Programming Hive Guide

User Manual:

Open the PDF directly: View PDF .
Page Count: 350 [warning: Documents this large are best viewed by clicking the View PDF Link!]

Table of Contents
Preface
- Conventions Used in This Book
- Using Code Examples
- Safari® Books Online
- How to Contact Us
- What Brought Us to Hive?
- Acknowledgments
Chapter 1. Introduction
- An Overview of Hadoop and MapReduce
  - MapReduce
- Hive in the Hadoop Ecosystem
- Java Versus Hive: The Word Count Algorithm
- What’s Next
Chapter 2. Getting Started
- Installing a Preconfigured Virtual Machine
- Detailed Installation
- What Is Inside Hive?
- Starting Hive
- Configuring Your Hadoop Environment
- The Hive Command
  - Command Options
- The Command-Line Interface
Chapter 3. Data Types and File Formats
- Primitive Data Types
- Collection Data Types
- Text File Encoding of Data Values
- Schema on Read
Chapter 4. HiveQL: Data Definition
- Databases in Hive
- Alter Database
- Creating Tables
  - Managed Tables
  - External Tables
- Partitioned, Managed Tables
  - External Partitioned Tables
  - Customizing Table Storage Formats
- Dropping Tables
- Alter Table
Chapter 5. HiveQL: Data Manipulation
- Loading Data into Managed Tables
- Inserting Data into Tables from Queries
  - Dynamic Partition Inserts
- Creating Tables and Loading Them in One Query
- Exporting Data
Chapter 6. HiveQL: Queries
- SELECT … FROM Clauses
- WHERE Clauses
- GROUP BY Clauses
  - HAVING Clauses
- JOIN Statements
- ORDER BY and SORT BY
- DISTRIBUTE BY with SORT BY
- CLUSTER BY
- Casting
  - Casting BINARY Values
- Queries that Sample Data
  - Block Sampling
  - Input Pruning for Bucket Tables
- UNION ALL
Chapter 7. HiveQL: Views
- Views to Reduce Query Complexity
- Views that Restrict Data Based on Conditions
- Views and Map Type for Dynamic Tables
- View Odds and Ends
Chapter 8. HiveQL: Indexes
- Creating an Index
  - Bitmap Indexes
- Rebuilding the Index
- Showing an Index
- Dropping an Index
- Implementing a Custom Index Handler
Chapter 9. Schema Design
- Table-by-Day
- Over Partitioning
- Unique Keys and Normalization
- Making Multiple Passes over the Same Data
- The Case for Partitioning Every Table
- Bucketing Table Data Storage
- Adding Columns to a Table
- Using Columnar Tables
  - Repeated Data
  - Many Columns
- (Almost) Always Use Compression!
Chapter 10. Tuning
- Using EXPLAIN
- EXPLAIN EXTENDED
- Limit Tuning
- Optimized Joins
- Local Mode
- Parallel Execution
- Strict Mode
- Tuning the Number of Mappers and Reducers
- JVM Reuse
- Indexes
- Dynamic Partition Tuning
- Speculative Execution
- Single MapReduce MultiGROUP BY
- Virtual Columns
Chapter 11. Other File Formats and Compression
- Determining Installed Codecs
- Choosing a Compression Codec
- Enabling Intermediate Compression
- Final Output Compression
- Sequence Files
- Compression in Action
- Archive Partition
- Compression: Wrapping Up
Chapter 12. Developing
- Changing Log4J Properties
- Connecting a Java Debugger to Hive
- Building Hive from Source
  - Running Hive Test Cases
