an outer join cannot drive from the outer (preserved) table to the inner
(optional) table, it cannot use a hash join or nested loops joins.

•

Full Outer Joins
A full outer join is a combination of the left and right outer joins.

•

Multiple Tables on the Left of an Outer Join
In Oracle Database 12c, multiple tables may exist on the left of an outer-joined
table. This enhancement enables Oracle Database to merge a view that contains
multiple tables and appears on the left of outer join.

9.3.2.1 Nested Loops Outer Joins
The database uses this operation to loop through an outer join between two tables.
The outer join returns the outer (preserved) table rows, even when no corresponding
rows are in the inner (optional) table.
In a standard nested loop, the optimizer chooses the order of tables—which is the
driving table and which the driven table—based on the cost. However, in a nested loop
outer join, the join condition determines the order of tables. The database uses the
outer, row-preserved table to drive to the inner table.
The optimizer uses nested loops joins to process an outer join in the following
circumstances:
•

It is possible to drive from the outer table to the inner table.

•

Data volume is low enough to make the nested loop method efficient.

For an example of a nested loop outer join, you can add the USE_NL hint to
Example 9-11 to instruct the optimizer to use a nested loop. For example:
SELECT /*+ USE_NL(c o) */ cust_last_name,
SUM(NVL2(o.customer_id,0,1)) "Count"
FROM customers c, orders o
WHERE c.credit_limit > 1000

9-33

Chapter 9

Join Types

AND
c.customer_id = o.customer_id(+)
GROUP BY cust_last_name;

9.3.2.2 Hash Join Outer Joins
The optimizer uses hash joins for processing an outer join when either the data
volume is large enough to make a hash join efficient, or it is impossible to drive from
the outer table to the inner table.
The cost determines the order of tables. The outer table, including preserved rows,
may be used to build the hash table, or it may be used to probe the hash table.
Example 9-11

Hash Join Outer Joins

This example shows a typical hash join outer join query, and its execution plan. In this
example, all the customers with credit limits greater than 1000 are queried. An outer
join is needed so that the query captures customers who have no orders.
•

The outer table is customers.

•

The inner table is orders.

•

The join preserves the customers rows, including those rows without a
corresponding row in orders.

You could use a NOT EXISTS subquery to return the rows. However, because you are
querying all the rows in the table, the hash join performs better (unless the NOT EXISTS
subquery is not nested).
SELECT cust_last_name, SUM(NVL2(o.customer_id,0,1)) "Count"
FROM customers c, orders o
WHERE c.credit_limit > 1000
AND
c.customer_id = o.customer_id(+)
GROUP BY cust_last_name;
-------------------------------------------------------------------------------| Id | Operation
| Name
| Rows | Bytes | Cost (%CPU)| Time
|
-------------------------------------------------------------------------------PLAN_TABLE_OUTPUT
-------------------------------------------------------------------------------| 0 | SELECT STATEMENT
|
|
|
|
7 (100)|
|
| 1 | HASH GROUP BY
|
| 168 | 3192 |
7 (29)| 00:00:01 |
|* 2 |
HASH JOIN OUTER
|
| 318 | 6042 |
6 (17)| 00:00:01 |
|* 3 |
TABLE ACCESS FULL| CUSTOMERS | 260 | 3900 |
3 (0)| 00:00:01 |
|* 4 |
TABLE ACCESS FULL| ORDERS
| 105 | 420 |
2 (0)| 00:00:01 |
-------------------------------------------------------------------------------Predicate Information (identified by operation id):
--------------------------------------------------2 - access("C"."CUSTOMER_ID"="O"."CUSTOMER_ID")
PLAN_TABLE_OUTPUT
------------------------------------------------------------------------------3 - filter("C"."CREDIT_LIMIT">1000)
4 - filter("O"."CUSTOMER_ID">0)

The query looks for customers which satisfy various conditions. An outer join returns
NULL for the inner table columns along with the outer (preserved) table rows when it

9-34

Chapter 9

Join Types

does not find any corresponding rows in the inner table. This operation finds all the
customers rows that do not have any orders rows.
In this case, the outer join condition is the following:
customers.customer_id = orders.customer_id(+)

The components of this condition represent the following:
Example 9-12

Outer Join to a Multitable View

In this example, the outer join is to a multitable view. The optimizer cannot drive into
the view like in a normal join or push the predicates, so it builds the entire row set of
the view.
SELECT c.cust_last_name, sum(revenue)
FROM customers c, v_orders o
WHERE c.credit_limit > 2000
AND
o.customer_id(+) = c.customer_id
GROUP BY c.cust_last_name;
---------------------------------------------------------------------------| Id | Operation
| Name
| Rows | Bytes | Cost (%CPU)|
---------------------------------------------------------------------------| 0 | SELECT STATEMENT
|
| 144 | 4608 |
16 (32)|
| 1 | HASH GROUP BY
|
| 144 | 4608 |
16 (32)|
|* 2 | HASH JOIN OUTER
|
| 663 | 21216 |
15 (27)|
|* 3 |
TABLE ACCESS FULL | CUSTOMERS
| 195 | 2925 |
6 (17)|
| 4 |
VIEW
| V_ORDERS
| 665 | 11305 |
|
| 5 |
HASH GROUP BY
|
| 665 | 15960 |
9 (34)|
|* 6 |
HASH JOIN
|
| 665 | 15960 |
8 (25)|
|* 7 |
TABLE ACCESS FULL| ORDERS
| 105 | 840 |
4 (25)|
| 8 |
TABLE ACCESS FULL| ORDER_ITEMS | 665 | 10640 |
4 (25)|
---------------------------------------------------------------------------Predicate Information (identified by operation id):
--------------------------------------------------2 - access("O"."CUSTOMER_ID"(+)="C"."CUSTOMER_ID")
3 - filter("C"."CREDIT_LIMIT">2000)
6 - access("O"."ORDER_ID"="L"."ORDER_ID")
7 - filter("O"."CUSTOMER_ID">0)

The view definition is as follows:
CREATE OR REPLACE view v_orders AS
SELECT l.product_id, SUM(l.quantity*unit_price) revenue,
o.order_id, o.customer_id
FROM orders o, order_items l
WHERE o.order_id = l.order_id
GROUP BY l.product_id, o.order_id, o.customer_id;

9.3.2.3 Sort Merge Outer Joins
When an outer join cannot drive from the outer (preserved) table to the inner (optional)
table, it cannot use a hash join or nested loops joins.
In this case, it uses the sort merge outer join.
The optimizer uses sort merge for an outer join in the following cases:

9-35

Chapter 9

Join Types

•

A nested loops join is inefficient. A nested loops join can be inefficient because of
data volumes.

•

The optimizer finds it is cheaper to use a sort merge over a hash join because of
sorts required by other operations.

9.3.2.4 Full Outer Joins
A full outer join is a combination of the left and right outer joins.
In addition to the inner join, rows from both tables that have not been returned in the
result of the inner join are preserved and extended with nulls. In other words, full outer
joins join tables together, yet show rows with no corresponding rows in the joined
tables.
Example 9-13

Full Outer Join

The following query retrieves all departments and all employees in each department,
but also includes:
•

Any employees without departments

•

Any departments without employees

SELECT d.department_id, e.employee_id
FROM employees e FULL OUTER JOIN departments d
ON
e.department_id = d.department_id
ORDER BY d.department_id;

The statement produces the following output:
DEPARTMENT_ID EMPLOYEE_ID
------------- ----------10
200
20
201
20
202
30
114
30
115
30
116
...
270
280
178
207
125 rows selected.

Example 9-14

Execution Plan for a Full Outer Join

Starting with Oracle Database 11g, Oracle Database automatically uses a native
execution method based on a hash join for executing full outer joins whenever
possible. When the database uses the new method to execute a full outer join, the
execution plan for the query contains HASH JOIN FULL OUTER. The query in
Example 9-13 uses the following execution plan:
-------------------------------------------------------------------------------| Id| Operation
| Name
|Rows |Bytes |Cost (%CPU)| Time
|
-------------------------------------------------------------------------------| 0 | SELECT STATEMENT
|
| 122 | 4758 | 6 (34)| 00:0 0:01 |
| 1 | SORT ORDER BY
|
| 122 | 4758 | 6 (34)| 00:0 0:01 |
| 2 | VIEW
| VW_FOJ_0 | 122 | 4758 | 5 (20)| 00:0 0:01 |

9-36

Chapter 9

Join Types

|*3 |
HASH JOIN FULL OUTER |
| 122 | 1342 | 5 (20)| 00:0 0:01 |
| 4 |
INDEX FAST FULL SCAN| DEPT_ID_PK | 27 | 108 | 2 (0)| 00:0 0:01 |
| 5 |
TABLE ACCESS FULL | EMPLOYEES | 107 | 749 | 2 (0)| 00:0 0:01 |
-------------------------------------------------------------------------------Predicate Information (identified by operation id):
--------------------------------------------------3 - access("E"."DEPARTMENT_ID"="D"."DEPARTMENT_ID")
HASH JOIN FULL OUTER is included in the preceding plan (Step 3), indicating that the

query uses the hash full outer join execution method. Typically, when the full outer join
condition between two tables is an equijoin, the hash full outer join execution method
is possible, and Oracle Database uses it automatically.
To instruct the optimizer to consider using the hash full outer join execution method,
apply the NATIVE_FULL_OUTER_JOIN hint. To instruct the optimizer not to consider using
the hash full outer join execution method, apply the NO_NATIVE_FULL_OUTER_JOIN hint.
The NO_NATIVE_FULL_OUTER_JOIN hint instructs the optimizer to exclude the native
execution method when joining each specified table. Instead, the full outer join is
executed as a union of left outer join and an antijoin.

9.3.2.5 Multiple Tables on the Left of an Outer Join
In Oracle Database 12c, multiple tables may exist on the left of an outer-joined table.
This enhancement enables Oracle Database to merge a view that contains multiple
tables and appears on the left of outer join.
In releases before Oracle Database 12c, a query such as the following was invalid,
and would trigger an ORA-01417 error message:
SELECT
FROM
WHERE
AND
AND

t1.d, t3.c
t1, t2, t3
t1.z = t2.z
t1.x = t3.x (+)
t2.y = t3.y (+);

Starting in Oracle Database 12c, the preceding query is valid.

9.3.3 Semijoins
A semijoin is a join between two data sets that returns a row from the first set when a
matching row exists in the subquery data set.
The database stops processing the second data set at the first match. Thus,
optimization does not duplicate rows from the first data set when multiple rows in the
second data set satisfy the subquery criteria.

Note:
Semijoins and antijoins are considered join types even though the SQL
constructs that cause them are subqueries. They are internal algorithms that
the optimizer uses to flatten subquery constructs so that they can be
resolved in a join-like way.

9-37

Chapter 9

Join Types

This section contains the following topics:
•

When the Optimizer Considers Semijoins
A semijoin avoids returning a huge number of rows when a query only needs to
determine whether a match exists.

•

How Semijoins Work
The semijoin optimization is implemented differently depending on what type of
join is used.

9.3.3.1 When the Optimizer Considers Semijoins
A semijoin avoids returning a huge number of rows when a query only needs to
determine whether a match exists.
With large data sets, this optimization can result in significant time savings over a
nested loops join that must loop through every record returned by the inner query for
every row in the outer query. The optimizer can apply the semijoin optimization to
nested loops joins, hash joins, and sort merge joins.
The optimizer may choose a semijoin in the following circumstances:
•

The statement uses either an IN or EXISTS clause.

•

The statement contains a subquery in the IN or EXISTS clause.

•

The IN or EXISTS clause is not contained inside an OR branch.

9.3.3.2 How Semijoins Work
The semijoin optimization is implemented differently depending on what type of join is
used.
The following pseudocode shows a semijoin for a nested loops join:
FOR ds1_row IN ds1 LOOP
match := false;
FOR ds2_row IN ds2_subquery LOOP
IF (ds1_row matches ds2_row) THEN
match := true;
EXIT -- stop processing second data set when a match is found
END IF
END LOOP
IF (match = true) THEN
RETURN ds1_row
END IF
END LOOP

In the preceding pseudocode, ds1 is the first data set, and ds2_subquery is the subquery
data set. The code obtains the first row from the first data set, and then loops through
the subquery data set looking for a match. The code exits the inner loop as soon as it
finds a match, and then begins processing the next row in the first data set.
Example 9-15

Semijoin Using WHERE EXISTS

The following query uses a WHERE EXISTS clause to list only the departments that
contain employees:
SELECT department_id, department_name
FROM departments
WHERE EXISTS (SELECT 1

9-38

Chapter 9

Join Types

FROM employees
WHERE employees.department_id = departments.department_id)

The execution plan reveals a NESTED LOOPS SEMI operation in Step 1:
-------------------------------------------------------------------------------| Id| Operation
| Name
|Rows|Bytes|Cost (%CPU)| Time
|
-------------------------------------------------------------------------------| 0 | SELECT STATEMENT |
|
|
|
2 (100)|
|
| 1 | NESTED LOOPS SEMI |
| 11 | 209 |
2 (0)| 00:00:01 |
| 2 | TABLE ACCESS FULL| DEPARTMENTS
| 27 | 432 |
2 (0)| 00:00:01 |
|*3 | INDEX RANGE SCAN | EMP_DEPARTMENT_IX | 44 | 132 |
0 (0)|
|
--------------------------------------------------------------------------------

For each row in departments, which forms the outer loop, the database obtains the
department ID, and then probes the employees.department_id index for matching
entries. Conceptually, the index looks as follows:
10,rowid
10,rowid
10,rowid
10,rowid
30,rowid
30,rowid
30,rowid
...

If the first entry in the departments table is department 30, then the database performs a
range scan of the index until it finds the first 30 entry, at which point it stops reading the
index and returns the matching row from departments. If the next row in the outer loop
is department 20, then the database scans the index for a 20 entry, and not finding any
matches, performs the next iteration of the outer loop. The database proceeds in this
way until all matching rows are returned.
Example 9-16

Semijoin Using IN

The following query uses a IN clause to list only the departments that contain
employees:
SELECT department_id, department_name
FROM departments
WHERE department_id IN
(SELECT department_id
FROM employees);

The execution plan reveals a NESTED LOOPS SEMI operation in Step 1:
-------------------------------------------------------------------------------| Id| Operation
| Name
|Rows|Bytes|Cost (%CPU)| Time
|
-------------------------------------------------------------------------------| 0 | SELECT STATEMENT |
|
|
|
2 (100)|
|
| 1 | NESTED LOOPS SEMI |
| 11 | 209 |
2 (0)| 00:00:01 |
| 2 | TABLE ACCESS FULL| DEPARTMENTS
| 27 | 432 |
2 (0)| 00:00:01 |
|*3 | INDEX RANGE SCAN | EMP_DEPARTMENT_IX | 44 | 132 |
0 (0)|
|
--------------------------------------------------------------------------------

The plan is identical to the plan in Example 9-15.

9-39

Chapter 9

Join Types

9.3.4 Antijoins
An antijoin is a join between two data sets that returns a row from the first set when a
matching row does not exist in the subquery data set.
Like a semijoin, an antijoin stops processing the subquery data set when the first
match is found. Unlike a semijoin, the antijoin only returns a row when no match is
found.
This section contains the following topics:
•

When the Optimizer Considers Antijoins
An antijoin avoids unnecessary processing when a query only needs to return a
row when a match does not exist.

•

How Antijoins Work
The antijoin optimization is implemented differently depending on what type of join
is used.

•

How Antijoins Handle Nulls
For semijoins, IN and EXISTS are functionally equivalent. However, NOT IN and NOT
EXISTS are not functionally equivalent because of nulls.

9.3.4.1 When the Optimizer Considers Antijoins
An antijoin avoids unnecessary processing when a query only needs to return a row
when a match does not exist.
With large data sets, this optimization can result in significant time savings over a
nested loops join. The latter join must loop through every record returned by the inner
query for every row in the outer query. The optimizer can apply the antijoin
optimization to nested loops joins, hash joins, and sort merge joins.
The optimizer may choose an antijoin in the following circumstances:
•

The statement uses either the NOT IN or NOT EXISTS clause.

•

The statement has a subquery in the NOT IN or NOT EXISTS clause.

•

The NOT IN or NOT EXISTS clause is not contained inside an OR branch.

•

The statement performs an outer join and applies an IS NULL condition to a join
column, as in the following example:
SET AUTOTRACE TRACEONLY EXPLAIN
SELECT emp.*
FROM emp, dept
WHERE emp.deptno = dept.deptno(+)
AND
dept.deptno IS NULL
Execution Plan
---------------------------------------------------------Plan hash value: 1543991079
--------------------------------------------------------------------------| Id | Operation
| Name | Rows | Bytes | Cost (%CPU)| Time
|
--------------------------------------------------------------------------| 0 | SELECT STATEMENT |
|
14 | 1400 |
5 (20)| 00:00:01 |
|* 1 | HASH JOIN ANTI
|
|
14 | 1400 |
5 (20)| 00:00:01 |
| 2 | TABLE ACCESS FULL| EMP |
14 | 1218 |
2 (0)| 00:00:01 |

9-40

Chapter 9

Join Types

| 3 | TABLE ACCESS FULL| DEPT |
4 |
52 |
2 (0)| 00:00:01 |
--------------------------------------------------------------------------Predicate Information (identified by operation id):
--------------------------------------------------1 - access("EMP"."DEPTNO"="DEPT"."DEPTNO")
Note
----- dynamic statistics used: dynamic sampling (level=2)

9.3.4.2 How Antijoins Work
The antijoin optimization is implemented differently depending on what type of join is
used.
The following pseudocode shows an antijoin for a nested loops join:
FOR ds1_row IN ds1 LOOP
match := true;
FOR ds2_row IN ds2 LOOP
IF (ds1_row matches ds2_row) THEN
match := false;
EXIT -- stop processing second data set when a match is found
END IF
END LOOP
IF (match = true) THEN
RETURN ds1_row
END IF
END LOOP

In the preceding pseudocode, ds1 is the first data set, and ds2 is the second data set.
The code obtains the first row from the first data set, and then loops through the
second data set looking for a match. The code exits the inner loop as soon as it finds a
match, and begins processing the next row in the first data set.
Example 9-17

Semijoin Using WHERE EXISTS

The following query uses a WHERE EXISTS clause to list only the departments that
contain employees:
SELECT department_id, department_name
FROM departments
WHERE EXISTS (SELECT 1
FROM employees
WHERE employees.department_id = departments.department_id)

The execution plan reveals a NESTED LOOPS SEMI operation in Step 1:
-------------------------------------------------------------------------------| Id| Operation
| Name
|Rows|Bytes |Cost(%CPU)| Time
|
-------------------------------------------------------------------------------| 0 | SELECT STATEMENT |
|
|
| 2 (100)|
|
| 1 | NESTED LOOPS SEMI |
| 11 | 209 | 2 (0)| 00:00:01 |
| 2 | TABLE ACCESS FULL| DEPARTMENTS
| 27 | 432 | 2 (0)| 00:00:01 |
|*3 | INDEX RANGE SCAN | EMP_DEPARTMENT_IX | 44 | 132 | 0 (0)|
|
--------------------------------------------------------------------------------

9-41

Chapter 9

Join Types

For each row in departments, which forms the outer loop, the database obtains the
department ID, and then probes the employees.department_id index for matching
entries. Conceptually, the index looks as follows:
10,rowid
10,rowid
10,rowid
10,rowid
30,rowid
30,rowid
30,rowid
...

If the first record in the departments table is department 30, then the database performs
a range scan of the index until it finds the first 30 entry, at which point it stops reading
the index and returns the matching row from departments. If the next row in the outer
loop is department 20, then the database scans the index for a 20 entry, and not finding
any matches, performs the next iteration of the outer loop. The database proceeds in
this way until all matching rows are returned.

9.3.4.3 How Antijoins Handle Nulls
For semijoins, IN and EXISTS are functionally equivalent. However, NOT IN and NOT
EXISTS are not functionally equivalent because of nulls.
If a null value is returned to a NOT IN operator, then the statement returns no records.
To see why, consider the following WHERE clause:
WHERE department_id NOT IN (null, 10, 20)

The database tests the preceding expression as follows:
WHERE (department_id != null)
AND (department_id != 10)
AND (department_id != 20)

For the entire expression to be true, each individual condition must be true. However,
a null value cannot be compared to another value, so the department_id !=null
condition cannot be true, and thus the whole expression is always false. The following
techniques enable a statement to return records even when nulls are returned to the
NOT IN operator:
•

Apply an NVL function to the columns returned by the subquery.

•

Add an IS NOT NULL predicate to the subquery.

•

Implement NOT NULL constraints.

In contrast to NOT IN, the NOT EXISTS clause only considers predicates that return the
existence of a match, and ignores any row that does not match or could not be
determined because of nulls. If at least one row in the subquery matches the row from
the outer query, then NOT EXISTS returns false. If no tuples match, then NOT EXISTS
returns true. The presence of nulls in the subquery does not affect the search for
matching records.
In releases earlier than Oracle Database 11g, the optimizer could not use an antijoin
optimization when nulls could be returned by a subquery. However, starting in Oracle
Database 11g, the ANTI NA (and ANTI SNA) optimizations described in the following
sections enable the optimizer to use an antijoin even when nulls are possible.

9-42

Chapter 9

Join Types

Example 9-18

Antijoin Using NOT IN

Suppose that a user issues the following query with a NOT IN clause to list the
departments that contain no employees:
SELECT department_id, department_name
FROM departments
WHERE department_id NOT IN
(SELECT department_id
FROM employees);

The preceding query returns no rows even though several departments contain no
employees. This result, which was not intended by the user, occurs because the
employees.department_id column is nullable.
The execution plan reveals a NESTED LOOPS ANTI SNA operation in Step 2:
-------------------------------------------------------------------------------| Id| Operation
| Name
|Rows|Bytes|Cost (%CPU)| Time |
-------------------------------------------------------------------------------| 0 | SELECT STATEMENT
|
|
|
| 4(100)|
|
|*1 | FILTER
|
|
|
|
|
|
NESTED LOOPS ANTI SNA|
| 2 |
| 17 | 323 | 4 (50)| 00:00:01 |
| 3 |
TABLE ACCESS FULL | DEPARTMENTS
| 27 | 432 | 2 (0)| 00:00:01 |
|*4 |
INDEX RANGE SCAN
| EMP_DEPARTMENT_IX | 41 | 123 | 0 (0)|
|
|*5 | TABLE ACCESS FULL
| EMPLOYEES
| 1 | 3 | 2 (0)| 00:00:01 |
-------------------------------------------------------------------------------PLAN_TABLE_OUTPUT
-------------------------------------------------------------------------------Predicate Information (identified by operation id):
--------------------------------------------------1 - filter( IS NULL)
4 - access("DEPARTMENT_ID"="DEPARTMENT_ID")
5 - filter("DEPARTMENT_ID" IS NULL)

The ANTI SNA stands for "single null-aware antijoin." ANTI NA stands for "null-aware
antijoin." The null-aware operation enables the optimizer to use the semijoin
optimization even on a nullable column. In releases earlier than Oracle Database 11g,
the database could not perform antijoins on NOT IN queries when nulls were possible.
Suppose that the user rewrites the query by applying an IS NOT NULL condition to the
subquery:
SELECT department_id, department_name
FROM departments
WHERE department_id NOT IN
(SELECT department_id
FROM employees
WHERE department_id IS NOT NULL);

The preceding query returns 16 rows, which is the expected result. Step 1 in the plan
shows a standard NESTED LOOPS ANTI join instead of an ANTI NA or ANTI SNA join
because the subquery cannot returns nulls:
-------------------------------------------------------------------------------|Id| Operation
| Name
| Rows| Bytes | Cost (%CPU)| Time |
-------------------------------------------------------------------------------| 0| SELECT STATEMENT |
|
|
| 2 (100)|
|
| 1| NESTED LOOPS ANTI |
| 17 | 323 | 2 (0)| 00:00:01 |

9-43

Chapter 9

Join Optimizations

| 2| TABLE ACCESS FULL| DEPARTMENTS
| 27 | 432 | 2 (0)| 00:00:01 |
|*3| INDEX RANGE SCAN | EMP_DEPARTMENT_IX | 41 | 123 | 0 (0)|
|
-------------------------------------------------------------------------------PLAN_TABLE_OUTPUT
-------------------------------------------------------------------------------Predicate Information (identified by operation id):
--------------------------------------------------3 - access("DEPARTMENT_ID"="DEPARTMENT_ID")
filter("DEPARTMENT_ID" IS NOT NULL)

Example 9-19

Antijoin Using NOT EXISTS

Suppose that a user issues the following query with a NOT EXISTS clause to list the
departments that contain no employees:
SELECT department_id, department_name
FROM departments d
WHERE NOT EXISTS
(SELECT null
FROM employees e
WHERE e.department_id = d.department_id)

The preceding query avoids the null problem for NOT IN clauses. Thus, even though
employees.department_id column is nullable, the statement returns the desired result.
Step 1 of the execution plan reveals a NESTED LOOPS ANTI operation, not the ANTI NA
variant, which was necessary for NOT IN when nulls were possible:
-------------------------------------------------------------------------------| Id| Operation
| Name
| Rows | Bytes | Cost (%CPU)|Time|
-------------------------------------------------------------------------------| 0 | SELECT STATEMENT |
|
|
| 2 (100)|
|
| 1 | NESTED LOOPS ANTI |
|
17 | 323 | 2 (0)|00:00:01|
| 2 | TABLE ACCESS FULL| DEPARTMENTS
|
27 | 432 | 2 (0)|00:00:01|
|*3 | INDEX RANGE SCAN | EMP_DEPARTMENT_IX |
41 | 123 | 0 (0)|
|
-------------------------------------------------------------------------------PLAN_TABLE_OUTPUT
-------------------------------------------------------------------------------Predicate Information (identified by operation id):
--------------------------------------------------3 - access("E"."DEPARTMENT_ID"="D"."DEPARTMENT_ID")

9.4 Join Optimizations
Join optimizations enable joins to be more efficient.
This section describes common join optimizations:
•

Bloom Filters
A Bloom filter, named after its creator Burton Bloom, is a low-memory data
structure that tests membership in a set.

•

Partition-Wise Joins
A partition-wise join is an optimization that divides a large join of two tables, one
of which must be partitioned on the join key, into several smaller joins.

9-44

Chapter 9

Join Optimizations

•

In-Memory Join Groups
A join group is a user-created object that lists two or more columns that can be
meaningfully joined.

9.4.1 Bloom Filters
A Bloom filter, named after its creator Burton Bloom, is a low-memory data structure
that tests membership in a set.
A Bloom filter correctly indicates when an element is not in a set, but can incorrectly
indicate when an element is in a set. Thus, false negatives are impossible but false
positives are possible.
This section contains the following topics:
•

Purpose of Bloom Filters
A Bloom filter tests one set of values to determine whether they are members
another set.

•

How Bloom Filters Work
A Bloom filter uses an array of bits to indicate inclusion in a set.

•

Bloom Filter Controls
The optimizer automatically determines whether to use Bloom filters.

•

Bloom Filter Metadata
V$ views contain metadata about Bloom filters.

•

Bloom Filters: Scenario
In this example, a parallel query joins the sales fact table to the products and times
dimension tables, and filters on fiscal week 18.

9.4.1.1 Purpose of Bloom Filters
A Bloom filter tests one set of values to determine whether they are members another
set.
For example, one set is (10,20,30,40) and the second set is (10,30,60,70). A Bloom
filter can determine that 60 and 70 are guaranteed to be excluded from the first set,
and that 10 and 30 are probably members. Bloom filters are especially useful when the
amount of memory needed to store the filter is small relative to the amount of data in
the data set, and when most data is expected to fail the membership test.
Oracle Database uses Bloom filters to various specific goals, including the following:
•

Reduce the amount of data transferred to slave processes in a parallel query,
especially when the database discards most rows because they do not fulfill a join
condition

•

Eliminate unneeded partitions when building a partition access list in a join, known
as partition pruning

•

Test whether data exists in the server result cache, thereby avoiding a disk read

•

Filter members in Exadata cells, especially when joining a large fact table and
small dimension tables in a star schema

Bloom filters can occur in both parallel and serial processing.

9-45

Chapter 9

Join Optimizations

9.4.1.2 How Bloom Filters Work
A Bloom filter uses an array of bits to indicate inclusion in a set.
For example, 8 elements (an arbitrary number used for this example) in an array are
initially set to 0:
e1 e2 e3 e4 e5 e6 e7 e8
0 0 0 0 0 0 0 0

This array represents a set. To represent an input value i in this array, three separate
hash functions (three is arbitrary) are applied to i, each generating a hash value
between 1 and 8:
f1(i) = h1
f2(i) = h2
f3(i) = h3

For example, to store the value 17 in this array, the hash functions set i to 17, and then
return the following hash values:
f1(17) = 5
f2(17) = 3
f3(17) = 5

In the preceding example, two of the hash functions happened to return the same
value of 5, known as a hash collision. Because the distinct hash values are 5 and 3, the
5th and 3rd elements in the array are set to 1:
e1 e2 e3 e4 e5 e6 e7 e8
0 0 1 0 1 0 0 0

Testing the membership of 17 in the set reverses the process. To test whether the set
excludes the value 17, element 3 or element 5 must contain a 0. If a 0 is present in
either element, then the set cannot contain 17. No false negatives are possible.
To test whether the set includes 17, both element 3 and element 5 must contain 1
values. However, if the test indicates a 1 for both elements, then it is still possible for
the set not to include 17. False positives are possible. For example, the following array
might represent the value 22, which also has a 1 for both element 3 and element 5:
e1 e2 e3 e4 e5 e6 e7 e8
1 0 1 0 1 0 0 0

9.4.1.3 Bloom Filter Controls
The optimizer automatically determines whether to use Bloom filters.
To override optimizer decisions, use the hints PX_JOIN_FILTER and NO_PX_JOIN_FILTER.

See Also:
Oracle Database SQL Language Reference to learn more about the bloom
filter hints

9-46

Chapter 9

Join Optimizations

9.4.1.4 Bloom Filter Metadata
V$ views contain metadata about Bloom filters.

You can query the following views:
•

V$SQL_JOIN_FILTER

This view shows the number of rows filtered out (FILTERED column) and tested
(PROBED column) by an active Bloom filter.
•

V$PQ_TQSTAT

This view displays the number of rows processed through each parallel execution
server at each stage of the execution tree. You can use it to monitor how much
Bloom filters have reduced data transfer among parallel processes.
In an execution plan, a Bloom filter is indicated by keywords JOIN FILTER in the
Operation column, and the prefix :BF in the Name column, as in the 9th step of the
following plan snippet:
---------------------------------------------------------------------------| Id | Operation
| Name
|
TQ |IN-OUT| PQ Distrib |
---------------------------------------------------------------------------...
| 9 |
JOIN FILTER CREATE
| :BF0000 | Q1,03 | PCWP |
|

In the Predicate Information section of the plan, filters that contain functions beginning
with the string SYS_OP_BLOOM_FILTER indicate use of a Bloom filter.

9.4.1.5 Bloom Filters: Scenario
In this example, a parallel query joins the sales fact table to the products and times
dimension tables, and filters on fiscal week 18.
SELECT
FROM
WHERE
AND
AND

/*+ parallel(s) */ p.prod_name, s.quantity_sold
sh.sales s, sh.products p, sh.times t
s.prod_id = p.prod_id
s.time_id = t.time_id
t.fiscal_week_number = 18;

Querying DBMS_XPLAN.DISPLAY_CURSOR provides the following output:
SELECT * FROM
TABLE(DBMS_XPLAN.DISPLAY_CURSOR(format => 'BASIC,+PARALLEL,+PREDICATE'));
EXPLAINED SQL STATEMENT:
-----------------------SELECT /*+ parallel(s) */ p.prod_name, s.quantity_sold FROM sh.sales s,
sh.products p, sh.times t WHERE s.prod_id = p.prod_id AND s.time_id =
t.time_id AND t.fiscal_week_number = 18
Plan hash value: 1183628457
---------------------------------------------------------------------------| Id | Operation
| Name
|
TQ |IN-OUT| PQ Distrib |
---------------------------------------------------------------------------| 0 | SELECT STATEMENT
|
|
|
|
|
| 1 | PX COORDINATOR
|
|
|
|
|
| 2 | PX SEND QC (RANDOM)
| :TQ10003 | Q1,03 | P->S | QC (RAND) |

9-47

Chapter 9

Join Optimizations

|* 3 |
HASH JOIN BUFFERED
|
| Q1,03 | PCWP |
|
| 4 |
PX RECEIVE
|
| Q1,03 | PCWP |
|
| 5 |
PX SEND BROADCAST
| :TQ10001 | Q1,01 | S->P | BROADCAST |
| 6 |
PX SELECTOR
|
| Q1,01 | SCWC |
|
| 7 |
TABLE ACCESS FULL | PRODUCTS | Q1,01 | SCWP |
|
|* 8 |
HASH JOIN
|
| Q1,03 | PCWP |
|
| 9 |
JOIN FILTER CREATE
| :BF0000 | Q1,03 | PCWP |
|
| 10 |
BUFFER SORT
|
| Q1,03 | PCWC |
|
| 11 |
PX RECEIVE
|
| Q1,03 | PCWP |
|
| 12 |
PX SEND HYBRID HASH| :TQ10000 |
| S->P | HYBRID HASH|
|* 13 |
TABLE ACCESS FULL | TIMES
|
|
|
|
| 14 |
PX RECEIVE
|
| Q1,03 | PCWP |
|
| 15 |
PX SEND HYBRID HASH | :TQ10002 | Q1,02 | P->P | HYBRID HASH|
| 16 |
JOIN FILTER USE
| :BF0000 | Q1,02 | PCWP |
|
| 17 |
PX BLOCK ITERATOR |
| Q1,02 | PCWC |
|
|* 18 |
TABLE ACCESS FULL | SALES
| Q1,02 | PCWP |
|
---------------------------------------------------------------------------Predicate Information (identified by operation id):
--------------------------------------------------3
8
13
18

-

access("S"."PROD_ID"="P"."PROD_ID")
access("S"."TIME_ID"="T"."TIME_ID")
filter("T"."FISCAL_WEEK_NUMBER"=18)
access(:Z>=:Z AND :Z<=:Z)
filter(SYS_OP_BLOOM_FILTER(:BF0000,"S"."TIME_ID"))

A single server process scans the times table (Step 13), and then uses a hybrid hash
distribution method to send the rows to the parallel execution servers (Step 12). The
processes in set Q1,03 create a bloom filter (Step 9). The processes in set Q1,02 scan
sales in parallel (Step 18), and then use the Bloom filter to discard rows from sales
(Step 16) before sending them on to set Q1,03 using hybrid hash distribution (Step 15).
The processes in set Q1,03 hash join the times rows to the filtered sales rows (Step 8).
The processes in set Q1,01 scan products (Step 7), and then send the rows to Q1,03
(Step 5). Finally, the processes in Q1,03 join the products rows to the rows generated
by the previous hash join (Step 3).
The following figure illustrates the basic process.
Figure 9-8

Bloom Filter
Q1, 03
Create
Bloom filter
:BF0000

Q1, 01

Q1, 02

9.4.2 Partition-Wise Joins
A partition-wise join is an optimization that divides a large join of two tables, one of
which must be partitioned on the join key, into several smaller joins.

9-48

Chapter 9

Join Optimizations

Partition-wise joins are either of the following:
•

Full partition-wise join
Both tables must be equipartitioned on their join keys, or use reference partitioning
(that is, be related by referential constraints). The database divides a large join
into smaller joins between two partitions from the two joined tables.

•

Partial partition-wise joins
Only one table is partitioned on the join key. The other table may or may not be
partitioned.

This section contains the following topics:
•

Purpose of Partition-Wise Joins
Partition-wise joins reduce query response time by minimizing the amount of data
exchanged among parallel execution servers when joins execute in parallel.

•

How Partition-Wise Joins Work
When the database serially joins two partitioned tables without using a partitionwise join, a single server process performs the join.

See Also:
Oracle Database VLDB and Partitioning Guide explains partition-wise joins in
detail

9.4.2.1 Purpose of Partition-Wise Joins
Partition-wise joins reduce query response time by minimizing the amount of data
exchanged among parallel execution servers when joins execute in parallel.
This technique significantly reduces response time and improves the use of CPU and
memory. In Oracle Real Application Clusters (Oracle RAC) environments, partitionwise joins also avoid or at least limit the data traffic over the interconnect, which is the
key to achieving good scalability for massive join operations.

9.4.2.2 How Partition-Wise Joins Work
When the database serially joins two partitioned tables without using a partition-wise
join, a single server process performs the join.
In the following illustration, the join is not partition-wise because the server process
joins every partition of table t1 to every partition of table t2.

9-49

Chapter 9

Join Optimizations

Figure 9-9

Join That Is Not Partition-Wise
Server
Process

t1

t2

This section contains the following topics:
•

How a Full Partition-Wise Join Works
The database performs a full partition-wise join either serially or in parallel.

•

How a Partial Partition-Wise Join Works
Partial partition-wise joins, unlike their full partition-wise counterpart, must execute
in parallel.

9.4.2.2.1 How a Full Partition-Wise Join Works
The database performs a full partition-wise join either serially or in parallel.
The following graphic shows a full partition-wise join performed in parallel. In this case,
the granule of parallelism is a partition. Each parallel execution server joins the
partitions in pairs. For example, the first parallel execution server joins the first partition
of t1 to the first partition of t2. The parallel execution coordinator then assembles the
result.

9-50

Chapter 9

Join Optimizations

Figure 9-10

Full Partition-Wise Join in Parallel

PE Coordinator
t1

t2

PE Server

PE Server

PE Server

PE Server

A full partition-wise join can also join partitions to subpartitions, which is useful when
the tables use different partitioning methods. For example, customers is partitioned by
hash, but sales is partitioned by range. If you subpartition sales by hash, then the
database can perform a full partition-wise join between the hash partitions of the
customers and the hash subpartitions of sales.
In the execution plan, the presence of a partition operation before the join signals the
presence of a full partition-wise join, as in the following snippet:
| 8 |
|* 9 |

PX PARTITION HASH ALL|
HASH JOIN
|

See Also:
Oracle Database VLDB and Partitioning Guide explains full partition-wise
joins in detail, and includes several examples

9.4.2.2.2 How a Partial Partition-Wise Join Works
Partial partition-wise joins, unlike their full partition-wise counterpart, must execute in
parallel.
The following graphic shows a partial partition-wise join between t1, which is
partitioned, and t2, which is not partitioned.

9-51

Chapter 9

Join Optimizations

Figure 9-11

Partial Partition-Wise Join

PE Coordinator
t1

t2

PE Server
PE Server

t1

t2

PE Server
PE Server

PE Server
PE Server

PE Server
PE Server
Dynamically created
partitions

Because t2 is not partitioned, a set of parallel execution servers must generate
partitions from t2 as needed. A different set of parallel execution servers then joins the
t1 partitions to the dynamically generated partitions. The parallel execution coordinator
assembles the result.
In the execution plan, the operation PX SEND PARTITION (KEY) signals a partial partitionwise join, as in the following snippet:
| 11 |

PX SEND PARTITION (KEY)

|

See Also:
Oracle Database VLDB and Partitioning Guide explains full partition-wise
joins in detail, and includes several examples

9.4.3 In-Memory Join Groups
A join group is a user-created object that lists two or more columns that can be
meaningfully joined.
In certain queries, join groups eliminate the performance overhead of decompressing
and hashing column values. Join groups require an In-Memory Column Store (IM
column store).

9-52

Chapter 9

Join Optimizations

See Also:
Oracle Database In-Memory Guide to learn how to optimize In-Memory
queries with join groups

9-53

Part V
Optimizer Statistics
The accuracy of an execution plan depends on the quality of the optimizer statistics.
This part contains the following chapters:
•

Optimizer Statistics Concepts
Oracle Database optimizer statistics describe details about the database and its
objects.

•

Histograms
A histogram is a special type of column statistic that provides more detailed
information about the data distribution in a table column. A histogram sorts values
into "buckets," as you might sort coins into buckets.

•

Configuring Options for Optimizer Statistics Gathering
This chapter explains what optimizer statistics collection is and how to set
statistics preferences.

•

Gathering Optimizer Statistics
This chapter explains how to use the DBMS_STATS.GATHER_*_STATS program units.

•

Managing Extended Statistics
DBMS_STATS enables you to collect extended statistics, which are statistics that
can improve cardinality estimates when multiple predicates exist on different
columns of a table, or when predicates use expressions.

•

Controlling the Use of Optimizer Statistics
Using DBMS_STATS, you can specify when and how the optimizer uses statistics.

•

Managing Historical Optimizer Statistics
This chapter how to retain, report on, and restore non-current statistics.

•

Transporting Optimizer Statistics
You can export and import optimizer statistics from the data dictionary to userdefined statistics tables. You can also copy statistics from one database to another
database.

•

Analyzing Statistics Using Optimizer Statistics Advisor
Optimizer Statistics Advisor analyzes how optimizer statistics are gathered, and
then makes recommendations.

10
Optimizer Statistics Concepts
Oracle Database optimizer statistics describe details about the database and its
objects.
This chapter includes the following topics:
•

Introduction to Optimizer Statistics
The optimizer cost model relies on statistics collected about the objects involved
in a query, and the database and host where the query runs.

•

About Optimizer Statistics Types
The optimizer collects statistics on different types of database objects and
characteristics of the database environment.

•

How the Database Gathers Optimizer Statistics
Oracle Database provides several mechanisms to gather statistics.

•

When the Database Gathers Optimizer Statistics
The database collects optimizer statistics at various times and from various
sources.

Related Topics
•

Query Optimizer Concepts
This chapter describes the most important concepts relating to the query
optimizer, including its principal components.

•

Histograms
A histogram is a special type of column statistic that provides more detailed
information about the data distribution in a table column. A histogram sorts values
into "buckets," as you might sort coins into buckets.

•

Gathering Optimizer Statistics
This chapter explains how to use the DBMS_STATS.GATHER_*_STATS program units.

•

Managing Historical Optimizer Statistics
This chapter how to retain, report on, and restore non-current statistics.

10.1 Introduction to Optimizer Statistics
The optimizer cost model relies on statistics collected about the objects involved in a
query, and the database and host where the query runs.
The optimizer uses statistics to get an estimate of the number of rows (and number of
bytes) retrieved from a table, partition, or index. The optimizer estimates the cost for
the access, determines the cost for possible plans, and then picks the execution plan
with the lowest cost.
Optimizer statistics include the following:
•

Table statistics
–

Number of rows

10-1

Chapter 10

Introduction to Optimizer Statistics

•

•

•

–

Number of blocks

–

Average row length

Column statistics
–

Number of distinct values (NDV) in a column

–

Number of nulls in a column

–

Data distribution (histogram)

–

Extended statistics

Index statistics
–

Number of leaf blocks

–

Number of levels

–

Index clustering factor

System statistics
–

I/O performance and utilization

–

CPU performance and utilization

As shown in Figure 10-1, the database stores optimizer statistics for tables, columns,
indexes, and the system in the data dictionary. You can access these statistics using
data dictionary views.

Note:
The optimizer statistics are different from the performance statistics visible
through V$ views.

10-2

Chapter 10

About Optimizer Statistics Types

Figure 10-1

Optimizer Statistics

Database

Optimizer

Data Dictionary
Optimizer Statistics
Index

Table

Column

System

PERSON
Table

PERSON_ID_IX

GB

ID Name
100 Kumar

HJ

Execution
Plan
HJ

CPU and I/O

10.2 About Optimizer Statistics Types
The optimizer collects statistics on different types of database objects and
characteristics of the database environment.
This section contains the following topics:
•

Table Statistics
In Oracle Database, table statistics include information about rows and blocks.

•

Column Statistics
Column statistics track information about column values and data distribution.

•

Index Statistics
The index statistics include information about the number of index levels, the
number of index blocks, and the relationship between the index and the data
blocks. The optimizer uses these statistics to determine the cost of index scans.

•

Session-Specific Statistics for Global Temporary Tables
A global temporary table is a special table that stores intermediate sessionprivate data for a specific duration.

10-3

Chapter 10

About Optimizer Statistics Types

•

System Statistics
The system statistics describe hardware characteristics such as I/O and CPU
performance and utilization.

•

User-Defined Optimizer Statistics
The extensible optimizer enables authors of user-defined functions and indexes
to create statistics collection, selectivity, and cost functions. The optimizer cost
model is extended to integrate information supplied by the user to assess CPU
and the I/O cost.

10.2.1 Table Statistics
In Oracle Database, table statistics include information about rows and blocks.
The optimizer uses these statistics to determine the cost of table scans and table joins.
DBMS_STATS can gather statistics for both permanent and temporary tables.

The database tracks all relevant statistics about permanent tables.
DBMS_STATS.GATHER_TABLE_STATS commits before gathering statistics on permanent
tables. For example, table statistics stored in DBA_TAB_STATISTICS track the following:
•

Number of rows and average row length
The database uses the row count stored in DBA_TAB_STATISTICS when determining
cardinality.

•

Number of data blocks
The optimizer uses the number of data blocks with the
DB_FILE_MULTIBLOCK_READ_COUNT initialization parameter to determine the base table
access cost.

Example 10-1

Table Statistics

This example queries some table statistics for the sh.customers table.
sys@PROD> SELECT NUM_ROWS, AVG_ROW_LEN, BLOCKS, LAST_ANALYZED
2 FROM DBA_TAB_STATISTICS
3 WHERE OWNER='SH'
4 AND
TABLE_NAME='CUSTOMERS';
NUM_ROWS AVG_ROW_LEN
BLOCKS LAST_ANAL
---------- ----------- ---------- --------55500
181
1486 14-JUN-10

See Also:
•

"About Optimizer Initialization Parameters"

•

"Gathering Schema and Table Statistics"

•

Oracle Database Reference for a description of the DBA_TAB_STATISTICS
view and the DB_FILE_MULTIBLOCK_READ_COUNT initialization parameter

10.2.2 Column Statistics
Column statistics track information about column values and data distribution.

10-4

Chapter 10

About Optimizer Statistics Types

The optimizer uses column statistics to generate accurate cardinality estimates and
make better decisions about index usage, join orders, join methods, and so on. For
example, statistics in DBA_TAB_COL_STATISTICS track the following:
•

Number of distinct values

•

Number of nulls

•

High and low values

•

Histogram-related information

The optimizer can use extended statistics, which are a special type of column
statistics. These statistics are useful for informing the optimizer of logical relationships
among columns.

See Also:
•

"Histograms "

•

"About Statistics on Column Groups"

•

Oracle Database Reference for a description of the
DBA_TAB_COL_STATISTICS view

10.2.3 Index Statistics
The index statistics include information about the number of index levels, the number
of index blocks, and the relationship between the index and the data blocks. The
optimizer uses these statistics to determine the cost of index scans.
This section contains the following topics:
•

Types of Index Statistics
The DBA_IND_STATISTICS view tracks index statistics.

•

Index Clustering Factor
For a B-tree index, the index clustering factor measures the physical grouping of
rows in relation to an index value, such as last name.

•

Effect of Index Clustering Factor on Cost: Example
This example illustrates how the index clustering factor can influence the cost of
table access.

10.2.3.1 Types of Index Statistics
The DBA_IND_STATISTICS view tracks index statistics.
Statistics include the following:
•

Levels
The BLEVEL column shows the number of blocks required to go from the root block
to a leaf block. A B-tree index has two types of blocks: branch blocks for searching
and leaf blocks that store values. See Oracle Database Concepts for a conceptual
overview of B-tree indexes.

•

Distinct keys

10-5

Chapter 10

About Optimizer Statistics Types

This columns tracks the number of distinct indexed values. If a unique constraint is
defined, and if no NOT NULL constraint is defined, then this value equals the number
of non-null values.
•

Average number of leaf blocks for each distinct indexed key

•

Average number of data blocks pointed to by each distinct indexed key

See Also:
Oracle Database Reference for a description of the DBA_IND_STATISTICS view

Example 10-2

Index Statistics

This example queries some index statistics for the cust_lname_ix and customers_pk
indexes on the sh.customers table (sample output included):
SELECT INDEX_NAME, BLEVEL, LEAF_BLOCKS AS "LEAFBLK", DISTINCT_KEYS AS "DIST_KEY",
AVG_LEAF_BLOCKS_PER_KEY AS "LEAFBLK_PER_KEY",
AVG_DATA_BLOCKS_PER_KEY AS "DATABLK_PER_KEY"
FROM DBA_IND_STATISTICS
WHERE OWNER = 'SH'
AND
INDEX_NAME IN ('CUST_LNAME_IX','CUSTOMERS_PK');
INDEX_NAME
BLEVEL LEAFBLK DIST_KEY LEAFBLK_PER_KEY DATABLK_PER_KEY
-------------- ------ ------- -------- --------------- --------------CUSTOMERS_PK
1
115
55500
1
1
CUST_LNAME_IX
1
141
908
1
10

10.2.3.2 Index Clustering Factor
For a B-tree index, the index clustering factor measures the physical grouping of
rows in relation to an index value, such as last name.
The index clustering factor helps the optimizer decide whether an index scan or full
table scan is more efficient for certain queries). A low clustering factor indicates an
efficient index scan.
A clustering factor that is close to the number of blocks in a table indicates that the
rows are physically ordered in the table blocks by the index key. If the database
performs a full table scan, then the database tends to retrieve the rows as they are
stored on disk sorted by the index key. A clustering factor that is close to the number
of rows indicates that the rows are scattered randomly across the database blocks in
relation to the index key. If the database performs a full table scan, then the database
would not retrieve rows in any sorted order by this index key.
The clustering factor is a property of a specific index, not a table. If multiple indexes
exist on a table, then the clustering factor for one index might be small while the factor
for another index is large. An attempt to reorganize the table to improve the clustering
factor for one index may degrade the clustering factor of the other index.
Example 10-3

Index Clustering Factor
This example shows how the optimizer uses the index clustering factor to determine
whether using an index is more effective than a full table scan.

10-6

Chapter 10

About Optimizer Statistics Types

1.

Start SQL*Plus and connect to a database as sh, and then query the number of
rows and blocks in the sh.customers table (sample output included):
SELECT table_name, num_rows, blocks
FROM
user_tables
WHERE table_name='CUSTOMERS';
TABLE_NAME
NUM_ROWS
BLOCKS
------------------------------ ---------- ---------CUSTOMERS
55500
1486

2.

Create an index on the customers.cust_last_name column.
For example, execute the following statement:
CREATE INDEX CUSTOMERS_LAST_NAME_IDX ON customers(cust_last_name);

3.

Query the index clustering factor of the newly created index.
The following query shows that the customers_last_name_idx index has a high
clustering factor because the clustering factor is significantly more than the
number of blocks in the table:
SELECT
FROM
WHERE
AND

index_name, blevel, leaf_blocks, clustering_factor
user_indexes
table_name='CUSTOMERS'
index_name= 'CUSTOMERS_LAST_NAME_IDX';

INDEX_NAME
BLEVEL LEAF_BLOCKS CLUSTERING_FACTOR
------------------------------ ---------- ----------- ----------------CUSTOMERS_LAST_NAME_IDX
1
141
9859
4.

Create a new copy of the customers table, with rows ordered by cust_last_name.
For example, execute the following statements:
DROP TABLE customers3 PURGE;
CREATE TABLE customers3 AS
SELECT *
FROM customers
ORDER BY cust_last_name;

5.

Gather statistics on the customers3 table.
For example, execute the GATHER_TABLE_STATS procedure as follows:
EXEC DBMS_STATS.GATHER_TABLE_STATS(null,'CUSTOMERS3');

6.

Query the number of rows and blocks in the customers3 table .
For example, enter the following query (sample output included):
SELECT
FROM
WHERE

TABLE_NAME, NUM_ROWS, BLOCKS
USER_TABLES
TABLE_NAME='CUSTOMERS3';

TABLE_NAME
NUM_ROWS
BLOCKS
------------------------------ ---------- ---------CUSTOMERS3
55500
1485
7.

Create an index on the cust_last_name column of customers3.
For example, execute the following statement:
CREATE INDEX CUSTOMERS3_LAST_NAME_IDX ON customers3(cust_last_name);

8.

Query the index clustering factor of the customers3_last_name_idx index.

10-7

Chapter 10

About Optimizer Statistics Types

The following query shows that the customers3_last_name_idx index has a lower
clustering factor:
SELECT
FROM
WHERE
AND

INDEX_NAME, BLEVEL, LEAF_BLOCKS, CLUSTERING_FACTOR
USER_INDEXES
TABLE_NAME = 'CUSTOMERS3'
INDEX_NAME = 'CUSTOMERS3_LAST_NAME_IDX';

INDEX_NAME
BLEVEL LEAF_BLOCKS CLUSTERING_FACTOR
------------------------------ ---------- ----------- ----------------CUSTOMERS3_LAST_NAME_IDX
1
141
1455

The table customers3 has the same data as the original customers table, but the
index on customers3 has a much lower clustering factor because the data in the
table is ordered by the cust_last_name. The clustering factor is now about 10 times
the number of blocks instead of 70 times.
9.

Query the customers table.
For example, execute the following query (sample output included):
SELECT cust_first_name, cust_last_name
FROM customers
WHERE cust_last_name BETWEEN 'Puleo' AND 'Quinn';
CUST_FIRST_NAME
-------------------Vida
Harriett
Madeleine
Caresse

CUST_LAST_NAME
---------------------------------------Puleo
Quinlan
Quinn
Puleo

10. Display the cursor for the query.

For example, execute the following query (partial sample output included):
SELECT * FROM TABLE(DBMS_XPLAN.DISPLAY_CURSOR());
------------------------------------------------------------------------------| Id | Operation
| Name
| Rows |Bytes|Cost (%CPU)| Time |
------------------------------------------------------------------------------| 0| SELECT STATEMENT
|
|
|
| 405 (100)|
|
|* 1| TABLE ACCESS STORAGE FULL| CUSTOMERS | 2335|35025| 405 (1)|00:00:01|
-------------------------------------------------------------------------------

The preceding plan shows that the optimizer did not use the index on the original
customers tables.
11. Query the customers3 table.

For example, execute the following query (sample output included):
SELECT cust_first_name, cust_last_name
FROM customers3
WHERE cust_last_name BETWEEN 'Puleo' AND 'Quinn';
CUST_FIRST_NAME
-------------------Vida
Harriett
Madeleine
Caresse

CUST_LAST_NAME
---------------------------------------Puleo
Quinlan
Quinn
Puleo

12. Display the cursor for the query.

10-8

Chapter 10

About Optimizer Statistics Types

For example, execute the following query (partial sample output included):
SELECT * FROM TABLE(DBMS_XPLAN.DISPLAY_CURSOR());
--------------------------------------------------------------------------------------|Id| Operation
| Name
|Rows|Bytes|Cost(%CPU)| Time|
--------------------------------------------------------------------------------------| 0| SELECT STATEMENT
|
|
|
|69(100)|
|
| 1| TABLE ACCESS BY INDEX ROWID|CUSTOMERS3
|2335|35025|69(0) |00:00:01|
|*2| INDEX RANGE SCAN
|CUSTOMERS3_LAST_NAME_IDX|2335|
|7(0) |00:00:01|
---------------------------------------------------------------------------------------

The result set is the same, but the optimizer chooses the index. The plan cost is
much less than the cost of the plan used on the original customers table.
13. Query customers with a hint that forces the optimizer to use the index.

For example, execute the following query (partial sample output included):
SELECT /*+ index (Customers CUSTOMERS_LAST_NAME_IDX) */ cust_first_name,
cust_last_name
FROM customers
WHERE cust_last_name BETWEEN 'Puleo' and 'Quinn';
CUST_FIRST_NAME
-------------------Vida
Caresse
Harriett
Madeleine

CUST_LAST_NAME
---------------------------------------Puleo
Puleo
Quinlan
Quinn

14. Display the cursor for the query.

For example, execute the following query (partial sample output included):
SELECT * FROM TABLE(DBMS_XPLAN.DISPLAY_CURSOR());
----------------------------------------------------------------------------------------| Id | Operation
| Name
|Rows|Bytes|Cost(%CPU)| Time |
----------------------------------------------------------------------------------------| 0| SELECT STATEMENT
|
|
|
| 422(100) |
|
| 1| TABLE ACCESS BY INDEX ROWID|CUSTOMERS
|335 |35025| 422(0) |00:00:01|
|*2| INDEX RANGE SCAN
|CUSTOMERS_LAST_NAME_IDX|2335|
| 7(0)
|00:00:01|
-----------------------------------------------------------------------------------------

The preceding plan shows that the cost of using the index on customers is higher
than the cost of a full table scan. Thus, using an index does not necessarily
improve performance. The index clustering factor is a measure of whether an
index scan is more effective than a full table scan.

10.2.3.3 Effect of Index Clustering Factor on Cost: Example
This example illustrates how the index clustering factor can influence the cost of table
access.
Consider the following scenario:
•

A table contains 9 rows that are stored in 3 data blocks.

•

The col1 column currently stores the values A, B, and C.

•

A nonunique index named col1_idx exists on col1 for this table.

10-9

Chapter 10

About Optimizer Statistics Types

Example 10-4

Collocated Data

Assume that the rows are stored in the data blocks as follows:
Block 1
------A A A

Block 2
------B B B

Block 3
------C C C

In this example, the index clustering factor for col1_idx is low. The rows that have the
same indexed column values for col1 are in the same data blocks in the table. Thus,
the cost of using an index range scan to return all rows with value A is low because
only one block in the table must be read.
Example 10-5

Scattered Data

Assume that the same rows are scattered across the data blocks as follows:
Block 1
------A B C

Block 2
------A C B

Block 3
------B A C

In this example, the index clustering factor for col1_idx is higher. The database must
read all three blocks in the table to retrieve all rows with the value A in col1.