  - Execution Hooks
- Setting Up Hive and Eclipse
- Hive in a Maven Project
- Unit Testing in Hive with hive_test
- The New Plugin Developer Kit
Chapter 13. Functions
- Discovering and Describing Functions
- Calling Functions
- Standard Functions
- Aggregate Functions
- Table Generating Functions
- A UDF for Finding a Zodiac Sign from a Day
- UDF Versus GenericUDF
- Permanent Functions
- User-Defined Aggregate Functions
  - Creating a COLLECT UDAF to Emulate GROUP_CONCAT
- User-Defined Table Generating Functions
- Accessing the Distributed Cache from a UDF
- Annotations for Use with Functions
- Macros
Chapter 14. Streaming
- Identity Transformation
- Changing Types
- Projecting Transformation
- Manipulative Transformations
- Using the Distributed Cache
- Producing Multiple Rows from a Single Row
- Calculating Aggregates with Streaming
- CLUSTER BY, DISTRIBUTE BY, SORT BY
- GenericMR Tools for Streaming to Java
- Calculating Cogroups
Chapter 15. Customizing Hive File and Record Formats
- File Versus Record Formats
- Demystifying CREATE TABLE Statements
- File Formats
- Record Formats: SerDes
- CSV and TSV SerDes
- ObjectInspector
- Think Big Hive Reflection ObjectInspector
- XML UDF
- XPath-Related Functions
- JSON SerDe
- Avro Hive SerDe
- Binary Output
Chapter 16. Hive Thrift Service
- Starting the Thrift Server
- Setting Up Groovy to Connect to HiveService
- Connecting to HiveServer
- Getting Cluster Status
- Result Set Schema
- Fetching Results
- Retrieving Query Plan
- Metastore Methods
  - Example Table Checker
    - Finding tables not marked as external
- Administrating HiveServer
  - Productionizing HiveService
  - Cleanup
- Hive ThriftMetastore
  - ThriftMetastore Configuration
  - Client Configuration
Chapter 17. Storage Handlers and NoSQL
- Storage Handler Background
- HiveStorageHandler
- HBase
- Cassandra
- DynamoDB
Chapter 18. Security
- Integration with Hadoop Security
- Authentication with Hive
- Authorization in Hive
Chapter 19. Locking
- Locking Support in Hive with Zookeeper
- Explicit, Exclusive Locks
Chapter 20. Hive Integration with Oozie
- Oozie Actions
  - Hive Thrift Service Action
- A Two-Query Workflow
- Oozie Web Console
- Variables in Workflows
- Capturing Output
- Capturing Output to Variables
Chapter 21. Hive and Amazon Web Services (AWS)
- Why Elastic MapReduce?
- Instances
- Before You Start
- Managing Your EMR Hive Cluster
- Thrift Server on EMR Hive
- Instance Groups on EMR
- Configuring Your EMR Cluster
- Persistence and the Metastore on EMR
- HDFS and S3 on EMR Cluster
- Putting Resources, Configs, and Bootstrap Scripts on S3
- Logs on S3
- Spot Instances
- Security Groups
- EMR Versus EC2 and Apache Hive
- Wrapping Up
Chapter 22. HCatalog
- Introduction
- MapReduce
  - Reading Data
  - Writing Data
- Command Line
- Security Model
- Architecture
Chapter 23. Case Studies
- m6d.com (Media6Degrees)
- Outbrain
- NASA’s Jet Propulsion Laboratory
- Photobucket
- SimpleReach
- Experiences and Needs from the Customer Trenches
Glossary
Appendix. References
Index

Programming Hive

Edward Capriolo, Dean Wampler, and Jason Rutherglen

Beijing

•

Cambridge

•

Farnham

•

Köln

•

Sebastopol

•

Tokyo

Downloa d f r o m W o w ! e B o o k < w w w.woweb o o k . c o m >

Programming Hive

by Edward Capriolo, Dean Wampler, and Jason Rutherglen

served.

Printed in the United States of America.

Published by O’Reilly Media, Inc., 1005 Gravenstein Highway North, Sebastopol, CA 95472.

O’Reilly books may be purchased for educational, business, or sales promotional use. Online editions

are also available for most titles (http://my.safaribooksonline.com). For more information, contact our

corporate/institutional sales department: 800-998-9938 or corporate@oreilly.com.

Editors: Mike Loukides and Courtney Nash

Production Editors: Iris Febres and Rachel Steely

Proofreaders: Stacie Arellano and Kiel Van Horn

Indexer: Bob Pfahler

Cover Designer: Karen Montgomery

Interior Designer: David Futato

Illustrator: Rebecca Demarest

October 2012: First Edition.

Revision History for the First Edition:

2012-09-17 First release

See http://oreilly.com/catalog/errata.csp?isbn=9781449319335 for release details.

Nutshell Handbook, the Nutshell Handbook logo, and the O’Reilly logo are registered trademarks of

O’Reilly Media, Inc. Programming Hive, the image of a hornet’s hive, and related trade dress are trade-

marks of O’Reilly Media, Inc.

Many of the designations used by manufacturers and sellers to distinguish their products are claimed as

trademarks. Where those designations appear in this book, and O’Reilly Media, Inc., was aware of a

trademark claim, the designations have been printed in caps or initial caps.

While every precaution has been taken in the preparation of this book, the publisher and authors assume

no responsibility for errors or omissions, or for damages resulting from the use of the information con-

tained herein.

ISBN: 978-1-449-31933-5

[LSI]

1347905436

Table of Contents

Preface .................................................................... xiii

1. Introduction ........................................................... 1

An Overview of Hadoop and MapReduce 3

Hive in the Hadoop Ecosystem 6

Pig 8

HBase 8

Cascading, Crunch, and Others 9

Java Versus Hive: The Word Count Algorithm 10

What’s Next 13

2. Getting Started ........................................................ 15

Installing a Preconfigured Virtual Machine 15

Detailed Installation 16

Installing Java 16

Installing Hadoop 18

Local Mode, Pseudodistributed Mode, and Distributed Mode 19

Testing Hadoop 20

Installing Hive 21

What Is Inside Hive? 22

Starting Hive 23

Configuring Your Hadoop Environment 24

Local Mode Configuration 24

Distributed and Pseudodistributed Mode Configuration 26

Metastore Using JDBC 28

The Hive Command 29

Command Options 29

The Command-Line Interface 30

CLI Options 31

Variables and Properties 31

Hive “One Shot” Commands 34

iii

Executing Hive Queries from Files 35

The .hiverc File 36

More on Using the Hive CLI

The CLI supports a number of other useful features.

36 | Chapter 2: Getting Started

Autocomplete

If you start typing and hit the Tab key, the CLI will autocomplete possible keywords

and function names. For example, if you type SELE and then the Tab key, the CLI will

complete the word SELECT.

If you type the Tab key at the prompt, you’ll get this reply:

hive>

Display all 407 possibilities? (y or n)

If you enter y, you’ll get a long list of all the keywords and built-in functions.

A common source of error and confusion when pasting statements into

the CLI occurs where some lines begin with a tab. You’ll get the prompt

about displaying all possibilities, and subsequent characters in the

stream will get misinterpreted as answers to the prompt, causing the

command to fail.

Command History

You can use the up and down arrow keys to scroll through previous commands. Ac-

tually, each previous line of input is shown separately; the CLI does not combine mul-

tiline commands and queries into a single history entry. Hive saves the last 100,00 lines

into a file $HOME/.hivehistory.

If you want to repeat a previous command, scroll to it and hit Enter. If you want to edit

the line before entering it, use the left and right arrow keys to navigate to the point

where changes are required and edit the line. You can hit Return to submit it without

returning to the end of the line.

Most navigation keystrokes using the Control key work as they do for

the bash shell (e.g., Control-A goes to the beginning of the line and

Control-E goes to the end of the line). However, similar “meta,” Option,

or Escape keys don’t work (e.g., Option-F to move forward a word at a

time). Similarly, the Delete key will delete the character to the left of the

cursor, but the Forward Delete key doesn’t delete the character under

the cursor.

Shell Execution

You don’t need to leave the hive CLI to run simple bash shell commands. Simply

type ! followed by the command and terminate the line with a semicolon (;):

hive> ! /bin/echo "what up dog";

"what up dog"

hive> ! pwd;

/home/me/hiveplay

The Command-Line Interface | 37

Don’t invoke interactive commands that require user input. Shell “pipes” don’t work

and neither do file “globs.” For example, ! ls *.hql; will look for a file named *.hql;,

rather than all files that end with the .hql extension.