See Also:
Oracle Database Reference for a description of the DBA_INDEXES view

10.2.4 Session-Specific Statistics for Global Temporary Tables
A global temporary table is a special table that stores intermediate session-private
data for a specific duration.
The ON COMMIT clause of CREATE GLOBAL TEMPORARY TABLE indicates whether the table is
transaction-specific (DELETE ROWS) or session-specific (PRESERVE ROWS). Thus, a
temporary table holds intermediate result sets for the duration of either a transaction or
a session.
When you create a global temporary table, you create a definition that is visible to all
sessions. No physical storage is allocated. When a session first puts data into the
table, the database allocates storage space. The data in the temporary table is only
visible to the current session.
This section contains the following topics:
•

Shared and Session-Specific Statistics for Global Temporary Tables
Starting in Oracle Database 12c, you can set the table-level preference
GLOBAL_TEMP_TABLE_STATS to make statistics on a global temporary table shared
(SHARED) or session-specific (SESSION).

•

Effect of DBMS_STATS on Transaction-Specific Temporary Tables
DBMS_STATS commits changes to session-specific global temporary tables, but not to
transaction-specific global temporary tables.

10-10

Chapter 10

About Optimizer Statistics Types

10.2.4.1 Shared and Session-Specific Statistics for Global Temporary Tables
Starting in Oracle Database 12c, you can set the table-level preference
GLOBAL_TEMP_TABLE_STATS to make statistics on a global temporary table shared (SHARED)
or session-specific (SESSION).

When set to session-specific, you can gather statistics for a global temporary table in
one session, and then use the statistics for this session only. Meanwhile, users can
continue to maintain a shared version of the statistics. During optimization, the
optimizer first checks whether a global temporary table has session-specific statistics.
If yes, the optimizer uses them. Otherwise, the optimizer uses shared statistics if they
exist.
Session-specific statistics have the following characteristics:
•

Dictionary views that track statistics show both the shared statistics and the
session-specific statistics in the current session.
The views are DBA_TAB_STATISTICS, DBA_IND_STATISTICS, DBA_TAB_HISTOGRAMS, and
DBA_TAB_COL_STATISTICS (each view has a corresponding USER_ and ALL_ version).
The SCOPE column shows whether statistics are session-specific or shared.

•

Other sessions do not share the cursor using the session-specific statistics.
Different sessions can share the cursor using shared statistics, as in releases
earlier than Oracle Database 12c. The same session can share the cursor using
session-specific statistics.

•

Pending statistics are not supported for session-specific statistics.

•

When the GLOBAL_TEMP_TABLE_STATS preference is set to SESSION, by default
GATHER_TABLE_STATS immediately invalidates previous cursors compiled in the same
session. However, this procedure does not invalidate cursors compiled in other
sessions.

10.2.4.2 Effect of DBMS_STATS on Transaction-Specific Temporary Tables
DBMS_STATS commits changes to session-specific global temporary tables, but not to
transaction-specific global temporary tables.

Before Oracle Database 12c, running DBMS_STATS.GATHER_TABLE_STATS on a transactionspecific temporary table (ON COMMIT DELETE ROWS) would delete all rows, making the
statistics show the table as empty. Starting in Oracle Database 12c, the following
procedures do not commit for transaction-specific temporary tables, so that rows in
these tables are not deleted:
•

GATHER_TABLE_STATS

•

DELETE_TABLE_STATS

•

DELETE_COLUMN_STATS

•

DELETE_INDEX_STATS

•

SET_TABLE_STATS

•

SET_COLUMN_STATS

•

SET_INDEX_STATS

•

GET_TABLE_STATS

10-11

Chapter 10

About Optimizer Statistics Types

•

GET_COLUMN_STATS

•

GET_INDEX_STATS

The preceding program units observe the GLOBAL_TEMP_TABLE_STATS preference. For
example, if the table preference is set to SESSION, then SET_TABLE_STATS sets the
session statistics, and GATHER_TABLE_STATS preserves all rows in a transaction-specific
temporary table. If the table preference is set to SHARED, then SET_TABLE_STATS sets the
shared statistics, and GATHER_TABLE_STATS deletes all rows from a transaction-specific
temporary table.

See Also:
•

"Gathering Schema and Table Statistics"

•

Oracle Database Concepts to learn about global temporary tables

•

Oracle Database PL/SQL Packages and Types Reference to learn about
the DBMS_STATS.GATHER_TABLE_STATS procedure

10.2.5 System Statistics
The system statistics describe hardware characteristics such as I/O and CPU
performance and utilization.
System statistics enable the query optimizer to more accurately estimate I/O and CPU
costs when choosing execution plans. The database does not invalidate previously
parsed SQL statements when updating system statistics. The database parses all new
SQL statements using new statistics.

See Also:
•

"Gathering System Statistics Manually"

•

Oracle Database Reference

10.2.6 User-Defined Optimizer Statistics
The extensible optimizer enables authors of user-defined functions and indexes to
create statistics collection, selectivity, and cost functions. The optimizer cost model is
extended to integrate information supplied by the user to assess CPU and the I/O cost.
Statistics types act as interfaces for user-defined functions that influence the choice of
execution plan. However, to use a statistics type, the optimizer requires a mechanism
to bind the type to a database object such as a column, standalone function, object
type, index, indextype, or package. The SQL statement ASSOCIATE STATISTICS allows
this binding to occur.
Functions for user-defined statistics are relevant for columns that use both standard
SQL data types and object types, and for domain indexes. When you associate a
statistics type with a column or domain index, the database calls the statistics
collection method in the statistics type whenever DBMS_STATS gathers statistics.

10-12

Chapter 10

How the Database Gathers Optimizer Statistics

See Also:
•

"Gathering Schema and Table Statistics"

•

Oracle Database Data Cartridge Developer's Guide to learn about the
extensible optimizer and user-defined statistics

10.3 How the Database Gathers Optimizer Statistics
Oracle Database provides several mechanisms to gather statistics.
This section contains the following topics:
•

DBMS_STATS Package
The DBMS_STATS PL/SQL package collects and manages optimizer statistics.

•

Supplemental Dynamic Statistics
By default, when optimizer statistics are missing, stale, or insufficient, the
database automatically gathers dynamic statistics during a parse. The database
uses recursive SQL to scan a small random sample of table blocks.

•

Online Statistics Gathering for Bulk Loads
Starting in Oracle Database 12c, the database can gather table statistics
automatically during the following types of bulk loads: INSERT INTO ... SELECT into
an empty table using a direct path insert, and CREATE TABLE AS SELECT .

Related Topics
•

Configuring Automatic Optimizer Statistics Collection
Oracle Database can gather optimizer statistics automatically.

•

Gathering Optimizer Statistics Manually
As an alternative or supplement to automatic statistics gathering, you can use the
DBMS_STATS package to gather statistics manually.

•

Locking and Unlocking Optimizer Statistics
You can lock statistics to prevent them from changing.

10.3.1 DBMS_STATS Package
The DBMS_STATS PL/SQL package collects and manages optimizer statistics.
This package enables you to control what and how statistics are collected, including
the degree of parallelism for statistics collection, sampling methods, granularity of
statistics collection in partitioned tables, and so on.

Note:
Do not use the COMPUTE and ESTIMATE clauses of the ANALYZE statement to
collect optimizer statistics. These clauses have been deprecated. Instead,
use DBMS_STATS.

10-13

Chapter 10

How the Database Gathers Optimizer Statistics

Statistics gathered with the DBMS_STATS package are required for the creation of
accurate execution plans. For example, table statistics gathered by DBMS_STATS include
the number of rows, number of blocks, and average row length.
By default, Oracle Database uses automatic optimizer statistics collection. In this case,
the database automatically runs DBMS_STATS to collect optimizer statistics for all schema
objects for which statistics are missing or stale. The process eliminates many manual
tasks associated with managing the optimizer, and significantly reduces the risks of
generating suboptimal execution plans because of missing or stale statistics. You can
also update and manage optimizer statistics by manually executing DBMS_STATS.

See Also:
•

"Configuring Automatic Optimizer Statistics Collection"

•

"Gathering Optimizer Statistics Manually"

•

Oracle Database Administrator’s Guide to learn more about automated
maintenance tasks

•

Oracle Database PL/SQL Packages and Types Reference to learn about
DBMS_STATS

10.3.2 Supplemental Dynamic Statistics
By default, when optimizer statistics are missing, stale, or insufficient, the database
automatically gathers dynamic statistics during a parse. The database uses
recursive SQL to scan a small random sample of table blocks.

Note:
Dynamic statistics augment statistics rather than providing an alternative to
them.

Dynamic statistics supplement optimizer statistics such as table and index block
counts, table and join cardinalities (estimated number of rows), join column statistics,
and GROUP BY statistics. This information helps the optimizer improve plans by making
better estimates for predicate cardinality.
Dynamic statistics are beneficial in the following situations:
•

An execution plan is suboptimal because of complex predicates.

•

The sampling time is a small fraction of total execution time for the query.

•

The query executes many times so that the sampling time is amortized.

Related Topics
•

When the Database Samples Data
Starting in Oracle Database 12c, the optimizer automatically decides whether
dynamic statistics are useful and which sample size to use for all SQL statements.
In earlier releases, dynamic statistics were called dynamic sampling.

10-14

Chapter 10

How the Database Gathers Optimizer Statistics

•

Guideline for Setting the Sample Size
In the context of optimizer statistics, sampling is the gathering of statistics from a
random subset of table rows. By enabling the database to avoid full table scans
and sorts of entire tables, sampling minimizes the resources necessary to gather
statistics.

•

Configuring Options for Dynamic Statistics
Dynamic statistics are an optimization technique in which the database uses
recursive SQL to scan a small random sample of the blocks in a table.

10.3.3 Online Statistics Gathering for Bulk Loads
Starting in Oracle Database 12c, the database can gather table statistics automatically
during the following types of bulk loads: INSERT INTO ... SELECT into an empty table
using a direct path insert, and CREATE TABLE AS SELECT .

Note:
By default, a parallel insert uses a direct path insert. You can force a direct
path insert by using the /*+APPEND*/ hint.

This section contains the following topics:
•

Purpose of Online Statistics Gathering for Bulk Loads
Data warehouses typically load large amounts of data into the database. For
example, a sales data warehouse might load sales data nightly.

•

Global Statistics During Inserts into Empty Partitioned Tables
When inserting rows into an empty partitioned table, the database gathers global
statistics during the insert.

•

Index Statistics and Histograms During Bulk Loads
While gathering online statistics, the database does not gather index statistics or
create histograms. If these statistics are required, then Oracle recommends
running DBMS_STATS.GATHER_TABLE_STATS with the options parameter set to GATHER
AUTO after the bulk load.

•

Restrictions for Online Statistics Gathering for Bulk Loads
In some situations, optimizer statistics gathering does not occur automatically for
bulk loads.

•

Hints for Online Statistics Gathering for Bulk Loads
By default, the database gathers statistics during bulk loads. You can disable the
feature at the statement level by using the NO_GATHER_OPTIMIZER_STATISTICS hint,
and enable the feature at the statement level by using the
GATHER_OPTIMIZER_STATISTICS hint.

See Also:
Oracle Database Data Warehousing Guide to learn more about bulk loads

10-15

Chapter 10

How the Database Gathers Optimizer Statistics

10.3.3.1 Purpose of Online Statistics Gathering for Bulk Loads
Data warehouses typically load large amounts of data into the database. For example,
a sales data warehouse might load sales data nightly.
In releases earlier than Oracle Database 12c, to avoid the possibility of a suboptimal
plan caused by stale statistics, you needed to gather statistics manually after a bulk
load. The ability to gather statistics automatically during bulk loads has the following
benefits:
•

Improved performance
Gathering statistics during the load avoids an additional table scan to gather table
statistics.

•

Improved manageability
No user intervention is required to gather statistics after a bulk load.

10.3.3.2 Global Statistics During Inserts into Empty Partitioned Tables
When inserting rows into an empty partitioned table, the database gathers global
statistics during the insert.
For example, if sales is an empty partitioned table, and if you run INSERT INTO sales
SELECT, then the database gathers global statistics for sales. However, the database
does not gather partition-level statistics.
Assume a different case in which you use extended syntax to insert rows into a
particular partition or subpartition, which is empty. The database gathers statistics on
the empty partition during the insert. However, the database does not gather global
statistics.
Assume that you run INSERT INTO sales PARTITION (sales_q4_2000) SELECT. If partition
sales_q4_2000 is empty before the insert (other partitions need not be empty), then the
database gathers statistics during the insert. Moreover, if the INCREMENTAL preference is
enabled for sales, then the database also gathers a synopsis for sales_q4_2000.
Statistics are immediately available after the INSERT statement. However, if you roll
back the transaction, then the database automatically deletes statistics gathered
during the bulk load.

See Also:
•

"Considerations for Incremental Statistics Maintenance"

•

Oracle Database SQL Language Reference for INSERT syntax and
semantics

10.3.3.3 Index Statistics and Histograms During Bulk Loads
While gathering online statistics, the database does not gather index statistics or
create histograms. If these statistics are required, then Oracle recommends running
DBMS_STATS.GATHER_TABLE_STATS with the options parameter set to GATHER AUTO after the
bulk load.

10-16

Chapter 10

How the Database Gathers Optimizer Statistics

For example, the following command gathers statistics for the bulk-loaded sh_ctas
table:
EXEC DBMS_STATS.GATHER_TABLE_STATS( user, 'SH_CTAS', options => 'GATHER AUTO' );

The preceding example only gathers missing or stale statistics. The database does not
gather table and basic column statistics collected during the bulk load.

Note:
You can set the table preference options to GATHER AUTO on the tables that
you plan to bulk load. In this way, you need not explicitly set the options
parameter when running GATHER_TABLE_STATS.

See Also:
•

"Gathering Schema and Table Statistics"

•

Oracle Database Data Warehousing Guide to learn more about bulk
loading

10.3.3.4 Restrictions for Online Statistics Gathering for Bulk Loads
In some situations, optimizer statistics gathering does not occur automatically for bulk
loads.
Specifically, bulk loads do not gather statistics automatically when any of the following
conditions applies to the target table, partition, or subpartition:
•

It is not empty, and you perform an INSERT INTO ... SELECT.
In this case, an OPTIMIZER STATISTICS GATHERING row source appears in the plan,
but this row source is only a pass-through. The database does not actually gather
optimizer statistics.

Note:
The DBA_TAB_COL_STATISTICS.NOTES column is set to STATS_ON_LOAD by a
bulk load into an empty table. However, subsequent bulk loads into the
non-empty table do not reset the NOTES column. One technique for
determining whether the database gathered statistics is to query the
USER_TAB_MODIFICATIONS.INSERTS column. If the query returns a row
indicating the number of rows loaded, then the most recent bulk load did
not gather statistics automatically.
•

It is loaded using an INSERT INTO ... SELECT, and neither of the following
conditions is true: all columns of the target table are specified, or a subset of the
target columns are specified and the unspecified columns have default values.

10-17

Chapter 10

How the Database Gathers Optimizer Statistics

Put differently, the database only gathers statistics automatically for bulk loads
when either all columns of the target table are specified, or a subset of the target
columns are specified and the unspecified columns have default values. For
example, the sales table has only columns c1, c2, c3, and c4. The column c4 does
not have a default value. You load sales_copy by executing INSERT /*+ APPEND */
INTO sales_copy SELECT c1, c2, c3 FROM sales. In this case, the database does not
gather online statistics for sales_copy. The database would gather statistics if c4
had a default value or if it were included in the SELECT list.
•

It is in an Oracle-owned schema such as SYS.

•

It is one of the following types of tables: nested table, index-organized table (IOT),
external table, or global temporary table defined as ON COMMIT DELETE ROWS.

•

It has a PUBLISH preference set to FALSE.

•

Its statistics are locked.

•

It is partitioned, INCREMENTAL is set to true, and partition-extended syntax is not
used.
For example, assume that you execute DBMS_STATS.SET_TABLE_PREFS(null, 'sales',
incremental', 'true'). In this case, the database does not gather statistics for
INSERT INTO sales SELECT, even when sales is empty. However, the database does
gather statistics automatically for INSERT INTO sales PARTITION (sales_q4_2000)
SELECT.

•

It is loaded using a multitable INSERT statement.

See Also:
•

"Gathering Schema and Table Statistics"

•

Oracle Database PL/SQL Packages and Types Reference to learn more
about DBMS_STATS.SET_TABLE_PREFS

10.3.3.5 Hints for Online Statistics Gathering for Bulk Loads
By default, the database gathers statistics during bulk loads. You can disable the
feature at the statement level by using the NO_GATHER_OPTIMIZER_STATISTICS hint, and
enable the feature at the statement level by using the GATHER_OPTIMIZER_STATISTICS
hint.
For example, the following statement disables online statistics gathering for bulk loads:
CREATE TABLE employees2 AS
SELECT /*+NO_GATHER_OPTIMIZER_STATISTICS*/ * FROM employees

See Also:
Oracle Database SQL Language Reference to learn about the
GATHER_OPTIMIZER_STATISTICS and NO_GATHER_OPTIMIZER_STATISTICS hints

10-18

Chapter 10

When the Database Gathers Optimizer Statistics

10.4 When the Database Gathers Optimizer Statistics
The database collects optimizer statistics at various times and from various sources.
This section contains the following topics:
•

Sources for Optimizer Statistics
The optimizer uses several different sources for optimizer statistics.

•

SQL Plan Directives
A SQL plan directive is additional information and instructions that the optimizer
can use to generate a more optimal plan.

•

When the Database Samples Data
Starting in Oracle Database 12c, the optimizer automatically decides whether
dynamic statistics are useful and which sample size to use for all SQL statements.
In earlier releases, dynamic statistics were called dynamic sampling.

•

How the Database Samples Data
At the beginning of optimization, when deciding whether a table is a candidate for
dynamic statistics, the optimizer checks for the existence of persistent SQL plan
directives on the table.

10.4.1 Sources for Optimizer Statistics
The optimizer uses several different sources for optimizer statistics.
The sources are as follows:
•

DBMS_STATS execution, automatic or manual

This PL/SQL package is the primary means of gathering optimizer statistics.
•

SQL compilation
During SQL compilation, the database can augment the statistics previously
gathered by DBMS_STATS. In this stage, the database runs additional queries to
obtain more accurate information on how many rows in the tables satisfy the WHERE
clause predicates in the SQL statement.

•

SQL execution
During execution, the database can further augment previously gathered statistics.
In this stage, Oracle Database collects the number of rows produced by every row
source during the execution of a SQL statement. At the end of execution, the
optimizer determines whether the estimated number of rows is inaccurate enough
to warrant reparsing at the next statement execution. If the cursor is marked for
reparsing, then the optimizer uses actual row counts from the previous execution
instead of estimates.

•

SQL profiles
A SQL profile is a collection of auxiliary statistics on a query. The profile stores
these supplemental statistics in the data dictionary. The optimizer uses SQL
profiles during optimization to determine the most optimal plan.

The database stores optimizer statistics in the data dictionary and updates or replaces
them as needed. You can query statistics in data dictionary views.

10-19

Chapter 10

When the Database Gathers Optimizer Statistics

See Also:
•

"When the Database Samples Data"

•

"About SQL Profiles"

•

Oracle Database PL/SQL Packages and Types Reference to learn about
the DBMS_STATS.GATHER_TABLE_STATS procedure

10.4.2 SQL Plan Directives
A SQL plan directive is additional information and instructions that the optimizer can
use to generate a more optimal plan.
The directive is a “note to self” by the optimizer that it is misestimating cardinalities of
certain types of predicates, and also a reminder to DBMS_STATS to gather statistics
needed to correct the misestimates in the future. For example, when joining two tables
that have a data skew in their join columns, a SQL plan directive can direct the
optimizer to use dynamic statistics to obtain a more accurate join cardinality estimate.
This section contains the following topics:
•

When the Database Creates SQL Plan Directives
The database creates SQL plan directives automatically based on information
learned during automatic reoptimization. If a cardinality misestimate occurs during
SQL execution, then the database creates SQL plan directives.

•

How the Database Uses SQL Plan Directives
When compiling a SQL statement, if the optimizer sees a directive, then it obeys
the directive by gathering additional information.

•

SQL Plan Directive Maintenance
The database automatically creates SQL plan directives. You cannot create them
manually.

•

How the Optimizer Uses SQL Plan Directives: Example
This example shows how the database automatically creates and uses SQL plan
directives for SQL statements.

•

How the Optimizer Uses Extensions and SQL Plan Directives: Example
The example shows how the database uses a SQL plan directive until the
optimizer verifies that an extension exists and the statistics are applicable.

10.4.2.1 When the Database Creates SQL Plan Directives
The database creates SQL plan directives automatically based on information learned
during automatic reoptimization. If a cardinality misestimate occurs during SQL
execution, then the database creates SQL plan directives.
For each new directive, the DBA_SQL_PLAN_DIRECTIVES.STATE column shows the value
USABLE. This value indicates that the database can use the directive to correct
misestimates.
The optimizer defines a SQL plan directive on a query expression, for example, filter
predicates on two columns being used together. A directive is not tied to a specific
SQL statement or SQL ID. For this reason, the optimizer can use directives for

10-20

Chapter 10

When the Database Gathers Optimizer Statistics

statements that are not identical. For example, directives can help the optimizer with
queries that use similar patterns, such as queries that are identical except for a select
list item.
The Notes section of the execution plan indicates the number of SQL plan directives
used for a statement. Obtain more information about the directives by querying the
DBA_SQL_PLAN_DIRECTIVES and DBA_SQL_PLAN_DIR_OBJECTS views.

See Also:
Oracle Database Reference to learn more about DBA_SQL_PLAN_DIRECTIVES

10.4.2.2 How the Database Uses SQL Plan Directives
When compiling a SQL statement, if the optimizer sees a directive, then it obeys the
directive by gathering additional information.
The optimizer uses directives in the following ways:
•

Dynamic statistics
The optimizer uses dynamic statistics whenever it does not have sufficient
statistics corresponding to the directive. For example, the cardinality estimates for
a query whose predicate contains a specific pair of columns may be significantly
wrong. A SQL plan directive indicates that the whenever a query that contains
these columns is parsed, the optimizer needs to use dynamic sampling to avoid a
serious cardinality misestimate.
Dynamic statistics have some performance overhead. Every time the optimizer
hard parses a query to which a dynamic statistics directive applies, the database
must perform the extra sampling.
Starting in Oracle Database 12c Release 2 (12.2), the database writes statistics
from adaptive dynamic sampling to the SQL plan directives store, making them
available to other queries.

•

Column groups
The optimizer examines the query corresponding to the directive. If there is a
missing column group, and if the DBMS_STATS preference AUTO_STAT_EXTENSIONS is set
to ON (the default is OFF) for the affected table, then the optimizer automatically
creates this column group the next time DBMS_STATS gathers statistics on the table.
Otherwise, the optimizer does not automatically create the column group.
If a column group exists, then the next time this statement executes, the optimizer
uses the column group statistics in place of the SQL plan directive when possible
(equality predicates, GROUP BY, and so on). In subsequent executions, the optimizer
may create additional SQL plan directives to address other problems in the plan,
such as join or GROUP BY cardinality misestimates.

Note:
Currently, the optimizer monitors only column groups. The optimizer
does not create an extension on expressions.

10-21

Chapter 10

When the Database Gathers Optimizer Statistics

When the problem that occasioned a directive is solved, either because a better
directive exists or because a histogram or extension exists, the
DBA_SQL_PLAN_DIRECTIVES.STATE value changes from USABLE to SUPERSEDED. More
information about the directive state is exposed in the DBA_SQL_PLAN_DIRECTIVES.NOTES
column.

See Also:
•

"Managing Extended Statistics"

•

"About Statistics on Column Groups"

•

Oracle Database PL/SQL Packages and Types Reference to learn more
about the AUTO_STAT_EXTENSIONS preference for
DBMS_STATS.SET_TABLE_STATS

10.4.2.3 SQL Plan Directive Maintenance
The database automatically creates SQL plan directives. You cannot create them
manually.
The database initially creates directives in the shared pool. The database periodically
writes the directives to the SYSAUX tablespace. The database automatically purges any
SQL plan directive that is not used after the specified number of weeks
(SPD_RETENTION_WEEKS), which is 53 by default.
You can manage directives by using the DBMS_SPD package. For example, you can:
•

Enable and disable SQL plan directives (ALTER_SQL_PLAN_DIRECTIVE)

•

Change the retention period for SQL plan directives (SET_PREFS)

•

Export a directive to a staging table (PACK_STGTAB_DIRECTIVE)

•

Drop a directive (DROP_SQL_PLAN_DIRECTIVE)

•

Force the database to write directives to disk (FLUSH_SQL_PLAN_DIRECTIVE)

See Also:
Oracle Database PL/SQL Packages and Types Reference to learn about the
DBMS_SPD package

10.4.2.4 How the Optimizer Uses SQL Plan Directives: Example
This example shows how the database automatically creates and uses SQL plan
directives for SQL statements.
Assumptions
You plan to run queries against the sh schema, and you have privileges on this
schema and on data dictionary and V$ views.

10-22

Chapter 10

When the Database Gathers Optimizer Statistics

To see how the database uses a SQL plan directive:
1.

Query the sh.customers table.
SELECT
FROM
WHERE
AND

/*+gather_plan_statistics*/ *
customers
cust_state_province='CA'
country_id='US';

The gather_plan_statistics hint shows the actual number of rows returned from
each operation in the plan. Thus, you can compare the optimizer estimates with
the actual number of rows returned.
2.

Query the plan for the preceding query.
The following example shows the execution plan (sample output included):

SELECT * FROM TABLE(DBMS_XPLAN.DISPLAY_CURSOR(FORMAT=>'ALLSTATS LAST'));
PLAN_TABLE_OUTPUT
------------------------------------SQL_ID b74nw722wjvy3, child number 0
------------------------------------select /*+gather_plan_statistics*/ * from customers where
CUST_STATE_PROVINCE='CA' and country_id='US'
Plan hash value: 1683234692
-------------------------------------------------------------------------------------------| Id| Operation
| Name
| Starts |E-Rows| A-Rows | A-Time |Buffers | Reads |
-------------------------------------------------------------------------------------------| 0 | SELECT STATEMENT |
|
1 |
|
29 |00:00:00.01 |
17 |
14 |
|*1 | TABLE ACCESS FULL| CUSTOMERS |
1 |
8 |
29 |00:00:00.01 |
17 |
14 |
-------------------------------------------------------------------------------------------Predicate Information (identified by operation id):
--------------------------------------------------1 - filter(("CUST_STATE_PROVINCE"='CA' AND "COUNTRY_ID"='US'))

The actual number of rows (A-Rows) returned by each operation in the plan varies
greatly from the estimates (E-Rows). This statement is a candidate for automatic
reoptimization.
3.

Check whether the customers query can be reoptimized.
The following statement queries the V$SQL.IS_REOPTIMIZABLE value (sample output
included):
SELECT SQL_ID, CHILD_NUMBER, SQL_TEXT, IS_REOPTIMIZABLE
FROM V$SQL
WHERE SQL_TEXT LIKE 'SELECT /*+gather_plan_statistics*/%';
SQL_ID
CHILD_NUMBER SQL_TEXT
I
------------- ------------ ----------- b74nw722wjvy3
0 select /*+g Y
ather_plan_
statistics*
/ * from cu
stomers whe
re CUST_STA
TE_PROVINCE
='CA' and c

10-23

Chapter 10

When the Database Gathers Optimizer Statistics

ountry_id='
US'

The IS_REOPTIMIZABLE column is marked Y, so the database will perform a hard
parse of the customers query on the next execution. The optimizer uses the
execution statistics from this initial execution to determine the plan. The database
persists the information learned from reoptimization as a SQL plan directive.
4.

Display the directives for the sh schema.
The following example uses DBMS_SPD to write the SQL plan directives to disk, and
then shows the directives for the sh schema only:

EXEC DBMS_SPD.FLUSH_SQL_PLAN_DIRECTIVE;
SELECT TO_CHAR(d.DIRECTIVE_ID) dir_id, o.OWNER AS "OWN", o.OBJECT_NAME AS "OBJECT",
o.SUBOBJECT_NAME col_name, o.OBJECT_TYPE, d.TYPE, d.STATE, d.REASON
FROM DBA_SQL_PLAN_DIRECTIVES d, DBA_SQL_PLAN_DIR_OBJECTS o
WHERE d.DIRECTIVE_ID=o.DIRECTIVE_ID
AND
o.OWNER IN ('SH')
ORDER BY 1,2,3,4,5;
DIR_ID
OWN OBJECT
COL_NAME
OBJECT TYPE
STATE REASON
------------------- --- --------- ----------- ------ ---------------- ------ -----------1484026771529551585 SH CUSTOMERS COUNTRY_ID COLUMN DYNAMIC_SAMPLING USABLE SINGLE TABLE
CARDINALITY
MISESTIMATE
1484026771529551585 SH CUSTOMERS CUST_STATE_ COLUMN DYNAMIC_SAMPLING USABLE SINGLE TABLE
PROVINCE
CARDINALITY
MISESTIMATE
1484026771529551585 SH CUSTOMERS
TABLE DYNAMIC_SAMPLING USABLE SINGLE TABLE
CARDINALITY
MISESTIMATE

Initially, the database stores SQL plan directives in memory, and then writes them
to disk every 15 minutes. Thus, the preceding example calls
DBMS_SPD.FLUSH_SQL_PLAN_DIRECTIVE to force the database to write the directives to
the SYSAUX tablespace.
Monitor directives using the views DBA_SQL_PLAN_DIRECTIVES and
DBA_SQL_PLAN_DIR_OBJECTS. Three entries appear in the views, one for the customers
table itself, and one for each of the correlated columns. Because the customers
query has the IS_REOPTIMIZABLE value of Y, if you reexecute the statement, then the
database will hard parse it again, and then generate a plan based on the previous
execution statistics.
5.

Query the customers table again.
For example, enter the following statement:
SELECT
FROM
WHERE
AND

6.

/*+gather_plan_statistics*/ *
customers
cust_state_province='CA'
country_id='US';

Query the plan in the cursor.
The following example shows the execution plan (sample output included):
SELECT * FROM TABLE(DBMS_XPLAN.DISPLAY_CURSOR(FORMAT=>'ALLSTATS LAST'));
PLAN_TABLE_OUTPUT
-------------------------------------

10-24

Chapter 10

When the Database Gathers Optimizer Statistics

SQL_ID b74nw722wjvy3, child number 1
------------------------------------select /*+gather_plan_statistics*/ * from customers where
CUST_STATE_PROVINCE='CA' and country_id='US'
Plan hash value: 1683234692
--------------------------------------------------------------------------|Id | Operation
|Name
|Start|E-Rows|A-Rows| A-Time |Buffers|
--------------------------------------------------------------------------| 0| SELECT STATEMENT |
|
1|
|
29|00:00:00.01|
17|
|* 1| TABLE ACCESS FULL|CUSTOMERS|
1|
29|
29|00:00:00.01|
17|
--------------------------------------------------------------------------Predicate Information (identified by operation id):
--------------------------------------------------1 - filter(("CUST_STATE_PROVINCE"='CA' AND "COUNTRY_ID"='US'))
Note
----- cardinality feedback used for this statement

The Note section indicates that the database used reoptimization for this
statement. The estimated number of rows (E-Rows) is now correct. The SQL plan
directive has not been used yet.
7.

Query the cursors for the customers query.
For example, run the following query (sample output included):
SELECT SQL_ID, CHILD_NUMBER, SQL_TEXT, IS_REOPTIMIZABLE
FROM V$SQL
WHERE SQL_TEXT LIKE 'SELECT /*+gather_plan_statistics*/%';
SQL_ID
CHILD_NUMBER SQL_TEXT
I
------------- ------------ ----------- b74nw722wjvy3
0 select /*+g Y
ather_plan_
statistics*
/ * from cu
stomers whe
re CUST_STA
TE_PROVINCE
='CA' and c
ountry_id='
US'
b74nw722wjvy3

1 select /*+g N
ather_plan_
statistics*
/ * from cu
stomers whe
re CUST_STA
TE_PROVINCE
='CA' and c
ountry_id='
US'

A new plan exists for the customers query, and also a new child cursor.
8.

Confirm that a SQL plan directive exists and is usable for other statements.

10-25

Chapter 10

When the Database Gathers Optimizer Statistics

For example, run the following query, which is similar but not identical to the
original customers query (the state is MA instead of CA):
SELECT
FROM
WHERE
AND
9.

/*+gather_plan_statistics*/ CUST_EMAIL
CUSTOMERS
CUST_STATE_PROVINCE='MA'
COUNTRY_ID='US';

Query the plan in the cursor.
The following statement queries the cursor (sample output included).:
SELECT * FROM TABLE(DBMS_XPLAN.DISPLAY_CURSOR(FORMAT=>'ALLSTATS LAST'));
PLAN_TABLE_OUTPUT
------------------------------------SQL_ID 3tk6hj3nkcs2u, child number 0
------------------------------------Select /*+gather_plan_statistics*/ cust_email From
cust_state_province='MA' And
country_id='US'

customers Where

Plan hash value: 1683234692
--------------------------------------------------------------------------|Id | Operation
| Name
|Starts|E-Rows|A-Rows| A-Time |Buffers|
--------------------------------------------------------------------------| 0 | SELECT STATEMENT |
|
1 |
| 2 |00:00:00.01| 16 |
|*1 | TABLE ACCESS FULL| CUSTOMERS |
1 |
2 | 2 |00:00:00.01| 16 |
--------------------------------------------------------------------------Predicate Information (identified by operation id):
--------------------------------------------------1 - filter(("CUST_STATE_PROVINCE"='MA' AND "COUNTRY_ID"='US'))
Note
----- dynamic sampling used for this statement (level=2)
- 1 Sql Plan Directive used for this statement

The Note section of the plan shows that the optimizer used the SQL directive for
this statement, and also used dynamic statistics.

See Also:
•

"Automatic Reoptimization"

•

"Managing SQL Plan Directives"

•

Oracle Database Reference to learn about DBA_SQL_PLAN_DIRECTIVES,
V$SQL, and other database views

•

Oracle Database Reference to learn about DBMS_SPD

10-26

Chapter 10

When the Database Gathers Optimizer Statistics

10.4.2.5 How the Optimizer Uses Extensions and SQL Plan Directives:
Example
The example shows how the database uses a SQL plan directive until the optimizer
verifies that an extension exists and the statistics are applicable.
At this point, the directive changes its status to SUPERSEDED. Subsequent compilations
use the statistics instead of the directive.
Assumptions
This example assumes you have already followed the steps in "How the Optimizer
Uses SQL Plan Directives: Example".
To see how the optimizer uses an extension and SQL plan directive:
1.

Gather statistics for the sh.customers table.
For example, execute the following PL/SQL program:
BEGIN
DBMS_STATS.GATHER_TABLE_STATS('SH','CUSTOMERS');
END;
/

2.

Check whether an extension exists on the customers table.
For example, execute the following query (sample output included):
SELECT
FROM
WHERE
AND

TABLE_NAME, EXTENSION_NAME, EXTENSION
DBA_STAT_EXTENSIONS
OWNER='SH'
TABLE_NAME='CUSTOMERS';

TABLE_NAM EXTENSION_NAME
EXTENSION
--------- ------------------------------ ----------------------CUSTOMERS SYS_STU#S#WF25Z#QAHIHE#MOFFMM_ ("CUST_STATE_PROVINCE",
"COUNTRY_ID")

The preceding output indicates that a column group extension exists on the
cust_state_province and country_id columns.
3.

Query the state of the SQL plan directive.
Example 10-6 queries the data dictionary for information about the directive.
Although column group statistics exist, the directive has a state of USABLE because
the database has not yet recompiled the statement. During the next compilation,
the optimizer verifies that the statistics are applicable. If they are applicable, then
the status of the directive changes to SUPERSEDED. Subsequent compilations use the
statistics instead of the directive.

4.

Query the sh.customers table.
SELECT
FROM
WHERE
AND

5.

/*+gather_plan_statistics*/ *
customers
cust_state_province='CA'
country_id='US';

Query the plan in the cursor.
Example 10-7 shows the execution plan (sample output included).

10-27

Chapter 10

When the Database Gathers Optimizer Statistics

The Note section shows that the optimizer used the directive and not the extended
statistics. During the compilation, the database verified the extended statistics.
6.

Query the state of the SQL plan directive.
Example 10-8 queries the data dictionary for information about the directive.
The state of the directive, which has changed to SUPERSEDED, indicates that the
corresponding column or groups have an extension or histogram, or that another
SQL plan directive exists that can be used for the directive.

7.

Query the sh.customers table again, using a slightly different form of the statement.
For example, run the following query:
SELECT
FROM
WHERE
AND

/*+gather_plan_statistics*/ /* force reparse */ *
customers
cust_state_province='CA'
country_id='US';

If the cursor is in the shared SQL area, then the database typically shares the
cursor. To force a reparse, this step changes the SQL text slightly by adding a
comment.
8.

Query the plan in the cursor.
Example 10-9 shows the execution plan (sample output included).
The absence of a Note shows that the optimizer used the extended statistics
instead of the SQL plan directive. If the directive is not used for 53 weeks, then the
database automatically purges it.

See Also:
•
•

"Managing SQL Plan Directives"
Oracle Database Reference to learn about DBA_SQL_PLAN_DIRECTIVES,
V$SQL, and other database views

•

Example 10-6

Oracle Database Reference to learn about DBMS_SPD

Display Directives for sh Schema

EXEC DBMS_SPD.FLUSH_SQL_PLAN_DIRECTIVE;
SELECT TO_CHAR(d.DIRECTIVE_ID) dir_id, o.OWNER, o.OBJECT_NAME,
o.SUBOBJECT_NAME col_name, o.OBJECT_TYPE, d.TYPE, d.STATE, d.REASON
FROM DBA_SQL_PLAN_DIRECTIVES d, DBA_SQL_PLAN_DIR_OBJECTS o
WHERE d.DIRECTIVE_ID=o.DIRECTIVE_ID
AND
o.OWNER IN ('SH')
ORDER BY 1,2,3,4,5;
DIR_ID
OWN OBJECT_NA COL_NAME OBJECT TYPE
STATE REASON
------------------- --- --------- ---------- ------- ---------------- ------ -----------1484026771529551585 SH CUSTOMERS COUNTRY_ID COLUMN DYNAMIC_SAMPLING USABLE SINGLE TABLE
CARDINALITY
MISESTIMATE
1484026771529551585 SH CUSTOMERS CUST_STATE_ COLUMN DYNAMIC_SAMPLING USABLE SINGLE TABLE
PROVINCE
CARDINALITY
MISESTIMATE

10-28

Chapter 10

When the Database Gathers Optimizer Statistics

1484026771529551585 SH CUSTOMERS

Example 10-7

TABLE DYNAMIC_SAMPLING USABLE SINGLE TABLE
CARDINALITY
MISESTIMATE

Execution Plan

SQL> SELECT * FROM TABLE(DBMS_XPLAN.DISPLAY_CURSOR(FORMAT=>'ALLSTATS LAST'));
PLAN_TABLE_OUTPUT
------------------------------------SQL_ID b74nw722wjvy3, child number 0
------------------------------------select /*+gather_plan_statistics*/ * from customers where
CUST_STATE_PROVINCE='CA' and country_id='US'
Plan hash value: 1683234692
----------------------------------------------------------------------------------------| Id | Operation
| Name
| Starts | E-Rows | A-Rows | A-Time | Buffers |
----------------------------------------------------------------------------------------| 0 | SELECT STATEMENT |
|
1 |
|
29 |00:00:00.01 |
16 |
|* 1 | TABLE ACCESS FULL| CUSTOMERS |
1 |
29 |
29 |00:00:00.01 |
16 |
----------------------------------------------------------------------------------------Predicate Information (identified by operation id):
--------------------------------------------------1 - filter(("CUST_STATE_PROVINCE"='CA' AND "COUNTRY_ID"='US'))
Note
----- dynamic sampling used for this statement (level=2)
- 1 Sql Plan Directive used for this statement

Example 10-8

Display Directives for sh Schema

EXEC DBMS_SPD.FLUSH_SQL_PLAN_DIRECTIVE;
SELECT TO_CHAR(d.DIRECTIVE_ID) dir_id, o.OWNER, o.OBJECT_NAME,
o.SUBOBJECT_NAME col_name, o.OBJECT_TYPE, d.TYPE, d.STATE, d.REASON
FROM DBA_SQL_PLAN_DIRECTIVES d, DBA_SQL_PLAN_DIR_OBJECTS o
WHERE d.DIRECTIVE_ID=o.DIRECTIVE_ID
AND
o.OWNER IN ('SH')
ORDER BY 1,2,3,4,5;
DIR_ID
OWN OBJECT_NA COL_NAME
OBJECT TYPE
STATE
------------------- --- --------- ---------- ------ -------- --------1484026771529551585 SH CUSTOMERS COUNTRY_ID COLUMN DYNAMIC_ SUPERSEDED
SAMPLING

REASON
-----------SINGLE TABLE
CARDINALITY
MISESTIMATE
1484026771529551585 SH CUSTOMERS CUST_STATE_ COLUMN DYNAMIC_ SUPERSEDED SINGLE TABLE
PROVINCE
SAMPLING
CARDINALITY
MISESTIMATE
1484026771529551585 SH CUSTOMERS
TABLE DYNAMIC_ SUPERSEDED SINGLE TABLE
SAMPLING
CARDINALITY
MISESTIMATE

Example 10-9

Execution Plan

SQL> SELECT * FROM TABLE(DBMS_XPLAN.DISPLAY_CURSOR(FORMAT=>'ALLSTATS LAST'));

10-29

Chapter 10

When the Database Gathers Optimizer Statistics

PLAN_TABLE_OUTPUT
------------------------------------SQL_ID b74nw722wjvy3, child number 0
------------------------------------select /*+gather_plan_statistics*/ * from customers where
CUST_STATE_PROVINCE='CA' and country_id='US'
Plan hash value: 1683234692
----------------------------------------------------------------------------------------| Id | Operation
| Name
| Starts | E-Rows | A-Rows | A-Time | Buffers |
----------------------------------------------------------------------------------------| 0 | SELECT STATEMENT |
|
1 |
|
29 |00:00:00.01 |
17 |
|* 1 | TABLE ACCESS FULL| CUSTOMERS |
1 |
29 |
29 |00:00:00.01 |
17 |
----------------------------------------------------------------------------------------Predicate Information (identified by operation id):
--------------------------------------------------1 - filter(("CUST_STATE_PROVINCE"='CA' AND "COUNTRY_ID"='US'))
19 rows selected.

10.4.3 When the Database Samples Data
Starting in Oracle Database 12c, the optimizer automatically decides whether dynamic
statistics are useful and which sample size to use for all SQL statements. In earlier
releases, dynamic statistics were called dynamic sampling.
The primary factor in the decision to use dynamic statistics is whether available
statistics are sufficient to generate an optimal plan. If statistics are insufficient, then the
optimizer uses dynamic statistics.
Automatic dynamic statistics are enabled when the OPTIMIZER_DYNAMIC_SAMPLING
initialization parameter is not set to 0. By default, the dynamic statistics level is set to 2.
In general, the optimizer uses default statistics rather than dynamic statistics to
compute statistics needed during optimizations on tables, indexes, and columns. The
optimizer decides whether to use dynamic statistics based on several factors,
including the following:
•

The SQL statement uses parallel execution.

•

A SQL plan directive exists.

The following diagram illustrates the process of gathering dynamic statistics.

10-30

Chapter 10

When the Database Gathers Optimizer Statistics

Figure 10-2

Dynamic Statistics

Optimizer
GB
Statistics missing?
Statistics insufficient?
SQL directive exists?
Parallel execution?

Execution
Plan

HJ
No

HJ

Yes
CLIENT
SQL
Determine sampling
size

SELECT ...
WHERE ...

Pass results to
optimizer for
use in plan
generation

SELECT ...
FROM sales
WHERE ...

Recursive
SQL

Sales

As shown in Figure 10-2, the optimizer automatically gathers dynamic statistics in the
following cases:
•

Missing statistics
When tables in a query have no statistics, the optimizer gathers basic statistics on
these tables before optimization. Statistics can be missing because the application
creates new objects without a follow-up call to DBMS_STATS to gather statistics, or
because statistics were locked on an object before statistics were gathered.
In this case, the statistics are not as high-quality or as complete as the statistics
gathered using the DBMS_STATS package. This trade-off is made to limit the impact
on the compile time of the statement.

•

Insufficient statistics
Statistics can be insufficient whenever the optimizer estimates the selectivity of
predicates (filter or join) or the GROUP BY clause without taking into account
correlation between columns, skew in the column data distribution, statistics on
expressions, and so on.
Extended statistics help the optimizer obtain accurate quality cardinality estimates
for complex predicate expressions. The optimizer can use dynamic statistics to
compensate for the lack of extended statistics or when it cannot use extended
statistics, for example, for non-equality predicates.

10-31

Chapter 10

When the Database Gathers Optimizer Statistics

Note:
The database does not use dynamic statistics for queries that contain the AS
OF clause.

See Also:
•

"Configuring Options for Dynamic Statistics"

•

"About Statistics on Column Groups"

•

Oracle Database Reference to learn about the
OPTIMIZER_DYNAMIC_SAMPLING initialization parameter

10.4.4 How the Database Samples Data
At the beginning of optimization, when deciding whether a table is a candidate for
dynamic statistics, the optimizer checks for the existence of persistent SQL plan
directives on the table.
For each directive, the optimizer registers a statistics expression that the optimizer
computes when determining the cardinality of a predicate involving the table. In
Figure 10-2, the database issues a recursive SQL statement to scan a small random
sample of the table blocks. The database applies the relevant single-table predicates
and joins to estimate predicate cardinalities.
The database persists the results of dynamic statistics as sharable statistics. The
database can share the results during the SQL compilation of one query with
recompilations of the same query. The database can also reuse the results for queries
that have the same patterns.

See Also:
•

"Configuring Options for Dynamic Statistics" to learn how to set the
dynamic statistics level

•

Oracle Database Reference for details about the
OPTIMIZER_DYNAMIC_SAMPLING initialization parameter

10-32

11
Histograms
A histogram is a special type of column statistic that provides more detailed
information about the data distribution in a table column. A histogram sorts values into
"buckets," as you might sort coins into buckets.
Based on the NDV and the distribution of the data, the database chooses the type of
histogram to create. (In some cases, when creating a histogram, the database
samples an internally predetermined number of rows.) The types of histograms are as
follows:
•

Frequency histograms and top frequency histograms

•

Height-Balanced histograms (legacy)

•

Hybrid histograms

This chapter contains the following topics:
•

Purpose of Histograms
By default the optimizer assumes a uniform distribution of rows across the distinct
values in a column.

•

When Oracle Database Creates Histograms
If DBMS_STATS gathers statistics for a table, and if queries have referenced the
columns in this table, then Oracle Database creates histograms automatically as
needed according to the previous query workload.

•

How Oracle Database Chooses the Histogram Type
Oracle Database uses several criteria to determine which histogram to create:
frequency, top frequency, height-balanced, or hybrid.

•

Cardinality Algorithms When Using Histograms
For histograms, the algorithm for cardinality depends on factors such as the
endpoint numbers and values, and whether column values are popular or
nonpopular.

•

Frequency Histograms
In a frequency histogram, each distinct column value corresponds to a single
bucket of the histogram. Because each value has its own dedicated bucket, some
buckets may have many values, whereas others have few.

•

Top Frequency Histograms
A top frequency histogram is a variation on a frequency histogram that ignores
nonpopular values that are statistically insignificant.

•

Height-Balanced Histograms (Legacy)
In a legacy height-balanced histogram, column values are divided into buckets so
that each bucket contains approximately the same number of rows.

•

Hybrid Histograms
A hybrid histogram combines characteristics of both height-based histograms
and frequency histograms. This "best of both worlds" approach enables the
optimizer to obtain better selectivity estimates in some situations.

11-1

Chapter 11

Purpose of Histograms

11.1 Purpose of Histograms
By default the optimizer assumes a uniform distribution of rows across the distinct
values in a column.
For columns that contain data skew (a nonuniform distribution of data within the
column), a histogram enables the optimizer to generate accurate cardinality estimates
for filter and join predicates that involve these columns.
For example, a California-based book store ships 95% of the books to California, 4%
to Oregon, and 1% to Nevada. The book orders table has 300,000 rows. A table
column stores the state to which orders are shipped. A user queries the number of
books shipped to Oregon. Without a histogram, the optimizer assumes an even
distribution of 300000/3 (the NDV is 3), estimating cardinality at 100,000 rows. With
this estimate, the optimizer chooses a full table scan. With a histogram, the optimizer
calculates that 4% of the books are shipped to Oregon, and chooses an index scan.
Related Topics
•

Introduction to Access Paths
A row source is a set of rows returned by a step in an execution plan. A row
source can be a table, view, or result of a join or grouping operation.

11.2 When Oracle Database Creates Histograms
If DBMS_STATS gathers statistics for a table, and if queries have referenced the columns
in this table, then Oracle Database creates histograms automatically as needed
according to the previous query workload.
The basic process is as follows:
1.

You run DBMS_STATS for a table with the METHOD_OPT parameter set to the default SIZE
AUTO.

2.

A user queries the table.

3.

The database notes the predicates in the preceding query and updates the data
dictionary table SYS.COL_USAGE$.

4.

You run DBMS_STATS again, causing DBMS_STATS to query SYS.COL_USAGE$ to determine
which columns require histograms based on the previous query workload.

Consequences of the AUTO feature include the following:
•

As queries change over time, DBMS_STATS may change which statistics it gathers.
For example, even if the data in a table does not change, queries and DBMS_STATS
operations can cause the plans for queries that reference these tables to change.

•

If you gather statistics for a table and do not query the table, then the database
does not create histograms for columns in this table. For the database to create
the histograms automatically, you must run one or more queries to populate the
column usage information in SYS.COL_USAGE$.

Example 11-1

Automatic Histogram Creation

Assume that sh.sh_ext is an external table that contains the same rows as the
sh.sales table. You create new table sales2 and perform a bulk load using sh_ext as a

11-2

Chapter 11

When Oracle Database Creates Histograms

source, which automatically creates statistics for sales2. You also create indexes as
follows:
SQL> CREATE TABLE sales2 AS SELECT * FROM sh_ext;
SQL> CREATE INDEX sh_12c_idx1 ON sales2(prod_id);
SQL> CREATE INDEX sh_12c_idx2 ON sales2(cust_id,time_id);

You query the data dictionary to determine whether histograms exist for the sales2
columns. Because sales2 has not yet been queried, the database has not yet created
histograms:
SQL> SELECT COLUMN_NAME, NOTES, HISTOGRAM
2 FROM USER_TAB_COL_STATISTICS
3 WHERE TABLE_NAME = 'SALES2';
COLUMN_NAME
------------AMOUNT_SOLD
QUANTITY_SOLD
PROMO_ID
CHANNEL_ID
TIME_ID
CUST_ID
PROD_ID

NOTES
-------------STATS_ON_LOAD
STATS_ON_LOAD
STATS_ON_LOAD
STATS_ON_LOAD
STATS_ON_LOAD
STATS_ON_LOAD
STATS_ON_LOAD

HISTOGRAM
--------NONE
NONE
NONE
NONE
NONE
NONE
NONE

You query sales2 for the number of rows for product 42, and then gather table statistics
using the GATHER AUTO option:
SQL> SELECT COUNT(*) FROM sales2 WHERE prod_id = 42;
COUNT(*)
---------12116
SQL> EXEC DBMS_STATS.GATHER_TABLE_STATS(USER,'SALES2',OPTIONS=>'GATHER AUTO');

A query of the data dictionary now shows that the database created a histogram on
the prod_id column based on the information gather during the preceding query:
SQL> SELECT COLUMN_NAME, NOTES, HISTOGRAM
2 FROM USER_TAB_COL_STATISTICS
3 WHERE TABLE_NAME = 'SALES2';
COLUMN_NAME
------------AMOUNT_SOLD
QUANTITY_SOLD
PROMO_ID
CHANNEL_ID
TIME_ID
CUST_ID
PROD_ID

NOTES
-------------STATS_ON_LOAD
STATS_ON_LOAD
STATS_ON_LOAD
STATS_ON_LOAD
STATS_ON_LOAD
STATS_ON_LOAD
HISTOGRAM_ONLY

HISTOGRAM
--------NONE
NONE
NONE
NONE
NONE
NONE
FREQUENCY

Related Topics
•

Online Statistics Gathering for Bulk Loads
Starting in Oracle Database 12c, the database can gather table statistics
automatically during the following types of bulk loads: INSERT INTO ... SELECT into
an empty table using a direct path insert, and CREATE TABLE AS SELECT .

11-3

Chapter 11

How Oracle Database Chooses the Histogram Type

11.3 How Oracle Database Chooses the Histogram Type
Oracle Database uses several criteria to determine which histogram to create:
frequency, top frequency, height-balanced, or hybrid.
The histogram formula uses the following variables:
•

NDV
This represents the number of distinct values in a column. For example, if a
column only contains the values 100, 200, and 300, then the NDV for this column is
3.

•

n
This variable represents the number of histogram buckets. The default is 254.

•

p
This variable represents an internal percentage threshold that is equal to (1–(1/n))
* 100. For example, if n = 254, then p is 99.6.

An additional criterion is whether the estimate_percent parameter in the DBMS_STATS
statistics gathering procedure is set to AUTO_SAMPLE_SIZE (default).
The following diagram shows the decision tree for histogram creation.

Figure 11-1

?

Decision Tree for Histogram Creation

NDV>n

Yes

ESTIMATE_PERCENT=
AUTO_SAMPLE_SIZE

No

Frequency
Histogram

Yes

Percentage
of rows for top n
frequent values >= p

No

Height-Balanced
Histogram

Yes

Top n
Frequency
Histogram

No

Hybrid
Histogram

NDV = Number of distinct values
n = Number of histogram buckets (default is 254)
p = (1-(1/n))*100

11.4 Cardinality Algorithms When Using Histograms
For histograms, the algorithm for cardinality depends on factors such as the endpoint
numbers and values, and whether column values are popular or nonpopular.
This section contains the following topics:

11-4

Chapter 11

Cardinality Algorithms When Using Histograms

•

Endpoint Numbers and Values
An endpoint number is a number that uniquely identifies a bucket. In frequency
and hybrid histograms, the endpoint number is the cumulative frequency of all
values included in the current and previous buckets.

•

Popular and Nonpopular Values
The popularity of a value in a histogram affects the cardinality estimate algorithm.

•

Bucket Compression
In some cases, to reduce the total number of buckets, the optimizer compresses
multiple buckets into a single bucket.

11.4.1 Endpoint Numbers and Values
An endpoint number is a number that uniquely identifies a bucket. In frequency and
hybrid histograms, the endpoint number is the cumulative frequency of all values
included in the current and previous buckets.
For example, a bucket with endpoint number 100 means the total frequency of values
in the current and all previous buckets is 100. In height-balanced histograms, the
optimizer numbers buckets sequentially, starting at 0 or 1. In all cases, the endpoint
number is the bucket number.
An endpoint value is the highest value in the range of values in a bucket. For example,
if a bucket contains only the values 52794 and 52795, then the endpoint value is 52795.

11.4.2 Popular and Nonpopular Values
The popularity of a value in a histogram affects the cardinality estimate algorithm.
Specifically, the cardinality estimate is affected as follows:
•

Popular values
A popular value occurs as an endpoint value of multiple buckets. The optimizer
determines whether a value is popular by first checking whether it is the endpoint
value for a bucket. If so, then for frequency histograms, the optimizer subtracts the
endpoint number of the previous bucket from the endpoint number of the current
bucket. Hybrid histograms already store this information for each endpoint
individually. If this value is greater than 1, then the value is popular.
The optimizer calculates its cardinality estimate for popular values using the
following formula:
cardinality of popular value =
(num of rows in table) *
(num of endpoints spanned by this value / total num of endpoints)

•

Nonpopular values
Any value that is not popular is a nonpopular value. The optimizer calculates the
cardinality estimates for nonpopular values using the following formula:
cardinality of nonpopular value =
(num of rows in table) * density

The optimizer calculates density using an internal algorithm based on factors such
as the number of buckets and the NDV. Density is expressed as a decimal number
between 0 and 1. Values close to 1 indicate that the optimizer expects many rows

11-5

Chapter 11

Frequency Histograms

to be returned by a query referencing this column in its predicate list. Values close
to 0 indicate that the optimizer expects few rows to be returned.

See Also:
Oracle Database Reference to learn about the
DBA_TAB_COL_STATISTICS.DENSITY column

11.4.3 Bucket Compression
In some cases, to reduce the total number of buckets, the optimizer compresses
multiple buckets into a single bucket.
For example, the following frequency histogram indicates that the first bucket number
is 1 and the last bucket number is 23:
ENDPOINT_NUMBER ENDPOINT_VALUE
--------------- -------------1
52792
6
52793
8
52794
9
52795
10
52796
12
52797
14
52798
23
52799

Several buckets are "missing." Originally, buckets 2 through 6 each contained a single
instance of value 52793. The optimizer compressed all of these buckets into the bucket
with the highest endpoint number (bucket 6), which now contains 5 instances of value
52793. This value is popular because the difference between the endpoint number of
the current bucket (6) and the previous bucket (1) is 5. Thus, before compression the
value 52793 was the endpoint for 5 buckets.
The following annotations show which buckets are compressed, and which values are
popular:
ENDPOINT_NUMBER ENDPOINT_VALUE
--------------- -------------1
52792 -> nonpopular
6
52793 -> buckets 2-6 compressed into 6; popular
8
52794 -> buckets 7-8 compressed into 8; popular
9
52795 -> nonpopular
10
52796 -> nonpopular
12
52797 -> buckets 11-12 compressed into 12; popular
14
52798 -> buckets 13-14 compressed into 14; popular
23
52799 -> buckets 15-23 compressed into 23; popular

11.5 Frequency Histograms
In a frequency histogram, each distinct column value corresponds to a single bucket
of the histogram. Because each value has its own dedicated bucket, some buckets
may have many values, whereas others have few.