Hadoop dfs Commands from Inside Hive

You can run the hadoop dfs ... commands from within the hive CLI; just drop the

hadoop word from the command and add the semicolon at the end:

hive> dfs -ls / ;

Found 3 items

drwxr-xr-x - root supergroup 0 2011-08-17 16:27 /etl

drwxr-xr-x - edward supergroup 0 2012-01-18 15:51 /flag

drwxrwxr-x - hadoop supergroup 0 2010-02-03 17:50 /users

This method of accessing hadoop commands is actually more efficient than using the

hadoop dfs ... equivalent at the bash shell, because the latter starts up a new JVM

instance each time, whereas Hive just runs the same code in its current process.

You can see a full listing of help on the options supported by dfs using this command:

hive> dfs -help;

See also http://hadoop.apache.org/common/docs/r0.20.205.0/file_system_shell.html or

similar documentation for your Hadoop distribution.

Comments in Hive Scripts

As of Hive v0.8.0, you can embed lines of comments that start with the string --, for

example:

-- This is the best Hive script evar!!

SELECT * FROM massive_table;

...

The CLI does not parse these comment lines. If you paste them into the

CLI, you’ll get errors. They only work when used in scripts executed

with hive -f script_name.

Query Column Headers

As a final example that pulls together a few things we’ve learned, let’s tell the CLI to

print column headers, which is disabled by default. We can enable this feature by setting

the hiveconf property hive.cli.print.header to true:

38 | Chapter 2: Getting Started

hive> set hive.cli.print.header=true;

hive> SELECT * FROM system_logs LIMIT 3;

tstamp severity server message

1335667117.337715 ERROR server1 Hard drive hd1 is 90% full!

1335667117.338012 WARN server1 Slow response from server2.

1335667117.339234 WARN server2 Uh, Dude, I'm kinda busy right now...

If you always prefer seeing the headers, put the first line in your $HOME/.hiverc file.

The Command-Line Interface | 39

CHAPTER 3

Data Types and File Formats

Hive supports many of the primitive data types you find in relational databases, as well

as three collection data types that are rarely found in relational databases, for reasons

we’ll discuss shortly.

A related concern is how these types are represented in text files, as well as alternatives

to text storage that address various performance and other concerns. A unique feature

of Hive, compared to most databases, is that it provides great flexibility in how data is

encoded in files. Most databases take total control of the data, both how it is persisted

to disk and its life cycle. By letting you control all these aspects, Hive makes it easier

to manage and process data with a variety of tools.

Primitive Data Types

Hive supports several sizes of integer and floating-point types, a Boolean type, and

character strings of arbitrary length. Hive v0.8.0 added types for timestamps and binary

fields.

Table 3-1 lists the primitive types supported by Hive.

Table 3-1. Primitive data types

Type Size Literal syntax examples

TINYINT 1 byte signed integer. 20

SMALLINT 2 byte signed integer. 20

INT 4 byte signed integer. 20

BIGINT 8 byte signed integer. 20

BOOLEAN Boolean true or false. TRUE

FLOAT Single precision floating point. 3.14159

DOUBLE Double precision floating point. 3.14159

Type Size Literal syntax examples

STRING Sequence of characters. The character

set can be specified. Single or double

quotes can be used.

'Now is the time', "for all

good men"

TIMESTAMP (v0.8.0+) Integer, float, or string. 1327882394 (Unix epoch seconds),

1327882394.123456789 (Unix ep-

och seconds plus nanoseconds), and

'2012-02-03

12:34:56.123456789' (JDBC-

compliant java.sql.Timestamp

format)

BINARY (v0.8.0+) Array of bytes. See discussion below

As for other SQL dialects, the case of these names is ignored.

It’s useful to remember that each of these types is implemented in Java, so the particular

behavior details will be exactly what you would expect from the corresponding Java

types. For example, STRING is implemented by the Java String, FLOAT is implemented

by Java float, etc.

Note that Hive does not support “character arrays” (strings) with maximum-allowed

lengths, as is common in other SQL dialects. Relational databases offer this feature as

a performance optimization; fixed-length records are easier to index, scan, etc. In the

“looser” world in which Hive lives, where it may not own the data files and has to be

flexible on file format, Hive relies on the presence of delimiters to separate fields. Also,

Hadoop and Hive emphasize optimizing disk reading and writing performance, where

fixing the lengths of column values is relatively unimportant.

Values of the new TIMESTAMP type can be integers, which are interpreted as seconds since

the Unix epoch time (Midnight, January 1, 1970), floats, which are interpreted as sec-

onds since the epoch time with nanosecond resolution (up to 9 decimal places), and

strings, which are interpreted according to the JDBC date string format convention,

YYYY-MM-DD hh:mm:ss.fffffffff.

TIMESTAMPS are interpreted as UTC times. Built-in functions for conversion to and from

timezones are provided by Hive, to_utc_timestamp and from_utc_timestamp, respec-

tively (see Chapter 13 for more details).

The BINARY type is similar to the VARBINARY type found in many relational databases.

It’s not like a BLOB type, since BINARY columns are stored within the record, not sepa-

rately like BLOBs. BINARY can be used as a way of including arbitrary bytes in a record

and preventing Hive from attempting to parse them as numbers, strings, etc.

Note that you don’t need BINARY if your goal is to ignore the tail end of each record. If

a table schema specifies three columns and the data files contain five values for each

record, the last two will be ignored by Hive.

42 | Chapter 3: Data Types and File Formats

What if you run a query that wants to compare a float column to a double column or

compare a value of one integer type with a value of a different integer type? Hive will

implicitly cast any integer to the larger of the two integer types, cast FLOAT to DOUBLE,

and cast any integer value to DOUBLE, as needed, so it is comparing identical types.

What if you run a query that wants to interpret a string column as a number? You can

explicitly cast one type to another as in the following example, where s is a string

column that holds a value representing an integer:

... cast(s AS INT) ...;

(To be clear, the AS INT are keywords, so lowercase would be fine.)

We’ll discuss data conversions in more depth in “Casting” on page 109.

Collection Data Types

Hive supports columns that are structs, maps, and arrays. Note that the literal syntax

examples in Table 3-2 are actually calls to built-in functions.

Table 3-2. Collection data types

Type Description Literal syntax examples

STRUCT Analogous to a C struct or an “object.” Fields can be accessed

using the “dot” notation. For example, if a column name is of

type STRUCT {first STRING; last STRING}, then

the first name field can be referenced using name.first.

struct('John', 'Doe')

MAP A collection of key-value tuples, where the fields are accessed

using array notation (e.g., ['key']). For example, if a column

name is of type MAP with key→value pairs

'first'→'John' and 'last'→'Doe', then the last

name can be referenced using name['last'].

map('first', 'John',

'last', 'Doe')

ARRAY Ordered sequences of the same type that are indexable using

zero-based integers. For example, if a column name is of type

ARRAY of strings with the value ['John', 'Doe'], then

the second element can be referenced using name[1].

array('John', 'Doe')

As for simple types, the case of the type name is ignored.

Most relational databases don’t support such collection types, because using them

tends to break normal form. For example, in traditional data models, structs might be

captured in separate tables, with foreign key relations between the tables, as

appropriate.

A practical problem with breaking normal form is the greater risk of data duplication,

leading to unnecessary disk space consumption and potential data inconsistencies, as

duplicate copies can grow out of sync as changes are made.

Collection Data Types | 43

However, in Big Data systems, a benefit of sacrificing normal form is higher processing

throughput. Scanning data off hard disks with minimal “head seeks” is essential when

processing terabytes to petabytes of data. Embedding collections in records makes re-

trieval faster with minimal seeks. Navigating each foreign key relationship requires

seeking across the disk, with significant performance overhead.

Hive doesn’t have the concept of keys. However, you can index tables,

as we’ll see in Chapter 7.

Here is a table declaration that demonstrates how to use these types, an employees table

in a fictitious Human Resources application:

CREATE TABLE employees (

name STRING,

salary FLOAT,

subordinates ARRAY<STRING>,

deductions MAP<STRING, FLOAT>,

address STRUCT<street:STRING, city:STRING, state:STRING, zip:INT>);

The name is a simple string and for most employees, a float is large enough for the salary.

The list of subordinates is an array of string values, where we treat the name as a “primary

key,” so each element in subordinates would reference another record in the table.