11-6

Chapter 11

Frequency Histograms

An analogy to a frequency histogram is sorting coins so that each individual coin
initially gets its own bucket. For example, the first penny is in bucket 1, the second
penny is in bucket 2, the first nickel is in bucket 3, and so on. You then consolidate all
the pennies into a single penny bucket, all the nickels into a single nickel bucket, and
so on with the remainder of the coins.
This section contains the following topics:
•

Criteria For Frequency Histograms
Frequency histograms depend on the number of requested histogram buckets.

•

Generating a Frequency Histogram
This scenario shows how to generate a frequency histogram using the sample
schemas.

11.5.1 Criteria For Frequency Histograms
Frequency histograms depend on the number of requested histogram buckets.
As shown in the logic diagram in "How Oracle Database Chooses the Histogram
Type", the database creates a frequency histogram when the following criteria are met:
•

NDV is less than or equal to n, where n is the number of histogram buckets
(default 254).
For example, the sh.countries.country_subregion_id column has 8 distinct values,
ranging sequentially from 52792 to 52799. If n is the default of 254, then the
optimizer creates a frequency histogram because 8 <= 254.

•

The estimate_percent parameter in the DBMS_STATS statistics gathering procedure is
set to either a user-specified value or to AUTO_SAMPLE_SIZE.

Starting in Oracle Database 12c, if the sampling size is the default of AUTO_SAMPLE_SIZE,
then the database creates frequency histograms from a full table scan. For all other
sampling percentage specifications, the database derives frequency histograms from a
sample. In releases earlier than Oracle Database 12c, the database gathered
histograms based on a small sample, which meant that low-frequency values often did
not appear in the sample. Using density in this case sometimes led the optimizer to
overestimate selectivity.

See Also:
Oracle Database PL/SQL Packages and Types Reference to learn about
AUTO_SAMPLE_SIZE

11.5.2 Generating a Frequency Histogram
This scenario shows how to generate a frequency histogram using the sample
schemas.
Assumptions
This scenario assumes that you want to generate a frequency histogram on the
sh.countries.country_subregion_id column. This table has 23 rows.

11-7

Chapter 11

Frequency Histograms

The following query shows that the country_subregion_id column contains 8 distinct
values (sample output included) that are unevenly distributed:
SELECT country_subregion_id, count(*)
FROM sh.countries
GROUP BY country_subregion_id
ORDER BY 1;
COUNTRY_SUBREGION_ID COUNT(*)
-------------------- ---------52792
1
52793
5
52794
2
52795
1
52796
1
52797
2
52798
2
52799
9

To generate a frequency histogram:
1.

Gather statistics for sh.countries and the country_subregion_id column, letting the
number of buckets default to 254.
For example, execute the following PL/SQL anonymous block:
BEGIN
DBMS_STATS.GATHER_TABLE_STATS (
ownname
=> 'SH'
, tabname
=> 'COUNTRIES'
, method_opt => 'FOR COLUMNS COUNTRY_SUBREGION_ID'
);
END;

2.

Query the histogram information for the country_subregion_id column.
For example, use the following query (sample output included):
SELECT
FROM
WHERE
AND

TABLE_NAME, COLUMN_NAME, NUM_DISTINCT, HISTOGRAM
USER_TAB_COL_STATISTICS
TABLE_NAME='COUNTRIES'
COLUMN_NAME='COUNTRY_SUBREGION_ID';

TABLE_NAME COLUMN_NAME
NUM_DISTINCT HISTOGRAM
---------- -------------------- ------------ --------------COUNTRIES COUNTRY_SUBREGION_ID
8 FREQUENCY

The optimizer chooses a frequency histogram because n or fewer distinct values
exist in the column, where n defaults to 254.
3.

Query the endpoint number and endpoint value for the country_subregion_id
column.
For example, use the following query (sample output included):
SELECT
FROM
WHERE
AND

ENDPOINT_NUMBER, ENDPOINT_VALUE
USER_HISTOGRAMS
TABLE_NAME='COUNTRIES'
COLUMN_NAME='COUNTRY_SUBREGION_ID';

ENDPOINT_NUMBER ENDPOINT_VALUE
--------------- -------------1
52792

11-8

Chapter 11

Frequency Histograms

6
8
9
10
12
14
23

52793
52794
52795
52796
52797
52798
52799

Figure 11-2 is a graphical illustration of the 8 buckets in the histogram. Each value
is represented as a coin that is dropped into a bucket.
Figure 11-2

Frequency Histogram
52793

52793

52792

52793 52793 52793

52794 52794

Endpoint Value

Endpoint Value

Endpoint Value

52792

52793

52794

Endpoint
Number: 1

Endpoint
Number: 6

Endpoint
Number: 8

52795

52796

52797 52797

Endpoint Value

Endpoint Value

Endpoint Value

52795

52796

52797

Endpoint
Number: 9

Endpoint
Number: 10

Endpoint
Number: 12

52799
52799 52799
52799 52799 52799
52798 52798

52799 52799 52799

Endpoint Value

Endpoint Value

52798

52799

Endpoint
Number: 14

Endpoint
Number: 23

11-9

Chapter 11

Top Frequency Histograms

As shown in Figure 11-2, each distinct value has its own bucket. Because this is a
frequency histogram, the endpoint number is the cumulative frequency of
endpoints. For 52793, the endpoint number 6 indicates that the value appears 5
times (6 - 1). For 52794, the endpoint number 8 indicates that the value appears 2
times (8 - 6).
Every bucket whose endpoint is at least 2 greater than the previous endpoint
contains a popular value. Thus, buckets 6, 8, 12, 14, and 23 contain popular values.
The optimizer calculates their cardinality based on endpoint numbers. For
example, the optimizer calculates the cardinality (C) of value 52799 using the
following formula, where the number of rows in the table is 23:
C = 23 * ( 9 / 23 )

Buckets 1, 9, and 10 contain nonpopular values. The optimizer estimates their
cardinality based on density.

See Also:
•

Oracle Database PL/SQL Packages and Types Reference to learn about
the DBMS_STATS.GATHER_TABLE_STATS procedure

•

Oracle Database Reference to learn about the USER_TAB_COL_STATISTICS
view

•

Oracle Database Reference to learn about the USER_HISTOGRAMS view

11.6 Top Frequency Histograms
A top frequency histogram is a variation on a frequency histogram that ignores
nonpopular values that are statistically insignificant.
For example, if a pile of 1000 coins contains only a single penny, then you can ignore
the penny when sorting the coins into buckets. A top frequency histogram can produce
a better histogram for highly popular values.
This section contains the following topics:
•

Criteria For Top Frequency Histograms
If a small number of values occupies most of the rows, then creating a frequency
histogram on this small set of values is useful even when the NDV is greater than
the number of requested histogram buckets. To create a better quality histogram
for popular values, the optimizer ignores the nonpopular values and creates a top
frequency histogram.

•

Generating a Top Frequency Histogram
This scenario shows how to generate a top frequency histogram using the sample
schemas.

11.6.1 Criteria For Top Frequency Histograms
If a small number of values occupies most of the rows, then creating a frequency
histogram on this small set of values is useful even when the NDV is greater than the
number of requested histogram buckets. To create a better quality histogram for

11-10

Chapter 11

Top Frequency Histograms

popular values, the optimizer ignores the nonpopular values and creates a top
frequency histogram.
As shown in the logic diagram in "How Oracle Database Chooses the Histogram
Type", the database creates a top frequency histogram when the following criteria are
met:
•

NDV is greater than n, where n is the number of histogram buckets (default 254).

•

The percentage of rows occupied by the top n frequent values is equal to or
greater than threshold p, where p is (1-(1/n))*100.

•

The estimate_percent parameter in the DBMS_STATS statistics gathering procedure is
set to AUTO_SAMPLE_SIZE.

See Also:
Oracle Database PL/SQL Packages and Types Reference to learn about
AUTO_SAMPLE_SIZE

11.6.2 Generating a Top Frequency Histogram
This scenario shows how to generate a top frequency histogram using the sample
schemas.
Assumptions
This scenario assumes that you want to generate a top frequency histogram on the
sh.countries.country_subregion_id column. This table has 23 rows.

The following query shows that the country_subregion_id column contains 8 distinct
values (sample output included) that are unevenly distributed:
SELECT country_subregion_id, count(*)
FROM sh.countries
GROUP BY country_subregion_id
ORDER BY 1;
COUNTRY_SUBREGION_ID COUNT(*)
-------------------- ---------52792
1
52793
5
52794
2
52795
1
52796
1
52797
2
52798
2
52799
9

To generate a top frequency histogram:
1.

Gather statistics for sh.countries and the country_subregion_id column, specifying
fewer buckets than distinct values.
For example, enter the following command to specify 7 buckets:

11-11

Chapter 11

Top Frequency Histograms

BEGIN
DBMS_STATS.GATHER_TABLE_STATS (
ownname
=> 'SH'
, tabname
=> 'COUNTRIES'
, method_opt => 'FOR COLUMNS COUNTRY_SUBREGION_ID SIZE 7'
);
END;
2.

Query the histogram information for the country_subregion_id column.
For example, use the following query (sample output included):
SELECT
FROM
WHERE
AND

TABLE_NAME, COLUMN_NAME, NUM_DISTINCT, HISTOGRAM
USER_TAB_COL_STATISTICS
TABLE_NAME='COUNTRIES'
COLUMN_NAME='COUNTRY_SUBREGION_ID';

TABLE_NAME COLUMN_NAME
NUM_DISTINCT HISTOGRAM
---------- -------------------- ------------ --------------COUNTRIES COUNTRY_SUBREGION_ID
7 TOP-FREQUENCY

The sh.countries.country_subregion_id column contains 8 distinct values, but the
histogram only contains 7 buckets, making n=7. In this case, the database can only
create a top frequency or hybrid histogram. In the country_subregion_id column,
the top 7 most frequent values occupy 95.6% of the rows, which exceeds the
threshold of 85.7%, generating a top frequency histogram.
3.

Query the endpoint number and endpoint value for the column.
For example, use the following query (sample output included):
SELECT
FROM
WHERE
AND

ENDPOINT_NUMBER, ENDPOINT_VALUE
USER_HISTOGRAMS
TABLE_NAME='COUNTRIES'
COLUMN_NAME='COUNTRY_SUBREGION_ID';

ENDPOINT_NUMBER ENDPOINT_VALUE
--------------- -------------1
52792
6
52793
8
52794
9
52796
11
52797
13
52798
22
52799

Figure 11-3 is a graphical illustration of the 7 buckets in the top frequency
histogram. The values are represented in the diagram as coins.

11-12

Chapter 11

Top Frequency Histograms

Figure 11-3

Top Frequency Histogram
52793

52793

52792

52793 52793 52793

52794 52794

Endpoint Value

Endpoint Value

Endpoint Value

52792

52793

52794

Endpoint
Number: 1

Endpoint
Number: 6

Endpoint
Number: 8

52796

52797 52797

Endpoint Value

Endpoint Value

52796

52797

Endpoint
Number: 9

Endpoint
Number: 11

52799
52799 52799
52799 52799 52799
52798 52798

52799 52799 52799

Endpoint Value

Endpoint Value

52798

52799

Endpoint
Number: 13

Endpoint
Number: 22

As shown in Figure 11-3, each distinct value has its own bucket except for 52795,
which is excluded from the histogram because it is nonpopular and statistically
insignificant. As in a standard frequency histogram, the endpoint number
represents the cumulative frequency of values.

11-13

Chapter 11

Height-Balanced Histograms (Legacy)

See Also:
•

"Criteria For Frequency Histograms"

•

Oracle Database PL/SQL Packages and Types Reference to learn about
the DBMS_STATS.GATHER_TABLE_STATS procedure

•

Oracle Database Reference to learn about the USER_TAB_COL_STATISTICS
view

•

Oracle Database Reference to learn about the USER_HISTOGRAMS view

11.7 Height-Balanced Histograms (Legacy)
In a legacy height-balanced histogram, column values are divided into buckets so that
each bucket contains approximately the same number of rows.
For example, if you have 99 coins to distribute among 4 buckets, each bucket contains
about 25 coins. The histogram shows where the endpoints fall in the range of values.
This section contains the following topics:
•

Criteria for Height-Balanced Histograms
Before Oracle Database 12c, the database created a height-balanced histogram
when the NDV was greater than n. This type of histogram was useful for range
predicates, and equality predicates on values that appear as endpoints in at least
two buckets.

•

Generating a Height-Balanced Histogram
This scenario shows how to generate a height-balanced histogram using the
sample schemas.

11.7.1 Criteria for Height-Balanced Histograms
Before Oracle Database 12c, the database created a height-balanced histogram when
the NDV was greater than n. This type of histogram was useful for range predicates,
and equality predicates on values that appear as endpoints in at least two buckets.
As shown in the logic diagram in "How Oracle Database Chooses the Histogram
Type", the database creates a height-balanced histogram when the following criteria
are met:
•

NDV is greater than n, where n is the number of histogram buckets (default 254).

•

The estimate_percent parameter in the DBMS_STATS statistics gathering procedure is
not set to AUTO_SAMPLE_SIZE.

It follows that if Oracle Database 12c creates new histograms, and if the sampling
percentage is AUTO_SAMPLE_SIZE, then the histograms are either top frequency or hybrid,
but not height-balanced.
If you upgrade Oracle Database 11g to Oracle Database 12c, then any height-based
histograms created before the upgrade remain in use. However, if you refresh
statistics on the table on which the histogram was created, then the database replaces
existing height-balanced histograms on this table. The type of replacement histogram
depends on both the NDV and the following criteria:

11-14

Chapter 11

Height-Balanced Histograms (Legacy)

•

If the sampling percentage is AUTO_SAMPLE_SIZE, then the database creates either
hybrid or frequency histograms.

•

If the sampling percentage is not AUTO_SAMPLE_SIZE, then the database creates
either height-balanced or frequency histograms.

11.7.2 Generating a Height-Balanced Histogram
This scenario shows how to generate a height-balanced histogram using the sample
schemas.
Assumptions
This scenario assumes that you want to generate a height-balanced histogram on the
sh.countries.country_subregion_id column. This table has 23 rows.
The following query shows that the country_subregion_id column contains 8 distinct
values (sample output included) that are unevenly distributed:
SELECT country_subregion_id, count(*)
FROM sh.countries
GROUP BY country_subregion_id
ORDER BY 1;
COUNTRY_SUBREGION_ID COUNT(*)
-------------------- ---------52792
1
52793
5
52794
2
52795
1
52796
1
52797
2
52798
2
52799
9

To generate a height-balanced histogram:
1.

Gather statistics for sh.countries and the country_subregion_id column, specifying
fewer buckets than distinct values.

Note:
To simulate Oracle Database 11g behavior, which is necessary to create
a height-based histogram, set estimate_percent to a nondefault value. If
you specify a nondefault percentage, then the database creates
frequency or height-balanced histograms.

For example, enter the following command:
BEGIN DBMS_STATS.GATHER_TABLE_STATS (
ownname
=> 'SH'
, tabname
=> 'COUNTRIES'
, method_opt
=> 'FOR COLUMNS COUNTRY_SUBREGION_ID SIZE 7'
, estimate_percent => 100
);
END;

11-15

Chapter 11

Height-Balanced Histograms (Legacy)

2.

Query the histogram information for the country_subregion_id column.
For example, use the following query (sample output included):
SELECT
FROM
WHERE
AND

TABLE_NAME, COLUMN_NAME, NUM_DISTINCT, HISTOGRAM
USER_TAB_COL_STATISTICS
TABLE_NAME='COUNTRIES'
COLUMN_NAME='COUNTRY_SUBREGION_ID';

TABLE_NAME COLUMN_NAME
NUM_DISTINCT HISTOGRAM
---------- -------------------- ------------ --------------COUNTRIES COUNTRY_SUBREGION_ID
8 HEIGHT BALANCED

The optimizer chooses a height-balanced histogram because the number of
distinct values (8) is greater than the number of buckets (7), and the
estimate_percent value is nondefault.
3.

Query the number of rows occupied by each distinct value.
For example, use the following query (sample output included):
SELECT COUNT(country_subregion_id) AS NUM_OF_ROWS, country_subregion_id
FROM countries
GROUP BY country_subregion_id
ORDER BY 2;
NUM_OF_ROWS COUNTRY_SUBREGION_ID
----------- -------------------1
52792
5
52793
2
52794
1
52795
1
52796
2
52797
2
52798
9
52799

4.

Query the endpoint number and endpoint value for the country_subregion_id
column.
For example, use the following query (sample output included):
SELECT
FROM
WHERE
AND

ENDPOINT_NUMBER, ENDPOINT_VALUE
USER_HISTOGRAMS
TABLE_NAME='COUNTRIES'
COLUMN_NAME='COUNTRY_SUBREGION_ID';

ENDPOINT_NUMBER ENDPOINT_VALUE
--------------- -------------0
52792
2
52793
3
52795
4
52798
7
52799

Figure 11-4 is a graphical illustration of the height-balanced histogram. The values
are represented in the diagram as coins.

11-16

Chapter 11

Height-Balanced Histograms (Legacy)

Figure 11-4

Height-Balanced Histogram
52793

52793

52794

52792

52793 52793 52793

52794 52795

Endpoint Value

Endpoint Value

Endpoint Value

52792

52793

52795

Endpoint
Number: 0

Endpoint
Number: 2

Endpoint
Number: 3

52799
52799 52799
52797 52797

52799 52799 52799

52798 52798

52799 52799 52799

Endpoint Value

Endpoint Value

52798

52799

Endpoint
Number: 4

Endpoint
Number: 7

The bucket number is identical to the endpoint number. The optimizer records the
value of the last row in each bucket as the endpoint value, and then checks to
ensure that the minimum value is the endpoint value of the first bucket, and the
maximum value is the endpoint value of the last bucket. In this example, the
optimizer adds bucket 0 so that the minimum value 52792 is the endpoint of a
bucket.
The optimizer must evenly distribute 23 rows into the 7 specified histogram
buckets, so each bucket contains approximately 3 rows. However, the optimizer
compresses buckets with the same endpoint. So, instead of bucket 1 containing 2
instances of value 52793, and bucket 2 containing 3 instances of value 52793, the
optimizer puts all 5 instances of value 52793 into bucket 2. Similarly, instead of
having buckets 5, 6, and 7 contain 3 values each, with the endpoint of each bucket
as 52799, the optimizer puts all 9 instances of value 52799 into bucket 7.
In this example, buckets 3 and 4 contain nonpopular values because the difference
between the current endpoint number and previous endpoint number is 1. The
optimizer calculates cardinality for these values based on density. The remaining
buckets contain popular values. The optimizer calculates cardinality for these
values based on endpoint numbers.

11-17

Chapter 11

Hybrid Histograms

See Also:
•

Oracle Database PL/SQL Packages and Types Reference to learn about
the DBMS_STATS.GATHER_TABLE_STATS procedure

•

Oracle Database Reference to learn about the USER_TAB_COL_STATISTICS
view

•

Oracle Database Reference to learn about the USER_HISTOGRAMS view

11.8 Hybrid Histograms
A hybrid histogram combines characteristics of both height-based histograms and
frequency histograms. This "best of both worlds" approach enables the optimizer to
obtain better selectivity estimates in some situations.
The height-based histogram sometimes produces inaccurate estimates for values that
are almost popular. For example, a value that occurs as an endpoint value of only one
bucket but almost occupies two buckets is not considered popular.
To solve this problem, a hybrid histogram distributes values so that no value occupies
more than one bucket, and then stores the endpoint repeat count value, which is the
number of times the endpoint value is repeated, for each endpoint (bucket) in the
histogram. By using the repeat count, the optimizer can obtain accurate estimates for
almost popular values.
This section contains the following topics:
•

How Endpoint Repeat Counts Work
The analogy of coins distributed among buckets illustrate show endpoint repeat
counts work.

•

Criteria for Hybrid Histograms
The only differentiating criterion for hybrid histograms as compared to top
frequency histograms is that the top n frequent values is less than internal
threshold p.

•

Generating a Hybrid Histogram
This scenario shows how to generate a hybrid histogram using the sample
schemas.

11.8.1 How Endpoint Repeat Counts Work
The analogy of coins distributed among buckets illustrate show endpoint repeat counts
work.
The following figure illustrates a coins column that sorts values from low to high.

Figure 11-5
1

1

1

Coins
5

5

5

10

10

25

25

25

25

50 100 100

11-18

Chapter 11

Hybrid Histograms

You gather statistics for this table, setting the method_opt argument of
DBMS_STATS.GATHER_TABLE_STATS to FOR ALL COLUMNS SIZE 3. In this case, the optimizer
initially groups the values in the coins column into three buckets, as shown in the
following figure.

Figure 11-6
1

Initial Distribution of Values

1
5

1
5

5

10
25

10
25

25

25

50

100 100

Endpoint Value

Endpoint Value

Endpoint Value

5

25

100

Bucket 1

Bucket 2

Bucket 3

If a bucket border splits a value so that some occurrences of the value are in one
bucket and some in another, then the optimizer shifts the bucket border (and all other
following bucket borders) forward to include all occurrences of the value. For example,
the optimizer shifts value 5 so that it is now wholly in the first bucket, and the value 25
is now wholly in the second bucket.

Figure 11-7
1

Redistribution of Values

1
5

1
5

5

10

10

50

25

25

100 100

25

25

Endpoint Value

Endpoint Value

Endpoint Value

5

25

100

Bucket 1

Bucket 2

Bucket 3

The endpoint repeat count measures the number of times that the corresponding
bucket endpoint, which is the value at the right bucket border, repeats itself. For
example, in the first bucket, the value 5 is repeated 3 times, so the endpoint repeat
count is 3.

11-19

Chapter 11

Hybrid Histograms

Figure 11-8
1

Endpoint Repeat Count

1
5

1
5

5

10

10

50

25

25

100 100

25

25

Endpoint Value

Endpoint Value

Endpoint Value

5

25

100

Repeat Count: 3

Repeat Count: 4

Repeat Count: 2

Bucket 1

Bucket 2

Bucket 3

Height-balanced histograms do not store as much information as hybrid histograms.
By using repeat counts, the optimizer can determine exactly how many occurrences of
an endpoint value exist. For example, the optimizer knows that the value 5 appears 3
times, the value 25 appears 4 times, and the value 100 appears 2 times. This frequency
information helps the optimizer to generate better cardinality estimates.

11.8.2 Criteria for Hybrid Histograms
The only differentiating criterion for hybrid histograms as compared to top frequency
histograms is that the top n frequent values is less than internal threshold p.
As shown in the logic diagram in "How Oracle Database Chooses the Histogram
Type", the database creates a hybrid histogram when the following criteria are met:
•

NDV is greater than n, where n is the number of histogram buckets (default is 254).

•

The criteria for top frequency histograms do not apply.
This is another way to stating that the percentage of rows occupied by the top n
frequent values is less than threshold p, where p is (1-(1/n))*100.

•

The estimate_percent parameter in the DBMS_STATS statistics gathering procedure is
set to AUTO_SAMPLE_SIZE.
If users specify their own percentage, then the database creates frequency or
height-balanced histograms.

See Also:
•

"Criteria For Top Frequency Histograms."

•

"Height-Balanced Histograms (Legacy)"

•

Oracle Database PL/SQL Packages and Types Reference to learn more
about the estimate_percent parameter

11-20

Chapter 11

Hybrid Histograms

11.8.3 Generating a Hybrid Histogram
This scenario shows how to generate a hybrid histogram using the sample schemas.
Assumptions
This scenario assumes that you want to generate a hybrid histogram on the
sh.products.prod_subcategory_id column. This table has 72 rows. The
prod_subcategory_id column contains 22 distinct values.
To generate a hybrid histogram:
1.

Gather statistics for sh.products and the prod_subcategory_id column, specifying
10 buckets.
For example, enter the following command:
BEGIN DBMS_STATS.GATHER_TABLE_STATS (
ownname
=> 'SH'
, tabname
=> 'PRODUCTS'
, method_opt => 'FOR COLUMNS PROD_SUBCATEGORY_ID SIZE 10'
);
END;

2.

Query the number of rows occupied by each distinct value.
For example, use the following query (sample output included):
SELECT COUNT(prod_subcategory_id) AS NUM_OF_ROWS, prod_subcategory_id
FROM products
GROUP BY prod_subcategory_id
ORDER BY 1 DESC;
NUM_OF_ROWS PROD_SUBCATEGORY_ID
----------- ------------------8
2014
7
2055
6
2032
6
2054
5
2056
5
2031
5
2042
5
2051
4
2036
3
2043
2
2033
2
2034
2
2013
2
2012
2
2053
2
2035
1
2022
1
2041
1
2044
1
2011
1
2021
1
2052
22 rows selected.

11-21

Chapter 11

Hybrid Histograms

The column contains 22 distinct values. Because the number of buckets (10) is
less than 22, the optimizer cannot create a frequency histogram. The optimizer
considers both hybrid and top frequency histograms. To qualify for a top frequency
histogram, the percentage of rows occupied by the top 10 most frequent values
must be equal to or greater than threshold p, where p is (1-(1/10))*100, or 90%.
However, in this case the top 10 most frequent values occupy 54 rows out of 72,
which is only 75% of the total. Therefore, the optimizer chooses a hybrid histogram
because the criteria for a top frequency histogram do not apply.
3.

Query the histogram information for the country_subregion_id column.
For example, use the following query (sample output included):
SELECT
FROM
WHERE
AND

TABLE_NAME, COLUMN_NAME, NUM_DISTINCT, HISTOGRAM
USER_TAB_COL_STATISTICS
TABLE_NAME='PRODUCTS'
COLUMN_NAME='PROD_SUBCATEGORY_ID';

TABLE_NAME COLUMN_NAME
NUM_DISTINCT HISTOGRAM
---------- ------------------- ------------ --------PRODUCTS PROD_SUBCATEGORY_ID 22
HYBRID
4.

Query the endpoint number, endpoint value, and endpoint repeat count for the
country_subregion_id column.
For example, use the following query (sample output included):
SELECT ENDPOINT_NUMBER, ENDPOINT_VALUE, ENDPOINT_REPEAT_COUNT
FROM USER_HISTOGRAMS
WHERE TABLE_NAME='PRODUCTS'
AND
COLUMN_NAME='PROD_SUBCATEGORY_ID'
ORDER BY 1;
ENDPOINT_NUMBER ENDPOINT_VALUE ENDPOINT_REPEAT_COUNT
--------------- -------------- --------------------1
2011
1
13
2014
8
26
2032
6
36
2036
4
45
2043
3
51
2051
5
52
2052
1
54
2053
2
60
2054
6
72
2056
5
10 rows selected.

In a height-based histogram, the optimizer would evenly distribute 72 rows into the
10 specified histogram buckets, so that each bucket contains approximately 7
rows. Because this is a hybrid histogram, the optimizer distributes the values so
that no value occupies more than one bucket. For example, the optimizer does not
put some instances of value 2036 into one bucket and some instances of this value
into another bucket: all instances are in bucket 36.
The endpoint repeat count shows the number of times the highest value in the
bucket is repeated. By using the endpoint number and repeat count for these
values, the optimizer can estimate cardinality. For example, bucket 36 contains
instances of values 2033, 2034, 2035, and 2036. The endpoint value 2036 has an
endpoint repeat count of 4, so the optimizer knows that 4 instances of this value

11-22

Chapter 11

Hybrid Histograms

exist. For values such as 2033, which are not endpoints, the optimizer estimates
cardinality using density.

See Also:
•

Oracle Database PL/SQL Packages and Types Reference to learn about
the DBMS_STATS.GATHER_TABLE_STATS procedure

•

Oracle Database Reference to learn about the USER_TAB_COL_STATISTICS
view

•

Oracle Database Reference to learn about the USER_HISTOGRAMS view

11-23

12
Configuring Options for Optimizer Statistics
Gathering
This chapter explains what optimizer statistics collection is and how to set statistics
preferences.
This chapter contains the following topics:
•

About Optimizer Statistics Collection
In Oracle Database, optimizer statistics collection is the gathering of optimizer
statistics for database objects, including fixed objects.

•

Setting Optimizer Statistics Preferences
This topic explains how to set optimizer statistics defaults using
DBMS_STATS.SET_*_PREFS procedures.

•

Configuring Options for Dynamic Statistics
Dynamic statistics are an optimization technique in which the database uses
recursive SQL to scan a small random sample of the blocks in a table.

•

Managing SQL Plan Directives
A SQL plan directive is additional information and instructions that the optimizer
can use to generate a more optimal plan.

12.1 About Optimizer Statistics Collection
In Oracle Database, optimizer statistics collection is the gathering of optimizer
statistics for database objects, including fixed objects.
The database can collect optimizer statistics automatically. You can also collect them
manually using the DBMS_STATS package.
This section contains the following topics:
•

Purpose of Optimizer Statistics Collection
The contents of tables and associated indexes change frequently, which can lead
the optimizer to choose suboptimal execution plan for queries. To avoid potential
performance issues, statistics must be kept current.

•

User Interfaces for Optimizer Statistics Management
You can manage optimizer statistics either through Oracle Enterprise Manager
Cloud Control (Cloud Control) or using PL/SQL on the command line.

Related Topics
•

Gathering Optimizer Statistics Manually
As an alternative or supplement to automatic statistics gathering, you can use the
DBMS_STATS package to gather statistics manually.

12-1

Chapter 12

About Optimizer Statistics Collection

12.1.1 Purpose of Optimizer Statistics Collection
The contents of tables and associated indexes change frequently, which can lead the
optimizer to choose suboptimal execution plan for queries. To avoid potential
performance issues, statistics must be kept current.
To minimize DBA involvement, Oracle Database automatically gathers optimizer
statistics at various times. Some automatic options are configurable, such enabling
AutoTask to run DBMS_STATS.

12.1.2 User Interfaces for Optimizer Statistics Management
You can manage optimizer statistics either through Oracle Enterprise Manager Cloud
Control (Cloud Control) or using PL/SQL on the command line.
This section contains the following topics:
•

Graphical Interface for Optimizer Statistics Management
The Manage Optimizer Statistics page in Cloud Control is a GUI that enables you
to manage optimizer statistics.

•

Command-Line Interface for Optimizer Statistics Management
The DBMS_STATS package performs most optimizer statistics tasks.

12.1.2.1 Graphical Interface for Optimizer Statistics Management
The Manage Optimizer Statistics page in Cloud Control is a GUI that enables you to
manage optimizer statistics.
This section contains the following topics:
•

Accessing the Database Home Page in Cloud Control
Oracle Enterprise Manager Cloud Control enables you to manage multiple
databases within a single GUI-based framework.

•

Accessing the Optimizer Statistics Console
You can perform most necessary tasks relating to optimizer statistics through
pages linked to by the Optimizer Statistics Console page.

12.1.2.1.1 Accessing the Database Home Page in Cloud Control
Oracle Enterprise Manager Cloud Control enables you to manage multiple databases
within a single GUI-based framework.
To access a database home page using Cloud Control:
1.

Log in to Cloud Control with the appropriate credentials.

2.

Under the Targets menu, select Databases.

3.

In the list of database targets, select the target for the Oracle Database instance
that you want to administer.

4.

If prompted for database credentials, then enter the minimum credentials
necessary for the tasks you intend to perform.

12-2

Chapter 12

Setting Optimizer Statistics Preferences

See Also:
Cloud Control online help

12.1.2.1.2 Accessing the Optimizer Statistics Console
You can perform most necessary tasks relating to optimizer statistics through pages
linked to by the Optimizer Statistics Console page.
To manage optimizer statistics using Cloud Control:
1.

In Cloud Control, access the Database Home page.

2.

From the Performance menu, select SQL, then Optimizer Statistics.
The Optimizer Statistics Console appears.

See Also:
Online Help for Oracle Enterprise Manager Cloud Control

12.1.2.2 Command-Line Interface for Optimizer Statistics Management
The DBMS_STATS package performs most optimizer statistics tasks.
To enable and disable automatic statistics gathering, use the DBMS_AUTO_TASK_ADMIN
PL/SQL package.

See Also:
Oracle Database PL/SQL Packages and Types Reference to learn how to
use DBMS_STATS and DBMS_AUTO_TASK_ADMIN

12.2 Setting Optimizer Statistics Preferences
This topic explains how to set optimizer statistics defaults using
DBMS_STATS.SET_*_PREFS procedures.

This section contains the following topics:
•

About Optimizer Statistics Preferences
The optimizer statistics preferences set the default values of the parameters
used by automatic statistics collection and the DBMS_STATS statistics gathering
procedures.

•

Setting Global Optimizer Statistics Preferences Using Cloud Control
A global preference applies to any object in the database that does not have an
existing table preference. You can set optimizer statistics preferences at the global
level using Cloud Control.

12-3

Chapter 12

Setting Optimizer Statistics Preferences

•

Setting Object-Level Optimizer Statistics Preferences Using Cloud Control
You can set optimizer statistics preferences at the database, schema, and table
level using Cloud Control.

•

Setting Optimizer Statistics Preferences from the Command Line
If you do not use Cloud Control to set optimizer statistics preferences, then you
can invoke the DBMS_STATS procedures from the command line.

12.2.1 About Optimizer Statistics Preferences
The optimizer statistics preferences set the default values of the parameters used
by automatic statistics collection and the DBMS_STATS statistics gathering procedures.
You can set optimizer statistics preferences at the table, schema, database (all
tables), and global level. A global preference refers to tables with no preferences and
any tables created in the future. The procedure names follow the form SET_*_PREFS.
This section contains the following topics:
•

Purpose of Optimizer Statistics Preferences
Preferences enable you to maintain optimizer statistics automatically when some
objects require settings that differ from the default.

•

DBMS_STATS Procedures for Setting Statistics Preferences
The DBMS_STATS.SET_*_PREFS procedures change the defaults of parameters used
by the DBMS_STATS.GATHER_*_STATS procedures. To query the current preferences,
use the DBMS_STATS.GET_PREFS function.

•

Statistics Preference Overrides
The preference_overrides_parameter statistics preference determines whether,
when gathering optimizer statistics, to override the input value of a parameter with
the statistics preference. In this way, you control when the database honors a
parameter value passed to the statistics gathering procedures.

•

Setting Statistics Preferences: Example
This example illustrates the relationship between SET_TABLE_PREFS,
SET_SCHEMA_STATS, and SET_DATABASE_PREFS.

12.2.1.1 Purpose of Optimizer Statistics Preferences
Preferences enable you to maintain optimizer statistics automatically when some
objects require settings that differ from the default.
Preferences give you more granular control over how Oracle Database gathers
statistics. You can set optimizer statistics preferences at the following levels:
•

Table

•

Schema

•

Database (all tables)

•

Global (tables with no preferences and any tables created in the future)

The DBMS_STATS procedures for setting preference have names of the form SET_*_PREFS.

12-4

Chapter 12

Setting Optimizer Statistics Preferences

12.2.1.2 DBMS_STATS Procedures for Setting Statistics Preferences
The DBMS_STATS.SET_*_PREFS procedures change the defaults of parameters used by
the DBMS_STATS.GATHER_*_STATS procedures. To query the current preferences, use the
DBMS_STATS.GET_PREFS function.
When setting statistics preferences, the order of precedence is:
1.

Table preference (set for a specific table, all tables in a schema, or all tables in the
database)

2.

Global preference

3.

Default preference

The following table summarizes the relevant DBMS_STATS procedures.
Table 12-1 DBMS_STATS Procedures for Setting Optimizer Statistics
Preferences
Procedure

Scope

SET_TABLE_PREFS

Specified table only.

SET_SCHEMA_PREFS

All existing tables in the specified schema.
This procedure calls SET_TABLE_PREFS for each table in the
specified schema. Calling SET_SCHEMA_PREFS does not affect
any new tables created after it has been run. New tables use the
GLOBAL_PREF values for all parameters.

SET_DATABASE_PREFS

All user-defined schemas in the database. You can include
system-owned schemas such as SYS and SYSTEM by setting the
ADD_SYS parameter to true.
This procedure calls SET_TABLE_PREFS for each table in the
specified schema. Calling SET_DATABASE_PREFS does not affect
any new objects created after it has been run. New objects use
the GLOBAL_PREF values for all parameters.

SET_GLOBAL_PREFS

Any table that does not have an existing table preference.
All parameters default to the global setting unless a table
preference is set or the parameter is explicitly set in the
DBMS_STATS.GATHER_*_STATS statement. Changes made by
SET_GLOBAL_PREFS affect any new objects created after it runs.
New objects use the SET_GLOBAL_PREFS values for all
parameters.
With SET_GLOBAL_PREFS, you can set a default value for the
parameter AUTOSTATS_TARGET. This additional parameter
controls which objects the automatic statistic gathering job
running in the nightly maintenance window affects. Possible
values for AUTOSTATS_TARGET are ALL, ORACLE, and AUTO
(default).
You can only set the CONCURRENT preference at the global level.
You cannot set the preference INCREMENTAL_LEVEL using
SET_GLOBAL_PREFS.

12-5

Chapter 12

Setting Optimizer Statistics Preferences

See Also:
•

"About Concurrent Statistics Gathering"

•

Oracle Database PL/SQL Packages and Types Reference to learn about
the DBMS_STATS procedures for setting optimizer statistics

12.2.1.3 Statistics Preference Overrides
The preference_overrides_parameter statistics preference determines whether, when
gathering optimizer statistics, to override the input value of a parameter with the
statistics preference. In this way, you control when the database honors a parameter
value passed to the statistics gathering procedures.
When preference_overrides_parameter is set to FALSE (default), the input values for
statistics gathering procedures are honored. When set to TRUE, the input values are
ignored.
Set the preference_overrides_parameter preference using the SET_TABLE_PREFS,
SET_SCHEMA_PREFS, or SET_GLOBAL_PREFS procedures in DBMS_STATS. Regardless of
whether preference_overrides_parameter is set, the database uses the same order of
precedence for setting statistics:
1.

Table preference (set for a specific table, all tables in a schema, or all tables in the
database)

2.

Global preference

3.

Default preference

Example 12-1

Overriding Statistics Preferences at the Table Level

In this example, legacy scripts set estimate_percent explicitly rather than using the
recommended AUTO_SAMPLE_SIZE. Your goal is to prevent users from using these scripts
to set preferences on the sh.costs table.
Table 12-2

Overriding Statistics Preferences at the Table Level

Action

Description

SQL> SELECT DBMS_STATS.GET_PREFS
('estimate_percent', 'sh','costs')
AS "STAT_PREFS" FROM DUAL;

No preference for estimate_percent is set for
sh.costs or at the global level, so the preference
defaults to AUTO_SAMPLE_SIZE.

STAT_PREFS
---------DBMS_STATS.AUTO_SAMPLE_SIZE
SQL> EXEC DBMS_STATS.SET_TABLE_PREFS
('sh', 'costs',
'preference_overrides_parameter', 'true');
PL/SQL procedure successfully completed.

By default, Oracle Database accepts preferences that
are passed to the GATHER_*_STATS procedures. To
override these parameters, you use SET_TABLE_PREFS
to set the preference_overrides_parameter
preference to true for the costs table only.

12-6

Chapter 12

Setting Optimizer Statistics Preferences

Table 12-2

(Cont.) Overriding Statistics Preferences at the Table Level

Action

Description

SQL> EXEC DBMS_STATS.GATHER_TABLE_STATS
('sh', 'costs', estimate_percent=>100);
PL/SQL procedure successfully completed.

Example 12-2

You attempt to set estimate_percent to 100 when
gathering statistics for sh.costs. However, because
preference_overrides_parameter is true for this
table, Oracle Database gathers statistics using
AUTO_SAMPLE_SIZE, which is the default.

Overriding Statistics Preferences at the Global Level

In this example, you set estimate_percent to 5 at the global level, which mean this
preference applies to every table in the database that does not have a table
preference set. You then set an override on the sh.sales table, which does not have a
table-level preference set, to prevent users from overriding the global setting in their
scripts.
Table 12-3

Overriding Statistics Preferences at the Global Level

Action

Description

SQL> EXEC DBMS_STATS.SET_GLOBAL_PREFS
('estimate_percent', '5');
PL/SQL procedure successfully completed.

SQL> EXEC DBMS_STATS.SET_TABLE_PREFS
('sh', 'sales',
'preference_overrides_parameter', 'true');

You use the SET_GLOBAL_PREFS procedure to set the
estimate_percent preference to 5 for every table in the
database that does not have a table preference set.
Because sh.costs does not have a preference set, the
global setting applies to this table.
You use SET_TABLE_PREFS to set the
preference_overrides_parameter preference to true
for the sh.sales table only.

PL/SQL procedure successfully completed.
SQL> EXEC DBMS_STATS.GATHER_TABLE_STATS
('sh', 'costs', estimate_percent=>10);
PL/SQL procedure successfully completed.

You attempt to set estimate_percent to 10 when
gathering statistics for sh.sales. However, because
preference_overrides_parameter is true for this
table, and because a global preference is defined,
Oracle Database gathers statistics using the global
setting of 5.

See Also:
Oracle Database PL/SQL Packages and Types Reference to learn about the
DBMS_STATS procedures for setting optimizer statistics

12.2.1.4 Setting Statistics Preferences: Example
This example illustrates the relationship between SET_TABLE_PREFS, SET_SCHEMA_STATS,
and SET_DATABASE_PREFS.

12-7

Chapter 12

Setting Optimizer Statistics Preferences

Table 12-4

Changing Preferences for Statistics Gathering Procedures

Action
SQL> SELECT DBMS_STATS.GET_PREFS
('incremental', 'sh','costs')
AS "STAT_PREFS" FROM DUAL;

Description
You query the INCREMENTAL preference for costs and
determine that it is set to true.

STAT_PREFS
---------TRUE
SQL> EXEC DBMS_STATS.SET_TABLE_PREFS
('sh', 'costs', 'incremental', 'false');

You use SET_TABLE_PREFS to set the INCREMENTAL
preference to false for the costs table only.

PL/SQL procedure successfully completed.
SQL> SELECT DBMS_STATS.GET_PREFS
('incremental', 'sh', 'costs')
AS "STAT_PREFS" FROM DUAL;

You query the INCREMENTAL preference for costs and
confirm that it is set to false.

STAT_PREFS
---------FALSE
SQL> EXEC DBMS_STATS.SET_SCHEMA_PREFS
('sh', 'incremental', 'true');

You use SET_SCHEMA_PREFS to set the INCREMENTAL
preference to true for every table in the sh schema,
including costs.

PL/SQL procedure successfully completed.
SQL> SELECT DBMS_STATS.GET_PREFS
('incremental', 'sh', 'costs')
AS "STAT_PREFS" FROM DUAL;

You query the INCREMENTAL preference for costs and
confirm that it is set to true.

STAT_PREFS
---------TRUE
SQL> EXEC DBMS_STATS.SET_DATABASE_PREFS
('incremental', 'false');

You use SET_DATABASE_PREFS to set the INCREMENTAL
preference for all tables in all user-defined schemas to
false.

PL/SQL procedure successfully completed.
SQL> SELECT DBMS_STATS.GET_PREFS
('incremental', 'sh', 'costs')
AS "STAT_PREFS" FROM DUAL;

You query the INCREMENTAL preference for costs and
confirm that it is set to false.

STAT_PREFS
---------FALSE

12-8

Chapter 12

Setting Optimizer Statistics Preferences

12.2.2 Setting Global Optimizer Statistics Preferences Using Cloud
Control
A global preference applies to any object in the database that does not have an
existing table preference. You can set optimizer statistics preferences at the global
level using Cloud Control.
To set global optimizer statistics preferences using Cloud Control:
1.

In Cloud Control, access the Database Home page.

2.

From the Performance menu, select SQL, then Optimizer Statistics.
The Optimizer Statistics Console appears.

3.

Click Global Statistics Gathering Options.
The Global Statistics Gathering Options page appears.

4.

Make your desired changes, and click Apply.

See Also:
Online Help for Oracle Enterprise Manager Cloud Control

12.2.3 Setting Object-Level Optimizer Statistics Preferences Using
Cloud Control
You can set optimizer statistics preferences at the database, schema, and table level
using Cloud Control.
To set object-level optimizer statistics preferences using Cloud Control:
1.

In Cloud Control, access the Database Home page.

2.

From the Performance menu, select SQL, then Optimizer Statistics.
The Optimizer Statistics Console appears.

3.

Click Object Level Statistics Gathering Preferences.
The Object Level Statistics Gathering Preferences page appears.

4.

To modify table preferences for a table that has preferences set at the table level,
do the following (otherwise, skip to the next step):
a.

Enter values in Schema and Table Name. Leave Table Name blank to see all
tables in the schema.
The page refreshes with the table names.

b.

Select the desired tables and click Edit Preferences.
The General subpage of the Edit Preferences page appears.

c.

Change preferences as needed and click Apply.

12-9

Chapter 12

Setting Optimizer Statistics Preferences

5.

To set preferences for a table that does not have preferences set at the table level,
do the following (otherwise, skip to the next step):
a.

Click Add Table Preferences.
The General subpage of the Add Table Preferences page appears.

6.

b.

In Table Name, enter the schema and table name.

c.

Change preferences as needed and click OK.

To set preferences for a schema, do the following:
a.

Click Set Schema Tables Preferences.
The General subpage of the Edit Schema Preferences page appears.

b.

In Schema, enter the schema name.

c.

Change preferences as needed and click OK.

See Also:
Online Help for Oracle Enterprise Manager Cloud Control

12.2.4 Setting Optimizer Statistics Preferences from the Command
Line
If you do not use Cloud Control to set optimizer statistics preferences, then you can
invoke the DBMS_STATS procedures from the command line.
Prerequisites
This task has the following prerequisites:
•

To set the global or database preferences, you must have SYSDBA privileges, or
both ANALYZE ANY DICTIONARY and ANALYZE ANY system privileges.

•

To set schema preferences, you must connect as owner, or have SYSDBA privileges,
or have the ANALYZE ANY system privilege.

•

To set table preferences, you must connect as owner of the table or have the
ANALYZE ANY system privilege.

To set optimizer statistics preferences from the command line:
1.

In SQL*Plus or SQL Developer, log in to the database as a user with the
necessary privileges.

2.

Optionally, call the DBMS_STATS.GET_PREFS procedure to see preferences set at the
object level, or at the global level if a specific table is not set.
For example, obtain the STALE_PERCENT parameter setting for the sh.sales table as
follows:
SELECT DBMS_STATS.GET_PREFS('STALE_PERCENT', 'SH', 'SALES')
FROM DUAL;

3.

Execute the appropriate procedure from Table 12-1, specifying the following
parameters:

12-10

Chapter 12

Configuring Options for Dynamic Statistics

•

ownname - Set schema name (SET_TAB_PREFS and SET_SCHEMA_PREFS only)

•

tabname - Set table name (SET_TAB_PREFS only)

•

pname - Set parameter name

•

pvalue - Set parameter value

•

add_sys - Include system tables (optional, SET_DATABASE_PREFS only)

The following example specifies that 13% of rows in sh.sales must change before
the statistics on that table are considered stale:
EXEC DBMS_STATS.SET_TABLE_PREFS('SH', 'SALES', 'STALE_PERCENT', '13');
4.

Optionally, query the *_TAB_STAT_PREFS view to confirm the change.
For example, query DBA_TAB_STAT_PREFS as follows:
COL OWNER FORMAT a5
COL TABLE_NAME FORMAT a15
COL PREFERENCE_NAME FORMAT a20
COL PREFERENCE_VALUE FORMAT a30
SELECT * FROM DBA_TAB_STAT_PREFS;

Sample output appears as follows:
OWNER
----OE
SH

TABLE_NAME
--------------CUSTOMERS
SALES

PREFERENCE_NAME
-------------------NO_INVALIDATE
STALE_PERCENT

PREFERENCE_VALUE
-----------------------------DBMS_STATS.AUTO_INVALIDATE
13

See Also:
Oracle Database PL/SQL Packages and Types Reference for descriptions of
the parameter names and values for program units

12.3 Configuring Options for Dynamic Statistics
Dynamic statistics are an optimization technique in which the database uses
recursive SQL to scan a small random sample of the blocks in a table.
The sample scan estimate predicate selectivities. Using these estimates, the database
determines better default statistics for unanalyzed segments, and verifies its
estimates. By default, when optimizer statistics are missing, stale, or insufficient,
dynamic statistics automatically run recursive SQL during parsing to scan a small
random sample of table blocks.
This section contains the following topics:
•

About Dynamic Statistics Levels
The dynamic statistics level controls both when the database gathers dynamic
statistics, and the size of the sample that the optimizer uses to gather the
statistics.

•

Setting Dynamic Statistics Levels Manually
Determining a database-level setting that would be beneficial to all SQL
statements can be difficult.

12-11

Chapter 12

Configuring Options for Dynamic Statistics

•

Disabling Dynamic Statistics
In general, the best practice is not to incur the cost of dynamic statistics for queries
whose compile times must be as fast as possible, for example, unrepeated OLTP
queries.

Related Topics
•

Supplemental Dynamic Statistics
By default, when optimizer statistics are missing, stale, or insufficient, the
database automatically gathers dynamic statistics during a parse. The database
uses recursive SQL to scan a small random sample of table blocks.

12.3.1 About Dynamic Statistics Levels
The dynamic statistics level controls both when the database gathers dynamic
statistics, and the size of the sample that the optimizer uses to gather the statistics.
Set the dynamic statistics level using either the OPTIMIZER_DYNAMIC_SAMPLING
initialization parameter or a statement hint.

Note:
Dynamic statistics were called dynamic sampling in releases earlier than
Oracle Database 12c Release 1 (12.1).

The following table describes the levels for dynamic statistics. Note the following:

Table 12-5

•

If OPTIMIZER_DYNAMIC_STATISTICS is TRUE, and if dynamic statistics are not disabled,
then the database may choose to use dynamic statistics when a SQL statement
uses parallel execution.

•

If OPTIMIZER_ADAPTIVE_STATISTICS is TRUE, then the optimizer uses dynamic statistics
when relevant SQL plan directives exist. The database maintains the resulting
statistics in the SQL plan directives store, making them available to other queries.

Dynamic Statistics Levels

Level

When the Optimizer Uses Dynamic Statistics

Sample Size (Blocks)

0

Do not use dynamic statistics.

n/a

1

Use dynamic statistics for all tables that do not have statistics, but
only if the following criteria are met:

32

•
•
•
2

At least one nonpartitioned table in the query does not have
statistics.
This table has no indexes.
This table has more blocks than the number of blocks that
would be used for dynamic statistics of this table.

Use dynamic statistics if at least one table in the statement has no
statistics. This is the default value.

64

12-12

Chapter 12

Configuring Options for Dynamic Statistics

Table 12-5

(Cont.) Dynamic Statistics Levels

Level

When the Optimizer Uses Dynamic Statistics

Sample Size (Blocks)

3

Use dynamic statistics if any of the following conditions is true:

64

•
•

4

At least one table in the statement has no statistics.
The statement has one or more expressions used in the WHERE
clause predicates, for example, WHERE SUBSTR(CUSTLASTNAME,
1,3).

Use dynamic statistics if any of the following conditions is true:
•
•

•

64

At least one table in the statement has no statistics.
The statement has one or more expressions used in the WHERE
clause predicates, for example, WHERE SUBSTR(CUSTLASTNAME,
1,3).
The statement uses complex predicates (an OR or AND operator
between multiple predicates on the same table).

5

The criteria are identical to level 4, but the database uses a
different sample size.

128

6

The criteria are identical to level 4, but the database uses a
different sample size.

256

7

The criteria are identical to level 4, but the database uses a
different sample size.

512

8

The criteria are identical to level 4, but the database uses a
different sample size.

1024

9

The criteria are identical to level 4, but the database uses a
different sample size.

4086

10

The criteria are identical to level 4, but the database uses a
different sample size.

All blocks

11

The database uses adaptive dynamic sampling automatically when Automatically determined
the optimizer deems it necessary.

See Also:
•
•

"When the Database Samples Data"
Oracle Database Reference to learn about the
OPTIMIZER_DYNAMIC_SAMPLING initialization parameter

12.3.2 Setting Dynamic Statistics Levels Manually
Determining a database-level setting that would be beneficial to all SQL statements
can be difficult.
When setting the level for dynamic statistics, Oracle recommends setting the
OPTIMIZER_DYNAMIC_SAMPLING initialization parameter at the session level.
Assumptions
This tutorial assumes the following:

12-13

Chapter 12

Configuring Options for Dynamic Statistics

•

You want correct selectivity estimates for the following query, which has WHERE
clause predicates on two correlated columns:
SELECT
FROM
WHERE
AND

*
sh.customers
cust_city='Los Angeles'
cust_state_province='CA';

•

The preceding query uses serial processing.

•

The sh.customers table contains 932 rows that meet the conditions in the query.

•

You have gathered statistics on the sh.customers table.

•

You created an index on the cust_city and cust_state_province columns.

•

The OPTIMIZER_DYNAMIC_SAMPLING initialization parameter is set to the default level of
2.

To set the dynamic statistics level manually:
1.

In SQL*Plus or SQL Developer, log in to the database as a user with the
necessary privileges.

2.

Explain the execution plan as follows:
EXPLAIN PLAN FOR
SELECT *
FROM sh.customers
WHERE cust_city='Los Angeles'
AND
cust_state_province='CA';

3.

Query the plan as follows:
SET LINESIZE 130
SET PAGESIZE 0
SELECT *
FROM TABLE(DBMS_XPLAN.DISPLAY);

The output appears below (the example has been reformatted to fit on the page):
------------------------------------------------------------------------------|Id| Operation
| Name
|Rows|Bytes|Cost | Time |
------------------------------------------------------------------------------| 0| SELECT STATEMENT
|
| 53| 9593|53(0)|00:00:01|
| 1| TABLE ACCESS BY INDEX ROWID|CUSTOMERS
| 53| 9593|53(0)|00:00:01|
|*2| INDEX RANGE SCAN
|CUST_CITY_STATE_IND| 53| 9593| 3(0)|00:00:01|
------------------------------------------------------------------------------Predicate Information (identified by operation id):
--------------------------------------------------2 - access("CUST_CITY"='Los Angeles' AND "CUST_STATE_PROVINCE"='CA')

The columns in the WHERE clause have a real-world correlation, but the optimizer is
not aware that Los Angeles is in California and assumes both predicates reduce
the number of rows returned. Thus, the table contains 932 rows that meet the
conditions, but the optimizer estimates 53, as shown in bold.
If the database had used dynamic statistics for this plan, then the Note section of
the plan output would have indicated this fact. The optimizer did not use dynamic
statistics because the statement executed serially, standard statistics exist, and
the parameter OPTIMIZER_DYNAMIC_SAMPLING is set to the default of 2.

12-14

Chapter 12

Configuring Options for Dynamic Statistics

4.

Set the dynamic statistics level to 4 in the session using the following statement:
ALTER SESSION SET OPTIMIZER_DYNAMIC_SAMPLING=4;

5.

Explain the plan again:
EXPLAIN PLAN FOR
SELECT *
FROM sh.customers
WHERE cust_city='Los Angeles'
AND
cust_state_province='CA';

The new plan shows a more accurate estimate of the number of rows, as shown
by the value 932 in bold:
PLAN_TABLE_OUTPUT
--------------------------------------------------------------------------Plan hash value: 2008213504
--------------------------------------------------------------------------| Id | Operation
| Name
|Rows | Bytes |Cost (%CPU)|Time
|
--------------------------------------------------------------------------| 0 | SELECT STATEMENT |
| 932 | 271K| 406 (1)| 00:00:05 |
|* 1 | TABLE ACCESS FULL| CUSTOMERS | 932 | 271K| 406 (1)| 00:00:05 |
--------------------------------------------------------------------------Predicate Information (identified by operation id):
--------------------------------------------------1 - filter("CUST_CITY"='Los Angeles' AND "CUST_STATE_PROVINCE"='CA')
Note
----- dynamic statistics used for this statement (level=4)

The note at the bottom of the plan indicates that the sampling level is 4. The
additional dynamic statistics made the optimizer aware of the real-world
relationship between the cust_city and cust_state_province columns, thereby
enabling it to produce a more accurate estimate for the number of rows: 932 rather
than 53.

See Also:
•
•

Oracle Database SQL Language Reference to learn about setting
sampling levels with the DYNAMIC_SAMPLING hint
Oracle Database Reference to learn about the
OPTIMIZER_DYNAMIC_SAMPLING initialization parameter

12-15

Chapter 12

Managing SQL Plan Directives

12.3.3 Disabling Dynamic Statistics
In general, the best practice is not to incur the cost of dynamic statistics for queries
whose compile times must be as fast as possible, for example, unrepeated OLTP
queries.
To disable dynamic statistics at the session level:
1.

Connect SQL*Plus to the database with the appropriate privileges.

2.

Set the dynamic statistics level to 0.
For example, run the following statement:
ALTER SESSION SET OPTIMIZER_DYNAMIC_SAMPLING=0;

See Also:
Oracle Database Reference to learn about the OPTIMIZER_DYNAMIC_SAMPLING
initialization parameter

12.4 Managing SQL Plan Directives
A SQL plan directive is additional information and instructions that the optimizer can
use to generate a more optimal plan.
A directive informs the database that the optimizer is misestimate cardinalities of
certain types of predicates, and alerts DBMS_STATS to gather additional statistics in the
future. Thus, directives have an effect on statistics gathering.
The database automatically creates and manages SQL plan directives in the SGA,
and then periodically writes them to the data dictionary. If the directives are not used
within 53 weeks, then the database automatically purges them.
You can use DBMS_SPD procedures and functions to alter, save, drop, and transport
directives manually. The following table lists some of the more commonly used
procedures and functions.
Table 12-6

DBMS_SPD Procedures

Procedure

Description

FLUSH_SQL_PLAN_DIRECTIVE

Forces the database to write directives from memory to
persistent storage in the SYSAUX tablespace.

12-16

Chapter 12

Managing SQL Plan Directives

Table 12-6

(Cont.) DBMS_SPD Procedures

Procedure

Description

DROP_SQL_PLAN_DIRECTIVE

Drops a SQL plan directive. If a directive that triggers
dynamic sampling is creating unacceptable performance
overhead, then you may want to remove it manually.
If a SQL plan directive is dropped manually or
automatically, then the database can re-create it. To
prevent its re-creation, you can use
DBMS_SPM.ALTER_SQL_PLAN_DIRECTIVE to do the
following:
•
•

Disable the directive by setting ENABLED to NO
Prevent the directive from being dropped by setting
AUTO_DROP to NO
To disable SQL plan directives, set
OPTIMIZER_ADAPTIVE_STATISTICS to FALSE.

Prerequisites
You must have the Administer SQL Management Object privilege to execute the
DBMS_SPD APIs.

Assumptions
This tutorial assumes that you want to do the following:
•

Write all directives for the sh schema to persistent storage.

•

Delete all directives for the sh schema.

To write and then delete all sh schema plan directives:
1.

In SQL*Plus or SQL Developer, log in to the database as a user with the
necessary privileges.

2.

Force the database to write the SQL plan directives to disk.
For example, execute the following DBMS_SPD program:
BEGIN
DBMS_SPD.FLUSH_SQL_PLAN_DIRECTIVE;
END;
/

3.

Query the data dictionary for information about existing directives in the sh
schema.
Example 12-3 queries the data dictionary for information about the directive.

4.

Delete the existing SQL plan directive for the sh schema.
The following PL/SQL program unit deletes the SQL plan directive with the ID
1484026771529551585:
BEGIN
DBMS_SPD.DROP_SQL_PLAN_DIRECTIVE ( directive_id => 1484026771529551585 );
END;
/

12-17

Chapter 12

Managing SQL Plan Directives

Example 12-3

Display Directives for sh Schema
This example shows SQL plan directives, and the results of SQL plan directive
dynamic sampling queries.

SELECT TO_CHAR(d.DIRECTIVE_ID) dir_id, o.OWNER, o.OBJECT_NAME,
o.SUBOBJECT_NAME col_name, o.OBJECT_TYPE object, d.TYPE,
d.STATE, d.REASON
FROM DBA_SQL_PLAN_DIRECTIVES d, DBA_SQL_PLAN_DIR_OBJECTS o
WHERE d.DIRECTIVE_ID=o.DIRECTIVE_ID
AND
o.OWNER IN ('SH')
ORDER BY 1,2,3,4,5;
DIR_ID
OWN OBJECT_NA COL_NAME OBJECT TYPE
STATE
------------------- --- --------- ---------- ------- -------- ---------1484026771529551585 SH CUSTOMERS COUNTRY_ID COLUMN DYNAMIC_ SUPERSEDED
SAMPLING
1484026771529551585 SH CUSTOMERS CUST_STATE_ COLUMN DYNAMIC_ SUPERSEDED
PROVINCE
SAMPLING
1484026771529551585 SH CUSTOMERS

9781501826140511330 SH dyg4msnst5

9872337207064898539 SH TIMES

9781501826140511330 SH 2nk1v0fdx0

TABLE DYNAMIC_ SUPERSEDED
SAMPLING
SQL STA DYNAMIC_
TEMENT SAMPLING
_RESULT
TABLE DYNAMIC_
SAMPLING
_RESULT
SQL STA DYNAMIC_
TEMENT SAMPLING
_RESULT

USABLE

USABLE

USABLE

REASON
-----------SINGLE TABLE
CARDINALITY
MISESTIMATE
SINGLE TABLE
CARDINALITY
MISESTIMATE
SINGLE TABLE
CARDINALITY
MISESTIMATE
VERIFY
CARDINALITY
ESTIMATE
VERIFY
CARDINALITY
ESTIMATE
VERIFY
CARDINALITY
ESTIMATE

See Also:
•

"SQL Plan Directives"

•

Oracle Database PL/SQL Packages and Types Reference for complete
syntax and semantics for the DBMS_SPD package.

•

Oracle Database Reference to learn about DBA_SQL_PLAN_DIRECTIVES

12-18

13
Gathering Optimizer Statistics
This chapter explains how to use the DBMS_STATS.GATHER_*_STATS program units.
This chapter contains the following topics:
•

Configuring Automatic Optimizer Statistics Collection
Oracle Database can gather optimizer statistics automatically.

•

Gathering Optimizer Statistics Manually
As an alternative or supplement to automatic statistics gathering, you can use the
DBMS_STATS package to gather statistics manually.

•

Gathering System Statistics Manually
System statistics describe hardware characteristics, such as I/O and CPU
performance and utilization, to the optimizer.

•

Running Statistics Gathering Functions in Reporting Mode
You can run the DBMS_STATS statistics gathering procedures in reporting mode.

See Also:
•

"Optimizer Statistics Concepts"

•

"Query Optimizer Concepts "

•

Oracle Database PL/SQL Packages and Types Reference to learn about
DBMS_STATS.GATHER_TABLE_STATS

13.1 Configuring Automatic Optimizer Statistics Collection
Oracle Database can gather optimizer statistics automatically.
This section contains the following topics:
•

About Automatic Optimizer Statistics Collection
The automated maintenance tasks infrastructure (known as AutoTask) schedules
tasks to run automatically in Oracle Scheduler windows known as maintenance
windows.

•

Configuring Automatic Optimizer Statistics Collection Using Cloud Control
You can enable and disable all automatic maintenance tasks, including automatic
optimizer statistics collection, using Cloud Control.

•

Configuring Automatic Optimizer Statistics Collection from the Command Line
If you do not use Cloud Control to configure automatic optimizer statistics
collection, then you must use the command line.

13-1

Chapter 13

Configuring Automatic Optimizer Statistics Collection

13.1.1 About Automatic Optimizer Statistics Collection
The automated maintenance tasks infrastructure (known as AutoTask) schedules
tasks to run automatically in Oracle Scheduler windows known as maintenance
windows.
By default, one window is scheduled for each day of the week. Automatic optimizer
statistics collection runs as part of AutoTask. By default, the collection runs in all
predefined maintenance windows.

Note:
Data visibility and privilege requirements may differ when using automatic
optimizer statistics collection with pluggable databases.

To collect the optimizer statistics, the database calls an internal procedure that
operates similarly to the GATHER_DATABASE_STATS procedure with the GATHER AUTO option.
Automatic statistics collection honors all preferences set in the database.
The principal difference between manual and automatic collection is that the latter
prioritizes database objects that need statistics. Before the maintenance window
closes, automatic collection assesses all objects and prioritizes objects that have no
statistics or very old statistics.

Note:
When gathering statistics manually, you can reproduce the object
prioritization of automatic collection by using the DBMS_AUTO_TASK_IMMEDIATE
package. This package runs the same statistics gathering job that is
executed during the automatic nightly statistics gathering job.

See Also:
Oracle Database Administrator’s Guide for a table that summarizes how
manageability features work in a container database (CDB)

13.1.2 Configuring Automatic Optimizer Statistics Collection Using
Cloud Control
You can enable and disable all automatic maintenance tasks, including automatic
optimizer statistics collection, using Cloud Control.
The default window timing works well for most situations. However, you may have
operations such as bulk loads that occur during the window. In such cases, to avoid
potential conflicts that result from operations occurring at the same time as automatic
statistics collection, Oracle recommends that you change the window accordingly.

13-2

Chapter 13

Configuring Automatic Optimizer Statistics Collection

Prerequisites
Access the Database Home page, as described in "Accessing the Database Home
Page in Cloud Control."
To control automatic optimizer statistics collection using Cloud Control:
1.

From the Administration menu, select Oracle Scheduler, then Automated
Maintenance Tasks.
The Automated Maintenance Tasks page appears.
This page shows the predefined tasks. To retrieve information about each task,
click the corresponding link for the task.

2.

Click Configure.
The Automated Maintenance Tasks Configuration page appears.
By default, automatic optimizer statistics collection executes in all predefined
maintenance windows in MAINTENANCE_WINDOW_GROUP.

3.

Perform the following steps:
a.

In the Task Settings section for Optimizer Statistics Gathering, select either
Enabled or Disabled to enable or disable an automated task.

Note:
Oracle strongly recommends that you not disable automatic statistics
gathering because it is critical for the optimizer to generate optimal
plans for queries against dictionary and user objects. If you disable
automatic collection, ensure that you have a good manual statistics
collection strategy for dictionary and user schemas.
b.

To disable statistics gathering for specific days in the week, check the
appropriate box next to the window name.

c.

To change the characteristics of a window group, click Edit Window Group.

d.

To change the times for a window, click the name of the window (for example,
MONDAY_WINDOW), and then in the Schedule section, click Edit.
The Edit Window page appears.

13-3

Chapter 13

Configuring Automatic Optimizer Statistics Collection

In this page, you can change the parameters such as duration and start time
for window execution.
e.

Click Apply.

See Also:
Online Help for Oracle Enterprise Manager Cloud Control

13.1.3 Configuring Automatic Optimizer Statistics Collection from the
Command Line
If you do not use Cloud Control to configure automatic optimizer statistics collection,
then you must use the command line.
You have the following options:
•

Run the ENABLE or DISABLE procedure in the DBMS_AUTO_TASK_ADMIN PL/SQL package.
This package is the recommended command-line technique. For both the ENABLE
and DISABLE procedures, you can specify a particular maintenance window with the
window_name parameter.

•

Set the STATISTICS_LEVEL initialization level to BASIC to disable collection of all
advisories and statistics, including Automatic SQL Tuning Advisor.

13-4

Chapter 13

Configuring Automatic Optimizer Statistics Collection

Note:
Because monitoring and many automatic features are disabled, Oracle
strongly recommends that you do not set STATISTICS_LEVEL to BASIC.

To control automatic statistics collection using DBMS_AUTO_TASK_ADMIN:
1.

In SQL*Plus or SQL Developer, log in to the database as a user with
administrative privileges.

2.

Do one of the following:
•

To enable the automated task, execute the following PL/SQL block:
BEGIN
DBMS_AUTO_TASK_ADMIN.ENABLE (
client_name => 'auto optimizer stats collection'
, operation
=> NULL
, window_name => NULL
);
END;
/

•

To disable the automated task, execute the following PL/SQL block:
BEGIN
DBMS_AUTO_TASK_ADMIN.DISABLE (
client_name => 'auto optimizer stats collection'
, operation
=> NULL
, window_name => NULL
);
END;
/

3.

Query the data dictionary to confirm the change.
For example, query DBA_AUTOTASK_CLIENT as follows:
COL CLIENT_NAME FORMAT a31
SELECT CLIENT_NAME, STATUS
FROM DBA_AUTOTASK_CLIENT
WHERE CLIENT_NAME = 'auto optimizer stats collection';

Sample output appears as follows:
CLIENT_NAME
STATUS
------------------------------- -------auto optimizer stats collection ENABLED

To change the window attributes for automatic statistics collection:
1.

Connect SQL*Plus to the database with administrator privileges.

2.

Change the attributes of the maintenance window as needed.
For example, to change the Monday maintenance window so that it starts at 5
a.m., execute the following PL/SQL program:
BEGIN
DBMS_SCHEDULER.SET_ATTRIBUTE (
'MONDAY_WINDOW'

13-5

Chapter 13

Gathering Optimizer Statistics Manually

, 'repeat_interval'
, 'freq=daily;byday=MON;byhour=05;byminute=0;bysecond=0'
);
END;
/

See Also:
•

Oracle Database PL/SQL Packages and Types Reference to learn about
the DBMS_AUTO_TASK_ADMIN package

•

Oracle Database Reference to learn about the STATISTICS_LEVEL
initialization parameter

13.2 Gathering Optimizer Statistics Manually
As an alternative or supplement to automatic statistics gathering, you can use the
DBMS_STATS package to gather statistics manually.
This section contains the following topics:
•

About Manual Statistics Collection with DBMS_STATS
Use the DBMS_STATS package to manipulate optimizer statistics. You can gather
statistics on objects and columns at various levels of granularity: object, schema,
and database. You can also gather statistics for the physical system.

•

Guidelines for Gathering Optimizer Statistics Manually
In most cases, automatic statistics collection is sufficient for database objects
modified at a moderate speed.

•

Determining When Optimizer Statistics Are Stale
Stale statistics on a table do not accurately reflect its data. To help you determine
when a database object needs new statistics, the database provides a table
monitoring facility.

•

Gathering Schema and Table Statistics
Use GATHER_TABLE_STATS to collect table statistics, and GATHER_SCHEMA_STATS to
collect statistics for all objects in a schema.

•

Gathering Statistics for Fixed Objects
Fixed objects are dynamic performance tables and their indexes. These objects
record current database activity.

•

Gathering Statistics for Volatile Tables Using Dynamic Statistics
Statistics for volatile tables, which are tables modified significantly during the day,
go stale quickly. For example, a table may be deleted or truncated, and then
rebuilt.

•

Gathering Optimizer Statistics Concurrently
Oracle Database can gather statistics on multiple tables or partitions concurrently.

•

Gathering Incremental Statistics on Partitioned Objects
Incremental statistics scan only changed partitions. When gathering statistics on
large partitioned table by deriving global statistics from partition-level statistics,
incremental statistics maintenance improves performance.

13-6

Chapter 13

Gathering Optimizer Statistics Manually

See Also:
•

"Configuring Automatic Optimizer Statistics Collection"

•

Oracle Database PL/SQL Packages and Types Reference to learn about
the DBMS_STATS package

13.2.1 About Manual Statistics Collection with DBMS_STATS
Use the DBMS_STATS package to manipulate optimizer statistics. You can gather
statistics on objects and columns at various levels of granularity: object, schema, and
database. You can also gather statistics for the physical system.
The following table summarizes the DBMS_STATS procedures for gathering optimizer
statistics. This package does not gather statistics for table clusters. However, you can
gather statistics on individual tables in a table cluster.
Table 13-1

DBMS_STATS Procedures for Gathering Optimizer Statistics

Procedure

Purpose

GATHER_INDEX_STATS

Collects index statistics

GATHER_TABLE_STATS

Collects table, column, and index statistics

GATHER_SCHEMA_STATS

Collects statistics for all objects in a schema

GATHER_DICTIONARY_STATS

Collects statistics for all system schemas, including SYS
and SYSTEM, and other optional schemas, such as
CTXSYS and DRSYS

GATHER_DATABASE_STATS

Collects statistics for all objects in a database

When the OPTIONS parameter is set to GATHER STALE or GATHER AUTO, the
GATHER_SCHEMA_STATS and GATHER_DATABASE_STATS procedures gather statistics for any
table that has stale statistics and any table that is missing statistics. If a monitored
table has been modified more than 10%, then the database considers these statistics
stale and gathers them again.

Note:
As explained in "Configuring Automatic Optimizer Statistics Collection", you
can configure a nightly job to gather statistics automatically.

See Also:
•

"Gathering System Statistics Manually"

•

Oracle Database PL/SQL Packages and Types Reference to learn more
about the DBMS_STATS package

13-7

Chapter 13

Gathering Optimizer Statistics Manually

13.2.2 Guidelines for Gathering Optimizer Statistics Manually
In most cases, automatic statistics collection is sufficient for database objects modified
at a moderate speed.
Automatic collection may sometimes be inadequate or unavailable, as shown in the
following table.
Table 13-2

Reasons for Gathering Statistics Manually

Issue

To Learn More

You perform certain types of bulk load and
cannot wait for the maintenance window to
collect statistics because queries must be
executed immediately.

"Online Statistics Gathering for Bulk Loads"

During a nonrepresentative workload,
"Gathering Statistics for Fixed Objects"
automatic statistics collection gathers statistics
for fixed tables.
Automatic statistics collection does not gather
system statistics.

"Gathering System Statistics Manually"

Volatile tables are being deleted or truncated,
and then rebuilt during the day.

"Gathering Statistics for Volatile Tables Using
Dynamic Statistics"

This section offers guidelines for typical situations in which you may choose to gather
statistically manually:
•

Guideline for Setting the Sample Size
In the context of optimizer statistics, sampling is the gathering of statistics from a
random subset of table rows. By enabling the database to avoid full table scans
and sorts of entire tables, sampling minimizes the resources necessary to gather
statistics.

•

Guideline for Gathering Statistics in Parallel
By default, the database gathers statistics with the parallelism degree specified at
the table or index level.

•

Guideline for Partitioned Objects
For partitioned tables and indexes, DBMS_STATS can gather separate statistics for
each partition and global statistics for the entire table or index.

•

Guideline for Frequently Changing Objects
When tables are frequently modified, gather statistics often enough so that they do
not go stale, but not so often that collection overhead degrades performance.

•

Guideline for External Tables
Because the database does not permit data manipulation against external tables,
the database never marks statistics on external tables as stale. If new statistics
are required for an external table, for example, because the underlying data files
change, then regather the statistics.

13-8

Chapter 13

Gathering Optimizer Statistics Manually

13.2.2.1 Guideline for Setting the Sample Size
In the context of optimizer statistics, sampling is the gathering of statistics from a
random subset of table rows. By enabling the database to avoid full table scans and
sorts of entire tables, sampling minimizes the resources necessary to gather statistics.
The database gathers the most accurate statistics when it processes all rows in the
table, which is a 100% sample. However, larger sample sizes increase the time of
statistics gathering operations. The challenge is determining a sample size that
provides accurate statistics in a reasonable time.
DBMS_STATS uses sampling when a user specifies the parameter ESTIMATE_PERCENT,
which controls the percentage of the rows in the table to sample. To maximize
performance gains while achieving necessary statistical accuracy, Oracle
recommends that the ESTIMATE_PERCENT parameter use the default setting of
DBMS_STATS.AUTO_SAMPLE_SIZE. In this case, Oracle Database chooses the sample size
automatically. This setting enables the use of the following:

•

A hash-based algorithm that is much faster than sampling
This algorithm reads all rows and produces statistics that are nearly as accurate
as statistics from a 100% sample. The statistics computed using this technique are
deterministic.

•

Incremental statistics

•

Concurrent statistics

•

New histogram types

The DBA_TABLES.SAMPLE_SIZE column indicates the actual sample size used to gather
statistics.

See Also:
•

"Hybrid Histograms"

•

Oracle Database PL/SQL Packages and Types Reference to learn more
about DBMS_STATS.AUTO_SAMPLE_SIZE

13.2.2.2 Guideline for Gathering Statistics in Parallel
By default, the database gathers statistics with the parallelism degree specified at the
table or index level.
You can override this setting with the degree argument to the DBMS_STATS gathering
procedures. Oracle recommends setting degree to DBMS_STATS.AUTO_DEGREE. This setting
enables the database to choose an appropriate degree of parallelism based on the
object size and the settings for the parallelism-related initialization parameters.
The database can gather most statistics serially or in parallel. However, the database
does not gather some index statistics in parallel, including cluster indexes, domain
indexes, and bitmap join indexes. The database can use sampling when gathering
parallel statistics.

13-9

Chapter 13

Gathering Optimizer Statistics Manually

Note:
Do not confuse gathering statistics in parallel with gathering statistics
concurrently.

See Also:
•

"About Concurrent Statistics Gathering"

•

Oracle Database PL/SQL Packages and Types Reference to learn more
about DBMS_STATS.AUTO_DEGREE

13.2.2.3 Guideline for Partitioned Objects
For partitioned tables and indexes, DBMS_STATS can gather separate statistics for each
partition and global statistics for the entire table or index.
Similarly, for composite partitioning, DBMS_STATS can gather separate statistics for
subpartitions, partitions, and the entire table or index.
To determine the type of partitioning statistics to be gathered, specify the granularity
argument to the DBMS_STATS procedures. Oracle recommends setting granularity to the
default value of AUTO to gather subpartition, partition, or global statistics, depending on
partition type. The ALL setting gathers statistics for all types.

See Also:
"Gathering Incremental Statistics on Partitioned Objects"

13.2.2.4 Guideline for Frequently Changing Objects
When tables are frequently modified, gather statistics often enough so that they do not
go stale, but not so often that collection overhead degrades performance.
You may only need to gather new statistics every week or month. The best practice is
to use a script or job scheduler to regularly run the DBMS_STATS.GATHER_SCHEMA_STATS and
DBMS_STATS.GATHER_DATABASE_STATS procedures.

13.2.2.5 Guideline for External Tables
Because the database does not permit data manipulation against external tables, the
database never marks statistics on external tables as stale. If new statistics are
required for an external table, for example, because the underlying data files change,
then regather the statistics.
For external tables, use the same DBMS_STATS procedures that you use for internal
tables. Note that the scanrate parameter of DBMS_STATS.SET_TABLE_STATS and
DBMS_STATS.GET_TABLE_STATS specifies the rate (in MB/s) at which Oracle Database

13-10

Chapter 13

Gathering Optimizer Statistics Manually

scans data in tables, and is relevant only for external tables. The SCAN_RATE column
appears in the DBA_TAB_STATISTICS and DBA_TAB_PENDING_STATS data dictionary views.

See Also:
•
•

"Creating Artificial Optimizer Statistics for Testing"
Oracle Database PL/SQL Packages and Types Reference to learn about
SET_TABLE_STATS and GET_TABLE_STATS

•

Oracle Database Reference to learn about the DBA_TAB_STATISTICS view

13.2.3 Determining When Optimizer Statistics Are Stale
Stale statistics on a table do not accurately reflect its data. To help you determine
when a database object needs new statistics, the database provides a table
monitoring facility.
Monitoring tracks the approximate number of DML operations on a table and whether
the table has been truncated since the most recent statistics collection. To check
whether statistics are stale, query the STALE_STATS column in DBA_TAB_STATISTICS and
DBA_IND_STATISTICS. This column is based on data in the DBA_TAB_MODIFICATIONS view
and the STALE_PERCENT preference for DBMS_STATS.

Note:
Starting in Oracle Database 12c Release 2 (12.2), you no longer need to use
DBMS_STATS.FLUSH_DATABASE_MONITORING_INFO to ensure that view metadata is
current. The statistics shown in the DBA_TAB_STATISTICS, DBA_IND_STATISTICS,
and DBA_TAB_MODIFICATIONS views are obtained from both disk and memory.

The STALE_STATS column has the following possible values:
•

YES

The statistics are stale.
•

NO

The statistics are not stale.
•

null
The statistics are not collected.

Executing GATHER_SCHEMA_STATS or GATHER_DATABASE_STATS with the GATHER AUTO option
collects statistics only for objects with no statistics or stale statistics.
To determine stale statistics:
1.

Start SQL*Plus, and then log in to the database as a user with the necessary
privileges.

2.

Query the data dictionary for stale statistics.

13-11

Chapter 13

Gathering Optimizer Statistics Manually

The following example queries stale statistics for the sh.sales table (partial output
included):
COL PARTITION_NAME FORMAT a15
SELECT PARTITION_NAME, STALE_STATS
FROM DBA_TAB_STATISTICS
WHERE TABLE_NAME = 'SALES'
AND
OWNER = 'SH'
ORDER BY PARTITION_NAME;
PARTITION_NAME
--------------SALES_1995
SALES_1996
SALES_H1_1997
SALES_H2_1997
SALES_Q1_1998
SALES_Q1_1999
.
.
.

STA
--NO
NO
NO
NO
NO
NO

See Also:
Oracle Database Reference to learn about the DBA_TAB_MODIFICATIONS view

13.2.4 Gathering Schema and Table Statistics
Use GATHER_TABLE_STATS to collect table statistics, and GATHER_SCHEMA_STATS to collect
statistics for all objects in a schema.
To gather schema statistics using DBMS_STATS:
1.

Start SQL*Plus, and connect to the database with the appropriate privileges for the
procedure that you intend to run.

2.

Run the GATHER_TABLE_STATS or GATHER_SCHEMA_STATS procedure, specifying the
desired parameters.
Typical parameters include:
•

Owner - ownname

•

Object name - tabname, indname, partname

•

Degree of parallelism - degree

Example 13-1

Gathering Statistics for a Table

This example uses the DBMS_STATS package to gather statistics on the sh.customers
table with a parallelism setting of 2.
BEGIN
DBMS_STATS.GATHER_TABLE_STATS (
ownname => 'sh'
, tabname => 'customers'
, degree => 2
);

13-12

Chapter 13

Gathering Optimizer Statistics Manually

END;
/

See Also:
Oracle Database PL/SQL Packages and Types Reference to learn about the
GATHER_TABLE_STATS procedure

13.2.5 Gathering Statistics for Fixed Objects
Fixed objects are dynamic performance tables and their indexes. These objects record
current database activity.
Unlike other database tables, the database does not automatically use dynamic
statistics for SQL statement referencing X$ tables when optimizer statistics are
missing. Instead, the optimizer uses predefined default values. These defaults may not
be representative and could potentially lead to a suboptimal execution plan. Thus, it is
important to keep fixed object statistics current.
Oracle Database automatically gathers fixed object statistics as part of automated
statistics gathering if they have not been previously collected. You can also manually
collect statistics on fixed objects by calling DBMS_STATS.GATHER_FIXED_OBJECTS_STATS.
Oracle recommends that you gather statistics when the database has representative
activity.
Prerequisites
You must have the SYSDBA or ANALYZE ANY DICTIONARY system privilege to execute this
procedure.
To gather schema statistics using GATHER_FIXED_OBJECTS_STATS:
1.

In SQL*Plus or SQL Developer, log in to the database as a user with the
necessary privileges.

2.

Run the DBMS_STATS.GATHER_FIXED_OBJECTS_STATS procedure, specifying the desired
parameters.
Typical parameters include:
•

Table identifier describing where to save the current statistics - stattab

•

Identifier to associate with these statistics within stattab (optional) - statid

•

Schema containing stattab (if different from current schema) - statown

Example 13-2

Gathering Statistics for a Table

This example uses the DBMS_STATS package to gather fixed object statistics.
BEGIN
DBMS_STATS.GATHER_FIXED_OBJECTS_STATS;
END;
/

13-13

Chapter 13

Gathering Optimizer Statistics Manually

See Also:
•

"Configuring Automatic Optimizer Statistics Collection"

•

Oracle Database PL/SQL Packages and Types Reference to learn about
the GATHER_TABLE_STATS procedure

13.2.6 Gathering Statistics for Volatile Tables Using Dynamic Statistics
Statistics for volatile tables, which are tables modified significantly during the day, go
stale quickly. For example, a table may be deleted or truncated, and then rebuilt.
When you set the statistics of a volatile object to null, Oracle Database dynamically
gathers the necessary statistics during optimization using dynamic statistics. The
OPTIMIZER_DYNAMIC_SAMPLING initialization parameter controls this feature.
Assumptions
This tutorial assumes the following:
•

The oe.orders table is extremely volatile.

•

You want to delete and then lock the statistics on the orders table to prevent the
database from gathering statistics on the table. In this way, the database can
dynamically gather necessary statistics as part of query optimization.

•

The oe user has the necessary privileges to query DBMS_XPLAN.DISPLAY_CURSOR.

To delete and the lock optimizer statistics:
1.

Connect to the database as user oe, and then delete the statistics for the oe table.
For example, execute the following procedure:
BEGIN
DBMS_STATS.DELETE_TABLE_STATS('OE','ORDERS');
END;
/

2.

Lock the statistics for the oe table.
For example, execute the following procedure:
BEGIN
DBMS_STATS.LOCK_TABLE_STATS('OE','ORDERS');
END;
/

3.

You query the orders table.
For example, use the following statement:
SELECT COUNT(order_id) FROM orders;

4.

You query the plan in the cursor.
You run the following commands (partial output included):
SET LINESIZE 150
SET PAGESIZE 0

13-14

Chapter 13

Gathering Optimizer Statistics Manually

SELECT * FROM TABLE(DBMS_XPLAN.DISPLAY_CURSOR);
SQL_ID aut9632fr3358, child number 0
------------------------------------SELECT COUNT(order_id) FROM orders
Plan hash value: 425895392
--------------------------------------------------------------------| Id | Operation
| Name | Rows | Cost (%CPU)| Time
|
--------------------------------------------------------------------| 0 | SELECT STATEMENT |
|
|
2 (100)|
|
| 1 | SORT AGGREGATE
|
|
1 |
|
|
| 2 | TABLE ACCESS FULL| ORDERS | 105 |
2 (0)| 00:00:01 |
--------------------------------------------------------------------Note
----- dynamic statistics used for this statement (level=2)

The Note in the preceding execution plan shows that the database used dynamic
statistics for the SELECT statement.

See Also:
•

"Configuring Options for Dynamic Statistics"

•

"Locking and Unlocking Optimizer Statistics" to learn how to gather
representative statistics and then lock them, which is an alternative
technique for preventing statistics for volatile tables from going stale

13.2.7 Gathering Optimizer Statistics Concurrently
Oracle Database can gather statistics on multiple tables or partitions concurrently.
This section contains the following topics:
•

About Concurrent Statistics Gathering
By default, each partition of a partition table is gathered sequentially.

•

Enabling Concurrent Statistics Gathering
To enable concurrent statistics gathering, use the DBMS_STATS.SET_GLOBAL_PREFS
procedure to set the CONCURRENT preference.

•

Monitoring Statistics Gathering Operations
You can monitor statistics gathering jobs using data dictionary views.

13.2.7.1 About Concurrent Statistics Gathering
By default, each partition of a partition table is gathered sequentially.
When concurrent statistics gathering mode is enabled, the database can
simultaneously gather optimizer statistics for multiple tables in a schema, or multiple
partitions or subpartitions in a table. Concurrency can reduce the overall time required
to gather statistics by enabling the database to fully use multiple processors.

13-15

Chapter 13

Gathering Optimizer Statistics Manually

Note:
Concurrent statistics gathering mode does not rely on parallel query
processing, but is usable with it.

This section contains the following topics:
•

How DBMS_STATS Gathers Statistics Concurrently
Oracle Database employs multiple tools and technologies to create and manage
multiple statistics gathering jobs concurrently.

•

Concurrent Statistics Gathering and Resource Management
The DBMS_STATS package does not explicitly manage resources used by concurrent
statistics gathering jobs that are part of a user-initiated statistics gathering call.

13.2.7.1.1 How DBMS_STATS Gathers Statistics Concurrently
Oracle Database employs multiple tools and technologies to create and manage
multiple statistics gathering jobs concurrently.
The database uses the following:
•

Oracle Scheduler

•

Oracle Database Advanced Queuing (AQ)

•

Oracle Database Resource Manager (the Resource Manager)

Enable concurrent statistics gathering by setting the CONCURRENT preference with
DBMS_STATS.SET_GLOBAL_PREF.

The database runs as many concurrent jobs as possible. The Job Scheduler decides
how many jobs to execute concurrently and how many to queue. As running jobs
complete, the scheduler dequeues and runs more jobs until the database has
gathered statistics on all tables, partitions, and subpartitions. The maximum number of
jobs is bounded by the JOB_QUEUE_PROCESSES initialization parameter and available
system resources.
In most cases, the DBMS_STATS procedures create a separate job for each table partition
or subpartition. However, if the partition or subpartition is empty or very small, then the
database may automatically batch the object with other small objects into a single job
to reduce the overhead of job maintenance.
The following figure illustrates the creation of jobs at different levels, where Table 3 is
a partitioned table, and the other tables are nonpartitioned. Job 3 acts as a coordinator
job for Table 3, and creates a job for each partition in that table, and a separate job for
the global statistics of Table 3. This example assumes that incremental statistics
gathering is disabled; if enabled, then the database derives global statistics from
partition-level statistics after jobs for partitions complete.

13-16

Chapter 13

Gathering Optimizer Statistics Manually

Figure 13-1

Concurrent Statistics Gathering Jobs

Gather Database/Schema/Dictionary Statistics

Level 1

Job 1

Job 2

Job 3

Job 4

Table 1
Global
Statistics

Table 2
Global
Statistics

Table 3
Coordinator
Job

Table 4
Global
Statistics

Level 2

Job 3.1

Job 3.2

Job 3.3

Table 3
Partition 1

Table 3
Partition 2

Table 3
Global
Statistics

See Also:
•

"Enabling Concurrent Statistics Gathering"

•

Oracle Database PL/SQL Packages and Types Reference to learn about
the DBMS_STATS package

•

Oracle Database Reference to learn about the JOB_QUEUE_PROCESSES
initialization parameter

13.2.7.1.2 Concurrent Statistics Gathering and Resource Management
The DBMS_STATS package does not explicitly manage resources used by concurrent
statistics gathering jobs that are part of a user-initiated statistics gathering call.
Thus, the database may use system resources fully during concurrent statistics
gathering. To address this situation, use the Resource Manager to cap resources
consumed by concurrent statistics gathering jobs. The Resource Manager must be
enabled to gather statistics concurrently.
The system-supplied consumer group ORA$AUTOTASK registers all statistics gathering
jobs. You can create a resource plan with proper resource allocations for ORA$AUTOTASK
to prevent concurrent statistics gathering from consuming all available resources. If
you lack your own resource plan, and if choose not to create one, then consider
activating the Resource Manager with the system-supplied DEFAULT_PLAN.

13-17

Chapter 13

Gathering Optimizer Statistics Manually

Note:
The ORA$AUTOTASK consumer group is shared with the maintenance tasks that
automatically run during the maintenance windows. Thus, when concurrency
is activated for automatic statistics gathering, the database automatically
manages resources, with no extra steps required.

See Also:
Oracle Database Administrator’s Guide to learn about the Resource
Manager

13.2.7.2 Enabling Concurrent Statistics Gathering
To enable concurrent statistics gathering, use the DBMS_STATS.SET_GLOBAL_PREFS
procedure to set the CONCURRENT preference.
Possible values are as follows:
•

MANUAL

Concurrency is enabled only for manual statistics gathering.
•

AUTOMATIC

Concurrency is enabled only for automatic statistics gathering.
•

ALL

Concurrency is enabled for both manual and automatic statistics gathering.
•

OFF

Concurrency is disabled for both manual and automatic statistics gathering. This is
the default value.
This tutorial in this section explains how to enable concurrent statistics gathering.
Prerequisites
This tutorial has the following prerequisites:
•

In addition to the standard privileges for gathering statistics, you must have the
following privileges:
–

CREATE JOB

–

MANAGE SCHEDULER

–

MANAGE ANY QUEUE

•

The SYSAUX tablespace must be online because the scheduler stores its internal
tables and views in this tablespace.

•

The JOB_QUEUE_PROCESSES initialization parameter must be set to at least 4.

•

The Resource Manager must be enabled.

13-18

Chapter 13

Gathering Optimizer Statistics Manually

By default, the Resource Manager is disabled. If you do not have a resource plan,
then consider enabling the Resource Manager with the system-supplied
DEFAULT_PLAN.
Assumptions
This tutorial assumes that you want to do the following:
•

Enable concurrent statistics gathering

•

Gather statistics for the sh schema

•

Monitor the gathering of the sh statistics

To enable concurrent statistics gathering:
1.

Connect SQL*Plus to the database with the appropriate privileges, and then
enable the Resource Manager.
The following example uses the default plan for the Resource Manager:
ALTER SYSTEM SET RESOURCE_MANAGER_PLAN = 'DEFAULT_PLAN';

2.

Set the JOB_QUEUE_PROCESSES initialization parameter to at least twice the number of
CPU cores.
In Oracle Real Application Clusters, the JOB_QUEUE_PROCESSES setting applies to
each node.
Assume that the system has 4 CPU cores. The following example sets the
parameter to 8 (twice the number of cores):
ALTER SYSTEM SET JOB_QUEUE_PROCESSES=8;

3.

Confirm that the parameter change took effect.
For example, enter the following command in SQL*Plus (sample output included):
SHOW PARAMETER PROCESSES;
NAME
-------------------------------_high_priority_processes
aq_tm_processes
db_writer_processes
gcs_server_processes
global_txn_processes
job_queue_processes
log_archive_max_processes
processes

4.

TYPE
----------string
integer
integer
integer
integer
integer
integer
integer

VALUE
----VKTM
1
1
0
1
8
4
100

Enable concurrent statistics.
For example, execute the following PL/SQL anonymous block:
BEGIN
DBMS_STATS.SET_GLOBAL_PREFS('CONCURRENT','ALL');
END;
/

5.

Confirm that the statistics were enabled.
For example, execute the following query (sample output included):
SELECT DBMS_STATS.GET_PREFS('CONCURRENT') FROM DUAL;

13-19

Chapter 13

Gathering Optimizer Statistics Manually

DBMS_STATS.GET_PREFS('CONCURRENT')
---------------------------------ALL
6.

Gather the statistics for the SH schema.
For example, execute the following procedure:
EXEC DBMS_STATS.GATHER_SCHEMA_STATS('SH');

7.

In a separate session, monitor the job progress by querying
DBA_OPTSTAT_OPERATION_TASKS.

For example, execute the following query (sample output included):
SET LINESIZE 1000
COLUMN
COLUMN
COLUMN
COLUMN

TARGET FORMAT a8
TARGET_TYPE FORMAT a25
JOB_NAME FORMAT a14
START_TIME FORMAT a40

SELECT TARGET, TARGET_TYPE, JOB_NAME,
TO_CHAR(START_TIME, 'dd-mon-yyyy hh24:mi:ss')
FROM DBA_OPTSTAT_OPERATION_TASKS
WHERE STATUS = 'IN PROGRESS'
AND
OPID = (SELECT MAX(ID)
FROM DBA_OPTSTAT_OPERATIONS
WHERE OPERATION = 'gather_schema_stats');
TARGET
--------SH.SALES
SH.SALES
8.

TARGET_TYPE
------------------------TABLE (GLOBAL STATS ONLY)
TABLE (COORDINATOR JOB)

JOB_NAME
-------------ST$T292_1_B29
ST$SD290_1_B10

TO_CHAR(START_TIME,'
-------------------30-nov-2012 14:22:47
30-nov-2012 14:22:08

In the original session, disable concurrent statistics gathering.
For example, execute the following query:
EXEC DBMS_STATS.SET_GLOBAL_PREFS('CONCURRENT','OFF');

See Also:
•

"Monitoring Statistics Gathering Operations"

•

Oracle Database Administrator’s Guide

•

Oracle Database PL/SQL Packages and Types Reference to learn how
to use the DBMS_STATS.SET_GLOBAL_PREFS procedure

13.2.7.3 Monitoring Statistics Gathering Operations
You can monitor statistics gathering jobs using data dictionary views.
The following views are relevant:
•

DBA_OPTSTAT_OPERATION_TASKS

This view contains the history of tasks that are performed or currently in progress
as part of statistics gathering operations (recorded in DBA_OPTSTAT_OPERATIONS).

13-20

Chapter 13

Gathering Optimizer Statistics Manually

Each task represents a target object to be processed in the corresponding parent
operation.
•

DBA_OPTSTAT_OPERATIONS

This view contains a history of statistics operations performed or currently in
progress at the table, schema, and database level using the DBMS_STATS package.
The TARGET column in the preceding views shows the target object for that statistics
gathering job in the following form:
OWNER.TABLE_NAME.PARTITION_OR_SUBPARTITION_NAME

All statistics gathering job names start with the string ST$.
To display currently running statistics tasks and jobs:
•

To list statistics gathering currently running tasks from all user sessions, use the
following SQL statement (sample output included):
SELECT OPID, TARGET, JOB_NAnME,
(SYSTIMESTAMP - START_TIME) AS elapsed_time
FROM DBA_OPTSTAT_OPERATION_TASKS
WHERE STATUS = 'IN PROGRESS';
OPID
---981
981

TARGET
------------------------SH.SALES.SALES_Q4_1998
SH.SALES

JOB_NAME
------------ST$T82_1_B29
ST$SD80_1_B10

ELAPSED_TIME
-------------------------+000000000 00:00:00.596321
+000000000 00:00:27.972033

To display completed statistics tasks and jobs:
•

To list only completed tasks and jobs from a particular operation, first identify the
operation ID from the DBA_OPTSTAT_OPERATIONS view based on the statistics
gathering operation name, target, and start time. After you identify the operation
ID, you can query the DBA_OPTSTAT_OPERATION_TASKS view to find the corresponding
tasks in that operation
For example, to list operations with the ID 981, use the following commands in
SQL*Plus (sample output included):
VARIABLE id NUMBER
EXEC :id := 981
SELECT
FROM
WHERE
AND

TARGET, JOB_NAME, (END_TIME - START_TIME) AS ELAPSED_TIME
DBA_OPTSTAT_OPERATION_TASKS
STATUS <> 'IN PROGRESS'
OPID = :id;

TARGET
------------------------SH.SALES_TRANSACTIONS_EXT
SH.CAL_MONTH_SALES_MV
SH.CHANNELS

JOB_NAME
ELAPSED_TIME
------------- -------------------------+000000000 00:00:45.479233
ST$SD88_1_B10 +000000000 00:00:45.382764
ST$SD88_1_B10 +000000000 00:00:45.307397

To display statistics gathering tasks and jobs that have failed:
•

Use the following SQL statement (partial sample output included):
SET LONG 10000
SELECT TARGET, JOB_NAME,

13-21

Chapter 13

Gathering Optimizer Statistics Manually

(END_TIME - START_TIME) AS ELAPSED_TIME, NOTES
FROM DBA_OPTSTAT_OPERATION_TASKS
WHERE STATUS = 'FAILED';
TARGET
JOB_NAME ELAPSED_TIME
NOTES
------------------ -------- -------------------------- ---------------------SYS.OPATCH_XML_INV
+000000007 02:36:31.130314 ORA-20011:
Approximate NDV failed:
ORA-29913: error in
.
.
.

See Also:
Oracle Database Reference to learn about the DBA_SCHEDULER_JOBS view

13.2.8 Gathering Incremental Statistics on Partitioned Objects
Incremental statistics scan only changed partitions. When gathering statistics on large
partitioned table by deriving global statistics from partition-level statistics, incremental
statistics maintenance improves performance.
This section contains the following topics:
•

Purpose of Incremental Statistics
In a typical case, an application loads data into a new partition of a rangepartitioned table. As applications add new partitions and load data, the database
must gather statistics on the new partition and keep global statistics up to date.

•

How DBMS_STATS Derives Global Statistics for Partitioned tables
When incremental statistics maintenance is enabled, DBMS_STATS gathers statistics
and creates synopses for changed partitions only. The database also
automatically merges partition-level synopses into a global synopsis, and derives
global statistics from the partition-level statistics and global synopses.

•

Gathering Statistics for a Partitioned Table: Basic Steps
This section explains how to gather optimizer statistics for a partitioned table.

•

Maintaining Incremental Statistics for Partition Maintenance Operations
A partition maintenance operation is a partition-related operation such as
adding, exchanging, merging, or splitting table partitions.

•

Maintaining Incremental Statistics for Tables with Stale or Locked Partition
Statistics
Starting in Oracle Database 12c, incremental statistics can automatically calculate
global statistics for a partitioned table even if the partition or subpartition statistics
are stale and locked.

13.2.8.1 Purpose of Incremental Statistics
In a typical case, an application loads data into a new partition of a range-partitioned
table. As applications add new partitions and load data, the database must gather
statistics on the new partition and keep global statistics up to date.

13-22

Chapter 13

Gathering Optimizer Statistics Manually

Typically, data warehouse applications access large partitioned tables. Often these
tables are partitioned on date columns, with only the recent partitions subject to
frequent DML changes. Without incremental statistics, statistics collection typically
uses a two-pass approach:
1.

The database scans the table to gather the global statistics.
The full scan of the table for global statistics collection can be very expensive,
depending on the size of the table. As the table adds partitions, the longer the
execution time for GATHER_TABLE_STATS because of the full table scan required for
the global statistics. The database must perform the scan of the entire table even if
only a small subset of partitions change.

2.

The database scans the changed partitions to gather their partition-level statistics.

Incremental maintenance provides a huge performance benefit for data warehouse
applications because of the following:
•

The database must scan the table only once to gather partition statistics and to
derive the global statistics by aggregating partition-level statistics. Thus, the
database avoids the two full scans that are required when not using incremental
statistics: one scan for the partition-level statistics, and one scan for the globallevel statistics.

•

In subsequent statistics gathering, the database only needs to scan the stale
partitions and update their statistics (including synopses). The database can
derive global statistics from the fresh partition statistics, which saves a full table
scan.

When using incremental statistics, the database must still gather statistics on any
partition that will change the global or table-level statistics. Incremental statistics
maintenance yields the same statistics as gathering table statistics from scratch, but
performs better.

13.2.8.2 How DBMS_STATS Derives Global Statistics for Partitioned tables
When incremental statistics maintenance is enabled, DBMS_STATS gathers statistics and
creates synopses for changed partitions only. The database also automatically merges
partition-level synopses into a global synopsis, and derives global statistics from the
partition-level statistics and global synopses.
The database avoids a full table scan when computing global statistics by deriving
some global statistics from the partition-level statistics. For example, the number of
rows at the global level is the sum of number of rows of partitions. Even global
histograms can be derived from partition histograms.
However, the database cannot derive all statistics from partition-level statistics,
including the NDV of a column. The following example shows the NDV for two
partitions in a table:
Table 13-3

NDV for Two Partitions

Object

Column Values

NDV

Partition 1

1,3,3,4,5

4

Partition 2

2,3,4,5,6

5

13-23

Chapter 13

Gathering Optimizer Statistics Manually

Calculating the NDV in the table by adding the NDV of the individual partitions
produces an NDV of 9, which is incorrect. Thus, a more accurate technique is
required: synopses.
This section contains the following topics:
•

Partition-Level Synopses
A synopsis is special type of statistic that tracks the number of distinct values
(NDV) for each column in a partition. You can consider a synopsis as an internal
management structure that samples distinct values.

•

NDV Algorithms: Adaptive Sampling and HyperLogLog
Starting in Oracle Database 12c Release 2 (12.2), the HyperLogLog algorithm can
improve NDV (number of distinct values) calculation performance, and also reduce
the storage space required for synopses.

•

Aggregation of Global Statistics Using Synopses: Example
In this example, the database gathers statistics for the initial six partitions of the
sales table, and then creates synopses for each partition (S1, S2, and so on). The
database creates global statistics by aggregating the partition-level statistics and
synopses.

13.2.8.2.1 Partition-Level Synopses
A synopsis is special type of statistic that tracks the number of distinct values (NDV)
for each column in a partition. You can consider a synopsis as an internal
management structure that samples distinct values.
The database can accurately derive the global-level NDV for each column by merging
partition-level synopses. In the example shown in Table 13-3, the database can use
synopses to calculate the NDV for the column as 6.
Each partition maintains a synopsis in incremental mode. When a new partition is
added to the table you only need to gather statistics for the new partition. The
database automatically updates the global statistics by aggregating the new partition
synopsis with the synopses for existing partitions. Subsequent statistics gathering
operations are faster than when synopses are not used.
The database stores synopses in data dictionary tables WRI$_OPTSTAT_SYNOPSIS_HEAD$
and WRI$_OPTSTAT_SYNOPSIS$ in the SYSAUX tablespace. The DBA_PART_COL_STATISTICS
dictionary view contains information of the column statistics in partitions. If the NOTES
column contains the keyword INCREMENTAL, then this column has synopses.

See Also:
Oracle Database Reference to learn more about DBA_PART_COL_STATISTICS

13.2.8.2.2 NDV Algorithms: Adaptive Sampling and HyperLogLog
Starting in Oracle Database 12c Release 2 (12.2), the HyperLogLog algorithm can
improve NDV (number of distinct values) calculation performance, and also reduce the
storage space required for synopses.
The legacy algorithm for calculating NDV uses adaptive sampling. A synopsis is a
sample of the distinct values. When calculating the NDV, the database initially stores

13-24

Chapter 13

Gathering Optimizer Statistics Manually

every distinct value in a hash table. Each distinct value occupies a distinct hash
bucket, so a column with 5000 distinct values has 5000 hash buckets. The database
then halves the number of hash buckets, and then continues to halve the result until a
small number of buckets remain. The algorithm is “adaptive” because the sampling
rate changes based on the number of hash table splits.
To calculate the NDV for the column, the database uses the following formula, where
B is the number of hash buckets remaining after all the splits have been performed,
and S is the number of splits:
NDV = B * 2^S

Adaptive sampling produces accurate NDV statistics, but has the following
consequences:
•

Synopses occupy significant disk space, especially when tables have many
columns and partitions, and the NDV in each column is high.
For example, a 60-column table might have 300,000 partitions, with an average
per-column NDV of 5,000. In this example, each partition has 300,000 entries (60
x 5000). In total, the synopses tables have 90 billion entries (300,000 squared),
which occupies at least 720 GB of storage space.

•

Bulk processing of synopses can negatively affect performance.
Before the database regathers statistics on the stale partitions, it must delete the
associated synopses. Bulk deletion can be slow because it generates significant
amounts of undo and redo data.

In contrast to dynamic sampling, the HyperLogLog algorithm uses a randomization
technique. Although the algorithm is complex, the foundational insight is that in a
stream of random values, n distinct values will be spaced on average 1/n apart.
Therefore, if you know the smallest value in the stream, then you can roughly estimate
the number of distinct values. For example, if the values range from 0 to 1, and if the
smallest value you observe is .2, then the numbers will on average be evenly spaced .
2 apart, so the NDV estimate is 5.
The HyperLogLog algorithm expands on and corrects the original estimate. The
database applies a hash function to every column value, resulting in a set of hash
values with the same cardinality as the column. For the base estimate, the NDV
equals 2n, where n is the maximum number of trailing zeroes observed in the binary
representation of the hash values. The database refines its NDV estimate by using
part of the output to split values into different hash buckets.
The advantages of the HyperLogLog algorithm over adaptive sampling are:
•

The accuracy of the new algorithm is similar to the original algorithm.

•

The memory required is significantly lower, which typically leads to huge
reductions in synopsis size.
Synopses can become large when many partitions exist, and they have many
columns with high NDV. Synopses that use the HyperLogLog algorithm are more
compact. Creating and deleting synopses affects batch run times. Any operational
procedures that manage partitions reduce run time.

The DBMS_STATS preference APPROXIMATE_NDV_ALGORITHM determines which algorithm the
database uses for NDV calculation.

13-25

Chapter 13

Gathering Optimizer Statistics Manually

See Also:
Oracle Database PL/SQL Packages and Types Reference to learn about the
APPROXIMATE_NDV_ALGORITHM preference

13.2.8.2.3 Aggregation of Global Statistics Using Synopses: Example
In this example, the database gathers statistics for the initial six partitions of the sales
table, and then creates synopses for each partition (S1, S2, and so on). The database
creates global statistics by aggregating the partition-level statistics and synopses.

Figure 13-2

Aggregating Statistics

Sales Table

1 The database gathers partition-level
statistics, and creates synopses

May 18 2012

S1

May 19 2012

S2

May 20 2012

S3

May 21 2012

S4

May 22 2012

S5

May 23 2012

S6

2 The database generates
1
global statistics by
aggregating partition-level
statistics and synopses
Global
Statistics

Sysaux
Tablespace

The following graphic shows a new partition, containing data for May 24, being added
to the sales table. The database gathers statistics for the newly added partition,
retrieves synopses for the other partitions, and then aggregates the synopses to
create global statistics.

13-26

Chapter 13

Gathering Optimizer Statistics Manually

Figure 13-3

Aggregating Statistics after Adding a Partition

Sales Table
May 18 2012

S1

May 19 2012

S2

May 20 2012

S3

May 21 2012

S4

May 22 2012

S5

May 23 2012

S6

May 24 2012

S7

3
1 The table adds a
new partition.

6
1 The database generates
global statistics by
aggregating partition-level
synopses.
Global
Statistics

1 The database retrieves
5
statistics and synopses
for other partitions.

4
1 The database gathers statistics
and synopses for the new partition.
Sysaux
Tablespace

13.2.8.3 Gathering Statistics for a Partitioned Table: Basic Steps
This section explains how to gather optimizer statistics for a partitioned table.
This section contains the following topics:
•

Considerations for Incremental Statistics Maintenance
Enabling incremental statistics maintenance has several consequences.

•

Enabling Incremental Statistics Using SET_TABLE_PREFS
To enable incremental statistics maintenance for a partitioned table, use
DBMS_STATS.SET_TABLE_PREFS to set the INCREMENTAL value to true. When INCREMENTAL
is set to false, which is the default, the database uses a full table scan to maintain
global statistics.

•

About the APPROXIMATE_NDV_ALGORITHM Settings
The DBMS_STATS.APPROXIMATE_NDV_ALGORITHM preference specifies the synopsis
generation algorithm, either HyperLogLog or adaptive sampling. The
INCREMENTAL_STALENESS preference controls when the database reformats synopses
that use the adaptive sampling format.

•

Configuring Synopsis Generation: Examples
These examples show different approaches, both conservative and aggressive, to
switching synopses to the new HyperLogLog format.

13.2.8.3.1 Considerations for Incremental Statistics Maintenance
Enabling incremental statistics maintenance has several consequences.
Specifically, note the following:

13-27

Chapter 13

Gathering Optimizer Statistics Manually

•

If a table uses composite partitioning, then the database only gathers statistics for
modified subpartitions. The database does not gather statistics at the subpartition
level for unmodified subpartitions. In this way, the database reduces work by
skipping unmodified partitions.

•

If a table uses incremental statistics, and if this table has a locally partitioned
index, then the database gathers index statistics at the global level and for
modified (not unmodified) index partitions. The database does not generate global
index statistics from the partition-level index statistics. Rather, the database
gathers global index statistics by performing a full index scan.

•

The SYSAUX tablespace consumes additional space to maintain global statistics for
partitioned tables.

See Also:
Oracle Database PL/SQL Packages and Types Reference to learn more
about DBMS_STATS

13.2.8.3.2 Enabling Incremental Statistics Using SET_TABLE_PREFS
To enable incremental statistics maintenance for a partitioned table, use
DBMS_STATS.SET_TABLE_PREFS to set the INCREMENTAL value to true. When INCREMENTAL is
set to false, which is the default, the database uses a full table scan to maintain global
statistics.
For the database to update global statistics incrementally by scanning only the
partitions that have changed, the following conditions must be met:
•

The PUBLISH value for the partitioned table is true.

•

The INCREMENTAL value for the partitioned table is true.

•

The statistics gathering procedure must specify AUTO_SAMPLE_SIZE for
ESTIMATE_PERCENT and AUTO for GRANULARITY.

Example 13-3

Enabling Incremental Statistics

Assume that the PUBLISH value for the partitioned table sh.sales is true. The following
program enables incremental statistics for this table:
EXEC DBMS_STATS.SET_TABLE_PREFS('sh', 'sales', 'INCREMENTAL', 'TRUE');

13.2.8.3.3 About the APPROXIMATE_NDV_ALGORITHM Settings
The DBMS_STATS.APPROXIMATE_NDV_ALGORITHM preference specifies the synopsis
generation algorithm, either HyperLogLog or adaptive sampling. The
INCREMENTAL_STALENESS preference controls when the database reformats synopses that
use the adaptive sampling format.
The APPROXIMATE_NDV_ALGORITHM preference has the following possible values:
•

REPEAT OR HYPERLOGLOG

This is the default. If INCREMENTAL is enabled on the table, then the database
preserves the format of any existing synopses that use the adaptive sampling
algorithm. However, the database creates any new synopses in HyperLogLog

13-28

Chapter 13

Gathering Optimizer Statistics Manually

format. This approach is attractive when existing performance is acceptable, and
you do not want to incur the performance cost of reformatting legacy content.
•

ADAPTIVE SAMPLING

The database uses the adaptive sampling algorithm for all synopses. This is the
most conservative option.
•

HYPERLOGLOG

The database uses the HyperLogLog algorithm for all new and stale synopses.
The INCREMENTAL_STALENESS preference controls when a synopsis is considered stale.
When the APPROXIMATE_NDV_ALGORITHM preference is set to HYPERLOGLOG, then the
following INCREMENTAL_STALENESS settings apply:
•

ALLOW_MIXED_FORMAT

This is the default. If this value is specified, and if the following conditions are met,
then the database does not consider existing adaptive sampling synopses as
stale:
–

The synopses are fresh.

–

You gather statistics manually.

Thus, synopses in both the legacy and HyperLogLog formats can co-exist.
However, over time the automatic statistics gathering job regathers statistics on
synopses that use the old format, and replaces them with synopses in
HyperLogLog format. In this way, the automatic statistics gather job gradually
phases out the old format. Manual statistics gathering jobs do not reformat
synopses that use the adaptive sampling format.
•

Null
Any partitions with the synopses in the legacy format are considered stale, which
immediately triggers the database to regather statistics for stale synopses. The
advantage is that the performance cost occurs only once. The disadvantage is that
regathering all statistics on large tables can be resource-intensive.

13.2.8.3.4 Configuring Synopsis Generation: Examples
These examples show different approaches, both conservative and aggressive, to
switching synopses to the new HyperLogLog format.
Example 13-4

Taking a Conservative Approach to Reformatting Synopses

In this example, you allow synopses in mixed formats to coexist for the sh.sales table.
Mixed formats yield less accurate statistics. However, you do not need to regather
statistics for all partitions of the table.
To ensure that all new and stale synopses use the HyperLogLog algorithm, set the
APPROXIMATE_NDV_ALGORITHM preference to HYPERLOGLOG. To ensure that the automatic
statistics gathering job reformats stale synopses gradually over time, set the
INCREMENTAL_STALENESS preference to ALLOW_MIXED_FORMAT.
BEGIN
DBMS_STATS.SET_TABLE_PREFS
( ownname => 'sh'
, tabname => 'sales'
, pname => 'approximate_ndv_algorithm'
, pvalue => 'hyperloglog' );

13-29

Chapter 13

Gathering Optimizer Statistics Manually

DBMS_STATS.SET_TABLE_PREFS
( ownname => 'sh'
, tabname => 'sales'
, pname
=> 'incremental_staleness'
, pvalue => 'allow_mixed_format' );
END;

Example 13-5

Taking an Aggressive Approach to Reformatting Synopses

In this example, you force all synopses to use the HyperLogLog algorithm for the
sh.sales table. In this case, the database must regather statistics for all partitions of
the table.
To ensure that all new and stale synopses use the HyperLogLog algorithm, set the
APPROXIMATE_NDV_ALGORITHM preference to HYPERLOGLOG. To force the database to
immediately regather statistics for all partitions in the table and store them in the new
format, set the INCREMENTAL_STALENESS preference to null.
BEGIN
DBMS_STATS.SET_TABLE_PREFS
( ownname => 'sh'
, tabname => 'sales'
, pname => 'approximate_ndv_algorithm'
, pvalue => 'hyperloglog' );
DBMS_STATS.SET_TABLE_PREFS
( ownname => 'sh'
, tabname => 'sales'
, pname
=> 'incremental_staleness'
, pvalue => 'null' );
END;

13.2.8.4 Maintaining Incremental Statistics for Partition Maintenance
Operations
A partition maintenance operation is a partition-related operation such as adding,
exchanging, merging, or splitting table partitions.
Oracle Database provides the following support for incremental statistics maintenance:
•

If a partition maintenance operation triggers statistics gathering, then the database
can reuse synopses that would previously have been dropped with the old
segments.