Employees without subordinates would have an empty array. In a traditional model,

the relationship would go the other way, from an employee to his or her manager. We’re

not arguing that our model is better for Hive; it’s just a contrived example to illustrate

the use of arrays.

The deductions is a map that holds a key-value pair for every deduction that will be

subtracted from the employee’s salary when paychecks are produced. The key is the

name of the deduction (e.g., “Federal Taxes”), and the key would either be a percentage

value or an absolute number. In a traditional data model, there might be separate tables

for deduction type (each key in our map), where the rows contain particular deduction

values and a foreign key pointing back to the corresponding employee record.

Finally, the home address of each employee is represented as a struct, where each field

is named and has a particular type.

Note that Java syntax conventions for generics are followed for the collection types. For

example, MAP<STRING, FLOAT> means that every key in the map will be of type STRING

and every value will be of type FLOAT. For an ARRAY<STRING>, every item in the array will

be a STRING. STRUCTs can mix different types, but the locations are fixed to the declared

position in the STRUCT.

44 | Chapter 3: Data Types and File Formats

Text File Encoding of Data Values

Let’s begin our exploration of file formats by looking at the simplest example, text files.

You are no doubt familiar with text files delimited with commas or tabs, the so-called

comma-separated values (CSVs) or tab-separated values (TSVs), respectively. Hive can

use those formats if you want and we’ll show you how shortly. However, there is a

drawback to both formats; you have to be careful about commas or tabs embedded in

text and not intended as field or column delimiters. For this reason, Hive uses various

control characters by default, which are less likely to appear in value strings. Hive uses

the term field when overriding the default delimiter, as we’ll see shortly. They are listed

in Table 3-3.

Table 3-3. Hive’s default record and field delimiters

Delimiter Description

\n For text files, each line is a record, so the line feed character separates records.

^A (“control” A) Separates all fields (columns). Written using the octal code \001 when explicitly

specified in CREATE TABLE statements.

^B Separate the elements in an ARRAY or STRUCT, or the key-value pairs in a MAP.

Written using the octal code \002 when explicitly specified in CREATE TABLE

statements.

^C Separate the key from the corresponding value in MAP key-value pairs. Written using

the octal code \003 when explicitly specified in CREATE TABLE statements.

Records for the employees table declared in the previous section would look like the

following example, where we use ^A, etc., to represent the field delimiters. A text editor

like Emacs will show the delimiters this way. Note that the lines have been wrapped in

the example because they are too long for the printed page. To clearly indicate the

division between records, we have added blank lines between them that would not

appear in the file:

John Doe^A100000.0^AMary Smith^BTodd Jones^AFederal Taxes^C.2^BState

Taxes^C.05^BInsurance^C.1^A1 Michigan Ave.^BChicago^BIL^B60600

Mary Smith^A80000.0^ABill King^AFederal Taxes^C.2^BState Taxes^C.

05^BInsurance^C.1^A100 Ontario St.^BChicago^BIL^B60601

Todd Jones^A70000.0^AFederal Taxes^C.15^BState Taxes^C.03^BInsurance^C.

1^A200 Chicago Ave.^BOak Park^BIL^B60700

Bill King^A60000.0^AFederal Taxes^C.15^BState Taxes^C.03^BInsurance^C.

1^A300 Obscure Dr.^BObscuria^BIL^B60100

This is a little hard to read, but you would normally let Hive do that for you, of course.

Let’s walk through the first line to understand the structure. First, here is what it would

Text File Encoding of Data Values | 45

look like in JavaScript Object Notation (JSON), where we have also inserted the names

from the table schema:

{

"name": "John Doe",

"salary": 100000.0,

"subordinates": ["Mary Smith", "Todd Jones"],

"deductions": {

"Federal Taxes": .2,

"State Taxes": .05,

"Insurance": .1

"address": {

"street": "1 Michigan Ave.",

"city": "Chicago",

"state": "IL",

"zip": 60600

}

You’ll note that maps and structs are effectively the same thing in JSON.

Now, here’s how the first line of the text file breaks down:

•John Doe is the name.

•100000.0 is the salary.

•Mary Smith^BTodd Jones are the subordinates “Mary Smith” and “Todd Jones.”