•

DBMS_STATS can create a synopsis on a nonpartitioned table. The synopsis enables
the database to maintain incremental statistics as part of a partition exchange
operation without having to explicitly gather statistics on the partition after the
exchange.

When the DBMS_STATS preference INCREMENTAL is set to true on a table, the
INCREMENTAL_LEVEL preference controls which synopses are collected and when. This
preference takes the following values:
•

TABLE
DBMS_STATS gathers table-level synopses on this table. You can only set
INCREMENTAL_LEVEL to TABLE at the table level, not at the schema, database, or

global level.
•

PARTITION (default)

13-30

Chapter 13

Gathering Optimizer Statistics Manually

DBMS_STATS only gathers synopsis at the partition level of partitioned tables.

When performing a partition exchange, to have synopses after the exchange for the
partition being exchanged, set INCREMENTAL to true and INCREMENTAL_LEVEL to TABLE on
the table to be exchanged with the partition.
Assumptions
This tutorial assumes the following:
•

You want to load empty partition p_sales_01_2010 in a sales table.

•

You create a staging table t_sales_01_2010, and then populate the table.

•

You want the database to maintain incremental statistics as part of the partition
exchange operation without having to explicitly gather statistics on the partition
after the exchange.

To maintain incremental statistics as part of a partition exchange operation:
1.

Set incremental statistics preferences for staging table t_sales_01_2010.
For example, run the following statement:
BEGIN
DBMS_STATS.SET_TABLE_PREFS (
ownname => 'sh'
,
tabname => 't_sales_01_2010'
,
pname
=> 'INCREMENTAL'
,
pvalue => 'true'
);
DBMS_STATS.SET_TABLE_PREFS (
ownname => 'sh'
,
tabname => 't_sales_01_2010'
,
pname
=> 'INCREMENTAL_LEVEL'
,
pvalue => 'table'
);
END;

2.

Gather statistics on staging table t_sales_01_2010.
For example, run the following PL/SQL code:
BEGIN
DBMS_STATS.GATHER_TABLE_STATS (
ownname => 'SH'
,
tabname => 'T_SALES_01_2010'
);
END;
/
DBMS_STATS gathers table-level synopses on t_sales_01_2010.

3.

Ensure that the INCREMENTAL preference is true on the sh.sales table.
For example, run the following PL/SQL code:
BEGIN
DBMS_STATS.SET_TABLE_PREFS (
ownname => 'sh'
,
tabname => 'sales'
,
pname
=> 'INCREMENTAL'
,
pvalue => 'true'
);

13-31

Chapter 13

Gathering Optimizer Statistics Manually

END;
/
4.

If you have never gathered statistics on sh.sales before with INCREMENTAL set to
true, then gather statistics on the partition to be exchanged.
For example, run the following PL/SQL code:
BEGIN
DBMS_STATS.GATHER_TABLE_STATS (
ownname => 'sh'
,
tabname => 'sales'
,
pname
=> 'p_sales_01_2010'
,
pvalue => granularity=>'partition'
);
END;
/

5.

Perform the partition exchange.
For example, use the following SQL statement:
ALTER TABLE sales EXCHANGE PARTITION p_sales_01_2010 WITH TABLE t_sales_01_2010;

After the exchange, the partitioned table has both statistics and a synopsis for
partition p_sales_01_2010.
In releases before Oracle Database 12c, the preceding statement swapped the
segment data and statistics of p_sales_01_2010 with t_sales_01_2010. The database
did not maintain synopses for nonpartitioned tables such as t_sales_01_2010. To
gather global statistics on the partitioned table, you needed to rescan the
p_sales_01_2010 partition to obtain its synopses.

See Also:
Oracle Database PL/SQL Packages and Types Reference to learn more
about DBMS_STATS.SET_TABLE_PREFS

13.2.8.5 Maintaining Incremental Statistics for Tables with Stale or Locked
Partition Statistics
Starting in Oracle Database 12c, incremental statistics can automatically calculate
global statistics for a partitioned table even if the partition or subpartition statistics are
stale and locked.
When incremental statistics are enabled in releases before Oracle Database 12c, if
any DML occurs on a partition, then the optimizer considers statistics on this partition
to be stale. Thus, DBMS_STATS must gather the statistics again to accurately aggregate
the global statistics. Furthermore, if DML occurs on a partition whose statistics are
locked, then DBMS_STATS cannot regather the statistics on the partition, so a full table
scan is the only means of gathering global statistics. Regathering statistics creates
performance overhead.
In Oracle Database 12c, the statistics preference INCREMENTAL_STALENESS controls how
the database determines whether the statistics on a partition or subpartition are stale.
This preference takes the following values:

13-32

Chapter 13

Gathering Optimizer Statistics Manually

•

USE_STALE_PERCENT

A partition or subpartition is not considered stale if DML changes are less than the
STALE_PERCENT preference specified for the table. The default value of STALE_PERCENT
is 10, which means that if DML causes more than 10% of row changes, then the
table is considered stale.
•

USE_LOCKED_STATS

Locked partition or subpartition statistics are not considered stale, regardless of
DML changes.
•

NULL (default)

A partition or subpartition is considered stale if it has any DML changes. This
behavior is identical to the Oracle Database 11g behavior. When the default value
is used, statistics gathered in incremental mode are guaranteed to be the same as
statistics gathered in nonincremental mode. When a nondefault value is used,
statistics gathered in incremental mode might be less accurate than those
gathered in nonincremental mode.
You can specify USE_STALE_PERCENT and USE_LOCKED_STATS together. For example, you
can write the following anonymous block:
BEGIN
DBMS_STATS.SET_TABLE_PREFS (
ownname
=> null
, table_name => 't'
, pname
=> 'incremental_staleness'
, pvalue
=> 'use_stale_percent,use_locked_stats'
);
END;

Assumptions
This tutorial assumes the following:
•

The STALE_PERCENT for a partitioned table is set to 10.

•

The INCREMENTAL value is set to true.

•

The table has had statistics gathered in INCREMENTAL mode before.

•

You want to discover how statistics gathering changes depending on the setting
for INCREMENTAL_STALENESS, whether the statistics are locked, and the percentage of
DML changes.

To test for tables with stale or locked partition statistics:
1.

Set INCREMENTAL_STALENESS to NULL.
Afterward, 5% of the rows in one partition change because of DML activity.

2.

Use DBMS_STATS to gather statistics on the table.
DBMS_STATS regathers statistics for the partition that had the 5% DML activity, and

incrementally maintains the global statistics.
3.

Set INCREMENTAL_STALENESS to USE_STALE_PERCENT.
Afterward, 5% of the rows in one partition change because of DML activity.

4.

Use DBMS_STATS to gather statistics on the table.

13-33

Chapter 13

Gathering System Statistics Manually

DBMS_STATS does not regather statistics for the partition that had DML activity
(because the changes are under the staleness threshold of 10%), and
incrementally maintains the global statistics.
5.

Lock the partition statistics.
Afterward, 20% of the rows in one partition change because of DML activity.

6.

Use DBMS_STATS to gather statistics on the table.
DBMS_STATS does not regather statistics for the partition because the statistics are

locked. The database gathers the global statistics with a full table scan.
Afterward, 5% of the rows in one partition change because of DML activity.
7.

Use DBMS_STATS to gather statistics on the table.
When you gather statistics on this table, DBMS_STATS does not regather statistics for
the partition because they are not considered stale. The database maintains global
statistics incrementally using the existing statistics for this partition.

8.

Set INCREMENTAL_STALENESS to USE_LOCKED_STATS and USE_STALE_PERCENT.
Afterward, 20% of the rows in one partition change because of DML activity.

9.

Use DBMS_STATS to gather statistics on the table.
Because USE_LOCKED_STATS is set, DBMS_STATS ignores the fact that the statistics are
stale and uses the locked statistics. The database maintains global statistics
incrementally using the existing statistics for this partition.

See Also:
Oracle Database PL/SQL Packages and Types Reference to learn more
about DBMS_STATS.SET_TABLE_PREFS

13.3 Gathering System Statistics Manually
System statistics describe hardware characteristics, such as I/O and CPU
performance and utilization, to the optimizer.
System statistics enable the optimizer to choose a more efficient execution plan.
Oracle recommends that you gather system statistics when a physical change occurs
in the environment, for example, the server has faster CPUs, more memory, or
different storage.
This section contains the following topics:
•

About Gathering System Statistics with DBMS_STATS
To gather system statistics, use DBMS_STATS.GATHER_SYSTEM_STATS.

•

Guidelines for Gathering System Statistics
The database automatically gathers essential parts of system statistics at startup.

•

Gathering Workload Statistics
Oracle recommends that you use DBMS_STATS.GATHER_SYSTEM_STATS to capture
statistics when the database has the most typical workload.

13-34

Chapter 13

Gathering System Statistics Manually

•

Gathering Noworkload Statistics
Noworkload statistics capture characteristics of the I/O system.

•

Deleting System Statistics
Use the DBMS_STATS.DELETE_SYSTEM_STATS function to delete system statistics.

13.3.1 About Gathering System Statistics with DBMS_STATS
To gather system statistics, use DBMS_STATS.GATHER_SYSTEM_STATS.
When the database gathers system statistics, it analyzes activity in a specified time
period (workload statistics) or simulates a workload (noworkload statistics). The input
arguments to DBMS_STATS.GATHER_SYSTEM_STATS are:
•

NOWORKLOAD

The optimizer gathers statistics based on system characteristics only, without
regard to the workload.
•

INTERVAL

After the specified number of minutes has passed, the optimizer updates system
statistics either in the data dictionary, or in an alternative table (specified by
stattab). Statistics are based on system activity during the specified interval.
•

START and STOP
START initiates gathering statistics. STOP calculates statistics for the elapsed period
(since START) and refreshes the data dictionary or an alternative table (specified by
stattab). The optimizer ignores INTERVAL.

•

EXADATA

The system statistics consider the unique capabilities provided by using Exadata,
such as large I/O size and high I/O throughput. The optimizer sets the multiblock
read count and I/O throughput statistics along with CPU speed.
The following table lists the optimizer system statistics gathered by DBMS_STATS and the
options for gathering or manually setting specific system statistics.
Table 13-4

Optimizer System Statistics in the DBMS_STAT Package

Parameter
Name

Description

Initialization

Options for
Gathering or
Setting
Statistics

cpuspeedNW

Represents
noworkload CPU
speed. CPU
speed is the
average number
of CPU cycles in
each second.

At system startup Set
gathering_mode
= NOWORKLOAD or
set statistics
manually.

Unit

Millions/sec.

13-35

Chapter 13

Gathering System Statistics Manually

Table 13-4

(Cont.) Optimizer System Statistics in the DBMS_STAT Package

Parameter
Name

Description

Initialization

Options for
Gathering or
Setting
Statistics

Unit

ioseektim

Represents the
At system startup
time it takes to
10 (default)
position the disk
head to read
data. I/O seek
time equals seek
time + latency
time + operating
system overhead
time.

Set
gathering_mode
= NOWORKLOAD or
set statistics
manually.

ms

iotfrspeed

Represents the
rate at which an
Oracle database
can read data in
the single read
request.

At system startup Set
gathering_mode
4096 (default)
= NOWORKLOAD or
set statistics
manually.

Bytes/ms

cpuspeed

Represents
workload CPU
speed. CPU
speed is the
average number
of CPU cycles in
each second.

None

Set
gathering_mode
= NOWORKLOAD,
INTERVAL, or
START|STOP, or
set statistics
manually.

Millions/sec.

maxthr

Maximum I/O
throughput is the
maximum
throughput that
the I/O
subsystem can
deliver.

None

Set
gathering_mode
= NOWORKLOAD,
INTERVAL, or
START|STOP, or
set statistics
manually.

Bytes/sec.

slavethr

Slave I/O
throughput is the
average parallel
execution server
I/O throughput.

None

Set
gathering_mode
= INTERVAL or
START|STOP, or
set statistics
manually.

Bytes/sec.

sreadtim

Single-block read None
time is the
average time to
read a single
block randomly.

Set
gathering_mode
= INTERVAL or
START|STOP, or
set statistics
manually.

ms

mreadtim

Multiblock read is None
the average time
to read a
multiblock
sequentially.

Set
gathering_mode
= INTERVAL or
START|STOP, or
set statistics
manually.

ms

13-36

Chapter 13

Gathering System Statistics Manually

Table 13-4

(Cont.) Optimizer System Statistics in the DBMS_STAT Package

Parameter
Name

Description

Initialization

Options for
Gathering or
Setting
Statistics

Unit

mbrc

Multiblock count
is the average
multiblock read
count
sequentially.

None

Set
gathering_mode
= INTERVAL or
START|STOP, or
set statistics
manually.

blocks

See Also:
Oracle Database PL/SQL Packages and Types Reference for detailed
information on the procedures in the DBMS_STATS package for implementing
system statistics

13.3.2 Guidelines for Gathering System Statistics
The database automatically gathers essential parts of system statistics at startup.
CPU and I/O characteristics tend to remain constant over time. Typically, these
characteristics only change when some aspect of the configuration is upgraded. For
this reason, Oracle recommends that you gather system statistics only when a
physical change occurs in your environment, for example, the server gets faster CPUs,
more memory, or different disk storage.
Note the following guidelines:
•

Oracle Database initializes noworkload statistics to default values at the first
instance startup. Oracle recommends that you gather noworkload statistics after
you create new tablespaces on storage that is not used by any other tablespace.

•

The best practice is to capture statistics in the interval of time when the system
has the most common workload. Gathering workload statistics does not generate
additional overhead.

13.3.3 Gathering Workload Statistics
Oracle recommends that you use DBMS_STATS.GATHER_SYSTEM_STATS to capture statistics
when the database has the most typical workload.
For example, database applications can process OLTP transactions during the day
and generate OLAP reports at night.
This section contains the following topics:
•

About Workload Statistics
Workload statistics analyze activity in a specified time period.

13-37

Chapter 13

Gathering System Statistics Manually

•

Starting and Stopping System Statistics Gathering
This tutorial explains how to set the workload interval with the START and STOP
parameters of GATHER_SYSTEM_STATS.

•

Gathering System Statistics During a Specified Interval
This tutorial explains how to set the workload interval with the INTERVAL parameter
of GATHER_SYSTEM_STATS.

13.3.3.1 About Workload Statistics
Workload statistics analyze activity in a specified time period.
Workload statistics include the following statistics listed in Table 13-4:
•

Single block (sreadtim) and multiblock (mreadtim) read times

•

Multiblock count (mbrc)

•

CPU speed (cpuspeed)

•

Maximum system throughput (maxthr)

•

Average parallel execution throughput (slavethr)

The database computes sreadtim, mreadtim, and mbrc by comparing the number of
physical sequential and random reads between two points in time from the beginning
to the end of a workload. The database implements these values through counters that
change when the buffer cache completes synchronous read requests.
Because the counters are in the buffer cache, they include not only I/O delays, but
also waits related to latch contention and task switching. Thus, workload statistics
depend on system activity during the workload window. If system is I/O bound (both
latch contention and I/O throughput), then the statistics promote a less I/O-intensive
plan after the database uses the statistics.
As shown in Figure 13-4, if you gather workload statistics, then the optimizer uses the
mbrc value gathered for workload statistics to estimate the cost of a full table scan.
Figure 13-4

Workload Statistics Counters
Optimizer

Database Buffer Cache
Counters for Workload
Statistics
mreadtim
mbrc
sreadtim
cpuspeed
maxthr
slavethr

Estimate
costs of
full table
scans
May not be
available if
no full table
scans occur

May use if mbrc and mreadtim
are not available

DB_FILE_MULTIBLOCK_READ_COUNT

When gathering workload statistics, the database may not gather the mbrc and
mreadtim values if no table scans occur during serial workloads, as is typical of OLTP
systems. However, full table scans occur frequently on DSS systems. These scans
may run parallel and bypass the buffer cache. In such cases, the database still gathers
the sreadtim because index lookups use the buffer cache.

13-38

Chapter 13

Gathering System Statistics Manually

If the database cannot gather or validate gathered mbrc or mreadtim values, but has
gathered sreadtim and cpuspeed, then the database uses only sreadtim and cpuspeed for
costing. In this case, the optimizer uses the value of the initialization parameter
DB_FILE_MULTIBLOCK_READ_COUNT to cost a full table scan. However, if
DB_FILE_MULTIBLOCK_READ_COUNT is 0 or is not set, then the optimizer uses a value of 8 for
calculating cost.
Use the DBMS_STATS.GATHER_SYSTEM_STATS procedure to gather workload statistics. The
GATHER_SYSTEM_STATS procedure refreshes the data dictionary or a staging table with
statistics for the elapsed period. To set the duration of the collection, use either of the
following techniques:
•

Specify START the beginning of the workload window, and then STOP at the end of
the workload window.

•

Specify INTERVAL and the number of minutes before statistics gathering
automatically stops. If needed, you can use GATHER_SYSTEM_STATS
(gathering_mode=>'STOP') to end gathering earlier than scheduled.

See Also:
Oracle Database Reference to learn about the
DB_FILE_MULTIBLOCK_READ_COUNT initialization parameter

13.3.3.2 Starting and Stopping System Statistics Gathering
This tutorial explains how to set the workload interval with the START and STOP
parameters of GATHER_SYSTEM_STATS.
Assumptions
This tutorial assumes the following:
•

The hour between 10 a.m. and 11 a.m. is representative of the daily workload.

•

You intend to collect system statistics directly in the data dictionary.

To gather workload statistics using START and STOP:
1.

Start SQL*Plus and connect to the database with administrator privileges.

2.

Start statistics collection.
For example, at 10 a.m., execute the following procedure to start collection:
EXECUTE DBMS_STATS.GATHER_SYSTEM_STATS( gathering_mode => 'START' );

3.

Generate the workload.

4.

End statistics collection.
For example, at 11 a.m., execute the following procedure to end collection:
EXECUTE DBMS_STATS.GATHER_SYSTEM_STATS( gathering_mode => 'STOP' );

The optimizer can now use the workload statistics to generate execution plans that
are effective during the normal daily workload.
5.

Optionally, query the system statistics.

13-39

Chapter 13

Gathering System Statistics Manually

For example, run the following query:
COL PNAME FORMAT a15
SELECT PNAME, PVAL1
FROM SYS.AUX_STATS$
WHERE SNAME = 'SYSSTATS_MAIN';

See Also:
Oracle Database PL/SQL Packages and Types Reference to learn about the
DBMS_STATS.GATHER_SYSTEM_STATS procedure

13.3.3.3 Gathering System Statistics During a Specified Interval
This tutorial explains how to set the workload interval with the INTERVAL parameter of
GATHER_SYSTEM_STATS.

Assumptions
This tutorial assumes the following:
•

The database application processes OLTP transactions during the day and runs
OLAP reports at night. To gather representative statistics, you collect them during
the day for two hours and then at night for two hours.

•

You want to store statistics in a table named workload_stats.

•

You intend to switch between the statistics gathered.

To gather workload statistics using INTERVAL:
1.

Start SQL*Plus and connect to the production database as administrator dba1.

2.

Create a table to hold the production statistics.
For example, execute the following PL/SQL program to create user statistics table
workload_stats:
BEGIN
DBMS_STATS.CREATE_STAT_TABLE (
ownname => 'dba1'
,
stattab => 'workload_stats'
);
END;
/

3.

Ensure that JOB_QUEUE_PROCESSES is not 0 so that DBMS_JOB jobs and Oracle
Scheduler jobs run.
ALTER SYSTEM SET JOB_QUEUE_PROCESSES = 1;

4.

Gather statistics during the day.
For example, gather statistics for two hours with the following program:
BEGIN
DBMS_STATS.GATHER_SYSTEM_STATS (
interval => 120
,
stattab => 'workload_stats'
,
statid
=> 'OLTP'

13-40

Chapter 13

Gathering System Statistics Manually

);
END;
/
5.

Gather statistics during the evening.
For example, gather statistics for two hours with the following program:
BEGIN
DBMS_STATS.GATHER_SYSTEM_STATS (
interval => 120
,
stattab => 'workload_stats'
,
statid
=> 'OLAP'
);
END;
/

6.

In the day or evening, import the appropriate statistics into the data dictionary.
For example, during the day you can import the OLTP statistics from the staging
table into the dictionary with the following program:
BEGIN
DBMS_STATS.IMPORT_SYSTEM_STATS (
stattab => 'workload_stats'
,
statid => 'OLTP'
);
END;
/

For example, during the night you can import the OLAP statistics from the staging
table into the dictionary with the following program:
BEGIN
DBMS_STATS.IMPORT_SYSTEM_STATS (
stattab => 'workload_stats'
,
statid => 'OLAP'
);
END;
/

See Also:
Oracle Database PL/SQL Packages and Types Reference to learn about the
DBMS_STATS.GATHER_SYSTEM_STATS procedure

13.3.4 Gathering Noworkload Statistics
Noworkload statistics capture characteristics of the I/O system.
By default, Oracle Database uses noworkload statistics and the CPU cost model. The
values of noworkload statistics are initialized to defaults at the first instance startup.
You can also use the DBMS_STATS.GATHER_SYSTEM_STATS procedure to gather noworkload
statistics manually.
Noworkload statistics include the following system statistics listed in Table 13-4:
•

I/O transfer speed (iotfrspeed)

13-41

Chapter 13

Gathering System Statistics Manually

•

I/O seek time (ioseektim)

•

CPU speed (cpuspeednw)

The major difference between workload statistics and noworkload statistics is in the
gathering method. Noworkload statistics gather data by submitting random reads
against all data files, whereas workload statistics uses counters updated when
database activity occurs. If you gather workload statistics, then Oracle Database uses
them instead of noworkload statistics.
To gather noworkload statistics, run DBMS_STATS.GATHER_SYSTEM_STATS with no
arguments or with the gathering mode set to noworkload. There is an overhead on the
I/O system during the gathering process of noworkload statistics. The gathering
process may take from a few seconds to several minutes, depending on I/O
performance and database size.
When you gather noworkload statistics, the database analyzes the information and
verifies it for consistency. In some cases, the values of noworkload statistics may
retain their default values. You can either gather the statistics again, or use
SET_SYSTEM_STATS to set the values manually to the I/O system specifications.
Assumptions
This tutorial assumes that you want to gather noworkload statistics manually.
To gather noworkload statistics manually:
1.

Start SQL*Plus and connect to the database with administrator privileges.

2.

Gather the noworkload statistics.
For example, run the following statement:
BEGIN
DBMS_STATS.GATHER_SYSTEM_STATS (
gathering_mode => 'NOWORKLOAD'
);
END;

3.

Optionally, query the system statistics.
For example, run the following query:
COL PNAME FORMAT a15
SELECT PNAME, PVAL1
FROM SYS.AUX_STATS$
WHERE SNAME = 'SYSSTATS_MAIN';

See Also:
Oracle Database PL/SQL Packages and Types Reference to learn about the
DBMS_STATS.GATHER_SYSTEM_STATS procedure

13-42

Chapter 13

Running Statistics Gathering Functions in Reporting Mode

13.3.5 Deleting System Statistics
Use the DBMS_STATS.DELETE_SYSTEM_STATS function to delete system statistics.
This procedure deletes workload statistics collected using the INTERVAL or START and
STOP options, and then resets the default to noworkload statistics. However, if the
stattab parameter specifies a table for storing statistics, then the subprogram deletes
all system statistics with the associated statid from the statistics table.
Assumptions
This tutorial assumes the following:
•

You gathered statistics for a specific intensive workload, but no longer want the
optimizer to use these statistics.

•

You stored workload statistics in the default location, not in a user-specified table.

To delete system statistics:
1.

In SQL*Plus, log in to the database as a user with administrative privileges.

2.

Delete the system statistics.
For example, run the following statement:
EXEC DBMS_STATS.DELETE_SYSTEM_STATS;

See Also:
Oracle Database PL/SQL Packages and Types Reference to learn about the
DBMS_STATS.DELETE_SYSTEM_STATS procedure

13.4 Running Statistics Gathering Functions in Reporting
Mode
You can run the DBMS_STATS statistics gathering procedures in reporting mode.
When you use the REPORT_* procedures, the optimizer does not actually gather
statistics. Rather, the package reports objects that would be processed if you were to
use a specified statistics gathering function.
The following table lists the DBMS_STATS.REPORT_GATHER_*_STATS functions. For all
functions, the input parameters are the same as for the corresponding GATHER_*_STATS
procedure, with the following additional parameters: detail_level and format.
Supported formats are XML, HTML, and TEXT.

13-43

Chapter 13

Running Statistics Gathering Functions in Reporting Mode

Table 13-5

DBMS_STATS Reporting Mode Functions

Function

Description

REPORT_GATHER_TABLE_STATS

Runs GATHER_TABLE_STATS in reporting mode. The
procedure does not collect statistics, but reports all
objects that would be affected by invoking
GATHER_TABLE_STATS.

REPORT_GATHER_SCHEMA_STATS

Runs GATHER_SCHEMA_STATS in reporting mode.
The procedure does not actually collect statistics,
but reports all objects that would be affected by
invoking GATHER_SCHEMA_STATS.

REPORT_GATHER_DICTIONARY_STATS

Runs GATHER_DICTIONARY_STATS in reporting
mode. The procedure does not actually collect
statistics, but reports all objects that would be
affected by invoking GATHER_DICTIONARY_STATS.

REPORT_GATHER_DATABASE_STATS

Runs GATHER_DATABASE_STATS in reporting mode.
The procedure does not actually collect statistics,
but reports all objects that would be affected by
invoking GATHER_DATABASE_STATS.

REPORT_GATHER_FIXED_OBJ_STATS

Runs GATHER_FIXED_OBJ_STATS in reporting mode.
The procedure does not actually collect statistics,
but reports all objects that would be affected by
invoking GATHER_FIXED_OBJ_STATS.

REPORT_GATHER_AUTO_STATS

Runs the automatic statistics gather job in
reporting mode. The procedure does not actually
collect statistics, but reports all objects that would
be affected by running the job.

Assumptions
This tutorial assumes that you want to generate an HTML report of the objects that
would be affected by running GATHER_SCHEMA_STATS on the oe schema.
To report on objects affected by running GATHER_SCHEMA_STATS:
1.

Start SQL*Plus and connect to the database with administrator privileges.

2.

Run the DBMS_STATS.REPORT_GATHER_SCHEMA_STATS function.
For example, run the following commands in SQL*Plus:
SET LINES 200 PAGES 0
SET LONG 100000
COLUMN REPORT FORMAT A200
VARIABLE my_report CLOB;
BEGIN
:my_report :=DBMS_STATS.REPORT_GATHER_SCHEMA_STATS(
ownname
=> 'OE'
,
detail_level => 'TYPICAL' ,
format
=> 'HTML'
);
END;
/

The following graphic shows a partial example report:

13-44

Chapter 13

Running Statistics Gathering Functions in Reporting Mode

See Also:
Oracle Database PL/SQL Packages and Types Reference to learn more
about DBMS_STATS

13-45

14
Managing Extended Statistics
DBMS_STATS enables you to collect extended statistics, which are statistics that can
improve cardinality estimates when multiple predicates exist on different columns of a
table, or when predicates use expressions.

An extension is either a column group or an expression. Column group statistics can
improve cardinality estimates when multiple columns from the same table occur
together in a SQL statement. Expression statistics improves optimizer estimates when
predicates use expressions, for example, built-in or user-defined functions.

Note:
You cannot create extended statistics on virtual columns.

This chapter contains the following topics:
•

Managing Column Group Statistics
A column group is a set of columns that is treated as a unit.

•

Managing Expression Statistics
The type of extended statistics known as expression statistics improve optimizer
estimates when a WHERE clause has predicates that use expressions.

See Also:
Oracle Database SQL Language Reference for a list of restrictions on virtual
columns

14.1 Managing Column Group Statistics
A column group is a set of columns that is treated as a unit.
Essentially, a column group is a virtual column. By gathering statistics on a column
group, the optimizer can more accurately determine the cardinality estimate when a
query groups these columns together.
The following sections provide an overview of column group statistics, and explain how
to manage them manually:
•

About Statistics on Column Groups
Individual column statistics are useful for determining the selectivity of a single
predicate in a WHERE clause.

14-1

Chapter 14

Managing Column Group Statistics

•

Detecting Useful Column Groups for a Specific Workload
You can use DBMS_STATS.SEED_COL_USAGE and REPORT_COL_USAGE to determine which
column groups are required for a table based on a specified workload.

•

Creating Column Groups Detected During Workload Monitoring
You can use the DBMS_STATS.CREATE_EXTENDED_STATS function to create column
groups that were detected previously by executing DBMS_STATS.SEED_COL_USAGE.

•

Creating and Gathering Statistics on Column Groups Manually
In some cases, you may know the column group that you want to create.

•

Displaying Column Group Information
To obtain the name of a column group, use the
DBMS_STATS.SHOW_EXTENDED_STATS_NAME function or a database view.

•

Dropping a Column Group
Use the DBMS_STATS.DROP_EXTENDED_STATS function to delete a column group from a
table.

See Also:
Oracle Database PL/SQL Packages and Types Reference to learn about the
DBMS_STATS package

14.1.1 About Statistics on Column Groups
Individual column statistics are useful for determining the selectivity of a single
predicate in a WHERE clause.
When the WHERE clause includes multiple predicates on different columns from the
same table, individual column statistics do not show the relationship between the
columns. This is the problem solved by a column group.
The optimizer calculates the selectivity of the predicates independently, and then
combines them. However, if a correlation between the individual columns exists, then
the optimizer cannot take it into account when determining a cardinality estimate,
which it creates by multiplying the selectivity of each table predicate by the number of
rows.
The following graphic contrasts two ways of gathering statistics on the
cust_state_province and country_id columns of the sh.customers table. The diagram
shows DBMS_STATS collecting statistics on each column individually and on the group.
The column group has a system-generated name.

14-2

Chapter 14

Managing Column Group Statistics

Figure 14-1
CUST_ID

Column Group Statistics
CUST_STATE_PROVINCE

101095
103105

CA
Sao Paulo

COUNTRY_ID

...

52790
52775

SYS_STU#S#WF25Z#QAHIHE#MOFFMM_

Statistics for
Column Group
Statistics for
CUST_STATE_PROVINCE

Statistics for
COUNTRY_ID
DBMS_STATS

Note:
The optimizer uses column group statistics for equality predicates, inlist
predicates, and for estimating the GROUP BY cardinality.

This section contains the following topics:
•

Why Column Group Statistics Are Needed: Example
This example demonstrates how column group statistics enable the optimizer to
give a more accurate cardinality estimate.

•

Automatic and Manual Column Group Statistics
Oracle Database can create column group statistics either automatically or
manually.

•

User Interface for Column Group Statistics
Several DBMS_STATS program units have preferences that are relevant for column
groups.

14.1.1.1 Why Column Group Statistics Are Needed: Example
This example demonstrates how column group statistics enable the optimizer to give a
more accurate cardinality estimate.
The following query of the DBA_TAB_COL_STATISTICS table shows information about
statistics that have been gathered on the columns cust_state_province and country_id
from the sh.customers table:
COL COLUMN_NAME FORMAT a20
COL NDV FORMAT 999
SELECT COLUMN_NAME, NUM_DISTINCT AS "NDV", HISTOGRAM
FROM DBA_TAB_COL_STATISTICS
WHERE OWNER = 'SH'

14-3

Chapter 14

Managing Column Group Statistics

AND
AND

TABLE_NAME = 'CUSTOMERS'
COLUMN_NAME IN ('CUST_STATE_PROVINCE', 'COUNTRY_ID');

Sample output is as follows:
COLUMN_NAME
NDV HISTOGRAM
-------------------- ---------- --------------CUST_STATE_PROVINCE
145 FREQUENCY
COUNTRY_ID
19 FREQUENCY

As shown in the following query, 3341 customers reside in California:
SELECT COUNT(*)
FROM sh.customers
WHERE cust_state_province = 'CA';
COUNT(*)
---------3341

Consider an explain plan for a query of customers in the state CA and in the country
with ID 52790 (USA):
EXPLAIN PLAN FOR
SELECT *
FROM sh.customers
WHERE cust_state_province = 'CA'
AND
country_id=52790;
Explained.
sys@PROD> SELECT * FROM TABLE(DBMS_XPLAN.DISPLAY);
PLAN_TABLE_OUTPUT
-------------------------------------------------------------------------------Plan hash value: 1683234692
------------------------------------------------------------------------------| Id | Operation
| Name
| Rows | Bytes | Cost (%CPU)| Time
|
------------------------------------------------------------------------------| 0 | SELECT STATEMENT |
| 128 | 24192 | 442 (7)| 00:00:06 |
|* 1 | TABLE ACCESS FULL| CUSTOMERS |
128 | 24192 | 442 (7)| 00:00:06 |
------------------------------------------------------------------------------Predicate Information (identified by operation id):
--------------------------------------------------PLAN_TABLE_OUTPUT
-------------------------------------------------------------------------------1 - filter("CUST_STATE_PROVINCE"='CA' AND "COUNTRY_ID"=52790)
13 rows selected.

Based on the single-column statistics for the country_id and cust_state_province
columns, the optimizer estimates that the query of California customers in the USA will
return 128 rows. In fact, 3341 customers reside in California, but the optimizer does
not know that the state of California is in the country of the USA, and so greatly
underestimates cardinality by assuming that both predicates reduce the number of
returned rows.

14-4

Chapter 14

Managing Column Group Statistics

You can make the optimizer aware of the real-world relationship between values in
country_id and cust_state_province by gathering column group statistics. These

statistics enable the optimizer to give a more accurate cardinality estimate.

See Also:
•

"Detecting Useful Column Groups for a Specific Workload"

•

"Creating Column Groups Detected During Workload Monitoring"

•

"Creating and Gathering Statistics on Column Groups Manually"

14.1.1.2 Automatic and Manual Column Group Statistics
Oracle Database can create column group statistics either automatically or manually.
The optimizer can use SQL plan directives to generate a more optimal plan. If the
DBMS_STATS preference AUTO_STAT_EXTENSIONS is set to ON (by default it is OFF), then a
SQL plan directive can automatically trigger the creation of column group statistics
based on usage of predicates in the workload. You can set AUTO_STAT_EXTENSIONS with
the SET_TABLE_PREFS, SET_GLOBAL_PREFS, or SET_SCHEMA_PREFS procedures.
When you want to manage column group statistics manually, then use DBMS_STATS as
follows:
•

Detect column groups

•

Create previously detected column groups

•

Create column groups manually and gather column group statistics

See Also:
•

"Detecting Useful Column Groups for a Specific Workload"

•

"Creating Column Groups Detected During Workload Monitoring"

•

"Creating and Gathering Statistics on Column Groups Manually"

•

Oracle Database PL/SQL Packages and Types Reference to learn about
the DBMS_STATS procedures for setting optimizer statistics

14.1.1.3 User Interface for Column Group Statistics
Several DBMS_STATS program units have preferences that are relevant for column
groups.

14-5

Chapter 14

Managing Column Group Statistics

Table 14-1

DBMS_STATS APIs Relevant for Column Groups

Program Unit or Preference

Description

SEED_COL_USAGE Procedure

Iterates over the SQL statements in the specified workload,
compiles them, and then seeds column usage information for
the columns that appear in these statements.
To determine the appropriate column groups, the database
must observe a representative workload. You do not need to
run the queries themselves during the monitoring period.
Instead, you can run EXPLAIN PLAN for some longer-running
queries in your workload to ensure that the database is
recording column group information for these queries.

REPORT_COL_USAGE Function

Generates a report that lists the columns that were seen in
filter predicates, join predicates, and GROUP BY clauses in the
workload.
You can use this function to review column usage information
recorded for a specific table.

CREATE_EXTENDED_STATS
Function

Creates extensions, which are either column groups or
expressions. The database gathers statistics for the
extension when either a user-generated or automatic
statistics gathering job gathers statistics for the table.

AUTO_STAT_EXTENSIONS
Preference

Controls the automatic creation of extensions, including
column groups, when optimizer statistics are gathered. Set
this preference using SET_TABLE_PREFS, SET_SCHEMA_PREFS,
or SET_GLOBAL_PREFS.
When AUTO_STAT_EXTENSIONS is set to OFF (default), the
database does not create column group statistics
automatically. To create extensions, you must execute the
CREATE_EXTENDED_STATS function or specify extended
statistics explicitly in the METHOD_OPT parameter in the
DBMS_STATS API.
When set to ON, a SQL plan directive can trigger the creation
of column group statistics automatically based on usage of
columns in the predicates in the workload.

See Also:
•

"Setting Artificial Optimizer Statistics for a Table"

•

Oracle Database PL/SQL Packages and Types Reference to learn about
the DBMS_STATS package

14.1.2 Detecting Useful Column Groups for a Specific Workload
You can use DBMS_STATS.SEED_COL_USAGE and REPORT_COL_USAGE to determine which
column groups are required for a table based on a specified workload.
This technique is useful when you do not know which extended statistics to create.
This technique does not work for expression statistics.

14-6

Chapter 14

Managing Column Group Statistics

Assumptions
This tutorial assumes the following:
•

Cardinality estimates have been incorrect for queries of the sh.customers_test
table (created from the customers table) that use predicates referencing the
columns country_id and cust_state_province.

•

You want the database to monitor your workload for 5 minutes (300 seconds).

•

You want the database to determine which column groups are needed
automatically.

To detect column groups:
1.

Start SQL*Plus or SQL Developer, and log in to the database as user sh.

2.

Create the customers_test table and gather statistics for it:
DROP TABLE customers_test;
CREATE TABLE customers_test AS SELECT * FROM customer;
EXEC DBMS_STATS.GATHER_TABLE_STATS(user, 'customers_test');

3.

Enable workload monitoring.
In a different SQL*Plus session, connect as SYS and run the following PL/SQL
program to enable monitoring for 300 seconds:
BEGIN
DBMS_STATS.SEED_COL_USAGE(null,null,300);
END;
/

4.

As user sh, run explain plans for two queries in the workload.
The following examples show the explain plans for two queries on the
customers_test table:
EXPLAIN PLAN FOR
SELECT *
FROM customers_test
WHERE cust_city = 'Los Angeles'
AND
cust_state_province = 'CA'
AND
country_id = 52790;
SELECT PLAN_TABLE_OUTPUT
FROM TABLE(DBMS_XPLAN.DISPLAY('plan_table', null,'basic rows'));
EXPLAIN PLAN FOR
SELECT country_id, cust_state_province, count(cust_city)
FROM
customers_test
GROUP BY country_id, cust_state_province;
SELECT PLAN_TABLE_OUTPUT
FROM TABLE(DBMS_XPLAN.DISPLAY('plan_table', null,'basic rows'));

Sample output appears below:
PLAN_TABLE_OUTPUT
--------------------------------------------------------------------------Plan hash value: 4115398853
----------------------------------------------------

14-7

Chapter 14

Managing Column Group Statistics

| Id | Operation
| Name
| Rows |
---------------------------------------------------| 0 | SELECT STATEMENT |
|
1 |
| 1 | TABLE ACCESS FULL| CUSTOMERS_TEST |
1 |
---------------------------------------------------8 rows selected.
PLAN_TABLE_OUTPUT
--------------------------------------------------------------------------Plan hash value: 3050654408
----------------------------------------------------| Id | Operation
| Name
| Rows |
----------------------------------------------------| 0 | SELECT STATEMENT |
| 1949 |
| 1 | HASH GROUP BY
|
| 1949 |
| 2 | TABLE ACCESS FULL| CUSTOMERS_TEST | 55500 |
----------------------------------------------------9 rows selected.

The first plan shows a cardinality of 1 row for a query that returns 932 rows. The
second plan shows a cardinality of 1949 rows for a query that returns 145 rows.
5.

Optionally, review the column usage information recorded for the table.
Call the DBMS_STATS.REPORT_COL_USAGE function to generate a report:
SET LONG 100000
SET LINES 120
SET PAGES 0
SELECT DBMS_STATS.REPORT_COL_USAGE(user, 'customers_test')
FROM DUAL;

The report appears below:
LEGEND:
.......
EQ
: Used in single table EQuality predicate
RANGE
: Used in single table RANGE predicate
LIKE
: Used in single table LIKE predicate
NULL
: Used in single table is (not) NULL predicate
EQ_JOIN
: Used in EQuality JOIN predicate
NONEQ_JOIN : Used in NON EQuality JOIN predicate
FILTER
: Used in single table FILTER predicate
JOIN
: Used in JOIN predicate
GROUP_BY : Used in GROUP BY expression
...........................................................................
###########################################################################
COLUMN USAGE REPORT FOR SH.CUSTOMERS_TEST
.........................................
1.
2.
3.
4.

COUNTRY_ID
CUST_CITY
CUST_STATE_PROVINCE
(CUST_CITY, CUST_STATE_PROVINCE,
COUNTRY_ID)

: EQ
: EQ
: EQ
: FILTER

14-8

Chapter 14

Managing Column Group Statistics

5. (CUST_STATE_PROVINCE, COUNTRY_ID) : GROUP_BY
###########################################################################

In the preceding report, the first three columns were used in equality predicates in
the first monitored query:
...
WHERE cust_city = 'Los Angeles'
AND
cust_state_province = 'CA'
AND
country_id = 52790;

All three columns appeared in the same WHERE clause, so the report shows them as
a group filter. In the second query, two columns appeared in the GROUP BY clause,
so the report labels them as GROUP_BY. The sets of columns in the FILTER and
GROUP_BY report are candidates for column groups.

See Also:
•

"Managing SQL Tuning Sets"

•

Oracle Database PL/SQL Packages and Types Reference to learn about
the DBMS_STATS package

14.1.3 Creating Column Groups Detected During Workload Monitoring
You can use the DBMS_STATS.CREATE_EXTENDED_STATS function to create column groups
that were detected previously by executing DBMS_STATS.SEED_COL_USAGE.
Assumptions
This tutorial assumes that you have performed the steps in "Detecting Useful Column
Groups for a Specific Workload".
To create column groups:
1.

Create column groups for the customers_test table based on the usage information
captured during the monitoring window.
For example, run the following query:
SELECT DBMS_STATS.CREATE_EXTENDED_STATS(user, 'customers_test') FROM DUAL;

Sample output appears below:
###########################################################################
EXTENSIONS FOR SH.CUSTOMERS_TEST
................................
1. (CUST_CITY, CUST_STATE_PROVINCE,
COUNTRY_ID)
:SYS_STUMZ$C3AIHLPBROI#SKA58H_N created
2. (CUST_STATE_PROVINCE, COUNTRY_ID):SYS_STU#S#WF25Z#QAHIHE#MOFFMM_ created
###########################################################################

The database created two column groups for customers_test: one column group
for the filter predicate and one group for the GROUP BY operation.
2.

Regather table statistics.

14-9

Chapter 14

Managing Column Group Statistics

Run GATHER_TABLE_STATS to regather the statistics for customers_test:
EXEC DBMS_STATS.GATHER_TABLE_STATS(user,'customers_test');
3.

As user sh, run explain plans for two queries in the workload.
Check the USER_TAB_COL_STATISTICS view to determine which additional statistics
were created by the database:
SELECT COLUMN_NAME, NUM_DISTINCT, HISTOGRAM
FROM USER_TAB_COL_STATISTICS
WHERE TABLE_NAME = 'CUSTOMERS_TEST'
ORDER BY 1;

Partial sample output appears below:
CUST_CITY
...
SYS_STU#S#WF25Z#QAHIHE#MOFFMM_
SYS_STUMZ$C3AIHLPBROI#SKA58H_N

620 HEIGHT BALANCED
145 NONE
620 HEIGHT BALANCED

This example shows the two column group names returned from the
DBMS_STATS.CREATE_EXTENDED_STATS function. The column group created on
CUST_CITY, CUST_STATE_PROVINCE, and COUNTRY_ID has a height-balanced histogram.
4.

Explain the plans again.
The following examples show the explain plans for two queries on the
customers_test table:
EXPLAIN PLAN FOR
SELECT *
FROM customers_test
WHERE cust_city = 'Los Angeles'
AND
cust_state_province = 'CA'
AND
country_id = 52790;
SELECT PLAN_TABLE_OUTPUT
FROM TABLE(DBMS_XPLAN.DISPLAY('plan_table', null,'basic rows'));
EXPLAIN PLAN FOR
SELECT country_id, cust_state_province, count(cust_city)
FROM
customers_test
GROUP BY country_id, cust_state_province;
SELECT PLAN_TABLE_OUTPUT
FROM TABLE(DBMS_XPLAN.DISPLAY('plan_table', null,'basic rows'));

The new plans show more accurate cardinality estimates:
---------------------------------------------------| Id | Operation
| Name
| Rows |
---------------------------------------------------| 0 | SELECT STATEMENT |
| 1093 |
| 1 | TABLE ACCESS FULL| CUSTOMERS_TEST | 1093 |
---------------------------------------------------8 rows selected.
Plan hash value: 3050654408
----------------------------------------------------| Id | Operation
| Name
| Rows |

14-10

Chapter 14

Managing Column Group Statistics

----------------------------------------------------| 0 | SELECT STATEMENT |
| 145 |
| 1 | HASH GROUP BY
|
|
145 |
| 2 | TABLE ACCESS FULL| CUSTOMERS_TEST | 55500 |
----------------------------------------------------9 rows selected.

See Also:
Oracle Database PL/SQL Packages and Types Reference to learn about the
DBMS_STATS package

14.1.4 Creating and Gathering Statistics on Column Groups Manually
In some cases, you may know the column group that you want to create.
The METHOD_OPT argument of the DBMS_STATS.GATHER_TABLE_STATS function can create and
gather statistics on a column group automatically. You can create a new column group
by specifying the group of columns using FOR COLUMNS.
Assumptions
This tutorial assumes the following:
•

You want to create a column group for the cust_state_province and country_id
columns in the customers table in sh schema.

•

You want to gather statistics (including histograms) on the entire table and the new
column group.

To create a column group and gather statistics for this group:
1.

In SQL*Plus, log in to the database as the sh user.

2.

Create the column group and gather statistics.
For example, execute the following PL/SQL program:

BEGIN
DBMS_STATS.GATHER_TABLE_STATS( 'sh','customers',
METHOD_OPT => 'FOR ALL COLUMNS SIZE SKEWONLY ' ||
'FOR COLUMNS SIZE SKEWONLY (cust_state_province,country_id)' );
END;
/

See Also:
Oracle Database PL/SQL Packages and Types Reference to learn about the
DBMS_STATS.GATHER_TABLE_STATS procedure

14-11

Chapter 14

Managing Column Group Statistics

14.1.5 Displaying Column Group Information
To obtain the name of a column group, use the DBMS_STATS.SHOW_EXTENDED_STATS_NAME
function or a database view.
You can also use views to obtain information such as the number of distinct values,
and whether the column group has a histogram.
Assumptions
This tutorial assumes the following:
•

You created a column group for the cust_state_province and country_id columns
in the customers table in sh schema.

•

You want to determine the column group name, the number of distinct values, and
whether a histogram has been created for a column group.

To monitor a column group:
1.

Start SQL*Plus and connect to the database as the sh user.

2.

To determine the column group name, do one of the following.
•

Execute the SHOW_EXTENDED_STATS_NAME function.
For example, run the following PL/SQL program:
SELECT SYS.DBMS_STATS.SHOW_EXTENDED_STATS_NAME( 'sh','customers',
'(cust_state_province,country_id)' ) col_group_name
FROM DUAL;

The output is similar to the following:
COL_GROUP_NAME
---------------SYS_STU#S#WF25Z#QAHIHE#MOFFMM_

•

Query the USER_STAT_EXTENSIONS view.
For example, run the following query:
SELECT EXTENSION_NAME, EXTENSION
FROM USER_STAT_EXTENSIONS
WHERE TABLE_NAME='CUSTOMERS';
EXTENSION_NAME
EXTENSION
----------------------------------------------------------------------SYS_STU#S#WF25Z#QAHIHE#MOFFMM_
("CUST_STATE_PROVINCE","COUNTRY_ID")

3.

Query the number of distinct values and find whether a histogram has been
created for a column group.
For example, run the following query:
SELECT
FROM
WHERE
AND
AND

e.EXTENSION col_group, t.NUM_DISTINCT, t.HISTOGRAM
USER_STAT_EXTENSIONS e, USER_TAB_COL_STATISTICS t
e.EXTENSION_NAME=t.COLUMN_NAME
e.TABLE_NAME=t.TABLE_NAME
t.TABLE_NAME='CUSTOMERS';

COL_GROUP

NUM_DISTINCT

HISTOGRAM

14-12

Chapter 14

Managing Expression Statistics

------------------------------------------------------------------("COUNTRY_ID","CUST_STATE_PROVINCE") 145
FREQUENCY

See Also:
Oracle Database PL/SQL Packages and Types Reference to learn about the
DBMS_STATS.SHOW_EXTENDED_STATS_NAME function

14.1.6 Dropping a Column Group
Use the DBMS_STATS.DROP_EXTENDED_STATS function to delete a column group from a
table.
Assumptions
This tutorial assumes the following:
•

You created a column group for the cust_state_province and country_id columns
in the customers table in sh schema.

•

You want to drop the column group.

To drop a column group:
1.

Start SQL*Plus and connect to the database as the sh user.

2.

Drop the column group.
For example, the following PL/SQL program deletes a column group from the
customers table:
BEGIN
DBMS_STATS.DROP_EXTENDED_STATS( 'sh', 'customers',
'(cust_state_province, country_id)' );
END;
/

See Also:
Oracle Database PL/SQL Packages and Types Reference to learn about the
DBMS_STATS.DROP_EXTENDED_STATS function

14.2 Managing Expression Statistics
The type of extended statistics known as expression statistics improve optimizer
estimates when a WHERE clause has predicates that use expressions.
This section contains the following topics:
•

About Expression Statistics
For an expression in the form (function(col)=constant) applied to a WHERE clause
column, the optimizer does not know how this function affects predicate cardinality

14-13

Chapter 14

Managing Expression Statistics

unless a function-based index exists. However, you can gather expression
statistics on the expression(function(col) itself.
•

Creating Expression Statistics
You can use DBMS_STATS to create statistics for a user-specified expression.

•

Displaying Expression Statistics
To obtain information about expression statistics, use the database view
DBA_STAT_EXTENSIONS and the DBMS_STATS.SHOW_EXTENDED_STATS_NAME function.

•

Dropping Expression Statistics
To delete a column group from a table, use the DBMS_STATS.DROP_EXTENDED_STATS
function.

14.2.1 About Expression Statistics
For an expression in the form (function(col)=constant) applied to a WHERE clause
column, the optimizer does not know how this function affects predicate cardinality
unless a function-based index exists. However, you can gather expression statistics on
the expression(function(col) itself.
The following graphic shows the optimizer using statistics to generate a plan for a
query that uses a function. The top shows the optimizer checking statistics for the
column. The bottom shows the optimizer checking statistics corresponding to the
expression used in the query. The expression statistics yield more accurate estimates.
Figure 14-2

Expression Statistics
SELECT * FROM sh.customers
WHERE LOWER (cust_state_province) = ‘ca’

Optimizer

Use
Expression
Statistics

Yes

Do
expression
statistics
exist?

No

Use Default
Column
Statistics

LOWER(cust_state_province)
Expression Statistics

cust_state_province
Column Statistics

Optimal
Estimate

Suboptimal
Estimate

As shown in Figure 14-2, when expression statistics are not available, the optimizer
can produce suboptimal plans.
This section contains the following topics:

14-14

Chapter 14

Managing Expression Statistics

•

When Expression Statistics Are Useful: Example

See Also:
Oracle Database SQL Language Reference to learn about SQL functions

14.2.1.1 When Expression Statistics Are Useful: Example
The following query of the sh.customers table shows that 3341 customers are in the
state of California:
sys@PROD> SELECT COUNT(*) FROM sh.customers WHERE cust_state_province='CA';
COUNT(*)
---------3341

Consider the plan for the same query with the LOWER() function applied:
sys@PROD> EXPLAIN PLAN FOR
2 SELECT * FROM sh.customers WHERE LOWER(cust_state_province)='ca';
Explained.
sys@PROD> select * from table(dbms_xplan.display);
PLAN_TABLE_OUTPUT
------------------------------------------------------------------------------Plan hash value: 2008213504
------------------------------------------------------------------------------| Id | Operation
| Name
| Rows | Bytes | Cost (%CPU)| Time
|
------------------------------------------------------------------------------| 0 | SELECT STATEMENT |
| 555 | 108K| 406 (1)| 00:00:05 |
|* 1 | TABLE ACCESS FULL| CUSTOMERS | 555 | 108K| 406 (1)| 00:00:05 |
------------------------------------------------------------------------------Predicate Information (identified by operation id):
--------------------------------------------------1 - filter(LOWER("CUST_STATE_PROVINCE")='ca')

Because no expression statistics exist for LOWER(cust_state_province)='ca', the
optimizer estimate is significantly off. You can use DBMS_STATS procedures to correct
these estimates.

14.2.2 Creating Expression Statistics
You can use DBMS_STATS to create statistics for a user-specified expression.
You can use either of the following program units:
•

GATHER_TABLE_STATS procedure

•

CREATE_EXTENDED_STATISTICS function followed by the GATHER_TABLE_STATS procedure

14-15

Chapter 14

Managing Expression Statistics

Assumptions
This tutorial assumes the following:
•

Selectivity estimates are inaccurate for queries of sh.customers that use the
UPPER(cust_state_province) function.

•

You want to gather statistics on the UPPER(cust_state_province) expression.

To create expression statistics:
1.

Start SQL*Plus and connect to the database as the sh user.

2.

Gather table statistics.
For example, run the following command, specifying the function in the method_opt
argument:
BEGIN
DBMS_STATS.GATHER_TABLE_STATS(
'sh'
, 'customers'
, method_opt => 'FOR ALL COLUMNS SIZE SKEWONLY ' ||
'FOR COLUMNS (LOWER(cust_state_province)) SIZE SKEWONLY'
);
END;

See Also:
Oracle Database PL/SQL Packages and Types Reference to learn about the
DBMS_STATS.GATHER_TABLE_STATS procedure

14.2.3 Displaying Expression Statistics
To obtain information about expression statistics, use the database view
DBA_STAT_EXTENSIONS and the DBMS_STATS.SHOW_EXTENDED_STATS_NAME function.
You can also use views to obtain information such as the number of distinct values,
and whether the column group has a histogram.
Assumptions
This tutorial assumes the following:
•

You created extended statistics for the LOWER(cust_state_province) expression.

•

You want to determine the column group name, the number of distinct values, and
whether a histogram has been created for a column group.

To monitor expression statistics:
1.

Start SQL*Plus and connect to the database as the sh user.

2.

Query the name and definition of the statistics extension.
For example, run the following query:

14-16

Chapter 14

Managing Expression Statistics

COL EXTENSION_NAME FORMAT a30
COL EXTENSION FORMAT a35
SELECT EXTENSION_NAME, EXTENSION
FROM USER_STAT_EXTENSIONS
WHERE TABLE_NAME='CUSTOMERS';

Sample output appears as follows:
EXTENSION_NAME
EXTENSION
------------------------------ -----------------------------SYS_STUBPHJSBRKOIK9O2YV3W8HOUE (LOWER("CUST_STATE_PROVINCE"))
3.

Query the number of distinct values and find whether a histogram has been
created for the expression.
For example, run the following query:
SELECT
FROM
WHERE
AND
AND

e.EXTENSION expression, t.NUM_DISTINCT, t.HISTOGRAM
USER_STAT_EXTENSIONS e, USER_TAB_COL_STATISTICS t
e.EXTENSION_NAME=t.COLUMN_NAME
e.TABLE_NAME=t.TABLE_NAME
t.TABLE_NAME='CUSTOMERS';

EXPRESSION
NUM_DISTINCT
HISTOGRAM
------------------------------------------------------------------(LOWER("CUST_STATE_PROVINCE"))
145
FREQUENCY

See Also:
•

Oracle Database PL/SQL Packages and Types Reference to learn about
the DBMS_STATS.SHOW_EXTENDED_STATS_NAME procedure

•

Oracle Database Reference to learn about the DBA_STAT_EXTENSIONS view

14.2.4 Dropping Expression Statistics
To delete a column group from a table, use the DBMS_STATS.DROP_EXTENDED_STATS
function.
Assumptions
This tutorial assumes the following:
•

You created extended statistics for the LOWER(cust_state_province) expression.

•

You want to drop the expression statistics.

To drop expression statistics:
1.

Start SQL*Plus and connect to the database as the sh user.

2.

Drop the column group.
For example, the following PL/SQL program deletes a column group from the
customers table:
BEGIN
DBMS_STATS.DROP_EXTENDED_STATS(

14-17

Chapter 14

Managing Expression Statistics

'sh'
, 'customers'
, '(LOWER(cust_state_province))'
);
END;
/

See Also:
Oracle Database PL/SQL Packages and Types Reference to learn about the
DBMS_STATS.DROP_EXTENDED_STATS procedure

14-18

15
Controlling the Use of Optimizer Statistics
Using DBMS_STATS, you can specify when and how the optimizer uses statistics.
This section contains the following topics:
•

Locking and Unlocking Optimizer Statistics
You can lock statistics to prevent them from changing.

•

Publishing Pending Optimizer Statistics
By default, the database automatically publishes statistics when the statistics
collection ends.

•

Creating Artificial Optimizer Statistics for Testing
To provide the optimizer with user-created statistics for testing purposes, you can
use the DBMS_STATS.SET_*_STATS procedures. These procedures provide the
optimizer with artificial values for the specified statistics.

15.1 Locking and Unlocking Optimizer Statistics
You can lock statistics to prevent them from changing.
After statistics are locked, you cannot make modifications to the statistics until the
statistics have been unlocked. Locking procedures are useful in a static environment
when you want to guarantee that the statistics and resulting plan never change. For
example, you may want to prevent new statistics from being gathered on a table or
schema by the DBMS_STATS_JOB process, such as highly volatile tables.
When you lock statistics on a table, all dependent statistics are locked. The locked
statistics include table statistics, column statistics, histograms, and dependent index
statistics. To overwrite statistics even when they are locked, you can set the value of
the FORCE argument in various DBMS_STATS procedures, for example, DELETE_*_STATS and
RESTORE_*_STATS, to true.
This section contains the following topics:
•

Locking Statistics
The DBMS_STATS package provides two procedures for locking statistics:
LOCK_SCHEMA_STATS and LOCK_TABLE_STATS.

•

Unlocking Statistics
The DBMS_STATS package provides two procedures for unlocking statistics:
UNLOCK_SCHEMA_STATS and UNLOCK_TABLE_STATS.

15.1.1 Locking Statistics
The DBMS_STATS package provides two procedures for locking statistics:
LOCK_SCHEMA_STATS and LOCK_TABLE_STATS.
Assumptions
This tutorial assumes the following:

15-1

Chapter 15

Locking and Unlocking Optimizer Statistics

•

You gathered statistics on the oe.orders table and on the hr schema.

•

You want to prevent the oe.orders table statistics and hr schema statistics from
changing.

To lock statistics:
1.

Start SQL*Plus and connect to the database as the oe user.

2.

Lock the statistics on oe.orders.
For example, execute the following PL/SQL program:
BEGIN
DBMS_STATS.LOCK_TABLE_STATS('OE','ORDERS');
END;
/

3.

Connect to the database as the hr user.

4.

Lock the statistics in the hr schema.
For example, execute the following PL/SQL program:
BEGIN
DBMS_STATS.LOCK_SCHEMA_STATS('HR');
END;
/

See Also:
Oracle Database PL/SQL Packages and Types Reference to learn about the
DBMS_STATS.LOCK_TABLE_STATS procedure

15.1.2 Unlocking Statistics
The DBMS_STATS package provides two procedures for unlocking statistics:
UNLOCK_SCHEMA_STATS and UNLOCK_TABLE_STATS.
Assumptions
This tutorial assumes the following:
•

You locked statistics on the oe.orders table and on the hr schema.

•

You want to unlock these statistics.

To unlock statistics:
1.

Start SQL*Plus and connect to the database as the oe user.

2.

Unlock the statistics on oe.orders.
For example, execute the following PL/SQL program:
BEGIN
DBMS_STATS.UNLOCK_TABLE_STATS('OE','ORDERS');
END;
/

3.

Connect to the database as the hr user.

15-2

Chapter 15

Publishing Pending Optimizer Statistics

4.

Unlock the statistics in the hr schema.
For example, execute the following PL/SQL program:
BEGIN
DBMS_STATS.UNLOCK_SCHEMA_STATS('HR');
END;
/

See Also:
Oracle Database PL/SQL Packages and Types Reference to learn about the
DBMS_STATS.UNLOCK_TABLE_STATS procedure

15.2 Publishing Pending Optimizer Statistics
By default, the database automatically publishes statistics when the statistics
collection ends.
Alternatively, you can use pending statistics to save the statistics and not publish them
immediately after the collection. This technique is useful for testing queries in a
session with pending statistics. When the test results are satisfactory, you can publish
the statistics to make them available for the entire database.
This section contains the following topics:
•

About Pending Optimizer Statistics
The database stores pending statistics in the data dictionary just as for published
statistics.

•

User Interfaces for Publishing Optimizer Statistics
You can use the DBMS_STATS package to perform operations relating to publishing
statistics.

•

Managing Published and Pending Statistics
This section explains how to use DBMS_STATS program units to change the
publishing behavior of optimizer statistics, and also to export and delete these
statistics.

15.2.1 About Pending Optimizer Statistics
The database stores pending statistics in the data dictionary just as for published
statistics.
By default, the optimizer uses published statistics. You can change the default
behavior by setting the OPTIMIZER_USE_PENDING_STATISTICS initialization parameter to
true (the default is false).
The top part of the following graphic shows the optimizer gathering statistics for the
sh.customers table and storing them in the data dictionary with pending status. The

bottom part of the diagram shows the optimizer using only published statistics to
process a query of sh.customers.

15-3

Chapter 15

Publishing Pending Optimizer Statistics

Figure 15-1

Published and Pending Statistics
Data Dictionary
GATHER_TABLE_STATS

Optimizer Statistics
1 0 0 1 1 1
0 1 0 0 0 1

Optimizer

0 0 1 0 0 0
1 1 0 0 1 0

Published
Statistics

Pending
Statistics

Publishing
preferences
set to false
Customers
Table

Data Dictionary
Optimizer Statistics
SELECT ...
FROM
customers

Optimizer

1 0 0 1 1 1
0 1 0 0 0 1
0 0 1 0 0 0
1 1 0 0 1 0

Published
Statistics

Pending
Statistics

OPTIMIZER_USE_PENDING_STATISTICS=false

In some cases, the optimizer can use a combination of published and pending
statistics. For example, the database stores both published and pending statistics for
the customers table. For the orders table, the database stores only published statistics.
If OPTIMIZER_USE_PENDING_STATS = true, then the optimizer uses pending statistics for
customers and published statistics for orders. If OPTIMIZER_USE_PENDING_STATS = false,
then the optimizer uses published statistics for customers and orders.

See Also:
Oracle Database Reference to learn about the
OPTIMIZER_USE_PENDING_STATISTICS initialization parameter

15-4

Chapter 15

Publishing Pending Optimizer Statistics

15.2.2 User Interfaces for Publishing Optimizer Statistics
You can use the DBMS_STATS package to perform operations relating to publishing
statistics.
The following table lists the relevant program units.
Table 15-1
Statistics

DBMS_STATS Program Units Relevant for Publishing Optimizer

Program Unit

Description

GET_PREFS

Check whether the statistics are automatically published
as soon as DBMS_STATS gathers them. For the parameter
PUBLISH, true indicates that the statistics must be
published when the database gathers them, whereas
false indicates that the database must keep the statistics
pending.

SET_TABLE_PREFS

Set the PUBLISH setting to true or false at the table
level.

SET_SCHEMA_PREFS

Set the PUBLISH setting to true or false at the schema
level.

PUBLISH_PENDING_STATS

Publish valid pending statistics for all objects or only
specified objects.

DELETE_PENDING_STATS

Delete pending statistics.

EXPORT_PENDING_STATS

Export pending statistics.

The initialization parameter OPTIMIZER_USE_PENDING_STATISTICS determines whether the
database uses pending statistics when they are available. The default value is false,
which means that the optimizer uses only published statistics. Set to true to specify
that the optimizer uses any existing pending statistics instead. The best practice is to
set this parameter at the session level rather than at the database level.
You can use access information about published statistics from data dictionary views.
Table 15-2 lists relevant views.
Table 15-2

Views Relevant for Publishing Optimizer Statistics

View

Description

USER_TAB_STATISTICS

Displays optimizer statistics for the tables accessible
to the current user.

USER_TAB_COL_STATISTICS

Displays column statistics and histogram information
extracted from ALL_TAB_COLUMNS.

USER_PART_COL_STATISTICS

Displays column statistics and histogram information
for the table partitions owned by the current user.

USER_SUBPART_COL_STATISTICS

Describes column statistics and histogram
information for subpartitions of partitioned objects
owned by the current user.

USER_IND_STATISTICS

Displays optimizer statistics for the indexes
accessible to the current user.

15-5

Chapter 15

Publishing Pending Optimizer Statistics

Table 15-2

(Cont.) Views Relevant for Publishing Optimizer Statistics

View

Description

USER_TAB_PENDING_STATS

Describes pending statistics for tables, partitions, and
subpartitions accessible to the current user.

USER_COL_PENDING_STATS

Describes the pending statistics of the columns
accessible to the current user.

USER_IND_PENDING_STATS

Describes the pending statistics for tables, partitions,
and subpartitions accessible to the current user
collected using the DBMS_STATS package.

See Also:
•

Oracle Database PL/SQL Packages and Types Reference to learn about
the DBMS_STATS package

•

Oracle Database Reference to learn about USER_TAB_PENDING_STATS and
related views

15.2.3 Managing Published and Pending Statistics
This section explains how to use DBMS_STATS program units to change the publishing
behavior of optimizer statistics, and also to export and delete these statistics.
Assumptions
This tutorial assumes the following:
•

You want to change the preferences for the sh.customers and sh.sales tables so
that newly collected statistics have pending status.

•

You want the current session to use pending statistics.

•

You want to gather and publish pending statistics on the sh.customers table.

•

You gather the pending statistics on the sh.sales table, but decide to delete them
without publishing them.

•

You want to change the preferences for the sh.customers and sh.sales tables so
that newly collected statistics are published.

To manage published and pending statistics:
1.

Start SQL*Plus and connect to the database as user sh.

2.

Query the global optimizer statistics publishing setting.
Run the following query (sample output included):
sh@PROD> SELECT DBMS_STATS.GET_PREFS('PUBLISH') PUBLISH FROM DUAL;
PUBLISH
------TRUE

15-6

Chapter 15

Publishing Pending Optimizer Statistics

The value true indicates that the database publishes statistics as it gathers them.
Every table uses this value unless a specific table preference has been set.
When using GET_PREFS, you can also specify a schema and table name. The
function returns a table preference if it is set. Otherwise, the function returns the
global preference.
3.

Query the pending statistics.
For example, run the following query (sample output included):
sh@PROD> SELECT * FROM USER_TAB_PENDING_STATS;
no rows selected

This example shows that the database currently stores no pending statistics for
the sh schema.
4.

Change the publishing preferences for the sh.customers table.
For example, execute the following procedure so that statistics are marked as
pending:
BEGIN
DBMS_STATS.SET_TABLE_PREFS('sh', 'customers', 'publish', 'false');
END;
/

Subsequently, when you gather statistics on the customers table, the database
does not automatically publish statistics when the gather job completes. Instead,
the database stores the newly gathered statistics in the USER_TAB_PENDING_STATS
table.
5.

Gather statistics for sh.customers.
For example, run the following program:
BEGIN
DBMS_STATS.GATHER_TABLE_STATS('sh','customers');
END;
/

6.

Query the pending statistics.
For example, run the following query (sample output included):
sh@PROD> SELECT TABLE_NAME, NUM_ROWS FROM USER_TAB_PENDING_STATS;
TABLE_NAME
NUM_ROWS
------------------------------ ---------CUSTOMERS
55500

This example shows that the database now stores pending statistics for the
sh.customers table.
7.

Instruct the optimizer to use the pending statistics in this session.
Set the initialization parameter OPTIMIZER_USE_PENDING_STATISTICS to true as
shown:
ALTER SESSION SET OPTIMIZER_USE_PENDING_STATISTICS = true;

8.

Run a workload.

15-7

Chapter 15

Publishing Pending Optimizer Statistics

The following example changes the email addresses of all customers named
Bruce Chalmers:
UPDATE
SET
WHERE
AND
COMMIT;

sh.customers
cust_email='ChalmersB@company.com'
cust_first_name = 'Bruce'
cust_last_name = 'Chalmers';

The optimizer uses the pending statistics instead of the published statistics when
compiling all SQL statements in this session.
9.

Publish the pending statistics for sh.customers.
For example, execute the following program:
BEGIN
DBMS_STATS.PUBLISH_PENDING_STATS('SH','CUSTOMERS');
END;
/

10. Change the publishing preferences for the sh.sales table.

For example, execute the following program:
BEGIN
DBMS_STATS.SET_TABLE_PREFS('sh', 'sales', 'publish', 'false');
END;
/

Subsequently, when you gather statistics on the sh.sales table, the database does
not automatically publish statistics when the gather job completes. Instead, the
database stores the statistics in the USER_TAB_PENDING_STATS table.
11. Gather statistics for sh.sales.

For example, run the following program:
BEGIN
DBMS_STATS.GATHER_TABLE_STATS('sh','sales');
END;
/
12. Delete the pending statistics for sh.sales.

Assume you change your mind and now want to delete pending statistics for
sh.sales. Run the following program:
BEGIN
DBMS_STATS.DELETE_PENDING_STATS('sh','sales');
END;
/
13. Change the publishing preferences for the sh.customers and sh.sales tables back

to their default setting.
For example, execute the following program:
BEGIN
DBMS_STATS.SET_TABLE_PREFS('sh', 'customers', 'publish', null);
DBMS_STATS.SET_TABLE_PREFS('sh', 'sales', 'publish', null);
END;
/

15-8

Chapter 15

Creating Artificial Optimizer Statistics for Testing

15.3 Creating Artificial Optimizer Statistics for Testing
To provide the optimizer with user-created statistics for testing purposes, you can use
the DBMS_STATS.SET_*_STATS procedures. These procedures provide the optimizer with
artificial values for the specified statistics.
This section contains the following topics:
•

About Artificial Optimizer Statistics
For testing purposes, you can manually create artificial statistics for a table, index,
or the system using the DBMS_STATS.SET_*_STATS procedures. These procedures
insert the artificial statistics into the data dictionary directly (when stattab is null) or
into a user-created table.

•

Setting Artificial Optimizer Statistics for a Table
This topic explains how to set artificial statistics for a table using
DBMS_STATS.SET_TABLE_STATS. The basic steps are the same for SET_INDEX_STATS and
SET_SYSTEM_STATS.

•

Setting Optimizer Statistics: Example
This example shows how to gather optimizer statistics for a table, set artificial
statistics, and then compare the plans that the optimizer chooses based on the
differing statistics.

15.3.1 About Artificial Optimizer Statistics
For testing purposes, you can manually create artificial statistics for a table, index, or
the system using the DBMS_STATS.SET_*_STATS procedures. These procedures insert the
artificial statistics into the data dictionary directly (when stattab is null) or into a usercreated table.

Caution:
The DBMS_STATS.SET_*_STATS procedures are intended for development testing
only. Do not use them in a production database. If you set statistics in the
data dictionary, then Oracle Database considers the set statistics as the
“real” statistics, which means that statistics gathering jobs may not re-gather
artificial statistics when they do not meet the criteria for staleness.
The most typical use cases for the DBMS_STATS.SET_*_STATS procedures are:
•

Showing how execution plans change as the numbers of rows or blocks in a table
change
For example, SET_TABLE_STATS can set number of rows and blocks in a small or
empty table to a large number. When you execute a query using the altered
statistics, the optimizer may change the execution plan. For example, the
increased row count may lead the optimizer to choose an index scan rather than a
full table scan. By experimenting with different values, you can see how the
optimizer will change its execution plan over time.

•

Creating realistic statistics for temporary tables

15-9

Chapter 15

Creating Artificial Optimizer Statistics for Testing

You may want to see what the optimizer does when a large temporary table is
referenced in multiple SQL statements. You can create a regular table, load
representative data, and then use GET_TABLE_STATS to retrieve the statistics. After
you create the temporary table, you can “deceive” the optimizer into using these
statistics by invoking SET_TABLE_STATS.
Optionally, you can specify a unique ID for statistics in a user-created table. The
SET_*_STATS procedures have corresponding GET_*_STATS procedures.
Table 15-3

DBMS_STATS Procedures for Setting Optimizer Statistics

DBMS_STATS Procedure

Description

SET_TABLE_STATS

Sets table or partition statistics using parameters such as
numrows, numblks, avgrlen, and so on.
If the database uses the In-Memory Column store, you can set
im_imcu_count to the number of IMCUs in the table or partition,
and im_block_count to the number of blocks in the table or
partition. For an external table, scanrate specifies the rate at
which data is scanned in MB/second.
The optimizer uses the cached data to estimate the number of
cached blocks for index or statistics table access. The total cost
is the I/O cost of reading data blocks from disk, the CPU cost of
reading cached blocks from the buffer cache, and the CPU cost
of processing the data.

SET_COLUMN_STATS

Sets column statistics using parameters such as distcnt,
density, nullcnt, and so on.
In the version of this procedure that deals with user-defined
statistics, use stattypname to specify the type of statistics to
store in the data dictionary.

SET_SYSTEM_STATS

Sets system statistics using parameters such as iotfrspeed,
sreadtim, cpuspeed, and so on.

SET_INDEX_STATS

Sets index statistics using parameters such as numrows,
numlblks, avglblk, clstfct, indlevel, and so on.
In the version of this procedure that deals with user-defined
statistics, use stattypname to specify the type of statistics to
store in the data dictionary.

See Also:
Oracle Database PL/SQL Packages and Types Reference to learn more
about DBMS_STATS.SET_TABLE_STATS and the other procedures for setting
optimizer statistics

15.3.2 Setting Artificial Optimizer Statistics for a Table
This topic explains how to set artificial statistics for a table using
DBMS_STATS.SET_TABLE_STATS. The basic steps are the same for SET_INDEX_STATS and
SET_SYSTEM_STATS.

Note the following task prerequisites:

15-10

Chapter 15

Creating Artificial Optimizer Statistics for Testing

•

For an object not owned by SYS, you must be the owner of the object, or have the
ANALYZE ANY privilege.

•

For an object owned by SYS, you must have the ANALYZE ANY DICTIONARY privilege or
the SYSDBA privilege.

•

When invoking GET_*_STATS for a table, column, or index, the referenced object
must exist.

This task assumes the following:
•

You have the required privileges to use DBMS_STATS.SET_TABLE_STATS for the
specified table.

•

You intend to store the statistics in the data dictionary.

1.

In SQL*Plus, log in to the database as a user with the required privileges.

2.

Run the DBMS_STATS.SET_TABLE_STATS procedure, specifying the appropriate
parameters for the statistics.
Typical parameters include the following:
•

ownname (not null)

This parameter specifies the name of the schema containing the table.
•

tabname (not null)

This parameter specifies the name of the table whose statistics you intend to
set.
•

partname

This parameter specifies the name of a partition of the table.
•

numrows

This parameter specifies the number of rows in the table.
•

numblks

This parameter specifies the number of blocks in the table.
3.

Query the table.

4.

Optionally, to determine how the statistics affected the optimizer, query the
execution plan.

5.

Optionally, to perform further testing, return to Step 2 and reset the optimizer
statistics.

15.3.3 Setting Optimizer Statistics: Example
This example shows how to gather optimizer statistics for a table, set artificial
statistics, and then compare the plans that the optimizer chooses based on the
differing statistics.
This example assumes:
•

You are logged in to the database as a user with DBA privileges.

•

You want to test when the optimizer chooses an index scan.

1.

Create a table called contractors, and index the salary column.

15-11

Chapter 15

Creating Artificial Optimizer Statistics for Testing

CREATE TABLE contractors (
con_id
NUMBER,
last_name VARCHAR2(50),
salary
NUMBER,
CONSTRAINT cond_id_pk PRIMARY KEY(con_id) );
CREATE INDEX salary_ix ON contractors(salary);
2.

Insert a single row into this table.
INSERT INTO contractors VALUES (8, 'JONES',1000);
COMMIT;

3.

Gather statistics for the table.
EXECUTE DBMS_STATS.GATHER_TABLE_STATS( user, tabname => 'CONTRACTORS' );

4.

Query the number of rows for the table and index (sample output included):
SQL> SELECT NUM_ROWS FROM USER_TABLES WHERE TABLE_NAME = 'CONTRACTORS';
NUM_ROWS
---------1
SQL> SELECT NUM_ROWS FROM USER_INDEXES WHERE INDEX_NAME = 'SALARY_IX';
NUM_ROWS
---------1

5.

Query contractors whose salary is 1000, using the dynamic_sampling hint to disable
dynamic sampling:
SELECT /*+ dynamic_sampling(contractors 0) */ *
FROM contractors
WHERE salary = 1000;

6.

Query the execution plan chosen by the optimizer (sample output included):
SQL> SELECT * FROM TABLE(DBMS_XPLAN.DISPLAY_CURSOR);
SQL_ID
cy0wzytc16g9n, child number 0
------------------------------------SELECT /*+ dynamic_sampling(contractors 0) */ * FROM contractors WHERE
salary = 1000
Plan hash value: 5038823
---------------------------------------------------------------------| Id | Operation
| Name
|Rows|Bytes|Cost (%CPU)| Time|
---------------------------------------------------------------------| 0 | SELECT STATEMENT |
| |
| 2 (100)|
|
|* 1 | TABLE ACCESS FULL| CONTRACTORS | 1 | 12 | 2 (0)| 00:00:01 |
---------------------------------------------------------------------Predicate Information (identified by operation id):
--------------------------------------------------1 - filter("SALARY"=1000)
19 rows selected.

15-12

Chapter 15

Creating Artificial Optimizer Statistics for Testing

Because only 1 row exists in the table, the optimizer chooses a full table scan over
an index range scan.
7.

Use SET_TABLE_STATS and SET_INDEX_STATS to simulate statistics for a table with
2000 rows stored in 10 data blocks:
BEGIN
DBMS_STATS.SET_TABLE_STATS(
ownname => user
, tabname => 'CONTRACTORS'
, numrows => 2000
, numblks => 10 );
END;
/
BEGIN
DBMS_STATS.SET_INDEX_STATS(
ownname => user
, indname => 'SALARY_IX'
, numrows => 2000 );
END;
/

8.

Query the number of rows for the table and index (sample output included):
SQL> SELECT NUM_ROWS FROM USER_TABLES WHERE TABLE_NAME = 'CONTRACTORS';
NUM_ROWS
---------2000
SQL> SELECT NUM_ROWS FROM USER_INDEXES WHERE INDEX_NAME = 'SALARY_IX';
NUM_ROWS
---------2000

Now the optimizer believes that the table contains 2000 rows in 10 blocks, even
though only 1 row actually exists in one block.
9.

Flush the shared pool to eliminate possibility of plan reuse, and then execute the
same query of contractors:
ALTER SYSTEM FLUSH SHARED_POOL;
SELECT /*+ dynamic_sampling(contractors 0) */ *
FROM contractors
WHERE salary = 1000;

10. Query the execution plan chosen by the optimizer based on the artificial statistics

(sample output included):
SQL> SELECT * FROM TABLE(DBMS_XPLAN.DISPLAY_CURSOR);
SQL_ID
cy0wzytc16g9n, child number 0
------------------------------------SELECT /*+ dynamic_sampling(contractors 0) */ * FROM contractors WHERE
salary = 1000
Plan hash value: 996794789
--------------------------------------------------------------------------------|Id| Operation
| Name
|Rows|Bytes|Cost(%CPU)|Time|

15-13

Chapter 15

Creating Artificial Optimizer Statistics for Testing

--------------------------------------------------------------------------------| 0| SELECT STATEMENT
|
|
|
|3(100)|
|
| 1| TABLE ACCESS BY INDEX ROWID BATCHED|CONTRACTORS|2000|24000|3 (34)|00:00:01|
|*2| INDEX RANGE SCAN
|SALARY_IX |2000|
|1 (0)|00:00:01|
--------------------------------------------------------------------------------Predicate Information (identified by operation id):
--------------------------------------------------2 - access("SALARY"=1000)
20 rows selected.

Based on the artificially generated statistics for the number of rows and block
distribution, the optimizer considers an index range scan more cost-effective.

15-14

16
Managing Historical Optimizer Statistics
This chapter how to retain, report on, and restore non-current statistics.
This chapter contains the following topics:
•

Restoring Optimizer Statistics
You can use DBMS_STATS to restore old versions of statistics that are stored in the
data dictionary.

•

Managing Optimizer Statistics Retention
By default, the database retains optimizer statistics for 31 days, after which time
the statistics are scheduled for purging.

•

Reporting on Past Statistics Gathering Operations
You can use DBMS_STATS functions to report on a specific statistics gathering
operation or on operations that occurred during a specified time.

16.1 Restoring Optimizer Statistics
You can use DBMS_STATS to restore old versions of statistics that are stored in the data
dictionary.
This topic contains the following topics:
•

About Restore Operations for Optimizer Statistics
Whenever statistics in the data dictionary are modified, the database automatically
saves old versions of statistics. If newly collected statistics lead to suboptimal
execution plans, then you may want to revert to the previous statistics.

•

Guidelines for Restoring Optimizer Statistics
Restoring statistics is similar to importing and exporting statistics.

•

Restrictions for Restoring Optimizer Statistics
When restoring previous versions of statistics, various limitations apply.

•

Restoring Optimizer Statistics Using DBMS_STATS
You can restore statistics using the DBMS_STATS.RESTORE_*_STATS procedures.

16.1.1 About Restore Operations for Optimizer Statistics
Whenever statistics in the data dictionary are modified, the database automatically
saves old versions of statistics. If newly collected statistics lead to suboptimal
execution plans, then you may want to revert to the previous statistics.
Restoring optimizer statistics can aid in troubleshooting suboptimal plans. The
following graphic illustrates a timeline for restoring statistics. In the graphic, statistics
collection occurs on August 10 and August 20. On August 24, the DBA determines
that the current statistics may be causing the optimizer to generate suboptimal plans.
On August 25, the administrator restores the statistics collected on August 10.

16-1

Chapter 16

Restoring Optimizer Statistics

Figure 16-1

Restoring Optimizer Statistics

8/10

8/20

8/24

8/25

AAAAA

BBBBB

BBBBB

AAAAA

Statistics
Gathered

Statistics
Gathered

Recent Statistics
May Be Causing
Suboptimal Plans

8/10 Statistics
Restored

16.1.2 Guidelines for Restoring Optimizer Statistics
Restoring statistics is similar to importing and exporting statistics.
In general, restore statistics instead of exporting them in the following situations:
•

You want to recover older versions of the statistics. For example, you want to
restore the optimizer behavior to an earlier date.

•

You want the database to manage the retention and purging of statistics histories.

Export statistics rather than restoring them in the following situations:
•

You want to experiment with multiple sets of statistics and change the values back
and forth.

•

You want to move the statistics from one database to another database. For
example, moving statistics from a production system to a test system.

•

You want to preserve a known set of statistics for a longer period than the desired
retention date for restoring statistics.

See Also:
Oracle Database PL/SQL Packages and Types Reference for an overview of
the procedures for restoring and importing statistics

16.1.3 Restrictions for Restoring Optimizer Statistics
When restoring previous versions of statistics, various limitations apply.
Limitations include the following:
•

DBMS_STATS.RESTORE_*_STATS procedures cannot restore user-defined statistics.

•

Old versions of statistics are not stored when the ANALYZE command has been
used for collecting statistics.

•

When you drop a table, workload information used by the auto-histogram
gathering feature and saved statistics history used by the RESTORE_*_STATS
procedures is lost. Without this data, these features do not function properly. To
remove all rows from a table, and to restore these statistics with DBMS_STATS, use
TRUNCATE instead of dropping and re-creating the same table.

16-2

Chapter 16

Restoring Optimizer Statistics

16.1.4 Restoring Optimizer Statistics Using DBMS_STATS
You can restore statistics using the DBMS_STATS.RESTORE_*_STATS procedures.
The procedures listed in the following table accept a timestamp as an argument and
restore statistics as of the specified time (as_of_timestamp).
Table 16-1

DBMS_STATS Restore Procedures

Procedure

Description

RESTORE_DICTIONARY_STATS

Restores statistics of all dictionary tables (tables of
SYS, SYSTEM, and RDBMS component schemas) as of a
specified timestamp.

RESTORE_FIXED_OBJECTS_STATS

Restores statistics of all fixed tables as of a specified
timestamp.

RESTORE_SCHEMA_STATS

Restores statistics of all tables of a schema as of a
specified timestamp.

RESTORE_SYSTEM_STATS

Restores system statistics as of a specified
timestamp.

RESTORE_TABLE_STATS

Restores statistics of a table as of a specified
timestamp. The procedure also restores statistics of
associated indexes and columns. If the table statistics
were locked at the specified timestamp, then the
procedure locks the statistics.

Dictionary views display the time of statistics modifications. You can use the following
views to determine the time stamp to be use for the restore operation:
•

The DBA_OPTSTAT_OPERATIONS view contain history of statistics operations performed
at schema and database level using DBMS_STATS.

•

The DBA_TAB_STATS_HISTORY views contains a history of table statistics
modifications.

Assumptions
This tutorial assumes the following:
•

After the most recent statistics collection for the oe.orders table, the optimizer
began choosing suboptimal plans for queries of this table.

•

You want to restore the statistics from before the most recent statistics collection
to see if the plans improve.

To restore optimizer statistics:
1.

Start SQL*Plus and connect to the database with administrator privileges.

2.

Query the statistics history for oe.orders.
For example, run the following query:
COL TABLE_NAME FORMAT a10
SELECT TABLE_NAME,
TO_CHAR(STATS_UPDATE_TIME,'YYYY-MM-DD:HH24:MI:SS') AS STATS_MOD_TIME
FROM DBA_TAB_STATS_HISTORY
WHERE TABLE_NAME='ORDERS'

16-3

Chapter 16

Managing Optimizer Statistics Retention

AND
OWNER='OE'
ORDER BY STATS_UPDATE_TIME DESC;

Sample output is as follows:
TABLE_NAME
---------ORDERS
ORDERS
3.

STATS_MOD_TIME
------------------2012-08-20:11:36:38
2012-08-10:11:06:20

Restore the optimizer statistics to the previous modification time.
For example, restore the oe.orders table statistics to August 10, 2012:
BEGIN
DBMS_STATS.RESTORE_TABLE_STATS( 'OE','ORDERS',
TO_TIMESTAMP('2012-08-10:11:06:20','YYYY-MM-DD:HH24:MI:SS') );
END;
/

You can specify any date between 8/10 and 8/20 because DBMS_STATS restores
statistics as of the specified time.

See Also:
Oracle Database PL/SQL Packages and Types Reference to learn more
about the DBMS_STATS.RESTORE_TABLE_STATS procedure

16.2 Managing Optimizer Statistics Retention
By default, the database retains optimizer statistics for 31 days, after which time the
statistics are scheduled for purging.
You can use the DBMS_STATS package to determine the retention period, change the
period, and manually purge old statistics.
This section contains the following topics:
•

Obtaining Optimizer Statistics History
You can use DBMS_STATS procedures to obtain historical information for optimizer
statistics.

•

Changing the Optimizer Statistics Retention Period
You can configure the retention period using the
DBMS_STATS.ALTER_STATS_HISTORY_RETENTION procedure. The default is 31 days.

•

Purging Optimizer Statistics
Automatic purging is enabled when the STATISTICS_LEVEL initialization parameter is
set to TYPICAL or ALL.

16-4

Chapter 16

Managing Optimizer Statistics Retention

16.2.1 Obtaining Optimizer Statistics History
You can use DBMS_STATS procedures to obtain historical information for optimizer
statistics.
Historical information is useful when you want to determine how long the database
retains optimizer statistics, and how far back these statistics can be restored. You can
use the following procedure to obtain information about the optimizer statistics history:
•

GET_STATS_HISTORY_RETENTION

This function can retrieve the current statistics history retention value.
•

GET_STATS_HISTORY_AVAILABILITY

This function retrieves the oldest time stamp when statistics history is available.
Users cannot restore statistics to a time stamp older than the oldest time stamp.
To obtain optimizer statistics history information:
1.

Start SQL*Plus and connect to the database with the necessary privileges.

2.

Execute the following PL/SQL program:
DECLARE
v_stats_retn NUMBER;
v_stats_date DATE;
BEGIN
v_stats_retn := DBMS_STATS.GET_STATS_HISTORY_RETENTION;
DBMS_OUTPUT.PUT_LINE('The retention setting is ' || v_stats_retn || '.');
v_stats_date := DBMS_STATS.GET_STATS_HISTORY_AVAILABILITY;
DBMS_OUTPUT.PUT_LINE('Earliest restore date is ' || v_stats_date || '.');
END;
/

See Also:
Oracle Database PL/SQL Packages and Types Reference to learn about the
DBMS_STATS.GET_STATS_HISTORY_RETENTION procedure

16.2.2 Changing the Optimizer Statistics Retention Period
You can configure the retention period using the
DBMS_STATS.ALTER_STATS_HISTORY_RETENTION procedure. The default is 31 days.

Prerequisites
To run this procedure, you must have either the SYSDBA privilege, or both the ANALYZE
ANY DICTIONARY and ANALYZE ANY system privileges.
Assumptions
This tutorial assumes the following:
•

The current retention period for optimizer statistics is 31 days.

16-5

Chapter 16

Managing Optimizer Statistics Retention

•

You run queries annually as part of an annual report. To keep the statistics history
for more than 365 days so that you have access to last year's plan (in case a
suboptimal plan occurs now), you set the retention period to 366 days.

•

You want to create a PL/SQL procedure set_opt_stats_retention that you can use
to change the optimizer statistics retention period.

To change the optimizer statistics retention period:
1.

Start SQL*Plus and connect to the database with the necessary privileges.

2.

Create a procedure that changes the retention period.
For example, create the following procedure:
CREATE OR REPLACE PROCEDURE set_opt_stats_retention
( p_stats_retn IN NUMBER )
IS
v_stats_retn NUMBER;
BEGIN
v_stats_retn := DBMS_STATS.GET_STATS_HISTORY_RETENTION;
DBMS_OUTPUT.PUT_LINE('Old retention setting is ' ||v_stats_retn|| '.');
DBMS_STATS.ALTER_STATS_HISTORY_RETENTION(p_stats_retn);
v_stats_retn := DBMS_STATS.GET_STATS_HISTORY_RETENTION;
DBMS_OUTPUT.PUT_LINE('New retention setting is ' ||v_stats_retn|| '.');
END;
/

3.

Change the retention period to 366 days.
For example, execute the procedure that you created in the previous step (sample
output included):
SQL> EXECUTE set_opt_stats_retention(366)
The old retention setting is 31.
The new retention setting is 366.
PL/SQL procedure successfully completed.

See Also:
Oracle Database PL/SQL Packages and Types Reference to learn about the
DBMS_STATS.ALTER_STATS_HISTORY_RETENTION procedure

16.2.3 Purging Optimizer Statistics
Automatic purging is enabled when the STATISTICS_LEVEL initialization parameter is set
to TYPICAL or ALL.
The database purges all history older than the older of (current time - the
ALTER_STATS_HISTORY_RETENTION setting) and (time of the most recent statistics gathering
- 1).
You can purge old statistics manually using the PURGE_STATS procedure. If you do not
specify an argument, then this procedure uses the automatic purging policy. If you
specify the before_timestamp parameter, then the database purges statistics saved
before the specified timestamp.

16-6

Chapter 16

Reporting on Past Statistics Gathering Operations

Prerequisites
To run this procedure, you must have either the SYSDBA privilege, or both the ANALYZE
ANY DICTIONARY and ANALYZE ANY system privileges.
Assumptions
This tutorial assumes that you want to purge statistics more than one week old.
To purge optimizer statistics:
1.

In SQL*Plus, log in to the database with the necessary privileges.

2.

Execute the DBMS_STATS.PURGE_STATS procedure.
For example, execute the procedure as follows:
EXEC DBMS_STATS.PURGE_STATS( SYSDATE-7 );

See Also:
Oracle Database PL/SQL Packages and Types Reference to learn about the
DBMS_STATS.PURGE_STATS procedure

16.3 Reporting on Past Statistics Gathering Operations
You can use DBMS_STATS functions to report on a specific statistics gathering operation
or on operations that occurred during a specified time.
Table 16-2 lists the functions.
Table 16-2

DBMS_STATS Reporting Functions

Function

Description

REPORT_STATS_OPERATIONS

Generates a report of all statistics operations that
occurred between two points in time. You can
narrow the scope of the report to include only
automatic statistics gathering runs. You can also
provide a set of pluggable database (PDB) IDs so
that the database reports only statistics operations
from the specified PDBs.

REPORT_SINGLE_STATS_OPERATION

Generates a report of the specified operation.
Optionally, you can specify a particular PDB ID in a
container database (CDB).

Assumptions
This tutorial assumes that you want to generate HTML reports of the following:
•

All statistics gathering operations within the last day

•

The most recent statistics gathering operation

16-7

Chapter 16

Reporting on Past Statistics Gathering Operations

To report on all operations in the past day:
1.

Start SQL*Plus and connect to the database with administrator privileges.

2.

Run the DBMS_STATS.REPORT_STATS_OPERATIONS function.
For example, run the following commands:
SET LINES 200 PAGES 0
SET LONG 100000
COLUMN REPORT FORMAT A200
VARIABLE my_report CLOB;
BEGIN
:my_report := DBMS_STATS.REPORT_STATS_OPERATIONS (
since
=> SYSDATE-1
,
until
=> SYSDATE
,
detail_level => 'TYPICAL'
,
format
=> 'HTML'
);
END;
/

The following graphic shows a sample report:

3.

Run the DBMS_STATS.REPORT_SINGLE_STATS_OPERATION function for an individual
operation.
For example, run the following program to generate a report of operation 848:
BEGIN
:my_report :=DBMS_STATS.REPORT_SINGLE_STATS_OPERATION (
OPID
=> 848
,
FORMAT => 'HTML'
);
END;

The following graphic shows a sample report:

16-8

Chapter 16

Reporting on Past Statistics Gathering Operations

See Also:
•

"Graphical Interface for Optimizer Statistics Management" to learn about
the Cloud Control GUI for statistics management

•

Oracle Database PL/SQL Packages and Types Reference to learn more
about DBMS_STATS

16-9

17
Transporting Optimizer Statistics
You can export and import optimizer statistics from the data dictionary to user-defined
statistics tables. You can also copy statistics from one database to another database.
This chapter contains the following topics:
•

About Transporting Optimizer Statistics
When you transport optimizer statistics between databases, you must use
DBMS_STATS to copy the statistics to and from a staging table, and tools to make the
table contents accessible to the destination database.

•

Transporting Optimizer Statistics to a Test Database: Tutorial
You can transport schema statistics from a production database to a test database
using Oracle Data Pump.

17.1 About Transporting Optimizer Statistics
When you transport optimizer statistics between databases, you must use DBMS_STATS
to copy the statistics to and from a staging table, and tools to make the table contents
accessible to the destination database.
Importing and exporting are especially useful for testing an application using
production statistics. You use DBMS_STATS.EXPORT_SCHEMA_STATS to export schema
statistics from a production database to a test database so that developers can tune
execution plans in a realistic environment before deploying applications.
The following figure illustrates the process using Oracle Data Pump and ftp.

Figure 17-1

Transporting Optimizer Statistics

Production
Database

Test
Database

Data Dictionary
EXPORT_SCHEMA_STATS

Data Dictionary
IMPORT_SCHEMA_STATS

Staging Table

Staging Table

Data Pump
Export

Data Pump
Import

.dmp
file

Transport ftp, nfs

.dmp
file

17-1

Chapter 17

Transporting Optimizer Statistics to a Test Database: Tutorial

As shown in Figure 17-1, the basic steps are as follows:
1.

In the production database, copy the statistics from the data dictionary to a staging
table using DBMS_STATS.EXPORT_SCHEMA_STATS.

2.

Export the statistics from the staging table to a .dmp file using Oracle Data Pump.

3.

Transfer the .dmp file from the production host to the test host using a transfer tool
such as ftp.

4.

In the test database, import the statistics from the .dmp file to a staging table using
Oracle Data Pump.

5.

Copy the statistics from the staging table to the data dictionary using
DBMS_STATS.IMPORT_SCHEMA_STATS.

17.2 Transporting Optimizer Statistics to a Test Database:
Tutorial
You can transport schema statistics from a production database to a test database
using Oracle Data Pump.
Prerequisites and Restrictions
When preparing to export optimizer statistics, note the following:
•

Before exporting statistics, you must create a table to hold the statistics. The
procedure DBMS_STATS.CREATE_STAT_TABLE creates the statistics table.

•

The optimizer does not use statistics stored in a user-owned table. The only
statistics used by the optimizer are the statistics stored in the data dictionary. To
make the optimizer use statistics in user-defined tables, import these statistics into
the data dictionary using the DBMS_STATS import procedure.

•

The Data Pump Export and Import utilities export and import optimizer statistics
from the database along with the table. When a column has system-generated
names, Original Export (exp) does not export statistics with the data, but this
restriction does not apply to Data Pump Export.

Note:
Exporting and importing statistics using DBMS_STATS is a distinct operation
from using Data Pump Export and Import.

Assumptions
This tutorial assumes the following:
•

You want to generate representative sh schema statistics on a production
database and use DBMS_STATS to import them into a test database.

•

Administrative user dba1 exists on both production and test databases.

•

You intend to create table opt_stats to store the schema statistics.

•

You intend to use Oracle Data Pump to export and import table opt_stats.

17-2

Chapter 17

Transporting Optimizer Statistics to a Test Database: Tutorial

To generate schema statistics and import them into a separate database:
1.

On the production host, start SQL*Plus and connect to the production database as
administrator dba1.

2.

Create a table to hold the production statistics.
For example, execute the following PL/SQL program to create user statistics table
opt_stats:
BEGIN
DBMS_STATS.CREATE_STAT_TABLE (
ownname => 'dba1'
, stattab => 'opt_stats'
);
END;
/

3.

Gather schema statistics.
For example, manually gather schema statistics as follows:
-- generate representative workload
EXEC DBMS_STATS.GATHER_SCHEMA_STATS('SH');

4.

Use DBMS_STATS to export the statistics.
For example, retrieve schema statistics and store them in the opt_stats table
created previously:
BEGIN
DBMS_STATS.EXPORT_SCHEMA_STATS (
ownname => 'dba1'
, stattab => 'opt_stats'
);
END;
/

5.

Use Oracle Data Pump to export the contents of the statistics table.
For example, run the expdp command at the operating schema prompt:
expdp dba1 DIRECTORY=dpump_dir1 DUMPFILE=stat.dmp TABLES=opt_stats

6.

Transfer the dump file to the test database host.

7.

Log in to the test host, and then use Oracle Data Pump to import the contents of
the statistics table.
For example, run the impdp command at the operating schema prompt:
impdp dba1 DIRECTORY=dpump_dir1 DUMPFILE=stat.dmp TABLES=opt_stats

8.

On the test host, start SQL*Plus and connect to the test database as administrator
dba1.

9.

Use DBMS_STATS to import statistics from the user statistics table and store them in
the data dictionary.
The following PL/SQL program imports schema statistics from table opt_stats into
the data dictionary:
BEGIN
DBMS_STATS.IMPORT_SCHEMA_STATS(
ownname => 'dba1'
, stattab => 'opt_stats'

17-3

Chapter 17

Transporting Optimizer Statistics to a Test Database: Tutorial

);
END;
/

See Also:
•

Oracle Database PL/SQL Packages and Types Reference to learn about
the DBMS_STATS.CREATE_STAT_TABLE function

•

Oracle Database PL/SQL Packages and Types Reference for an
overview of the statistics transfer functions

•

Oracle Database Utilities to learn about Oracle Data Pump

17-4

18
Analyzing Statistics Using Optimizer
Statistics Advisor
Optimizer Statistics Advisor analyzes how optimizer statistics are gathered, and then
makes recommendations.
This chapter contains the following topics:
•

About Optimizer Statistics Advisor
Optimizer Statistics Advisor is built-in diagnostic software that analyzes the quality
of statistics and statistics-related tasks.

•

Basic Tasks for Optimizer Statistics Advisor
This section explains the basic workflow for using Optimizer Statistics Advisor. All
procedures and functions are in the DBMS_STATS package.

18.1 About Optimizer Statistics Advisor
Optimizer Statistics Advisor is built-in diagnostic software that analyzes the quality of
statistics and statistics-related tasks.
The advisor task runs automatically in the maintenance window, but you can also run it
on demand. You can then view the advisor report. If the advisor makes
recommendations, then in some cases you can run system-generated scripts to
implement them.
The following figure provides a conceptual overview of Optimizer Statistics Advisor.

18-1

Chapter 18

About Optimizer Statistics Advisor

Figure 18-1

Optimizer Statistics Advisor

Optimizer
Automatic
Tuning
Optimizer

Filter Options

DBA

Optimizer
Statistics
Advisor

Findings
Recommendations

Task

Actions

Rules

Data Dictionary
and V$ Views

DBA_OPSTAT_OPERATIONS

This section contains the following topics:
•

Purpose of Optimizer Statistics Advisor
Optimizer Statistics Advisor inspects how optimizer statistics are gathered.

•

Optimizer Statistics Advisor Concepts
Optimizer Statistics Advisor uses the same advisor framework as Automatic
Database Diagnostic Monitor (ADDM), SQL Performance Analyzer, and other
advisors.

•

Command-Line Interface to Optimizer Statistics Advisor
Perform Optimizer Statistics Advisor tasks using the DBMS_STATS PL/SQL package.

18.1.1 Purpose of Optimizer Statistics Advisor
Optimizer Statistics Advisor inspects how optimizer statistics are gathered.
The advisor automatically diagnoses problems in the existing practices for gathering
statistics. The advisor does not gather a new or alternative set of optimizer statistics.
The output of the advisor is a report of findings and recommendations, which helps
you follow best practices for gathering statistics.

18-2

Chapter 18

About Optimizer Statistics Advisor

Optimizer statistics play a significant part in determining the execution plan for queries.
Therefore, it is critical for the optimizer to gather and maintain accurate and up-to-date
statistics. The optimizer provides the DBMS_STATS package, which evolves from release
to release, for this purpose. Typically, users develop their own strategies for gathering
statistics based on specific workloads, and then use homegrown scripts to implement
these strategies.
This section contains the following topics:
•

Problems with a Traditional Script-Based Approach
The advantage of the scripted approach is that the scripts are typically tested and
reviewed. However, the owner of suboptimal legacy scripts may not change them
for fear of causing plan changes.

•

Advantages of Optimizer Statistics Advisor
An advisor-based approach offers better scalability and maintainability than the
traditional approach.

18.1.1.1 Problems with a Traditional Script-Based Approach
The advantage of the scripted approach is that the scripts are typically tested and
reviewed. However, the owner of suboptimal legacy scripts may not change them for
fear of causing plan changes.
The traditional approach has the following problems:
•

Legacy scripts may not keep pace with new best practices, which can change from
release to release.
Frequently, successive releases add enhancements to histograms, sampling,
workload monitoring, concurrency, and other optimizer-related features. For
example, in Oracle Database 12c, Oracle recommends setting AUTO_SAMPLE_SIZE
instead of a percentage. However, legacy scripts typically specify a sampling
percentage, which may lead to suboptimal execution plans.

•

Resources are wasted on unnecessary statistics gathering.
A script may gather statistics multiple times each day on the same table.

•

Automatic statistics gathering jobs do not guarantee accurate and up-to-date
statistics.
For example, sometimes the automatic statistics gathering job is not running
because an initialization parameter combination disables it, or the job is
terminated. Moreover, sometimes the automatic job maintenance window is
insufficient because of resource constraints, or because too many objects require
statistics collection. Jobs that stop running before gathering all statistics cause
either no statistics or stale statistics for some objects, which can in turn cause
suboptimal plans.

•

Statistics can sometimes be missing, stale, or incorrect.
For example, statistics may be inconsistent between a table and its index, or
between tables with a primary key-foreign key relationship. Alternatively, a
statistics gathering job may have been disabled by accident, or you may be
unaware that a script has failed.

•

Lack of knowledge of the problem can be time-consuming and resource-intensive.
For example, a service request might seek a resolution to a problem, unaware that
the problem is caused by suboptimal statistics. The diagnosis might require a

18-3

Chapter 18

About Optimizer Statistics Advisor

great deal of time emailing scripts of the problematic queries, enabling traces, and
investigating traces.
•

Recommended fixes may not be feasible.
Performance engineers may recommend changing the application code that
maintains statistics. In some organizations, this requirement may be difficult or
impossible to satisfy.

18.1.1.2 Advantages of Optimizer Statistics Advisor
An advisor-based approach offers better scalability and maintainability than the
traditional approach.
If best practices change in a new release, then Optimizer Statistics Advisor encodes
these practices in its rules. In this way, the advisor always provides the most up-todate recommendations.
The advisor analyzes how you are currently gathering statistics (using manual scripts,
explicitly setting parameters, and so on), the effectiveness of existing statistics
gathering jobs, and the quality of the gathered statistics. Optimizer Statistics Advisor
does not gather a new or alternative set of optimizer statistics, and so does not affect
the workload. Rather, Optimizer Statistics Advisor analyzes information stored in the
data dictionary, and then stores the findings and recommendations in the database.
Optimizer Statistics Advisor provides the following advantages over the traditional
approach:
•

Provides easy-to-understand reports
The advisor applies rules to generate findings, recommendations, and actions.

•

Supplies scripts to implement necessary fixes without requiring changes to
application code
When you implement a recommended action, benefit accrues to every execution
of the improved statements. For example, if you set a global preference so that the
sample size is AUTO_SAMPLE_SIZE rather than a suboptimal percentage, then every
plan based on the improved statistics can benefit from this change.

•

Runs a predefined task named AUTO_STATS_ADVISOR_TASK once every day in the
maintenance window
For the automated job to run, the STATISTICS_LEVEL initialization parameter must be
set to TYPICAL or ALL.

•

Supplies an API in the DBMS_STATS package that enables you to create and run
tasks manually, store findings and recommendations in data dictionary views,
generate reports for the tasks, and implement corrections when necessary

•

Integrates with existing tools
The advisor integrates with SQL Tuning Advisor and AWR, which summarize the
Optimizer Statistics Advisor results.

18.1.2 Optimizer Statistics Advisor Concepts
Optimizer Statistics Advisor uses the same advisor framework as Automatic Database
Diagnostic Monitor (ADDM), SQL Performance Analyzer, and other advisors.
This section contains the following topics:

18-4

Chapter 18

About Optimizer Statistics Advisor

•

Components of Optimizer Statistics Advisor
The Optimizer Statistics Optimizer framework stores its metadata in data
dictionary and dynamic performance views.

•

Operational Modes for Optimizer Statistics Advisor
Optimizer Statistics Advisor supports both an automated and manual mode.

18.1.2.1 Components of Optimizer Statistics Advisor
The Optimizer Statistics Optimizer framework stores its metadata in data dictionary
and dynamic performance views.
The following Venn diagram shows the relationships among rules, findings,
recommendations, and actions for Optimizer Statistics Advisor. For example, all
findings are derived from rules, but not all rules generate findings.

Figure 18-2

Optimizer Statistics Advisor Components

Rules
Findings
Recommendations
Actions

This section contains the following topics:
•

Rules for Optimizer Statistics Advisor
An Optimizer Statistics Advisor rule is an Oracle-supplied standard by which
Optimizer Statistics Advisor performs its checks.

•

Findings for Optimizer Statistics Advisor
A finding results when Optimizer Statistics Advisor examines the evidence stored
in the database and concludes that the rules were not followed.

•

Recommendations for Optimizer Statistics Advisor
Based on each finding, Optimizer Statistics Advisor makes recommendations on
how to achieve better statistics.

•

Actions for Optimizer Statistics Advisor
An Optimizer Statistics Advisor action is a SQL or PL/SQL script that implements
recommendations. When feasible, recommendations have corresponding actions.
The advisor stores actions in DBA_ADVISOR_ACTIONS.

18.1.2.1.1 Rules for Optimizer Statistics Advisor
An Optimizer Statistics Advisor rule is an Oracle-supplied standard by which
Optimizer Statistics Advisor performs its checks.

18-5

Chapter 18

About Optimizer Statistics Advisor

The rules embody Oracle best practices based on the current feature set. If the best
practices change from release to release, then the Optimizer Statistics Advisor rules
also change.
The advisor organizes rules into the following classes:
•

System
This class checks the preferences for statistics collection, status of the automated
statistics gathering job, use of SQL plan directives, and so on. Rules in this class
have the value SYSTEM in V$STATS_ADVISOR_RULES.RULE_TYPE.

•

Operation
This class checks whether statistics collection uses the defaults, test statistics are
created using the SET_*_STATS procedures, and so on. Rules in this class have the
value OPERATION in V$STATS_ADVISOR_RULES.RULE_TYPE.

•

Object
This class checks for the quality of the statistics, staleness of statistics,
unnecessary collection of statistics, and so on. Rules in this class have the value
OBJECT in V$STATS_ADVISOR_RULES.RULE_TYPE.

The rules check for the following problems:
•

How to gather statistics
For example, one rule might specify the recommended setting for an initialization
parameter. Another rule might specify that statistics should be gathered at the
schema level.

•

When to gather statistics
For example, the advisor may recommend that the maintenance window for the
automatic statistics gathering job should be enabled, or that the window should be
extended.

•

How to improve the efficiency of statistics gathering
For example, a rule might specify that default parameters should be used in
DBMS_STATS, or that statistics should not be set manually.

In V$STATS_ADVISOR_RULES, each rule has a unique string ID that is usable in the
DBMS_STATS procedures and reports. You can use a rule filter to specify rules that
Optimizer Statistics Advisor should check. However, you cannot write new rules.
Example 18-1

Listing Rules in V$STATS_ADVISOR_RULES
The following query, with sample output, lists a subset of the rules in
V$STATS_ADVISOR_RULES. The rules may change from release to release.

SET
SET
COL
COL
COL

LINESIZE 208
PAGESIZE 100
ID FORMAT 99
NAME FORMAT a33
DESCRIPTION FORMAT a62

SELECT RULE_ID AS ID, NAME, RULE_TYPE, DESCRIPTION
FROM V$STATS_ADVISOR_RULES
WHERE RULE_ID BETWEEN 1 AND 12
ORDER BY RULE_ID;
ID NAME

RULE_TYPE DESCRIPTION

18-6

Chapter 18

About Optimizer Statistics Advisor

-1
2
3
4
5
6
7
8
9
10
11
12

------------------------------UseAutoJob
CompleteAutoJob
MaintainStatsHistory
UseConcurrent
UseDefaultPreference
TurnOnSQLPlanDirective
AvoidSetProcedures
UseDefaultParams
UseGatherSchemaStats
AvoidInefficientStatsOprSeq
AvoidUnnecessaryStatsCollection
AvoidStaleStats

--------SYSTEM
SYSTEM
SYSTEM
SYSTEM
SYSTEM
SYSTEM
OPERATION
OPERATION
OPERATION
OPERATION
OBJECT
OBJECT

------------------------------------------------------Use Auto Job for Statistics Collection
Auto Statistics Gather Job should complete successfully
Maintain Statistics History
Use Concurrent preference for Statistics Collection
Use Default Preference for Stats Collection
SQL Plan Directives should not be disabled
Avoid Set Statistics Procedures
Use Default Parameters in Statistics Collection Proc.
Use gather_schema_stats procedure
Avoid inefficient statistics operation sequences
Avoid unnecessary statistics collection
Avoid objects with stale or no statistics

12 rows selected.

See Also:
Oracle Database Reference to learn more about V$STATS_ADVISOR_RULES

18.1.2.1.2 Findings for Optimizer Statistics Advisor
A finding results when Optimizer Statistics Advisor examines the evidence stored in
the database and concludes that the rules were not followed.
To generate findings, Optimizer Statistics Advisor executes a task, which is invoked
either automatically or manually. This task analyzes the statistics history stored in the
data dictionary, the statistics operation log, and the current statistics footprint that
exists in SYSAUX. For example, the advisor queries DBA_TAB_STATISTICS and
DBA_IND_STATISTICS to determine whether statistics are stale, or whether a discrepancy
exists between the numbers of rows.
Typically, Optimizer Statistics Advisor generates a finding when a specific rule is not
followed or is violated, although some findings—such as object staleness—provide
only information. For example, a finding may show that DBMS_STATS.GATHER_TABLE_STATS
has used ESTIMATE_PERCENT=>0.01, which violates the
ESTIMATE_PERCENT=>AUTO_SAMPLE_SIZE rule.
A finding corresponds to exactly one rule. However, a rule can generate many
findings.

See Also:
•

Oracle Database PL/SQL Packages and Types Reference to learn more
about DBMS_STATS

•

Oracle Database Reference to learn more about ALL_TAB_STATISTICS

18.1.2.1.3 Recommendations for Optimizer Statistics Advisor
Based on each finding, Optimizer Statistics Advisor makes recommendations on how
to achieve better statistics.

18-7

Chapter 18

About Optimizer Statistics Advisor

For example, the advisor might discover a violation to the rule of not using sampling
when gathering statistics, and recommend specifying AUTO_SAMPLE_SIZE instead. The
advisor stores the recommendations in DBA_ADVISOR_RECOMMENDATIONS.
Multiple recommendations may exist for a single finding. In this case, you must
investigate to determine which recommendation to follow. Each recommendation
includes one or more rationales that explain why Optimizer Statistics Advisor makes its
recommendation. In some cases, findings may not generate recommendations.

See Also:
•

"Guideline for Setting the Sample Size" to learn the guideline for the
sample size

•

Oracle Database Reference to learn about DBA_ADVISOR_RECOMMENDATIONS

18.1.2.1.4 Actions for Optimizer Statistics Advisor
An Optimizer Statistics Advisor action is a SQL or PL/SQL script that implements
recommendations. When feasible, recommendations have corresponding actions. The
advisor stores actions in DBA_ADVISOR_ACTIONS.
For example, Optimizer Statistics Advisor executes a task that performs the following
steps:
1.

Checks rules
The advisor checks conformity to the rule that stale statistics should be avoided.

2.

Generates finding
The advisor discovers that a number of objects have no statistics.

3.

Generates recommendation
The advisor recommends gathering statistics on the objects with no statistics.

4.

Generates action
The advisor generates a PL/SQL script that executes
DBMS_STATS.GATHER_DATABASE_STATS, supplying a list of objects that need to have

statistics gathered.