•Federal Taxes^C.2^BState Taxes^C.05^BInsurance^C.1 are the deductions, where

20% is deducted for “Federal Taxes,” 5% is deducted for “State Taxes,” and 10%

is deducted for “Insurance.”

•1 Michigan Ave.^BChicago^BIL^B60600 is the address, “1 Michigan Ave., Chicago,

60600.”

You can override these default delimiters. This might be necessary if another applica-

tion writes the data using a different convention. Here is the same table declaration

again, this time with all the format defaults explicitly specified:

CREATE TABLE employees (

name STRING,

salary FLOAT,

subordinates ARRAY<STRING>,

deductions MAP<STRING, FLOAT>,

address STRUCT<street:STRING, city:STRING, state:STRING, zip:INT>

)

ROW FORMAT DELIMITED

FIELDS TERMINATED BY '\001'

COLLECTION ITEMS TERMINATED BY '\002'

MAP KEYS TERMINATED BY '\003'

LINES TERMINATED BY '\n'

STORED AS TEXTFILE;

46 | Chapter 3: Data Types and File Formats

The ROW FORMAT DELIMITED sequence of keywords must appear before any of the other

clauses, with the exception of the STORED AS … clause.

The character \001 is the octal code for ^A. The clause ROW FORMAT DELIMITED FIELDS

TERMINATED BY '\001' means that Hive will use the ^A character to separate fields.

Similarly, the character \002 is the octal code for ^B. The clause ROW FORMAT DELIMITED

COLLECTION ITEMS TERMINATED BY '\002' means that Hive will use the ^B character to

separate collection items.

Finally, the character \003 is the octal code for ^C. The clause ROW FORMAT DELIMITED

MAP KEYS TERMINATED BY '\003' means that Hive will use the ^C character to separate

map keys from values.

The clause LINES TERMINATED BY '…' and STORED AS … do not require the ROW FORMAT

DELIMITED keywords.

Actually, it turns out that Hive does not currently support any character for LINES

TERMINATED BY … other than '\n'. So this clause has limited utility today.

You can override the field, collection, and key-value separators and still use the default

text file format, so the clause STORED AS TEXTFILE is rarely used. For most of this book,

we will use the default TEXTFILE file format.

There are other file format options, but we’ll defer discussing them until Chapter 15.

A related issue is compression of files, which we’ll discuss in Chapter 11.

So, while you can specify all these clauses explicitly, using the default separators most

of the time, you normally only provide the clauses for explicit overrides.

These specifications only affect what Hive expects to see when it reads

files. Except in a few limited cases, it’s up to you to write the data files

in the correct format.

For example, here is a table definition where the data will contain comma-delimited

fields.

CREATE TABLE some_data (

first FLOAT,

second FLOAT,

third FLOAT

)

ROW FORMAT DELIMITED

FIELDS TERMINATED BY ',';

Use '\t' for tab-delimited fields.

Text File Encoding of Data Values | 47

This example does not properly handle the general case of files in CSV

(comma-separated values) and TSV (tab-separated values) formats. They

can include a header row with column names and column string values

might be quoted and they might contain embedded commas or tabs,

respectively. See Chapter 15 for details on handling these file types more

generally.

This powerful customization feature makes it much easier to use Hive with files created

by other tools and various ETL (extract, transform, and load) processes.

Schema on Read

When you write data to a traditional database, either through loading external data,

writing the output of a query, doing UPDATE statements, etc., the database has total

control over the storage. The database is the “gatekeeper.” An important implication

of this control is that the database can enforce the schema as data is written. This is

called schema on write.

Hive has no such control over the underlying storage. There are many ways to create,

modify, and even damage the data that Hive will query. Therefore, Hive can only en-

force queries on read. This is called schema on read.

So what if the schema doesn’t match the file contents? Hive does the best that it can to

read the data. You will get lots of null values if there aren’t enough fields in each record

to match the schema. If some fields are numbers and Hive encounters nonnumeric

strings, it will return nulls for those fields. Above all else, Hive tries to recover from all

errors as best it can.