See Also:
•
•

"Statistics Preference Overrides" to learn how to override statistics
gathering preferences
"Guideline for Setting the Sample Size" to learn more about
AUTO_SAMPLE_SIZE

•

Oracle Database Reference to learn about DBA_ADVISOR_ACTIONS

18-8

Chapter 18

About Optimizer Statistics Advisor

18.1.2.2 Operational Modes for Optimizer Statistics Advisor
Optimizer Statistics Advisor supports both an automated and manual mode.
•

Automated
The predefined task AUTO_STATS_ADVISOR_TASK runs automatically in the
maintenance window once per day. The task runs as part of the automatic
optimizer statistics collection client. The automated task generates findings and
recommendations, but does not implement actions automatically.
As for any other task, you can configure the automated task, and generate reports.
If the report recommends actions, then you can implement the actions manually.

•

Manual
You can create your own task using the DBMS_STATS.CREATE_ADVISOR_TASK function,
and then run it at any time using the EXECUTE_ADVISOR_TASK procedure.
Unlike the automated task, the manual task can implement actions automatically.
Alternatively, you can configure the task to generate a PL/SQL script, which you
can then run manually.

See Also:
•

"Configuring Automatic Optimizer Statistics Collection"

•

Oracle Database PL/SQL Packages and Types Reference to learn more
about DBMS_STATS.CREATE_ADVISOR_TASK

18.1.3 Command-Line Interface to Optimizer Statistics Advisor
Perform Optimizer Statistics Advisor tasks using the DBMS_STATS PL/SQL package.
Table 18-1

DBMS_STATS APIs for Task Creation and Deletion

PL/SQL Procedure or
Function

Description

CREATE_ADVISOR_TASK

Creates an advisor task for Optimizer Statistics Advisor. If the
task name is already specified, then the advisor uses the
specified task name; otherwise, the advisor automatically
generates a new task name.

DROP_ADVISOR_TASK

Deletes an Optimizer Statistics Advisor task and all its result
data.

Table 18-2

DBMS_STATS APIs for Task Execution

PL/SQL Procedure or
Function

Description

EXECUTE_ADVISOR_TASK

Executes a previously created Optimizer Statistics Advisor
task.

18-9

Chapter 18

About Optimizer Statistics Advisor

Table 18-2

(Cont.) DBMS_STATS APIs for Task Execution

PL/SQL Procedure or
Function

Description

INTERRUPT_ADVISOR_TASK

Interrupts a currently executing Optimizer Statistics Advisor
task. The task ends its operations as it would in a normal exit,
enabling you to access intermediate results. You can resume
the task later.

CANCEL_ADVISOR_TASK

Cancels an Optimizer Statistics Advisor task execution, and
removes all intermediate results of the current execution.

RESET_ADVISOR_TASK

Resets an Optimizer Statistics Advisor task execution to its
initial state. Call this procedure on a task that is not currently
executing.

RESUME_ADVISOR_TASK

Resumes the Optimizer Statistics Advisor task execution that
was most recently interrupted.

Table 18-3

DBMS_STATS APIs for Advisor Reports

PL/SQL Procedure or Function

Description

REPORT_STATS_ADVISOR_TASK

Reports the results of an Optimizer Statistics Advisor
task.

GET_ADVISOR_RECS

Generates a recommendation report on the given item.

Table 18-4

DBMS_STATS APIs for Task and Filter Configuration

PL/SQL Procedure or Function

Description

CONFIGURE_ADVISOR_TASK

Configures the Optimizer Statistics Advisor lists for
the execution, reporting, script generation, and
implementation of an advisor task.

GET_ADVISOR_OPR_FILTER

Creates an operation filter for a statistics operation.

CONFIGURE_ADVISOR_RULE_FILTER

Configures the rule filter for an Optimizer Statistics
Advisor task.

CONFIGURE_ADVISOR_OPR_FILTER

Configures the operation filter for an Optimizer
Statistics Advisor task.

CONFIGURE_ADVISOR_OBJ_FILTER

Configures the object filter for an Optimizer
Statistics Advisor task.

SET_ADVISOR_TASK_PARAMETER

Updates the value of an Optimizer Statistics
Advisor task parameter. Valid parameters are
TIME_LIMIT and OP_START_TIME.

Table 18-5

DBMS_STATS APIs for Implementation of Recommended Actions

PL/SQL Procedure or
Function

Description

SCRIPT_ADVISOR_TASK

Gets the script that implements the recommended actions for
the problems found by the advisor. You can check this script,
and then choose which actions to execute.

18-10

Chapter 18

Basic Tasks for Optimizer Statistics Advisor

Table 18-5
Actions

(Cont.) DBMS_STATS APIs for Implementation of Recommended

PL/SQL Procedure or
Function

Description

IMPLEMENT_ADVISOR_TASK

Implements the actions recommended by the advisor based
on results from a specified Optimizer Statistics Advisor
execution.

See Also:
Oracle Database PL/SQL Packages and Types Reference to learn about the
DBMS_STATS package

18.2 Basic Tasks for Optimizer Statistics Advisor
This section explains the basic workflow for using Optimizer Statistics Advisor. All
procedures and functions are in the DBMS_STATS package.
The following figure shows the automatic and manual paths in the workflow. If
AUTO_STATS_ADVISOR_TASK runs automatically in the maintenance window, then your
workflow begins by querying the report. In the manual workflow, you must use PL/SQL
to create and execute the tasks.

18-11

Chapter 18

Basic Tasks for Optimizer Statistics Advisor

Figure 18-3

Basic Tasks for Optimizer Statistics Advisor

Manual Mode

Automatic Mode

Create an advisor task

CREATE_ADVISOR_TASK

SELECT . . . FROM
DBA_ADVISOR_EXECUTIONS

Optionally,
list tasks
Optionally,
alter the
scope of
the advisor
checks

CONFIGURE_ADVISOR_*_FILTER

Execute the advisor
task

EXECUTE_ADVISOR_TASK

Generate a report of
findings and
recommendations

REPORT_ADVISOR_TASK

IMPLEMENT_ADVISOR_TASK

SCRIPT_ADVISOR_TASK

Generate
modifiable
PL/SQL
script

Edit PL/SQL
script

Implement all advisor
recommendations

Run PL/SQL
script

Typically, you perform Optimizer Statistics Advisor steps in the sequence shown in the
following table.
Table 18-6

Optimizer Statistics Advisor Workflow

Step

Description

To Learn More

1

Create an Optimizer Advisor task using
DBMS_STATS.CREATE_ADVISOR_TASK
(manual workflow only).

"Creating an Optimizer Statistics Advisor
Task"

2

Optionally, list executions of advisor
tasks by querying
DBA_ADVISOR_EXECUTIONS.

"Listing Optimizer Statistics Advisor Tasks"

3

Optionally, configure a filter for the task
using the
DBMS_STATS.CONFIGURE_ADVISOR_*_FIL
TER procedures.

"Creating Filters for an Optimizer Advisor
Task"

18-12

Chapter 18

Basic Tasks for Optimizer Statistics Advisor

Table 18-6

(Cont.) Optimizer Statistics Advisor Workflow

Step

Description

To Learn More

4

Execute the advisor task using
DBMS_STATS.EXECUTE_ADVISOR_TASK
(manual workflow only).

"Executing an Optimizer Statistics Advisor
Task"

5

Generate an advisor report.

"Generating a Report for an Optimizer
Statistics Advisor Task"

6

Implement the recommendations in
either of following ways:

"Implementing Actions Recommended by
Optimizer Statistics Advisor" and
"Generating a Script Using Optimizer
Statistics Advisor"

•

•

Implement all recommendations
automatically using
DBMS_STATS.IMPLEMENT_ADVISOR_TA
SK.
Generate a PL/SQL script that
implements recommendations
using
DBMS_STATS.SCRIPT_ADVISOR_TASK,
edit this script, and then run it
manually.

Example 18-2

Basic Script for Optimizer Statistics Advisor in Manual Workflow

This script illustrates a basic Optimizer Statistics Advisor session. It creates a task,
executes it, generates a report, and then implements the recommendations.
DECLARE
v_tname VARCHAR2(128) := 'my_task';
v_ename VARCHAR2(128) := NULL;
v_report CLOB := null;
v_script CLOB := null;
v_implementation_result CLOB;
BEGIN
-- create a task
v_tname := DBMS_STATS.CREATE_ADVISOR_TASK(v_tname);
-- execute the task
v_ename := DBMS_STATS.EXECUTE_ADVISOR_TASK(v_tname);
-- view the task report
v_report := DBMS_STATS.REPORT_ADVISOR_TASK(v_tname);
DBMS_OUTPUT.PUT_LINE(v_report);
-- implement all recommendations
v_implementation_result := DBMS_STATS.IMPLEMENT_ADVISOR_TASK(v_tname);
END;

This section contains the following topics:
•

Creating an Optimizer Statistics Advisor Task
The DBMS_STATS.CREATE_ADVISOR_TASK function creates a task for Optimizer Statistics
Advisor. If you do not specify a task name, then Optimizer Statistics Advisor
generates one automatically.

18-13

Chapter 18

Basic Tasks for Optimizer Statistics Advisor

•

Listing Optimizer Statistics Advisor Tasks
The DBA_ADVISOR_EXECUTIONS view lists executions of Optimizer Statistics Advisor
tasks.

•

Creating Filters for an Optimizer Advisor Task
Filters enable you to include or exclude objects, rules, and operations from
Optimizer Statistics Advisor tasks.

•

Executing an Optimizer Statistics Advisor Task
The DBMS_STATS.EXECUTE_ADVISOR_TASK function executes a task for Optimizer
Statistics Advisor. If you do not specify an execution name, then Optimizer
Statistics Advisor generates one automatically.

•

Generating a Report for an Optimizer Statistics Advisor Task
The DBMS_STATS.REPORT_ADVISOR_TASK function generates a report for an Optimizer
Statistics Advisor task.

•

Implementing Optimizer Statistics Advisor Recommendations
You can either implement all recommendations automatically using
DBMS_STATS.IMPLEMENT_ADVISOR_TASK, or generate an editable script using
DBMS_STATS.SCRIPT_ADVISOR_TASK.

See Also:
Oracle Database PL/SQL Packages and Types Reference to learn about the
DBMS_STATS package

18.2.1 Creating an Optimizer Statistics Advisor Task
The DBMS_STATS.CREATE_ADVISOR_TASK function creates a task for Optimizer Statistics
Advisor. If you do not specify a task name, then Optimizer Statistics Advisor generates
one automatically.
Prerequisites
To execute this subprogram, you must have the ADVISOR privilege.

Note:
This subprogram executes using invoker's rights.

To create an Optimizer Statistics Advisor task:
1.

In SQL*Plus, log in to the database as a user with the necessary privileges.

2.

Execute the DBMS_STATS.CREATE_ADVISOR_TASK function in the following form, where
tname is the name of the task and ret is the variable that contains the returned
output:
EXECUTE ret := DBMS_STATS.CREATE_ADVISOR_TASK('tname');

For example, to create the task opt_adv_task1, use the following code:

18-14

Chapter 18

Basic Tasks for Optimizer Statistics Advisor

DECLARE
v_tname VARCHAR2(32767);
v_ret VARCHAR2(32767);
BEGIN
v_tname := 'opt_adv_task1';
v_ret := DBMS_STATS.CREATE_ADVISOR_TASK(v_tname);
END;
/
3.

Optionally, query USER_ADVISOR_TASKS:
SELECT TASK_NAME, ADVISOR_NAME, CREATED, STATUS FROM USER_ADVISOR_TASKS;

Sample output appears below:
TASK_NAME
ADVISOR_NAME
CREATED STATUS
--------------- -------------------- --------- ----------OPT_ADV_TASK1 Statistics Advisor 05-SEP-16 INITIAL

See Also:
Oracle Database PL/SQL Packages and Types Reference to learn more
about CREATE_ADVISOR_TASK

18.2.2 Listing Optimizer Statistics Advisor Tasks
The DBA_ADVISOR_EXECUTIONS view lists executions of Optimizer Statistics Advisor tasks.
To list Optimizer Statistics Advisor tasks:
1.

In SQL*Plus, log in to the database as a user with administrator privileges.

2.

Query DBA_ADVISOR_EXECUTIONS as follows:
COL EXECUTION_NAME FORMAT a14
SELECT EXECUTION_NAME, EXECUTION_END, STATUS
FROM DBA_ADVISOR_EXECUTIONS
WHERE TASK_NAME = 'AUTO_STATS_ADVISOR_TASK'
ORDER BY 2;

The following sample output shows 8 executions:
EXECUTION_NAME
-------------EXEC_1
EXEC_17
EXEC_42
EXEC_67
EXEC_92
EXEC_117
EXEC_142
EXEC_167

EXECUTION
--------27-AUG-16
28-AUG-16
29-AUG-16
30-AUG-16
01-SEP-16
02-SEP-16
03-SEP-16
04-SEP-16

STATUS
----------COMPLETED
COMPLETED
COMPLETED
COMPLETED
COMPLETED
COMPLETED
COMPLETED
COMPLETED

8 rows selected.

18-15

Chapter 18

Basic Tasks for Optimizer Statistics Advisor

See Also:
Oracle Database Reference to learn more about DBA_ADVISOR_EXECUTIONS

18.2.3 Creating Filters for an Optimizer Advisor Task
Filters enable you to include or exclude objects, rules, and operations from Optimizer
Statistics Advisor tasks.
This section contains the following topics:
•

About Filters for Optimizer Statistics Advisor
A filter is the use of DBMS_STATS to restrict an Optimizer Statistics Advisor task to a
user-specified set of rules, schemas, or operations.

•

Creating an Object Filter for an Optimizer Advisor Task
The DBMS_STATS.CONFIGURE_ADVISOR_OBJ_FILTER function creates a rule filter for a
specified Optimizer Statistics Advisor task. The function returns a CLOB that
contains the updated values of the filter.

•

Creating a Rule Filter for an Optimizer Advisor Task
The DBMS_STATS.CONFIGURE_ADVISOR_RULE_FILTER function creates a rule filter for a
specified Optimizer Statistics Advisor task. The function returns a CLOB that
contains the updated values of the filter.

•

Creating an Operation Filter for an Optimizer Advisor Task
The DBMS_STATS.CONFIGURE_ADVISOR_OPR_FILTER function creates an operation filter
for a specified Optimizer Statistics Advisor task. The function returns a CLOB that
contains the updated values of the filter.

18.2.3.1 About Filters for Optimizer Statistics Advisor
A filter is the use of DBMS_STATS to restrict an Optimizer Statistics Advisor task to a userspecified set of rules, schemas, or operations.
Filters are useful for including or excluding a specific set of results. For example, you
can configure an advisor task to include only recommendations for the sh schema.
Also, you can exclude all violations of the rule for stale statistics. The primary
advantage of filters is the ability to ignore recommendations that you are not interested
in, and reduce the overhead of the advisor task.
The simplest way to create filters is to use the following DBMS_STATS procedures either
individually or in combination:
•

CONFIGURE_ADVISOR_OBJ_FILTER

Use this procedure to include or exclude the specified database schemas or
objects. The object filter takes in an owner name and an object name, with
wildcards (%) supported.
•

CONFIGURE_ADVISOR_RULE_FILTER

Use this procedure to include or exclude the specified rules. Obtain the names of
rules by querying V$STATS_ADVISOR_RULES.
•

CONFIGURE_ADVISOR_OPR_FILTER

18-16

Chapter 18

Basic Tasks for Optimizer Statistics Advisor

Use this procedure to include or exclude the specified DBMS_STATS operations.
Obtain the IDs and names for operations by querying DBA_OPTSTAT_OPERATIONS.
For the preceding functions, you can specify the type of operation to which the filter
applies: EXECUTE, REPORT, SCRIPT, and IMPLEMENT. You can also combine types, as in
EXECUTE + REPORT. Null indicates that the filter applies to all types of advisor operations.

See Also:
•

Oracle Database Reference to learn more about V$STATS_ADVISOR_RULES

•

Oracle Database PL/SQL Packages and Types Reference to learn more
about DBMS_STATS

18.2.3.2 Creating an Object Filter for an Optimizer Advisor Task
The DBMS_STATS.CONFIGURE_ADVISOR_OBJ_FILTER function creates a rule filter for a
specified Optimizer Statistics Advisor task. The function returns a CLOB that contains
the updated values of the filter.
You can use either of the following basic strategies:
•

Include findings for all objects (by default, all objects are considered), and then
exclude findings for specified objects.

•

Exclude findings for all objects, and then include findings only for specified objects.

Prerequisites
To use the DBMS_STATS.CONFIGURE_ADVISOR_OBJ_FILTER function, you must meet the
following prerequisites:
•

To execute this subprogram, you must have the ADVISOR privilege.

•

You must be the owner of the task.

Note:
This subprogram executes using invoker's rights.

To create an object filter:
1.

In SQL*Plus or SQL Developer, log in to the database as a user with the
necessary privileges.

2.

Either exclude or include objects for a specified task using the
DBMS_STATS.CONFIGURE_ADVISOR_OBJ_FILTER function.
Invoke the function in the following form, where the placeholders are defined as
follows:
•

report is the CLOB variable that contains the returned XML.

•

tname is the name of the task.

18-17

Chapter 18

Basic Tasks for Optimizer Statistics Advisor

•

opr_type is the type of operation to perform.

•

rule is the name of the rule.

•

owner is the schema for the objects.

•

table is the name of the table.

•

action is the name of the action: ENABLE, DISABLE, DELETE, or SHOW.

BEGIN
report := DBMS_STATS.CONFIGURE_ADVISOR_OBJ_FILTER(
task_name
=> 'tname'
, stats_adv_opr_type => 'opr_type'
, rule_name
=> 'rule'
, ownname
=> 'owner'
, tabname
=> 'table'
, action
=> 'action' );
END;

Example 18-3

Including Only Objects in a Single Schema

In this example, for the task named opt_adv_task1, your goal is to disable
recommendations for all objects except those in the sh schema. User account sh has
been granted ADVISOR and READ ANY TABLE privileges. You perform the following steps:
1.

Log in to the database as sh.

2.

Drop any existing task named opt_adv_task1.
DECLARE
v_tname VARCHAR2(32767);
BEGIN
v_tname := 'opt_adv_task1';
DBMS_STATS.DROP_ADVISOR_TASK(v_tname);
END;
/

3.

Create a procedure named sh_obj_filter that restricts a specified task to objects
in the sh schema.
CREATE OR REPLACE PROCEDURE sh_obj_filter(p_tname IN VARCHAR2) IS
v_retc CLOB;
BEGIN
-- Filter out all objects that are not in the sh schema
v_retc := DBMS_STATS.CONFIGURE_ADVISOR_OBJ_FILTER(
task_name
=> p_tname
, stats_adv_opr_type => 'EXECUTE'
, rule_name
=> NULL
, ownname
=> NULL
, tabname
=> NULL
, action
=> 'DISABLE' );
v_retc := DBMS_STATS.CONFIGURE_ADVISOR_OBJ_FILTER(
task_name
=> p_tname
, stats_adv_opr_type => 'EXECUTE'
, rule_name
=> NULL
, ownname
=> 'SH'
, tabname
=> NULL
, action
=> 'ENABLE' );
END;
/
SHOW ERRORS

18-18

Chapter 18

Basic Tasks for Optimizer Statistics Advisor

4.

Create a task named opt_adv_task1, and then execute the sh_obj_filter procedure
for this task.
DECLARE
v_tname VARCHAR2(32767);
v_ret VARCHAR2(32767);
BEGIN
v_tname := 'opt_adv_task1';
v_ret := DBMS_STATS.CREATE_ADVISOR_TASK(v_tname);
sh_obj_filter(v_tname);
END;
/

5.

Execute the task opt_adv_task1.
DECLARE
v_tname
v_ret
begin
v_tname
v_ret
END;
/

VARCHAR2(32767);
VARCHAR2(32767);
:= 'opt_adv_task1';
:= DBMS_STATS.EXECUTE_ADVISOR_TASK(v_tname);

See Also:
Oracle Database PL/SQL Packages and Types Reference to learn more
about DBMS_STATS.CONFIGURE_ADVISOR_OBJ_FILTER

18.2.3.3 Creating a Rule Filter for an Optimizer Advisor Task
The DBMS_STATS.CONFIGURE_ADVISOR_RULE_FILTER function creates a rule filter for a
specified Optimizer Statistics Advisor task. The function returns a CLOB that contains
the updated values of the filter.
You can use either of the following basic strategies:
•

Enable all rules (by default, all rules are enabled), and then disable specified rules.

•

Disable all rules, and then enable only specified rules.

Prerequisites
To use the DBMS_STATS.CONFIGURE_ADVISOR_RULE_FILTER function, you must meet the
following prerequisites:
•

To execute this subprogram, you must have the ADVISOR privilege.

•

You must be the owner of the task.

Note:
This subprogram executes using invoker's rights.

18-19

Chapter 18

Basic Tasks for Optimizer Statistics Advisor

To create a rule filter:
1.

In SQL*Plus or SQL Developer, log in to the database as a user with the
necessary privileges.

2.

Obtain the names of the advisor rules by querying V$STATS_ADVISOR_RULES.
For example, query the view as follows (partial sample output included):

SET
SET
COL
COL
COL

LINESIZE 200
PAGESIZE 100
ID FORMAT 99
NAME FORMAT a27
DESCRIPTION FORMAT a54

SELECT RULE_ID AS ID, NAME, RULE_TYPE, DESCRIPTION
FROM V$STATS_ADVISOR_RULES
ORDER BY RULE_ID;
ID NAME
-- --------------------------1 UseAutoJob
2 CompleteAutoJob
3 MaintainStatsHistory
4 UseConcurrent
...
3.

RULE_TYPE
--------SYSTEM
SYSTEM
SYSTEM
SYSTEM

DESCRIPTION
------------------------------------------------------Use Auto Job for Statistics Collection
Auto Statistics Gather Job should complete successfully
Maintain Statistics History
Use Concurrent preference for Statistics Collection

Either exclude or include rules for a specified task using the
DBMS_STATS.CONFIGURE_ADVISOR_RULE_FILTER function.

Invoke the function in the following form, where the placeholders are defined as
follows:
•

tname is the name of the task.

•

report is the CLOB variable that contains the returned XML.

•

opr_type is the type of operation to perform.

•

rule is the name of the rule.

•

action is the name of the action: ENABLE, DISABLE, DELETE, or SHOW.

BEGIN
report := DBMS_STATS.DBMS_STATS.CONFIGURE_ADVISOR_RULE_FILTER(
task_name
=> 'tname'
, stats_adv_opr_type => 'opr_type'
, rule_name
=> 'rule'
, action
=> 'action' );
END;

Example 18-4

Excluding the Rule for Stale Statistics

In this example, you know that statistics are stale because the automated statistics job
did not run. You want to generate a report for the task named opt_adv_task1, but do not
want to clutter it with recommendations about stale statistics.
1.

You query V$STATS_ADVISOR_RULES for rules that deal with stale statistics (sample
output included):
COL NAME FORMAT a15
SELECT RULE_ID AS ID, NAME, RULE_TYPE, DESCRIPTION
FROM V$STATS_ADVISOR_RULES
WHERE DESCRIPTION LIKE '%tale%'

18-20

Chapter 18

Basic Tasks for Optimizer Statistics Advisor

ORDER BY RULE_ID;
ID NAME
RULE_TYPE DESCRIPTION
--- --------------- --------- ----------------------------------------12 AvoidStaleStats OBJECT
Avoid objects with stale or no statistics
2.

You configure a filter using CONFIGURE_ADVISOR_RULE_FILTER, specifying that task
execution should exclude the rule AvoidStaleStats, but honor all other rules:
VARIABLE b_ret CLOB
BEGIN
:b_ret := DBMS_STATS.CONFIGURE_ADVISOR_RULE_FILTER(
task_name
=> 'opt_adv_task1'
,
stats_adv_opr_type => 'EXECUTE'
,
rule_name
=> 'AvoidStaleStats'
,
action
=> 'DISABLE' );
END;
/

Example 18-5

Including Only the Rule for Avoiding Stale Statistics

This example is the inverse of the preceding example. You want to generate a report
for the task named opt_adv_task1, but want to see only recommendations about stale
statistics.
1.

Query V$STATS_ADVISOR_RULES for rules that deal with stale statistics (sample output
included):
COL NAME FORMAT a15
SELECT RULE_ID AS ID, NAME, RULE_TYPE, DESCRIPTION
FROM V$STATS_ADVISOR_RULES
WHERE DESCRIPTION LIKE '%tale%'
ORDER BY RULE_ID;
ID NAME
RULE_TYPE DESCRIPTION
--- --------------- --------- ----------------------------------------12 AvoidStaleStats OBJECT
Avoid objects with stale or no statistics

2.

Configure a filter using CONFIGURE_ADVISOR_RULE_FILTER, specifying that task
execution should exclude all rules:
VARIABLE b_ret CLOB
BEGIN
:b_ret := DBMS_STATS.CONFIGURE_ADVISOR_RULE_FILTER(
task_name
=> 'opt_adv_task1'
,
stats_adv_opr_type => 'EXECUTE'
,
rule_name
=> null
,
action
=> 'DISABLE' );
END;
/

3.

Configure a filter that enables only the AvoidStaleStats rule:
BEGIN
:b_ret := DBMS_STATS.CONFIGURE_ADVISOR_RULE_FILTER(
task_name
=> 'opt_adv_task1'
,
stats_adv_opr_type => 'EXECUTE'
,
rule_name
=> 'AvoidStaleStats'
,
action
=> 'ENABLE' );
END;
/

18-21

Chapter 18

Basic Tasks for Optimizer Statistics Advisor

See Also:
•

Oracle Database Reference to learn more about V$STATS_ADVISOR_RULES

•

Oracle Database PL/SQL Packages and Types Reference to learn more
about CONFIGURE_ADVISOR_RULE_FILTER

18.2.3.4 Creating an Operation Filter for an Optimizer Advisor Task
The DBMS_STATS.CONFIGURE_ADVISOR_OPR_FILTER function creates an operation filter for a
specified Optimizer Statistics Advisor task. The function returns a CLOB that contains
the updated values of the filter.
You can use either of the following basic strategies:
•

Disable all operations, and then enable only specified operations.

•

Enable all operations (by default, all operations are enabled), and then disable
specified operations.

The DBA_OPTSTAT_OPERATIONS view contains the IDs of statistics-related operations.
Prerequisites
To use DBMS_STATS.CONFIGURE_ADVISOR_OPR_FILTER function, you must meet the following
prerequisites:
•

To execute this subprogram, you must have the ADVISOR privilege.

Note:
This subprogram executes using invoker's rights.
•

You must be the owner of the task.

•

To query the DBA_OPTSTAT_OPERATIONS view, you must have the SELECT ANY TABLE
privilege.

To create an operation filter:
1.

In SQL*Plus or SQL Developer, log in to the database as a user with the
necessary privileges.

2.

Query the types of operations.
For example, list all distinct operations in DBA_OPTSTAT_OPERATIONS (sample output
included):
SQL> SELECT DISTINCT(OPERATION) FROM DBA_OPTSTAT_OPERATIONS ORDER BY OPERATION;
OPERATION
----------------------gather_dictionary_stats
gather_index_stats
gather_schema_stats
gather_table_stats

18-22

Chapter 18

Basic Tasks for Optimizer Statistics Advisor

purge_stats
set_system_stats
3.

Obtain the IDs of the operations to be filtered by querying DBA_OPTSTAT_OPERATIONS.
For example, to obtain IDs for all statistics gathering operations for tables and
indexes in the SYS and sh schemas, use the following query:
SELECT ID
FROM DBA_OPTSTAT_OPERATIONS
WHERE ( OPERATION = 'gather_table_stats'
OR OPERATION = 'gather_index_stats')
AND
( TARGET LIKE 'SH.%'
OR TARGET LIKE 'SYS.%');

4.

Exclude or include rules for a specified task using the
DBMS_STATS.CONFIGURE_ADVISOR_OPR_FILTER function, specifying the IDs obtained in
the previous step.
Invoke the function in the following form, where the placeholders are defined as
follows:
•

report is the CLOB variable that contains the returned XML.

•

tname is the name of the task.

•

opr_type is the type of operation to perform. This value cannot be null.

•

rule is the name of the rule.

•

opr_id is the ID (from DBA_OPTSTAT_OPERATIONS.ID) of the operation to perform.

This value cannot be null.
•

action is the name of the action: ENABLE, DISABLE, DELETE, or SHOW.

BEGIN
report := DBMS_STATS.CONFIGURE_ADVISOR_OPR_FILTER(
task_name
=> 'tname'
, stats_adv_opr_type => 'opr_type'
, rule_name
=> 'rule'
, operation_id
=> 'op_id'
, action
=> 'action' );
END;

Example 18-6

Excluding Operations for Gathering Table Statistics

In this example, your goal is to exclude operations that gather table statistics in the hr
schema. User account stats has been granted the DBA role, ADVISOR privilege, and
SELECT ON DBA_OPTSTAT_OPERATIONS privilege. You perform the following steps:
1.

Log in to the database as stats.

2.

Drop any existing task named opt_adv_task1.
DECLARE
v_tname VARCHAR2(32767);
BEGIN
v_tname := 'opt_adv_task1';
DBMS_STATS.DROP_ADVISOR_TASK(v_tname);
END;
/

3.

Create a procedure named opr_filter that configures a task to advise on all
operations except those that gather statistics for tables in the hr schema.

18-23

Chapter 18

Basic Tasks for Optimizer Statistics Advisor

CREATE OR REPLACE PROCEDURE opr_filter(p_tname IN VARCHAR2) IS
v_retc CLOB;
BEGIN
-- For all rules, prevent the advisor from operating
-- on the operations selected in the following query
FOR rec IN
(SELECT ID FROM DBA_OPTSTAT_OPERATIONS WHERE OPERATION =
'gather_table_stats' AND TARGET LIKE 'HR.%')
LOOP
v_retc := DBMS_STATS.CONFIGURE_ADVISOR_OPR_FILTER(
task_name
=> p_tname
, stats_adv_opr_type => NULL
, rule_name
=> NULL
, operation_id
=> rec.id
, action
=> 'DISABLE');
END LOOP;
END;
/
SHOW ERRORS
4.

Create a task named opt_adv_task1, and then execute the opr_filter procedure
for this task.
DECLARE
v_tname VARCHAR2(32767);
v_ret VARCHAR2(32767);
BEGIN
v_tname := 'opt_adv_task1';
v_ret := DBMS_STATS.CREATE_ADVISOR_TASK(v_tname);
opr_filter(v_tname);
END;
/

5.

Execute the task opt_adv_task1.
DECLARE
v_tname
v_ret
begin
v_tname
v_ret
END;
/

6.

VARCHAR2(32767);
VARCHAR2(32767);
:= 'opt_adv_task1';
:= DBMS_STATS.EXECUTE_ADVISOR_TASK(v_tname);

Print the report.
SPOOL /tmp/rep.txt
SET LONG 1000000
COLUMN report FORMAT A200
SET LINESIZE 250
SET PAGESIZE 1000
SELECT DBMS_STATS.REPORT_ADVISOR_TASK(
task_name
=> 'opt_adv_task1'
, execution_name => NULL
, type
=> 'TEXT'
, section
=> 'ALL'
) AS report
FROM DUAL;
SPOOL OFF

18-24

Chapter 18

Basic Tasks for Optimizer Statistics Advisor

See Also:
•

Oracle Database Reference to learn more about DBA_OPTSTAT_OPERATIONS

•

Oracle Database PL/SQL Packages and Types Reference to learn more
about CONFIGURE_ADVISOR_OPR_FILTER

18.2.4 Executing an Optimizer Statistics Advisor Task
The DBMS_STATS.EXECUTE_ADVISOR_TASK function executes a task for Optimizer Statistics
Advisor. If you do not specify an execution name, then Optimizer Statistics Advisor
generates one automatically.
The results of performing this task depend on the privileges of the executing user:
•

SYSTEM level

Only users with both the ANALYZE ANY and ANALYZE ANY DICTIONARY privileges can
perform this task on system-level rules.
•

Operation level
The results depend on the following privileges:

•

–

Users with both the ANALYZE ANY and ANALYZE ANY DICTIONARY privileges can
perform this task for all statistics operations.

–

Users with the ANALYZE ANY privilege but not the ANALYZE ANY DICTIONARY
privilege can perform this task for statistics operations related to any schema
except SYS.

–

Users with the ANALYZE ANY DICTIONARY privilege but not the ANALYZE ANY
privilege can perform this task for statistics operations related to their own
schema and the SYS schema.

–

Users with neither the ANALYZE ANY nor the ANALYZE ANY DICTIONARY privilege
can only perform this operation for statistics operations relating to their own
schema.

Object level
Users can perform this task for any object for which they have statistics collection
privileges.

Prerequisites
This task has the following prerequisites:
•

To execute this subprogram, you must have the ADVISOR privilege.

•

You must be the owner of the task.

•

If you specify an execution name, then this name must not conflict with an existing
execution.

18-25

Chapter 18

Basic Tasks for Optimizer Statistics Advisor

Note:
This subprogram executes using invoker's rights.

To execute an Optimizer Statistics Advisor task:
1.

In SQL*Plus or SQL Developer, log in to the database as a user with the
necessary privileges.

2.

Execute the DBMS_STATS.EXECUTE_ADVISOR_TASK function in the following form, where
tname is the name of the task, execname is the optional name of the execution, and
ret is the variable that contains the returned output:
EXECUTE ret := DBMS_STATS.EXECUTE_ADVISOR_TASK('tname','execname');

For example, to execute the task opt_adv_task1, use the following code:
DECLARE
v_tname VARCHAR2(32767);
v_ret VARCHAR2(32767);
BEGIN
v_tname := 'opt_adv_task1';
v_ret := DBMS_STATS.EXECUTE_ADVISOR_TASK(v_tname);
END;
/
3.

Optionally, obtain details about the execution by querying USER_ADVISOR_EXECUTIONS:
SELECT TASK_NAME, EXECUTION_NAME,
EXECUTION_END, EXECUTION_TYPE AS TYPE, STATUS
FROM USER_ADVISOR_EXECUTIONS;

Sample output appears below:
TASK_NAME
EXECUTION_NAME
EXECUTION TYPE
STATUS
--------------- -------------------- --------- ---------- ----------OPT_ADV_TASK1 EXEC_136
23-NOV-15 STATISTICS COMPLETED

See Also:
Oracle Database PL/SQL Packages and Types Reference to learn more
about EXECUTE_ADVISOR_TASK

18.2.5 Generating a Report for an Optimizer Statistics Advisor Task
The DBMS_STATS.REPORT_ADVISOR_TASK function generates a report for an Optimizer
Statistics Advisor task.
The report contains the following sections:
•

General information
This section describes the task name, execution name, creation date, and
modification date.

•

Summary

18-26

Chapter 18

Basic Tasks for Optimizer Statistics Advisor

This section summarizes the findings and rules violated by the findings.
•

Findings
Each finding section lists the relevant rule and findings. If the advisor has a
recommendation, then the recommendation is described. In some cases, a
recommendation also has a rationale.

The name of the automated Optimizer Statistics Advisor task is
AUTO_STATS_ADVISOR_TASK. If you follow the automated workflow, then you only need to
query the automatically generated report.
Prerequisites
To generate a report with the DBMS_STATS.REPORT_ADVISOR_TASK function, you must meet
the following prerequisites:
•

To execute this subprogram, you must have the ADVISOR privilege.

•

You must be the owner of the task.

Note:
This subprogram executes using invoker's rights.

The results of performing this task depend on the privileges of the executing user:
•

SYSTEM level

Only users with both the ANALYZE ANY and ANALYZE ANY DICTIONARY privileges can
perform this task on system-level rules.
•

Operation level
The results depend on the following privileges:

•

–

Users with both the ANALYZE ANY and ANALYZE ANY DICTIONARY privileges can
perform this task for all statistics operations.

–

Users with the ANALYZE ANY privilege but not the ANALYZE ANY DICTIONARY
privilege can perform this task for statistics operations related to any schema
except SYS.

–

Users with the ANALYZE ANY DICTIONARY privilege but not the ANALYZE ANY
privilege can perform this task for statistics operations related to their own
schema and the SYS schema.

–

Users with neither the ANALYZE ANY nor the ANALYZE ANY DICTIONARY privilege
can only perform this operation for statistics operations relating to their own
schema.

Object level
Users can perform this task for any object for which they have statistics collection
privileges.

To generate an Optimizer Statistics Advisor report:
1.

In SQL*Plus, log in to the database as a user with ADVISOR privileges.

2.

Query the DBMS_STATS.REPORT_ADVISOR_TASK function output.

18-27

Chapter 18

Basic Tasks for Optimizer Statistics Advisor

Use the following query, where the placeholders have the following definitions:
•

tname is the name of the task.

•

exec is the name of the execution.

•

type is the type of output: TEXT, HTML, or XML.

•

sect is the section of the report: SUMMARY, FINDINGS, ERRORS, and ALL.

•

lvl is the format of the report: BASIC, TYPICAL, ALL, or SHOW_HIDDEN.

SET
SET
SET
SET

LINESIZE 3000
LONG 500000
PAGESIZE 0
LONGCHUNKSIZE 100000

SELECT DBMS_STATS.REPORT_ADVISOR_TASK('tname', 'exec', 'type', 'sect', 'lvl')
AS REPORT
FROM DUAL;

For example, to print a report for AUTO_STATS_ADVISOR_TASK, use the following query:
SELECT
DBMS_STATS.REPORT_ADVISOR_TASK('AUTO_STATS_ADVISOR_TASK',NULL,'TEXT','ALL','ALL')
AS REPORT
FROM DUAL;

The following sample report shows four findings:
GENERAL INFORMATION
------------------------------------------------------------------------------Task Name
: AUTO_STATS_ADVISOR_TASK
Execution Name : EXEC_136
Created
: 09-05-16 02:52:34
Last Modified
: 09-05-16 12:31:24
------------------------------------------------------------------------------SUMMARY
------------------------------------------------------------------------------For execution EXEC_136 of task AUTO_STATS_ADVISOR_TASK, the Statistics Advisor
has 4 findings. The findings are related to the following rules:
AVOIDSETPROCEDURES, USEDEFAULTPARAMS, USEGATHERSCHEMASTATS, NOTUSEINCREMENTAL.
Please refer to the finding section for detailed information.
------------------------------------------------------------------------------FINDINGS
------------------------------------------------------------------------------Rule Name:
AvoidSetProcedures
Rule Description: Avoid Set Statistics Procedures
Finding: There are 5 SET_[COLUMN|INDEX|TABLE|SYSTEM]_STATS procedures being
used for statistics gathering.
Recommendation: Do not use SET_[COLUMN|INDEX|TABLE|SYSTEM]_STATS procedures.
Gather statistics instead of setting them.
Rationale: SET_[COLUMN|INDEX|TABLE|SYSTEM]_STATS will cause bad plans due to
wrong or inconsistent statistics.
---------------------------------------------------Rule Name:
UseDefaultParams
Rule Description: Use Default Parameters in Statistics Collection Procedures
Finding: There are 367 statistics operations using nondefault parameters.
Recommendation: Use default parameters for statistics operations.
Example:
-- Gathering statistics for 'SH' schema using all default parameter values:

18-28

Chapter 18

Basic Tasks for Optimizer Statistics Advisor

BEGIN dbms_stats.gather_schema_stats('SH'); END;
Rationale: Using default parameter values for statistics gathering operations
is more efficient.
---------------------------------------------------Rule Name:
UseGatherSchemaStats
Rule Description: Use gather_schema_stats procedure
Finding: There are 318 uses of GATHER_TABLE_STATS.
Recommendation: Use GATHER_SCHEMA_STATS instead of GATHER_TABLE_STATS.
Example:
-- Gather statistics for 'SH' schema:
BEGIN dbms_stats.gather_schema_stats('SH'); END;
Rationale: GATHER_SCHEMA_STATS has more options available, including checking
for staleness and gathering statistics concurrently. Also it is
more maintainable for new tables added to the schema. If you only
want to gather statistics for certain tables in the schema, specify
them in the obj_filter_list parameter of GATHER_SCHEMA_STATS.
---------------------------------------------------Rule Name:
NotUseIncremental
Rule Description: Statistics should not be maintained incrementally when it is
not
Finding: Incremental option has been turned on for 10 tables, which will not benefit
from using the incremental option.
Schema:
SH
Objects:
CAL_MONTH_SALES_MV
CAL_MONTH_SALES_MV
CHANNELS
COUNTRIES
CUSTOMERS
DIMENSION_EXCEPTIONS
FWEEK_PSCAT_SALES_MV
FWEEK_PSCAT_SALES_MV
PRODUCTS
PROMOTIONS
SUPPLEMENTARY_DEMOGRAPHICS
TIMES
Recommendation: Do not use the incremental option for statistics gathering on these
objects.
Example:
-Turn off the incremental option for 'SH.SALES':
dbms_stats.set_table_prefs('SH', 'SALES', 'INCREMENTAL', 'FALSE');
Rationale: The overhead of using the incremental option on these tables
outweighs the benefit of using the incremental option.

See Also:
Oracle Database PL/SQL Packages and Types Reference to learn more
about REPORT_ADVISOR_TASK

18-29

Chapter 18

Basic Tasks for Optimizer Statistics Advisor

18.2.6 Implementing Optimizer Statistics Advisor Recommendations
You can either implement all recommendations automatically using
DBMS_STATS.IMPLEMENT_ADVISOR_TASK, or generate an editable script using
DBMS_STATS.SCRIPT_ADVISOR_TASK.
This section contains the following topics:
•

Implementing Actions Recommended by Optimizer Statistics Advisor
The DBMS_STATS.IMPLEMENT_ADVISOR_TASK function implements the recommendations
for a specified Optimizer Statistics Advisor task. If you do not specify an execution
name, then Optimizer Statistics Advisor uses the most recent execution.

•

Generating a Script Using Optimizer Statistics Advisor
The DBMS_STATS.SCRIPT_ADVISOR_TASK function generates an editable script with
recommendations for a specified Optimizer Statistics Advisor task.

18.2.6.1 Implementing Actions Recommended by Optimizer Statistics Advisor
The DBMS_STATS.IMPLEMENT_ADVISOR_TASK function implements the recommendations for
a specified Optimizer Statistics Advisor task. If you do not specify an execution name,
then Optimizer Statistics Advisor uses the most recent execution.
The simplest means of implementing recommendations is using
DBMS_STATS.IMPLEMENT_ADVISOR_TASK. In this case, no generation of a script is necessary.
You can specify that the advisor ignore the existing filters (level=>'ALL') or use the
default, which honors the existing filters (level=>'TYPICAL').
Prerequisites
To use DBMS_STATS.IMPLEMENT_ADVISOR_TASK, you must meet the following prerequisites:
•

To execute this subprogram, you must have the ADVISOR privilege.

•

You must be the owner of the task.

Note:
This subprogram executes using invoker's rights.

The results of performing this task depend on the privileges of the executing user:
•

SYSTEM level

Only users with both the ANALYZE ANY and ANALYZE ANY DICTIONARY privileges can
perform this task on system-level rules.
•

Operation level
The results depend on the following privileges:
–

Users with both the ANALYZE ANY and ANALYZE ANY DICTIONARY privileges can
perform this task for all statistics operations.

18-30

Chapter 18

Basic Tasks for Optimizer Statistics Advisor

•

–

Users with the ANALYZE ANY privilege but not the ANALYZE ANY DICTIONARY
privilege can perform this task for statistics operations related to any schema
except SYS.

–

Users with the ANALYZE ANY DICTIONARY privilege but not the ANALYZE ANY
privilege can perform this task for statistics operations related to their own
schema and the SYS schema.

–

Users with neither the ANALYZE ANY nor the ANALYZE ANY DICTIONARY privilege
can only perform this operation for statistics operations relating to their own
schema.

Object level
Users can perform this task for any object for which they have statistics collection
privileges.

To implement advisor actions:
1.

In SQL*Plus, log in to the database as a user with the necessary privileges.

2.

Execute the DBMS_STATS.IMPLEMENT_ADVISOR_TASK function in the following form,
where the placeholders have the following definitions:
•

tname is the name of the task.

•

result is the CLOB variable that contains a list of the recommendations that

have been successfully implemented.
•

filter_lvl is the level of implementation: TYPICAL (existing filters honored) or
ALL (filters ignored).

EXECUTE result := DBMS_STATS.IMPLEMENT_ADVISOR_TASK('tname', level =>
filter_lvl);

For example, to implement all recommendations for the task opt_adv_task1, use
the following code:
VARIABLE b_ret CLOB
DECLARE
v_tname VARCHAR2(32767);
BEGIN
v_tname := 'opt_adv_task1';
:b_ret := DBMS_STATS.IMPLEMENT_ADVISOR_TASK(v_tname);
END;
/
3.

Optionally, print the XML output to confirm the implemented actions.
For example, to print the XML returned in the previous step, use the following code
(sample output included):
SET LONG 10000
SELECT XMLType(:b_ret) AS imp_results FROM DUAL;
IMP_RESULTS
-----------------------------------

yes


yes


18-31

Chapter 18

Basic Tasks for Optimizer Statistics Advisor


yes


yes


no


yes


yes



See Also:
Oracle Database PL/SQL Packages and Types Reference to learn more
about DBMS_STATS.IMPLEMENT_ADVISOR_TASK

18.2.6.2 Generating a Script Using Optimizer Statistics Advisor
The DBMS_STATS.SCRIPT_ADVISOR_TASK function generates an editable script with
recommendations for a specified Optimizer Statistics Advisor task.
Unlike IMPLEMENT_ADVISOR_TASK, the SCRIPT_ADVISOR_TASK generates a script that you can
edit before execution. The output script contains both comments and executable code.
As with IMPLEMENT_ADVISOR_TASK, you can specify that the advisor ignore the existing
filters (level=>'ALL') or use the default, which honors the existing filters
(level=>'TYPICAL'). You can specify that the function returns the script as a CLOB and
file, or only a CLOB.
Prerequisites
To use the DBMS_STATS.SCRIPT_ADVISOR_TASK function, you must meet the following
prerequisites:
•

To execute this subprogram, you must have the ADVISOR privilege.

•

You must be the owner of the task.

Note:
This subprogram executes using invoker's rights.

The results of performing this task depend on the privileges of the executing user:
•

SYSTEM level

Only users with both the ANALYZE ANY and ANALYZE ANY DICTIONARY privileges can
perform this task on system-level rules.

18-32

Chapter 18

Basic Tasks for Optimizer Statistics Advisor

•

Operation level
The results depend on the following privileges:

•

–

Users with both the ANALYZE ANY and ANALYZE ANY DICTIONARY privileges can
perform this task for all statistics operations.

–

Users with the ANALYZE ANY privilege but not the ANALYZE ANY DICTIONARY
privilege can perform this task for statistics operations related to any schema
except SYS.

–

Users with the ANALYZE ANY DICTIONARY privilege but not the ANALYZE ANY
privilege can perform this task for statistics operations related to their own
schema and the SYS schema.

–

Users with neither the ANALYZE ANY nor the ANALYZE ANY DICTIONARY privilege
can only perform this operation for statistics operations relating to their own
schema.

Object level
Users can perform this task for any object for which they have statistics collection
privileges.

To generate an advisor script:
1.

In SQL*Plus, log in to the database as a user with ADVISOR privileges.

2.

Execute the DBMS_STATS.SCRIPT_ADVISOR_TASK function in the following form, where
the placeholders have the following definitions:
•

tname is the name of the task.

•

exec is the name of the execution (default is null).

•

dir is the name of the directory (default is null).

•

result is the CLOB variable that contains a list of the recommendations that

have been successfully implemented.
•

filter_lvl is the level of implementation: TYPICAL (existing filters honored) or
ALL (filters ignored).

EXEC result := DBMS_STATS.SCRIPT_ADVISOR_TASK('tname', execution_name => 'exec',
dir_name => 'dir', level => 'filter_lvl');

For example, to generate a script that contains recommendations for the task
opt_adv_task1, use the following code:
VARIABLE b_script CLOB
DECLARE
v_tname VARCHAR2(32767);
BEGIN
v_tname := 'opt_adv_task1';
:b_script := DBMS_STATS.SCRIPT_ADVISOR_TASK(v_tname);
END;
/

18-33

Chapter 18

Basic Tasks for Optimizer Statistics Advisor

Note:
If you do not specify an execution name, then Optimizer Statistics
Advisor uses the most recent execution.
3.

Print the script.
For example, to print the script returned in the previous step, use the following
code (sample output included):
DECLARE
v_len
NUMBER(10);
v_offset NUMBER(10) :=1;
v_amount NUMBER(10) :=10000;
BEGIN
v_len := DBMS_LOB.getlength(:b_script);
WHILE (v_offset < v_len)
LOOP
DBMS_OUTPUT.PUT_LINE(DBMS_LOB.SUBSTR(:b_script,v_amount,v_offset));
v_offset := v_offset + v_amount;
END LOOP;
END;
/

The following example shows a sample script:
-- Script generated for the recommendations from execution EXEC_23
-- in the statistics advisor task OPT_ADV_TASK1
-- Script version 12.2
------------------------------

No scripts will be provided for
for more details.
No scripts will be provided for
for more details.
No scripts will be provided for
report for more details.
No scripts will be provided for
the report for more details.
No scripts will be provided for
report for more details.
No scripts will be provided for
for more details.
No scripts will be provided for
for more details.
No scripts will be provided for
for more details.
No scripts will be provided for
for more details.
No scripts will be provided for
for more details.
No scripts will be provided for
report for more details.
No scripts will be provided for
the report for more details.
No scripts will be provided for
report for more details.
No scripts will be provided for
for more details.
No scripts will be provided for

the rule AVOIDSETPROCEDURES. Please check the report
the rule USEGATHERSCHEMASTATS. Please check the report
the rule AVOIDINEFFICIENTSTATSOPRSEQ. Please check the
the rule AVOIDUNNECESSARYSTATSCOLLECTION. Please check
the rule GATHERSTATSAFTERBULKDML. Please check the
the rule AVOIDDROPRECREATE. Please check the report
the rule AVOIDOUTOFRANGE. Please check the report
the rule AVOIDANALYZETABLE. Please check the report
the rule AVOIDSETPROCEDURES. Please check the report
the rule USEGATHERSCHEMASTATS. Please check the report
the rule AVOIDINEFFICIENTSTATSOPRSEQ. Please check the
the rule AVOIDUNNECESSARYSTATSCOLLECTION. Please check
the rule GATHERSTATSAFTERBULKDML. Please check the
the rule AVOIDDROPRECREATE. Please check the report
the rule AVOIDOUTOFRANGE. Please check the report

18-34

Chapter 18

Basic Tasks for Optimizer Statistics Advisor

-- for more details.
-- No scripts will be provided for the rule AVOIDANALYZETABLE. Please check the report
-- for more details.
-- Scripts for rule USEDEFAULTPARAMS
-- Rule Description: Use Default Parameters in Statistics Collection Procedures
-- Use the default preference value for parameters
begin dbms_stats.set_global_prefs('PREFERENCE_OVERRIDES_PARAMETER', 'TRUE'); end;
/
---------

Scripts for rule USEDEFAULTOBJECTPREFERENCE
Rule Description: Use Default Object Preference for statistics collection
Setting object-level preferences to default values
setting CASCADE to default value for object level preference
setting ESTIMATE_PERCENT to default value for object level preference
setting METHOD_OPT to default value for object level preference
setting GRANULARITY to default value for object level preference
setting NO_INVALIDATE to default value for object levelpreference

-- Scripts for rule USEINCREMENTAL
-- Rule Description: Statistics should be maintained incrementally when it is beneficial.
-- Turn on the incremental option for those objects for which using incremental is helpful.
-- Scripts for rule UNLOCKNONVOLATILETABLE
-- Rule Description: Statistics for objects with non-volatile should not be locked
-- Unlock statistics for objects that are not volatile.
-- Scripts for rule LOCKVOLATILETABLE
-- Rule Description: Statistics for objects with volatile data should be locked
-- Lock statistics for volatile objects.
-- Scripts for rule NOTUSEINCREMENTAL
-- Rule Description: Statistics should not be maintained incrementally when it is not
beneficial
-- Turn off incremental option for those objects for which using incremental is not helpful.
begin
/
begin
/
begin
/
begin
/
begin
/
begin
/
begin
/
begin
/
begin
/
begin
/

dbms_stats.set_table_prefs('SH', 'CAL_MONTH_SALES_MV', 'INCREMENTAL', 'FALSE'); end;
dbms_stats.set_table_prefs('SH', 'CHANNELS', 'INCREMENTAL', 'FALSE'); end;
dbms_stats.set_table_prefs('SH', 'COUNTRIES', 'INCREMENTAL', 'FALSE'); end;
dbms_stats.set_table_prefs('SH', 'CUSTOMERS', 'INCREMENTAL', 'FALSE'); end;
dbms_stats.set_table_prefs('SH', 'DIMENSION_EXCEPTIONS', 'INCREMENTAL', 'FALSE'); end;
dbms_stats.set_table_prefs('SH', 'FWEEK_PSCAT_SALES_MV', 'INCREMENTAL', 'FALSE'); end;
dbms_stats.set_table_prefs('SH', 'PRODUCTS', 'INCREMENTAL', 'FALSE'); end;
dbms_stats.set_table_prefs('SH', 'PROMOTIONS', 'INCREMENTAL', 'FALSE'); end;
dbms_stats.set_table_prefs('SH', 'SUPPLEMENTARY_DEMOGRAPHICS', 'INCREMENTAL', 'FALSE'); end;
dbms_stats.set_table_prefs('SH', 'TIMES', 'INCREMENTAL', 'FALSE'); end;

-- Scripts for rule USEAUTODEGREE
-- Rule Description: Use Auto Degree for statistics collection
-- Turn on auto degree for those objects for which using auto degree is helpful.

18-35

Chapter 18

Basic Tasks for Optimizer Statistics Advisor

-- Scripts for rule AVOIDSTALESTATS
-- Rule Description: Avoid objects with stale or no statistics
-- Gather statistics for those objcts that are missing or have no statistics.
-- Scripts for rule MAINTAINSTATSCONSISTENCY
-- Rule Description: Statistics of dependent objects should be consistent
-- Gather statistics for those objcts that are missing or have no statistics.

See Also:
Oracle Database PL/SQL Packages and Types Reference to learn more
about DBMS_STATS.SCRIPT_ADVISOR_TASK

18-36

Part VI
Optimizer Controls
You can use hints and initialization parameter to influence optimizer decisions and
behavior.
This part contains the following chapters:
•

Influencing the Optimizer
Optimizer defaults are adequate for most operations, but not all.

•

Improving Real-World Performance Through Cursor Sharing
Cursor sharing can improve database application performance by orders of
magnitude.

19
Influencing the Optimizer
Optimizer defaults are adequate for most operations, but not all.
In some cases you may have information unknown to the optimizer, or need to tune
the optimizer for a specific type of statement or workload. In such cases, influencing
the optimizer may provide better performance.
This chapter contains the following topics:
•

Techniques for Influencing the Optimizer
You can influence the optimizer using several techniques, including SQL profiles,
SQL Plan Management, initialization parameters, and hints.

•

Influencing the Optimizer with Initialization Parameters
This chapter explains which initialization parameters affect optimization, and how
to set them.

•

Influencing the Optimizer with Hints
Optimizer hints are special comments in a SQL statement that pass instructions to
the optimizer.

19.1 Techniques for Influencing the Optimizer
You can influence the optimizer using several techniques, including SQL profiles, SQL
Plan Management, initialization parameters, and hints.
The following figure shows the principal techniques for influencing the optimizer.

19-1

Chapter 19

Techniques for Influencing the Optimizer

Figure 19-1

Techniques for Influencing the Optimizer
User

DBMS_STATS

Initialization Parameters

SQL Plan Management
SQL Profiles
Hints

Optimizer

The overlapping squares in the preceding diagram show that SQL plan management
uses both initialization parameters and hints. SQL profiles also technically include
hints.

Note:
A stored outline is a legacy technique that serve a similar purpose to SQL
plan baselines.

You can use the following techniques to influence the optimizer:
Table 19-1

Optimizer Techniques

Technique

Description

To Learn More

Initialization
parameters

Parameters influence many types of
optimizer behavior at the database
instance and session level.

"Influencing the Optimizer with
Initialization Parameters"

Hints

A hint is a commented instruction in a SQL "Influencing the Optimizer with
statement. Hints control a wide range of
Hints"
behavior.

DBMS_STATS

This package updates and manages
"Gathering Optimizer Statistics"
optimizer statistics. The more accurate the
statistics, the better the optimizer
estimates. This chapter does not cover
DBMS_STATS.

19-2

Chapter 19

Influencing the Optimizer with Initialization Parameters

Table 19-1

(Cont.) Optimizer Techniques

Technique

Description

To Learn More

SQL profiles

A SQL profile is a database object that
contains auxiliary statistics specific to a
SQL statement. Conceptually, a SQL
profile is to a SQL statement what a set of
object-level statistics is to a table or index.
A SQL profile can correct suboptimal
optimizer estimates discovered during
SQL tuning.

"Managing SQL Profiles"

SQL plan
management
and stored
outlines

SQL plan management is a preventative
"Managing SQL Plan
mechanism that enables the optimizer to
Baselines"
automatically manage execution plans,
ensuring that the database uses only
known or verified plans. This chapter does
not cover SQL plan management.

In some cases, multiple techniques optimize the same behavior. For example, you can
set optimizer goals using both initialization parameters and hints.

See Also:
"Migrating Stored Outlines to SQL Plan Baselines" to learn how to migrate
stored outlines to SQL plan baselines

19.2 Influencing the Optimizer with Initialization Parameters
This chapter explains which initialization parameters affect optimization, and how to
set them.
This section contains the following topics:
•

About Optimizer Initialization Parameters
Oracle Database provides initialization parameters to influence various aspects of
optimizer behavior, including cursor sharing, adaptive optimization, and the
optimizer mode.

•

Enabling Optimizer Features
The OPTIMIZER_FEATURES_ENABLE initialization parameter (or hint) controls a set of
optimizer-related features, depending on the database release.

•

Choosing an Optimizer Goal
The optimizer goal is the prioritization of resource usage by the optimizer.

•

Controlling Adaptive Optimization
In Oracle Database, adaptive query optimization is the process by which the
optimizer adapts an execution plan based on statistics collected at run time.

19-3

Chapter 19

Influencing the Optimizer with Initialization Parameters

19.2.1 About Optimizer Initialization Parameters
Oracle Database provides initialization parameters to influence various aspects of
optimizer behavior, including cursor sharing, adaptive optimization, and the optimizer
mode.
The following table lists some of the most important optimizer parameters.
Table 19-2

Initialization Parameters That Control Optimizer Behavior

Initialization Parameter

Description

APPROX_FOR_AGGREGATION

Uses approximate query processing for all aggregation and
analytic queries. Approximate processing is useful when you
want to obtain faster query results and avoid writes to a
temporary tablespaces. This optimizer uses a nondeterministic
algorithm to make its estimations, which means that different
queries can obtain different results.
You can set this parameter to TRUE at the system or session
level.
This parameter changes the optimizer environment. It does not
force the optimizer to change the SQL text for an affected query,
but it does force the optimizer to reparse the query and create a
new child cursor.

APPROX_FOR_COUNT_DISTINCT

Replaces queries that contain COUNT (DISTINCT expr) queries
with APPROX_COUNT_DISTINCT. Approximate counts are useful
when a column has a higher number of distinct values, and you
want to obtain faster query results and avoid writes to a
temporary tablespaces. Only use approximation when your
application can tolerate a nonzero error rate.
This parameter changes the optimizer environment. It does not
force the optimizer to change the SQL text for an affected query,
but it does force the optimizer to reparse the query and create a
new child cursor.

APPROX_FOR_PERCENTILE

Converts exact percentile functions to their approximate
percentile function counterparts.
Approximate percentile function queries are faster than their
exact percentile function query counterparts, so they can be
useful in situations where a tolerable amount of error is
acceptable in order to obtain faster query results.
Set to PERCENTILE_CONT to convert PERCENTILE_CONT functions
to APPROX_PERCENTILE, and PERCENTILE_DISC to convert
PERCENTILE_DISC functions to APPROX_PERCENTILE (or ALL to
convert both).
This parameter changes the optimizer environment. It does not
force the optimizer to change the SQL text for an affected query,
but it does force the optimizer to reparse the query and create a
new child cursor.

19-4

Chapter 19

Influencing the Optimizer with Initialization Parameters

Table 19-2

(Cont.) Initialization Parameters That Control Optimizer Behavior

Initialization Parameter

Description

CURSOR_INVALIDATION

Provides the default cursor invalidation level for DDL
statements.
IMMEDIATE sets the same cursor invalidation behavior for DDL
as in releases before Oracle Database 12c Release 2 (12.2).
This is the default.
DEFERRED allows an application to take advantage of the reduced
cursor invalidation for DDL without making any application
changes. Deferred invalidation reduces the number of cursor
invalidations and spreads the recompilation workload over time.
For this reason, a cursor may run with a suboptimal plan until it
is recompiled, and may incur small execution-time overhead.
You can set this parameter at the SYSTEM or SESSION level. See
"About the Life Cycle of Shared Cursors".

CURSOR_SHARING

Converts literal values in SQL statements to bind variables.
Converting the values improves cursor sharing and can affect
the execution plans of SQL statements. The optimizer generates
the execution plan based on the presence of the bind variables
and not the actual literal values.
Set to FORCE to enable the creation of a new cursor when
sharing an existing cursor, or when the cursor plan is not
optimal. Set to EXACT to allow only statements with identical text
to share the same cursor.

DB_FILE_MULTIBLOCK_READ_COUNT

Specifies the number of blocks that are read in a single I/O
during a full table scan or index fast full scan. The optimizer
uses the value of this parameter to calculate the cost of full table
scans and index fast full scans. Larger values result in a lower
cost for full table scans, which may result in the optimizer
choosing a full table scan over an index scan.
The default value of this parameter corresponds to the
maximum I/O size that the database can perform efficiently. This
value is platform-dependent and is 1 MB for most platforms.
Because the parameter is expressed in blocks, it is set to a
value equal to the maximum I/O size that can be performed
efficiently divided by the standard block size. If the number of
sessions is extremely large, then the multiblock read count value
decreases to avoid the buffer cache getting flooded with too
many table scan buffers.

OPTIMIZER_ADAPTIVE_PLANS

Controls adaptive plans. An adaptive plan has alternative
choices. The optimizer decides on a plan at run time based on
statistics collected as the query executes.
By default, this parameter is true, which means adaptive plans
are enabled. Setting to this parameter to false disables the
following features:
•
Nested loops and hash join selection
•
Star transformation bitmap pruning
•
Adaptive parallel distribution method
See "About Adaptive Query Plans".

19-5

Chapter 19

Influencing the Optimizer with Initialization Parameters

Table 19-2

(Cont.) Initialization Parameters That Control Optimizer Behavior

Initialization Parameter

Description

OPTIMIZER_ADAPTIVE_REPORTING_ONLY

Controls the reporting mode for automatic reoptimization and
adaptive plans (see "Adaptive Query Plans"). By default,
reporting mode is off (false), which means that adaptive
optimizations are enabled.
If set to true, then adaptive optimizations run in reporting-only
mode. In this case, the database gathers information required
for an adaptive optimization, but takes no action to change the
plan. For example, an adaptive plan always choose the default
plan, but the database collects information about which plan the
database would use if the parameter were set to false. You can
view the report by using DBMS_XPLAN.DISPLAY_CURSOR.

OPTIMIZER_ADAPTIVE_STATISTICS

Controls adaptive statistics. The optimizer can use adaptive
statistics when query predicates are too complex to rely on base
table statistics alone.
By default, OPTIMIZER_ADAPTIVE_STATISTICS is false, which
means that the following features are disabled:
•
SQL plan directives
•
Statistics feedback
•
Performance feedback
•
Adaptive dynamic sampling
See "Adaptive Statistics".

OPTIMIZER_MODE

Sets the optimizer mode at database instance startup. Possible
values are ALL_ROWS, FIRST_ROWS_n, and FIRST_ROWS.

OPTIMIZER_INDEX_CACHING

Controls the cost analysis of an index probe with a nested loop.
The range of values 0 to 100 indicates percentage of index
blocks in the buffer cache, which modifies optimizer
assumptions about index caching for nested loops and IN-list
iterators. A value of 100 infers that 100% of the index blocks are
likely to be found in the buffer cache, so the optimizer adjusts
the cost of an index probe or nested loop accordingly. Use
caution when setting this parameter because execution plans
can change in favor of index caching.

OPTIMIZER_INDEX_COST_ADJ

Adjusts the cost of index probes. The range of values is 1 to
10000. The default value is 100, which means that the optimizer
evaluates indexes as an access path based on the normal cost
model. A value of 10 means that the cost of an index access
path is one-tenth the normal cost of an index access path.

OPTIMIZER_INMEMORY_AWARE

This parameter enables (TRUE) or disables (FALSE) all Oracle
Database In-Memory (Database In-Memory) optimizer features,
including the cost model for the IM column store, table
expansion, Bloom filters, and so on. Setting the parameter to
FALSE causes the optimizer to ignore the INMEMORY property of
tables during the optimization of SQL statements.

OPTIMIZER_USE_INVISIBLE_INDEXES

Enables or disables the use of invisible indexes.

19-6

Chapter 19

Influencing the Optimizer with Initialization Parameters

Table 19-2

(Cont.) Initialization Parameters That Control Optimizer Behavior

Initialization Parameter

Description

RESULT_CACHE_MODE

Controls whether the database uses the SQL query result cache
for all queries, or only for the queries that are annotated with the
result cache hint. When set to MANUAL (the default), you must
use the RESULT_CACHE hint to specify that a specific result is to
be stored in the cache. When set to FORCE, the database stores
all results in the cache.
When setting this parameter, consider how the result cache
handles PL/SQL functions. The database invalidates query
results in the result cache using the same mechanism that
tracks data dependencies for PL/SQL functions, but otherwise
permits caching of queries that contain PL/SQL functions.
Because PL/SQL function result cache invalidation does not
track all kinds of dependencies (such as on sequences,
SYSDATE, SYS_CONTEXT, and package variables), indiscriminate
use of the query result cache on queries calling such functions
can result in changes to results, that is, incorrect results. Thus,
consider correctness and performance when choosing to enable
the result cache, especially when setting RESULT_CACHE_MODE to
FORCE.

RESULT_CACHE_MAX_SIZE

Changes the memory allocated to the result cache. If you set
this parameter to 0, then the result cache is disabled. The value
of this parameter is rounded to the largest multiple of 32 KB that
is not greater than the specified value. If the rounded value is 0,
then the feature is disabled.

RESULT_CACHE_MAX_RESULT

Specifies the maximum amount of cache memory that any
single result can use. The default value is 5%, but you can
specify any percentage value between 1 and 100.

RESULT_CACHE_REMOTE_EXPIRATION

Specifies the number of minutes for which a result that depends
on remote database objects remains valid. The default is 0,
which implies that the database should not cache results using
remote objects. Setting this parameter to a nonzero value can
produce stale answers, such as if a remote database modifies a
table that is referenced in a result.

STAR_TRANSFORMATION_ENABLED

Enables the optimizer to cost a star transformation for star
queries (if true). The star transformation combines the bitmap
indexes on the various fact table columns.

19-7

Chapter 19

Influencing the Optimizer with Initialization Parameters

See Also:
•

Oracle Database Reference for complete information about the
preceding initialization parameters

•

Oracle Database Performance Tuning Guide to learn how to tune the
query result cache

•

Oracle Database Data Warehousing Guide
to learn more about star transformations

•

Oracle Database In-Memory Guide to learn more about Database InMemory features

19.2.2 Enabling Optimizer Features
The OPTIMIZER_FEATURES_ENABLE initialization parameter (or hint) controls a set of
optimizer-related features, depending on the database release.
The parameter accepts one of a list of valid string values corresponding to the release
numbers, such as 11.2.0.2 or 12.2.0.1. You can use this parameter to preserve the old
behavior of the optimizer after a database upgrade. For example, if you upgrade
Oracle Database 12c Release 1 (12.1.0.2) to Oracle Database 12c Release 2
(12.2.0.1), then the default value of the OPTIMIZER_FEATURES_ENABLE parameter changes
from 12.1.0.2 to 12.2.0.1.
For backward compatibility, you may not want the execution plans to change because
of new optimizer features in a new release. In such cases, you can set
OPTIMIZER_FEATURES_ENABLE to an earlier version. If you upgrade to a new release, and if
you want to enable the features in the new release, then you do not need to explicitly
set the OPTIMIZER_FEATURES_ENABLE initialization parameter.

Caution:
Oracle does not recommend explicitly setting the OPTIMIZER_FEATURES_ENABLE
initialization parameter to an earlier release. To avoid SQL performance
regression that may result from execution plan changes, consider using SQL
plan management instead.

Assumptions
This tutorial assumes the following:
•

You recently upgraded the database from Oracle Database 12c Release 1 (12
1.0.2) to Oracle Database 12c Release 2 (12.2.0.1).

•

You want to preserve the optimizer behavior from the earlier release.

To enable query optimizer features for a specific release:
1.

Log in to the database with the appropriate privileges, and then query the current
optimizer features settings.

19-8

Chapter 19

Influencing the Optimizer with Initialization Parameters

For example, run the following SQL*Plus command:
SQL> SHOW PARAMETER optimizer_features_enable
NAME
TYPE
VALUE
------------------------------------ ----------- -------optimizer_features_enable
string
12.2.0.1
2.

Set the optimizer features setting at the instance or session level.
For example, run the following SQL statement to set the optimizer version to
12.1.0.2:
SQL> ALTER SYSTEM SET OPTIMIZER_FEATURES_ENABLE='12.1.0.2';

The preceding statement restores the optimizer functionality that existed in Oracle
Database 12c Release 1 (12.1.0.2).

See Also:
•

"Managing SQL Plan Baselines"

•

Oracle Database Reference to learn about optimizer features enabled
when you set OPTIMIZER_FEATURES_ENABLE to different release values

19.2.3 Choosing an Optimizer Goal
The optimizer goal is the prioritization of resource usage by the optimizer.
Using the OPTIMIZER_MODE initialization parameter, you can set the following optimizer
goals:
•

Best throughput (default)
When you set the OPTIMIZER_MODE value to ALL_ROWS, the database uses the least
amount of resources necessary to process all rows that the statement accessed.
For batch applications such as Oracle Reports, optimize for best throughput.
Usually, throughput is more important in batch applications because the user is
only concerned with the time necessary for the application to complete. Response
time is less important because the user does not examine the results of individual
statements while the application is running.

•

Best response time
When you set the OPTIMIZER_MODE value to FIRST_ROWS_n, the database optimizes
with a goal of best response time to return the first n rows, where n equals 1, 10,
100, or 1000.
For interactive applications in Oracle Forms or SQL*Plus, optimize for response
time. Usually, response time is important because the interactive user is waiting to
see the first row or rows that the statement accessed.

Assumptions
This tutorial assumes the following:

19-9

Chapter 19

Influencing the Optimizer with Initialization Parameters

•

The primary application is interactive, so you want to set the optimizer goal for the
database instance to minimize response time.

•

For the current session only, you want to run a report and optimize for throughput.

To enable query optimizer features for a specific release:
1.

Connect SQL*Plus to the database with the appropriate privileges, and then query
the current optimizer mode.
For example, run the following SQL*Plus command:
dba1@PROD> SHOW PARAMETER OPTIMIZER_MODE
NAME
TYPE
VALUE
------------------------------------ ----------- -------optimizer_mode
string
ALL_ROWS

2.

At the instance level, optimize for response time.
For example, run the following SQL statement to configure the system to retrieve
the first 10 rows as quickly as possible:
SQL> ALTER SYSTEM SET OPTIMIZER_MODE='FIRST_ROWS_10';

3.

At the session level only, optimize for throughput before running a report.
For example, run the following SQL statement to configure only this session to
optimize for throughput:
SQL> ALTER SESSION SET OPTIMIZER_MODE='ALL_ROWS';

See Also:
Oracle Database Reference to learn about the OPTIMIZER_MODE initialization
parameter

19.2.4 Controlling Adaptive Optimization
In Oracle Database, adaptive query optimization is the process by which the
optimizer adapts an execution plan based on statistics collected at run time.
Adaptive plans are enabled when the following initialization parameters are set:
•

OPTIMIZER_ADAPTIVE_PLANS is TRUE (default)

•

OPTIMIZER_FEATURES_ENABLE is 12.1.0.1 or later

•

OPTIMIZER_ADAPTIVE_REPORTING_ONLY is FALSE (default)

If OPTIMIZER_ADAPTIVE_REPORTING_ONLY is set to true, then adaptive optimization runs in
reporting-only mode. In this case, the database gathers information required for
adaptive optimization, but does not change the plans. An adaptive plan always
chooses the default plan, but the database collects information about the execution as
if the parameter were set to false.
Adaptive statistics are enabled when the following initialization parameters are set:
•

OPTIMIZER_ADAPTIVE_STATISTICS is TRUE (the default is FALSE)

19-10

Chapter 19

Influencing the Optimizer with Hints

•

OPTIMIZER_FEATURES_ENABLE is 12.1.0.1 or later

Assumptions
This tutorial assumes the following:
•

The OPTIMIZER_FEATURES_ENABLE initialization parameter is set to 12.1.0.1 or later.

•

The OPTIMIZER_ADAPTIVE_REPORTING_ONLY initialization parameter is set to false
(default).

•

You want to disable adaptive plans for testing purposes so that the database
generates only reports.

To disable adaptive plans:
1.

Connect SQL*Plus to the database as SYSTEM, and then query the current settings.
For example, run the following SQL*Plus command:
SHOW PARAMETER OPTIMIZER_ADAPTIVE_REPORTING_ONLY
NAME
TYPE
VALUE
------------------------------------ ----------- ----optimizer_adaptive_reporting_only
boolean
FALSE

2.

At the session level, set the OPTIMIZER_ADAPTIVE_REPORTING_ONLY initialization
parameter to true.
For example, in SQL*Plus run the following SQL statement:
ALTER SESSION SET OPTIMIZER_ADAPTIVE_REPORTING_ONLY=true;

3.

Run a query.

4.

Run DBMS_XPLAN.DISPLAY_CURSOR with the +REPORT parameter.
When the +REPORT parameter is set, the report shows the plan the optimizer would
have picked if automatic reoptimization had been enabled.

See Also:
•
•

"About Adaptive Query Optimization"
Oracle Database Reference to learn about the
OPTIMIZER_ADAPTIVE_REPORTING_ONLY initialization parameter

•

Oracle Database PL/SQL Packages and Types Reference to learn about
the +REPORT parameter of the DBMS_XPLAN.DISPLAY_CURSOR function

19.3 Influencing the Optimizer with Hints
Optimizer hints are special comments in a SQL statement that pass instructions to the
optimizer.
The optimizer uses hints to choose an execution plan for the statement unless
prevented by some condition.

19-11

Chapter 19

Influencing the Optimizer with Hints

Note:
Oracle Database SQL Language Reference contains a complete reference
for all SQL hints

This section contains the following topics:
•

About Optimizer Hints
Use hints to influence the optimizer mode, query transformation, access path, join
order, and join methods.

•

Guidelines for Join Order Hints
The join order can have a significant effect on the performance of a SQL
statement. In some cases, you can specify join order hints in a SQL statement so
that it does not access rows that have no effect on the result.

19.3.1 About Optimizer Hints
Use hints to influence the optimizer mode, query transformation, access path, join
order, and join methods.
For example, The following figure shows how you can use a hint to tell the optimizer to
use a specific index for a specific statement.

Figure 19-2

Optimizer Hint

SELECT /*+ INDEX (employees emp_dep_ix)*/ ...

Optimizer

Generate Plan
Id
0
1
*2

Operation

Name

SELECT STATEMENT
TABLE ACCESS BY INDEX ROWID
INDEX UNIQUE SCAN

EMPLOYEES
EMP_DEP_IX

The advantage of hints is that they enable you to make decisions normally made by
the optimizer. In a test environment, hints are useful for testing the performance of a
specific access path. For example, you may know that an index is more selective for
certain queries, as in Figure 19-2. In this case, the hint may cause the optimizer to
generate a better plan.
The disadvantage of hints is the extra code that you must manage, check, and control.
Hints were introduced in Oracle7, when users had little recourse if the optimizer
generated suboptimal plans. Because changes in the database and host environment

19-12

Chapter 19

Influencing the Optimizer with Hints

can make hints obsolete or have negative consequences, a good practice is to test
using hints, but use other techniques to manage execution plans.
Oracle provides several tools, including SQL Tuning Advisor, SQL plan management,
and SQL Performance Analyzer, to address performance problems not solved by the
optimizer. Oracle strongly recommends that you use these tools instead of hints
because they provide fresh solutions as the data and database environment change.
This section contains the following topics:
•

Types of Hints
You can use hints for tables, query blocks, and statements.

•

Scope of Hints
When you specify a hint, it optimizes only the statement block in which it appears,
overriding any instance-level or session-level parameters.

•

Guidelines for Hints
You must enclose hints within a SQL comment.

See Also:
Oracle Database SQL Language Reference for the most common hints by
functional category.

19.3.1.1 Types of Hints
You can use hints for tables, query blocks, and statements.
Hints fall into the following types:
•

Single-table
Single-table hints are specified on one table or view. INDEX and USE_NL are
examples of single-table hints. The following statement uses a single-table hint:
SELECT /*+ INDEX (employees emp_department_ix)*/ employee_id, department_id
FROM employees
WHERE department_id > 50;

•

Multi-table
Multi-table hints are like single-table hints except that the hint can specify multiple
tables or views. LEADING is an example of a multi-table hint. The following
statement uses a multi-table hint:
SELECT
FROM
WHERE
AND

/*+ LEADING(e j) */ *
employees e, departments d, job_history j
e.department_id = d.department_id
e.hire_date = j.start_date;

Note:
USE_NL(table1 table2) is not considered a multi-table hint because it is a
shortcut for USE_NL(table1) and USE_NL(table2).

19-13

Chapter 19

Influencing the Optimizer with Hints

•

Query block
Query block hints operate on single query blocks. STAR_TRANSFORMATION and UNNEST
are examples of query block hints. The following statement uses a query block
hint:
SELECT /*+ STAR_TRANSFORMATION */ s.time_id, s.prod_id, s.channel_id
FROM sales s, times t, products p, channels c
WHERE s.time_id = t.time_id AND s.prod_id = p.prod_id
AND s.channel_id = c.channel_id AND c.channel_desc = 'Tele Sales';

•

Statement
Statement hints apply to the entire SQL statement. ALL_ROWS is an example of a
statement hint. The following statement uses a statement hint:
SELECT /*+ ALL_ROWS */ * FROM sales;

19.3.1.2 Scope of Hints
When you specify a hint, it optimizes only the statement block in which it appears,
overriding any instance-level or session-level parameters.
A statement block is one of the following:
•

A simple MERGE, SELECT, INSERT, UPDATE, or DELETE statement

•

A parent statement or a subquery of a complex statement

•

A part of a query using set operators (UNION, MINUS, INTERSECT)

Example 19-1

Query Using a Set Operator

The following query consists of two component queries and the UNION operator:
SELECT /*+ FIRST_ROWS(10) */ prod_id, time_id FROM 2010_sales
UNION ALL
SELECT /*+ ALL_ROWS */ prod_id, time_id FROM current_year_sales;

The preceding statement has two blocks, one for each component query. Hints in the
first component query apply only to its optimization, not to the optimization of the
second component query. For example, in the first week of 2015 you query current
year and last year sales. You apply FIRST_ROWS(10) to the query of last year's (2014)
sales and the ALL_ROWS hint to the query of this year's (2015) sales.

See Also:
Oracle Database SQL Language Reference for an overview of hints

19.3.1.3 Guidelines for Hints
You must enclose hints within a SQL comment.
The hint comment must immediately follow the first keyword of a SQL statement block.
You can use either style of comment: a slash-star (/*) or pair of dashes (--). The plussign (+) hint delimiter must come immediately after the comment delimiter, as in the
following fragment:
SELECT /*+ hint_text */ ...

19-14

Chapter 19

Influencing the Optimizer with Hints

The database ignores incorrectly specified hints. The database also ignores
combinations of conflicting hints, even if these hints are correctly specified. If one hint
is incorrectly specified, but a hint in the same comment is correctly specified, then the
database considers the correct hint.

Caution:
The database does not issue error messages for hints that it ignores.

A statement block can have only one comment containing hints, but it can contain
many space-separated hints. For example, a complex query may include multiple table
joins. If you specify only the INDEX hint for a specified table, then the optimizer must
determine the remaining access paths and corresponding join methods. The optimizer
may not use the INDEX hint because the join methods and access paths prevent it.
Example 19-2 uses multiple hints to specify the exact join order.
Example 19-2

Multiple Hints

SELECT

/*+ LEADING(e2 e1) USE_NL(e1) INDEX(e1 emp_emp_id_pk)
USE_MERGE(j) FULL(j) */
e1.first_name, e1.last_name, j.job_id, sum(e2.salary) total_sal
FROM
employees e1, employees e2, job_history j
WHERE
e1.employee_id = e2.manager_id
AND
e1.employee_id = j.employee_id
AND
e1.hire_date = j.start_date
GROUP BY e1.first_name, e1.last_name, j.job_id
ORDER BY total_sal;

See Also:
Oracle Database SQL Language Reference to learn about the syntax rules
for comments and hints

19.3.2 Guidelines for Join Order Hints
The join order can have a significant effect on the performance of a SQL statement.
In some cases, you can specify join order hints in a SQL statement so that it does not
access rows that have no effect on the result.
The driving table in a join is the table to which other tables are joined. In general, the
driving table contains the filter condition that eliminates the highest percentage of rows
in the table.
Consider the following guidelines:
•

Avoid a full table scan when an index retrieves the requested rows more
efficiently.

•

Avoid using an index that fetches many rows from the driving table when you can
use a different index that fetches a small number of rows.

•

Choose the join order so that you join fewer rows to tables later in the join order.

19-15

Chapter 19

Influencing the Optimizer with Hints

The following example shows how to tune join order effectively:
SELECT *
FROM taba a,
tabb b,
tabc c
WHERE a.acol BETWEEN 100 AND 200
AND
b.bcol BETWEEN 10000 AND 20000
AND
c.ccol BETWEEN 10000 AND 20000
AND
a.key1 = b.key1
AND
a.key2 = c.key2;
1.

Choose the driving table and the driving index (if any).
Each of the first three conditions in the previous example is a filter condition that
applies to a single table. The last two conditions are join conditions.
Filter conditions dominate the choice of driving table and index. In general, the
driving table contains the filter condition that eliminates the highest percentage of
rows. Thus, because the range of 100 to 200 is narrow compared with the range of
acol, but the ranges of 10000 and 20000 are relatively large, taba is the driving
table, all else being equal.
With nested loops joins, the joins occur through the join indexes, which are the
indexes on the primary or foreign keys used to connect that table to an earlier
table in the join tree. Rarely do you use the indexes on the non-join conditions,
except for the driving table. Thus, after taba is chosen as the driving table, use the
indexes on b.key1 and c.key2 to drive into tabb and tabc, respectively.

2.

Choose the best join order, driving to the best unused filters earliest.
You can reduce the work of the following join by first joining to the table with the
best still-unused filter. Thus, if bcol BETWEEN is more restrictive (rejects a higher
percentage of the rows) than ccol BETWEEN, then the last join becomes easier (with
fewer rows) if tabb is joined before tabc.

3.

You can use the ORDERED or STAR hint to force the join order.

See Also:
Oracle Database Reference to learn about OPTIMIZER_MODE

19-16

20
Improving Real-World Performance
Through Cursor Sharing
Cursor sharing can improve database application performance by orders of
magnitude.
This chapter contains the following topics:
•

Overview of Cursor Sharing
Oracle Database can share cursors, which are pointers to private SQL areas in the
shared pool.

•

CURSOR_SHARING and Bind Variable Substitution
This topic explains what the CURSOR_SHARING initialization parameter is, and how
setting it to different values affects how Oracle Database uses bind variables.

•

Adaptive Cursor Sharing
The adaptive cursor sharing feature enables a single statement that contains
bind variables to use multiple execution plans.

•

Real-World Performance Guidelines for Cursor Sharing
The Real-World Performance team has created guidelines for how to optimize
cursor sharing in Oracle database applications.

20.1 Overview of Cursor Sharing
Oracle Database can share cursors, which are pointers to private SQL areas in the
shared pool.
This section contains the following topics:
•

About Cursors
A private SQL area holds information about a parsed SQL statement and other
session-specific information for processing.

•

About Cursors and Parsing
If an application issues a statement, and if Oracle Database cannot reuse a
cursor, then it must build a new executable version of the application code. This
operation is known as a hard parse.

•

About Literals and Bind Variables
Bind variables are essential to cursor sharing in Oracle database applications.

•

About the Life Cycle of Shared Cursors
The database allocates a new shared SQL area when the optimizer parses a new
SQL statement that is not DDL. The amount of memory required depends on the
statement complexity.

20.1.1 About Cursors
A private SQL area holds information about a parsed SQL statement and other
session-specific information for processing.

20-1

Chapter 20

Overview of Cursor Sharing

When a server process executes SQL or PL/SQL code, the process uses the private
SQL area to store bind variable values, query execution state information, and query
execution work areas. The private SQL areas for each execution of a statement are
not shared and may contain different values and data.
A cursor is a name or handle to a specific private SQL area. The cursor contains
session-specific state information such as bind variable values and result sets.
As shown in the following graphic, you can think of a cursor as a pointer on the client
side and as a state on the server side. Because cursors are closely associated with
private SQL areas, the terms are sometimes used interchangeably.

Figure 20-1

Cursor
PGA

SQL Work Areas
Session Memory

Server
Process

Private SQL Area

Cursor
Data Area
Pointer

Client
Process

This section contains the following topics:
•

Private and Shared SQL Areas
A cursor in the private SQL area points to a shared SQL area in the library cache.

•

Parent and Child Cursors
Every parsed SQL statement has a parent cursor and one or more child cursors.

20.1.1.1 Private and Shared SQL Areas
A cursor in the private SQL area points to a shared SQL area in the library cache.
Unlike the private SQL area, which contains session state information, the shared SQL
area contains the parse tree and execution plan for the statement. For example, an
execution of SELECT * FROM employees has a plan and parse tree stored in one shared
SQL area. An execution of SELECT * FROM departments, which differs both syntactically
and semantically, has a plan and parse tree stored in a separate shared SQL area.
Multiple private SQL areas in the same or different sessions can reference a single
shared SQL area, a phenomenon known as cursor sharing. For example, an
execution of SELECT * FROM employees in one session and an execution of the SELECT *
FROM employees (accessing the same table) in a different session can use the same
parse tree and plan. A shared SQL area that is accessed by multiple statements is
known as a shared cursor.

20-2

Chapter 20

Overview of Cursor Sharing

Figure 20-2

Cursor Sharing

Instance
System Global Area (SGA)
Shared Pool
Library Cache
Shared SQL Area
SELECT * FROM
employees

Data
Dictionary
Cache

Server
Result
Cache

Private
SQL Area
(Shared
Server Only)

Other

Reserved
Pool

PGA
Server
Process

SQL Work Areas
Session Memory

Private SQL Area

PGA
Server
Process

SELECT * FROM employees

Client
Process

SQL Work Areas
Session Memory

Private SQL Area

SELECT * FROM employees

Client
Process

Oracle Database automatically determines whether the SQL statement or PL/SQL
block being issued is textually identical to another statement currently in the library
cache, using the following steps:
1.

The text of the statement is hashed.

2.

The database looks for a matching hash value for an existing SQL statement in
the shared pool. The following options are possible:
•

No matching hash value exists.
In this case, the SQL statement does not currently exist in the shared pool, so
the database performs a hard parse. This ends the shared pool check.

•

A matching hash value exists.
In this case, the database proceeds to the next step, which is a text match.

3.

The database compares the text of the matched statement to the text of the
hashed statement to determine whether they are identical. The following options
are possible:
•

The textual match fails.
In this case, the text match process stops, resulting in a hard parse.

20-3

Chapter 20

Overview of Cursor Sharing

•

The textual match succeeds.
In this case, the database proceeds to the next step: determining whether the
SQL can share an existing parent cursor.
For a textual match to occur, the text of the SQL statements or PL/SQL blocks
must be character-for-character identical, including spaces, case, and
comments. For example, the following statements cannot use the same
shared SQL area:
SELECT * FROM employees;
SELECT * FROM Employees;
SELECT * FROM employees;

Usually, SQL statements that differ only in literals cannot use the same shared
SQL area. For example, the following statements do not resolve to the same
SQL area:
SELECT count(1) FROM employees WHERE manager_id = 121;
SELECT count(1) FROM employees WHERE manager_id = 247;

The only exception to this rule is when the parameter CURSOR_SHARING has been
set to FORCE, in which case similar statements can share SQL areas.

See Also:
•

"Parent and Child Cursors"

•

"Do Not Use CURSOR_SHARING = FORCE as a Permanent Fix" to
learn about the costs involved in using CURSOR_SHARING

•

Oracle Database Reference to learn more about the CURSOR_SHARING
initialization parameter

20.1.1.2 Parent and Child Cursors
Every parsed SQL statement has a parent cursor and one or more child cursors.
The parent cursor stores the text of the SQL statement. If the text of two statements is
identical, then the statements share the same parent cursor. If the text is different,
however, then the database creates a separate parent cursor.
Example 20-1

Parent Cursors

In this example, the first two statements are syntactically different (the letter “c” is
lowercase in the first statement and uppercase in the second statement), but
semantically identical. Because of the syntactic difference, these statements have
different parent cursors. The third statement is syntactically identical to the first
statement (lowercase “c”), but semantically different because it refers to a customers
table in a different schema. Because of the syntactic identity, the third statement can
share a parent cursor with the first statement.
SQL> CONNECT oe@inst1
Enter password: *******
Connected.
SQL> SELECT COUNT(*) FROM customers;

20-4

Chapter 20

Overview of Cursor Sharing

COUNT(*)
---------319
SQL> SELECT COUNT(*) FROM Customers;
COUNT(*)
---------319
SQL> CONNECT sh@inst1
Enter password: *******
Connected.
SQL> SELECT COUNT(*) FROM customers;
COUNT(*)
---------155500

The following query of V$SQL indicates the two parents. The statement with the SQL ID
of 8h916vv2yw400, which is the lowercase “c” version of the statement, has one parent
cursor and two child cursors: child 0 and child 1. The statement with the SQL ID of
5rn2uxjtpz0wd, which is the uppercase “c” version of the statement, has a different
parent cursor and only one child cursor: child 0.
SQL> CONNECT SYSTEM@inst1
Enter password: *******
Connected.
SQL> COL SQL_TEXT FORMAT a30
SQL> COL CHILD# FORMAT 99999
SQL> COL EXEC FORMAT 9999
SQL> COL SCHEMA FORMAT a6
SQL> SELECT SQL_ID, PARSING_SCHEMA_NAME AS SCHEMA, SQL_TEXT,
2 CHILD_NUMBER AS CHILD#, EXECUTIONS AS EXEC FROM V$SQL
3 WHERE SQL_TEXT LIKE '%ustom%' AND SQL_TEXT NOT LIKE '%SQL_TEXT%' ORDER BY
SQL_ID;
SQL_ID
------------5rn2uxjtpz0wd
8h916vv2yw400
8h916vv2yw400

SCHEMA
-----OE
OE
SH

SQL_TEXT
CHILD# EXEC
------------------------------ ------ ----SELECT COUNT(*) FROM Customers
0
1
SELECT COUNT(*) FROM customers
0
1
SELECT COUNT(*) FROM customers
1
1

This section contains the following topics:
•

Parent Cursors and V$SQLAREA
The V$SQLAREA view contains one row for every parent cursor.

•

Child Cursors and V$SQL
Every parent cursor has one or more child cursors.

•

Cursor Mismatches and V$SQL_SHARED_CURSOR
If a parent cursor has multiple children, then the V$SQL_SHARED_CURSOR view provides
information about why the cursor was not shared. For several types of
incompatibility, the TRANSLATION_MISMATCH column indicates a mismatch with the
value Y or N.

20-5

Chapter 20

Overview of Cursor Sharing

20.1.1.2.1 Parent Cursors and V$SQLAREA
The V$SQLAREA view contains one row for every parent cursor.
In the following example, a query of V$SQLAREA shows two parent cursors, each
identified with a different SQL_ID. The VERSION_COUNT indicates the number of child
cursors.
COL SQL_TEXT FORMAT a30
SELECT SQL_TEXT, SQL_ID, VERSION_COUNT, HASH_VALUE
FROM V$SQLAREA
WHERE SQL_TEXT LIKE '%mployee%'
AND
SQL_TEXT NOT LIKE '%SQL_TEXT%';
SQL_TEXT
-----------------------------SELECT * FROM Employees
SELECT * FROM employees

SQL_ID
VERSION_COUNT HASH_VALUE
------------- ------------- ---------5bzhzpaa0wy9m
1 2483976499
4959aapufrm1k
2 1961610290

In the preceding output, the VERSION_COUNT of 2 for SELECT * FROM employees indicates
multiple child cursors, which were necessary because the statement was executed
against two different objects. In contrast, the statement SELECT * FROM Employees (note
the capital "E") was executed once, and so has one parent cursor, and one child
cursor (VERSION_COUNT of 1).

20.1.1.2.2 Child Cursors and V$SQL
Every parent cursor has one or more child cursors.
A child cursor contains the execution plan, bind variables, metadata about objects
referenced in the query, optimizer environment, and other information. In contrast to
the parent cursor, the child cursor does not store the text of the SQL statement.
If a statement is able to reuse a parent cursor, then the database checks whether the
statement can reuse an existing child cursor. The database performs several checks,
including the following:
•

The database compares objects referenced in the issued statement to the objects
referenced by the statement in the pool to ensure that they are all identical.
References to schema objects in the SQL statements or PL/SQL blocks must
resolve to the same object in the same schema. For example, if two users issue
the following SQL statement, and if each user has its own employees table, then the
following statement is not identical because the statement references different
employees tables for each user:
SELECT * FROM employees;

•

The database determines whether the optimizer mode is identical.
For example, SQL statements must be optimized using the same optimizer goal
(see "Choosing an Optimizer Goal").

Example 20-2

Multiple Child Cursors

V$SQL describes the statements that currently reside in the library cache. It contains

one row for every child cursor, as shown in the following example:
SELECT SQL_TEXT, SQL_ID, USERNAME AS USR, CHILD_NUMBER AS CHILD#,
HASH_VALUE, PLAN_HASH_VALUE AS PLAN_HASHV

20-6

Chapter 20

Overview of Cursor Sharing

FROM
WHERE
AND
AND

V$SQL s, DBA_USERS d
SQL_TEXT LIKE '%mployee%'
SQL_TEXT NOT LIKE '%SQL_TEXT%'
d.USER_ID = s.PARSING_USER_ID;

SQL_TEXT
----------------------SELECT * FROM Employees
SELECT * FROM employees
SELECT * FROM employees

SQL_ID
------------5bzhzpaa0wy9m
4959aapufrm1k
4959aapufrm1k

USR CHILD# HASH_VALUE PLAN_HASHV
--- ------ ---------- ---------HR
0 2483976499 1445457117
HR
0 1961610290 1445457117
SH
1 1961610290 1445457117

In the preceding results, the CHILD# of the bottom two statements is different (0 and 1),
even though the SQL_ID is the same. This means that the statements have the same
parent cursor, but different child cursors. In contrast, the statement with the SQL_ID of
5bzhzpaa0wy9m has one parent and one child (CHILD# of 0). All three SQL statements use
the same execution plan, as indicated by identical values in the PLAN_HASH_VALUE
column.

20.1.1.2.3 Cursor Mismatches and V$SQL_SHARED_CURSOR
If a parent cursor has multiple children, then the V$SQL_SHARED_CURSOR view provides
information about why the cursor was not shared. For several types of incompatibility,
the TRANSLATION_MISMATCH column indicates a mismatch with the value Y or N.
Example 20-3

Translation Mismatch

In this example, the TRANSLATION_MISMATCH column shows that the two statements
(SELECT * FROM employees) referenced different objects, resulting in a
TRANSLATION_MISMATCH value of Y for the last statement. Because sharing was not
possible, each statement had a separate child cursor, as indicated by CHILD_NUMBER of 0
and 1.
SELECT S.SQL_TEXT, S.CHILD_NUMBER, s.CHILD_ADDRESS,
C.TRANSLATION_MISMATCH
FROM V$SQL S, V$SQL_SHARED_CURSOR C
WHERE SQL_TEXT LIKE '%employee%'
AND
SQL_TEXT NOT LIKE '%SQL_TEXT%'
AND S.CHILD_ADDRESS = C.CHILD_ADDRESS;
SQL_TEXT
CHILD_NUMBER
CHILD_ADDRESS T
------------------------------ ------------ ---------------- SELECT * FROM employees
0 0000000081EE8690 N
SELECT * FROM employees
1 0000000081F22508 Y

20.1.2 About Cursors and Parsing
If an application issues a statement, and if Oracle Database cannot reuse a cursor,
then it must build a new executable version of the application code. This operation is
known as a hard parse.
A soft parse is any parse that is not a hard parse, and occurs when the database can
reuse existing code. Some soft parses are less resource-intensive than others. For
example, if a parent cursor for the statement already exists, then Oracle Database can
perform various optimizations, and then store the child cursor in the shared SQL area.
If a parent cursor does not exist, however, then Oracle Database must also store the
parent cursor in the shared SQL area, which creates additional memory overhead.

20-7

Chapter 20

Overview of Cursor Sharing

Effectively, a hard parse recompiles a statement before running it. Hard parsing a SQL
statement before every execution is analogous to recompiling a C program before
every execution. A hard parse performs operations such as the following:
•

Checking the syntax of the SQL statement

•

Checking the semantics of the SQL statement

•

Checking the access rights of the user issuing the statement

•

Creating an execution plan

•

Accessing the library cache and data dictionary cache numerous times to check
the data dictionary

An especially resource-intensive aspect of hard parsing is accessing the library cache
and data dictionary cache numerous times to check the data dictionary. When the
database accesses these areas, it uses a serialization device called a latch on
required objects so that their definition does not change during the check. Latch
contention increases statement execution time and decreases concurrency.
For all of the preceding reasons, the CPU and memory overhead of hard parses can
create serious performance problems. The problems are especially evident in web
applications that accept user input from a form, and then generate SQL statements
dynamically. The Real-World Performance group strongly recommends reducing hard
parsing as much as possible.

Video:
Video

Example 20-4

Finding Parse Information Using V$SQL

You can use various techniques to monitor hard and soft parsing. This example
queries the session statistics to determine whether repeated executions of a DBA_JOBS
query increase the hard parse count. The first execution of the statement increases
the hard parse count to 49, but the second execution does not change the hard parse
count, which means that Oracle Database reused application code.
SQL> ALTER SYSTEM FLUSH SHARED_POOL;
System altered.
SQL> COL NAME FORMAT a18
SQL>
2
3
4

SELECT
FROM
WHERE
AND

s.NAME, m.VALUE
V$STATNAME s, V$MYSTAT m
s.STATISTIC# = m.STATISTIC#
s.NAME LIKE '%(hard%';

NAME
VALUE
------------------ ---------parse count (hard)
48
SQL> SELECT COUNT(*) FROM DBA_JOBS;
COUNT(*)
----------

20-8

Chapter 20

Overview of Cursor Sharing

0
SQL>
2
3
4

SELECT
FROM
WHERE
AND

s.NAME, m.VALUE
V$STATNAME s, V$MYSTAT m
s.STATISTIC# = m.STATISTIC#
s.NAME LIKE '%(hard%';

NAME
VALUE
------------------ ---------parse count (hard)
49
SQL> SELECT COUNT(*) FROM DBA_JOBS;
COUNT(*)
---------0
SQL>
2
3
4

SELECT
FROM
WHERE
AND

s.NAME, m.VALUE
V$STATNAME s, V$MYSTAT m
s.STATISTIC# = m.STATISTIC#
s.NAME LIKE '%(hard%';

NAME
VALUE
------------------ ---------parse count (hard)
49

Example 20-5

Finding Parse Information Using Trace Files

This example uses SQL Trace and the TKPROF utility to find parse information. You
log in to the database with administrator privileges, and then query the directory
location of the trace files (sample output included):
SET LINESIZE 120
COLUMN value FORMAT A80
SELECT value
FROM v$diag_info
WHERE name = 'Default Trace File';
VALUE
-------------------------------------------------------------------------------/disk1/oracle/log/diag/rdbms/orcl/orcl/trace/orcl_ora_23054.trc

You enable tracing, use the TRACEFILE_IDENTIFIER initialization parameter to give the
trace file a meaningful name, and then query hr.employees:
EXEC DBMS_MONITOR.SESSION_TRACE_ENABLE(waits=>TRUE, binds=>TRUE);
ALTER SESSION SET TRACEFILE_IDENTIFIER = "emp_stmt";
SELECT * FROM hr.employees;
EXIT;

Search the default trace file directory for the trace file that you generated:
% ls *emp_stmt.trc
orcl_ora_17950_emp_stmt.trc

Use TKPROF to format the trace file, and then open the formatted file:
% tkprof orcl_ora_17950_emp_stmt.trc emp.out; vi emp.out

The formatted trace file contains the parse information for the query of hr.employees.

20-9

Chapter 20

Overview of Cursor Sharing

SQL ID: brmjpfs7dcnub Plan Hash: 1445457117
SELECT *
FROM
hr.employees

call
count
cpu
elapsed
disk
query
current
rows
------- ------ -------- ---------- ---------- ---------- ---------- ---------Parse
1
0.07
0.08
0
0
0
0
Execute
1
0.00
0.00
0
0
0
0
Fetch
9
0.00
0.00
3
12
0
107
------- ------ -------- ---------- ---------- ---------- ---------- ---------total
11
0.07
0.08
3
12
0
107
Misses in library cache during parse: 1
Optimizer mode: ALL_ROWS
Parsing user id: SYSTEM
Number of plan statistics captured: 1
Rows (1st) Rows (avg) Rows (max) Row Source Operation
---------- ---------- ---------- --------------------------------------------------107
107
107 TABLE ACCESS FULL EMPLOYEES (cr=12 pr=3 pw=0
time=497
us starts=1 cost=2 size=7383 card=107)

A library cache miss indicates a hard parse. Performing the same steps for a second
execution of the same statement produces the following trace output, which shows no
library cache misses:
SQL ID: brmjpfs7dcnub Plan Hash: 1445457117
SELECT *
FROM
hr.employees

call
count
cpu
elapsed
disk
query
current
rows
------- ------ -------- ---------- ---------- ---------- ---------- ---------Parse
1
0.00
0.00
0
0
0
0
Execute
1
0.00
0.00
0
0
0
0
Fetch
9
0.00
0.00
3
12
0
107
------- ------ -------- ---------- ---------- ---------- ---------- ---------total
11
0.00
0.00
3
12
0
107
Misses in library cache during parse: 0
Optimizer mode: ALL_ROWS
Parsing user id: SYSTEM
Number of plan statistics captured: 1
Rows (1st) Rows (avg) Rows (max) Row Source Operation
---------- ---------- ---------- --------------------------------------------------107
107
107 TABLE ACCESS FULL EMPLOYEES (cr=12 pr=3 pw=0
time=961
us starts=1 cost=2 size=7383 card=107)

20-10

Chapter 20

Overview of Cursor Sharing

See Also:
"Shared Pool Check"

20.1.3 About Literals and Bind Variables
Bind variables are essential to cursor sharing in Oracle database applications.
This section contains the following topics:
•

Literals and Cursors
When constructing SQL statements, some Oracle applications use literals instead
of bind variables.

•

Bind Variables and Cursors
You can develop Oracle applications to use bind variables instead of literals.

•

Bind Variable Peeking
In bind variable peeking (also known as bind peeking), the optimizer looks at the
value in a bind variable when the database performs a hard parse of a statement.

20.1.3.1 Literals and Cursors
When constructing SQL statements, some Oracle applications use literals instead of
bind variables.
For example, the statement SELECT SUM(salary) FROM hr.employees WHERE employee_id
< 101 uses the literal value 101 for the employee ID. By default, when similar
statements do not use bind variables, Oracle Database cannot take advantage of
cursor sharing. Thus, Oracle Database sees a statement that is identical except for the
value 102, or any other random value, as a completely new statement, requiring a hard
parse.
The Real-World Performance group has determined that applications that use literals
are a frequent cause of performance, scalability, and security problems. In the real
world, it is not uncommon for applications to be written quickly, without considering
cursor sharing. A classic example is a “screen scraping” application that copies the
contents out of a web form, and then concatenates strings to construct the SQL
statement dynamically.
Major problems that result from using literal values include the following:
•

Applications that concatenate literals input by an end user are prone to SQL
injection attacks. Only rewriting the applications to use bind variables eliminates
this threat.

•

If every statement is hard parsed, then cursors are not shared, and so the
database must consume more memory to create the cursors.

•

Oracle Database must latch the shared pool and library cache when hard parsing.
As the number of hard parses increases, so does the number of processes waiting
to latch the shared pool. This situation decreases concurrency and increases
contention.

20-11

Chapter 20

Overview of Cursor Sharing

Video:
Video

Example 20-6

Literals and Cursor Sharing

Consider an application that executes the following statements, which differ only in
literals:
SELECT SUM(salary) FROM hr.employees WHERE employee_id < 101;
SELECT SUM(salary) FROM hr.employees WHERE employee_id < 120;
SELECT SUM(salary) FROM hr.employees WHERE employee_id < 165;

The following query of V$SQLAREA shows that the three statements require three
different parent cursors. As shown by VERSION_COUNT, each parent cursor requires its
own child cursor.
COL SQL_TEXT FORMAT a30
SELECT SQL_TEXT, SQL_ID, VERSION_COUNT, HASH_VALUE
FROM V$SQLAREA
WHERE SQL_TEXT LIKE '%mployee%'
AND
SQL_TEXT NOT LIKE '%SQL_TEXT%';
SQL_TEXT
-----------------------------SELECT SUM(salary) FROM hr.emp
loyees WHERE employee_id < 165
SELECT SUM(salary) FROM hr.emp
loyees WHERE employee_id < 101
SELECT SUM(salary) FROM hr.emp
loyees WHERE employee_id < 120

SQL_ID VERSION_COUNT HASH_VALUE
------------- ------------- ---------b1tvfcc5qnczb
1 191509483
cn5250y0nqpym

1 2169198547

au8nag2vnfw67

1 3074912455

See Also:
"Do Not Use CURSOR_SHARING = FORCE as a Permanent Fix" to learn
about SQL injection

20.1.3.2 Bind Variables and Cursors
You can develop Oracle applications to use bind variables instead of literals.
A bind variable is a placeholder in a query. For example, the statement SELECT
SUM(salary) FROM hr.employees WHERE employee_id < :emp_id uses the bind
variable:emp_id for the employee ID.
The Real-World Performance group has found that applications that use bind variables
perform better, scale better, and are more secure. Major benefits that result from using
bind variables include the following:
•

Applications that use bind variables are not vulnerable to the same SQL injection
attacks as applications that use literals.

20-12

Chapter 20

Overview of Cursor Sharing

•

When identical statements use bind variables, Oracle Database can take
advantage of cursor sharing, and share the plan and other information when
different values are bound to the same statement.

•

Oracle Database avoids the overhead of latching the shared pool and library
cache required for hard parsing.

Video:
Video

Example 20-7

Bind Variables and Shared Cursors

The following example uses the VARIABLE command in SQL*Plus to create the emp_id
bind variable, and then executes a query using three different bind values (101, 120,
and 165):
VARIABLE emp_id NUMBER
EXEC :emp_id := 101;
SELECT SUM(salary) FROM hr.employees WHERE employee_id < :emp_id;
EXEC :emp_id := 120;
SELECT SUM(salary) FROM hr.employees WHERE employee_id < :emp_id;
EXEC :emp_id := 165;
SELECT SUM(salary) FROM hr.employees WHERE employee_id < :emp_id;

The following query of V$SQLAREA shows one unique SQL statement:
COL SQL_TEXT FORMAT a34
SELECT SQL_TEXT, SQL_ID, VERSION_COUNT, HASH_VALUE
FROM V$SQLAREA
WHERE SQL_TEXT LIKE '%mployee%'
AND
SQL_TEXT NOT LIKE '%SQL_TEXT%';
SQL_TEXT
SQL_ID
VERSION_COUNT HASH_VALUE
---------------------------------- ------------- ------------- ---------SELECT SUM(salary) FROM hr.employe 4318cbskba8yh
1 615850960
es WHERE employee_id < :emp_id

The VERSION_COUNT value of 1 indicates that the database reused the same child cursor
rather than creating three separate child cursors. Using a bind variable made this
reuse possible.

20.1.3.3 Bind Variable Peeking
In bind variable peeking (also known as bind peeking), the optimizer looks at the
value in a bind variable when the database performs a hard parse of a statement.
The optimizer does not look at the bind variable values before every parse. Rather, the
optimizer peeks only when the optimizer is first invoked, which is during the hard
parse.
When a query uses literals, the optimizer can use the literal values to find the best
plan. However, when a query uses bind variables, the optimizer must select the best
plan without the presence of literals in the SQL text. This task can be extremely
difficult. By peeking at bind values during the initial hard parse, the optimizer can

20-13

Chapter 20

Overview of Cursor Sharing

determine the cardinality of a WHERE clause condition as if literals had been used,
thereby improving the plan.
Because the optimizer only peeks at the bind value during the hard parse, the plan
may not be optimal for all possible bind values. The following examples illustrate this
principle.
Example 20-8

Literals Result in Different Execution Plans

Assume that you execute the following statements, which execute three different
statements using different literals (101, 120, and 165), and then display the execution
plans for each:
SET LINESIZE 167
SET PAGESIZE 0
SELECT SUM(salary) FROM hr.employees WHERE employee_id < 101;
SELECT * FROM TABLE(DBMS_XPLAN.DISPLAY_CURSOR());
SELECT SUM(salary) FROM hr.employees WHERE employee_id < 120;
SELECT * FROM TABLE(DBMS_XPLAN.DISPLAY_CURSOR());
SELECT SUM(salary) FROM hr.employees WHERE employee_id < 165;
SELECT * FROM TABLE(DBMS_XPLAN.DISPLAY_CURSOR());

The database hard parsed all three statements, which were not identical. The
DISPLAY_CURSOR output, which has been edited for clarity, shows that the optimizer

chose the same index range scan plan for the first two statements, but a full table scan
plan for the statement using literal 165:
SQL_ID
cn5250y0nqpym, child number 0
------------------------------------SELECT SUM(salary) FROM hr.employees WHERE employee_id < 101
Plan hash value: 2410354593
------------------------------------------------------------------------------------|Id| Operation
| Name
|Rows|Bytes|Cost(%CPU)|Time|
------------------------------------------------------------------------------------| 0| SELECT STATEMENT
|
| | |2 (100)|
|
| 1| SORT AGGREGATE
|
|1 | 8 |
|
|
| 2| TABLE ACCESS BY INDEX ROWID BATCHED| EMPLOYEES
|1 | 8 |2 (0) | 00:00:01 |
|*3|
INDEX RANGE SCAN
| EMP_EMP_ID_PK |1 | |1 (0) | 00:00:01 |
------------------------------------------------------------------------------------Predicate Information (identified by operation id):
--------------------------------------------------3 - access("EMPLOYEE_ID"<101)
SQL_ID
au8nag2vnfw67, child number 0
------------------------------------SELECT SUM(salary) FROM hr.employees WHERE employee_id < 120
Plan hash value: 2410354593
------------------------------------------------------------------------------------|Id| Operation
| Name
|Rows|Bytes|Cost(%CPU)|Time|
------------------------------------------------------------------------------------| 0| SELECT STATEMENT
|
| | |2 (100)|
|
| 1| SORT AGGREGATE
|
|1 | 8 |
|
|
| 2| TABLE ACCESS BY INDEX ROWID BATCHED| EMPLOYEES
|20|160|2 (0) | 00:00:01 |
|*3|
INDEX RANGE SCAN
| EMP_EMP_ID_PK |20| |1 (0) | 00:00:01 |
-------------------------------------------------------------------------------------

20-14

Chapter 20

Overview of Cursor Sharing

Predicate Information (identified by operation id):
--------------------------------------------------3 - access("EMPLOYEE_ID"<120)
SQL_ID
b1tvfcc5qnczb, child number 0
------------------------------------SELECT SUM(salary) FROM hr.employees WHERE employee_id < 165
Plan hash value: 1756381138
------------------------------------------------------------------------| Id | Operation
| Name
|Rows| Bytes |Cost(%CPU)| Time |
------------------------------------------------------------------------| 0 | SELECT STATEMENT |
|
|
| 2 (100)|
|
| 1 | SORT AGGREGATE
|
| 1 |
8 |
|
|
|* 2 | TABLE ACCESS FULL| EMPLOYEES | 66 | 528 | 2 (0)| 00:00:01 |
------------------------------------------------------------------------Predicate Information (identified by operation id):
--------------------------------------------------2 - filter("EMPLOYEE_ID"<165)

The preceding output shows that the optimizer considers a full table scan more
efficient than an index scan for the query that returns more rows.
Example 20-9

Bind Variables Result in Cursor Reuse

This example rewrites the queries executed in Example 20-8 to use bind variables
instead of literals. You bind the same values (101, 120, and 165) to the bind
variable :emp_id, and then display the execution plans for each:
VAR emp_id NUMBER
EXEC :emp_id := 101;
SELECT SUM(salary) FROM hr.employees WHERE employee_id < :emp_id;
SELECT * FROM TABLE(DBMS_XPLAN.DISPLAY_CURSOR());
EXEC :emp_id := 120;
SELECT SUM(salary) FROM hr.employees WHERE employee_id < :emp_id;
SELECT * FROM TABLE(DBMS_XPLAN.DISPLAY_CURSOR());
EXEC :emp_id := 165;
SELECT SUM(salary) FROM hr.employees WHERE employee_id < :emp_id;
SELECT * FROM TABLE(DBMS_XPLAN.DISPLAY_CURSOR());

The DISPLAY_CURSOR output shows that the optimizer chose exactly the same plan for all
three statements:
SELECT SUM(salary) FROM hr.employees WHERE employee_id < :emp_id
Plan hash value: 2410354593
------------------------------------------------------------------------------------| Id | Operation
| Name
|Rows|Bytes|Cost (%CPU)|Time|
------------------------------------------------------------------------------------| 0 | SELECT STATEMENT
|
| | |2 (100)|
|
| 1 | SORT AGGREGATE
|
|1|8 |
|
|
| 2 | TABLE ACCESS BY INDEX ROWID BATCHED| EMPLOYEES
|1|8 | 2 (0)| 00:00:01 |
|* 3 |
INDEX RANGE SCAN
| EMP_EMP_ID_PK |1| | 1 (0)| 00:00:01 |
-------------------------------------------------------------------------------------

20-15

Chapter 20

Overview of Cursor Sharing

Predicate Information (identified by operation id):
--------------------------------------------------3 - access("EMPLOYEE_ID"<:EMP_ID)

In contrast, when the preceding statements were executed with literals, the optimizer
chose a lower-cost full table scan when the employee ID value was 165. This is the
problem solved by adaptive cursor sharing.

See Also:
"Adaptive Cursor Sharing"

20.1.4 About the Life Cycle of Shared Cursors
The database allocates a new shared SQL area when the optimizer parses a new SQL
statement that is not DDL. The amount of memory required depends on the statement
complexity.
The database can remove a shared SQL area from the shared pool even if this area
corresponds to an open cursor that has been unused for a long time.