48 | Chapter 3: Data Types and File Formats

CHAPTER 4

HiveQL: Data Definition

HiveQL is the Hive query language. Like all SQL dialects in widespread use, it doesn’t

fully conform to any particular revision of the ANSI SQL standard. It is perhaps closest

to MySQL’s dialect, but with significant differences. Hive offers no support for row-

level inserts, updates, and deletes. Hive doesn’t support transactions. Hive adds ex-

tensions to provide better performance in the context of Hadoop and to integrate with

custom extensions and even external programs.

Still, much of HiveQL will be familiar. This chapter and the ones that follow discuss

the features of HiveQL using representative examples. In some cases, we will briefly

mention details for completeness, then explore them more fully in later chapters.

This chapter starts with the so-called data definition language parts of HiveQL, which

are used for creating, altering, and dropping databases, tables, views, functions, and

indexes. We’ll discuss databases and tables in this chapter, deferring the discussion of

views until Chapter 7, indexes until Chapter 8, and functions until Chapter 13.

We’ll also discuss the SHOW and DESCRIBE commands for listing and describing items as

we go.

Subsequent chapters explore the data manipulation language parts of HiveQL that are

used to put data into Hive tables and to extract data to the filesystem, and how to

explore and manipulate data with queries, grouping, filtering, joining, etc.

Databases in Hive

The Hive concept of a database is essentially just a catalog or namespace of tables.

However, they are very useful for larger clusters with multiple teams and users, as a

way of avoiding table name collisions. It’s also common to use databases to organize

production tables into logical groups.

If you don’t specify a database, the default database is used.

The simplest syntax for creating a database is shown in the following example:

hive> CREATE DATABASE financials;

Hive will throw an error if financials already exists. You can suppress these warnings

with this variation:

hive> CREATE DATABASE IF NOT EXISTS financials;

While normally you might like to be warned if a database of the same name already

exists, the IF NOT EXISTS clause is useful for scripts that should create a database on-

the-fly, if necessary, before proceeding.

You can also use the keyword SCHEMA instead of DATABASE in all the database-related

commands.

At any time, you can see the databases that already exist as follows:

hive> SHOW DATABASES;

default

financials

hive> CREATE DATABASE human_resources;

hive> SHOW DATABASES;

default

financials

human_resources

If you have a lot of databases, you can restrict the ones listed using a regular expres-

sion, a concept we’ll explain in “LIKE and RLIKE” on page 96, if it is new to you. The

following example lists only those databases that start with the letter h and end with

any other characters (the .* part):

hive> SHOW DATABASES LIKE 'h.*';

human_resources

hive> ...

Hive will create a directory for each database. Tables in that database will be stored in

subdirectories of the database directory. The exception is tables in the default database,

which doesn’t have its own directory.

The database directory is created under a top-level directory specified by the property

hive.metastore.warehouse.dir, which we discussed in “Local Mode Configura-

tion” on page 24 and “Distributed and Pseudodistributed Mode Configura-

tion” on page 26. Assuming you are using the default value for this property, /user/hive/

warehouse, when the financials database is created, Hive will create the directory /user/

hive/warehouse/financials.db. Note the .db extension.

You can override this default location for the new directory as shown in this example:

hive> CREATE DATABASE financials

> LOCATION '/my/preferred/directory';

You can add a descriptive comment to the database, which will be shown by the

DESCRIBE DATABASE <database> command.

50 | Chapter 4: HiveQL: Data Definition

hive> CREATE DATABASE financials

> COMMENT 'Holds all financial tables';

hive> DESCRIBE DATABASE financials;

financials Holds all financial tables

hdfs://master-server/user/hive/warehouse/financials.db

Note that DESCRIBE DATABASE also shows the directory location for the database. In this

example, the URI scheme is hdfs. For a MapR installation, it would be maprfs. For an

Amazon Elastic MapReduce (EMR) cluster, it would also be hdfs, but you could set

hive.metastore.warehouse.dir to use Amazon S3 explicitly (i.e., by specifying s3n://

bucketname/… as the property value). You could use s3 as the scheme, but the newer

s3n is preferred.

In the output of DESCRIBE DATABASE, we’re showing master-server to indicate the URI

authority, in this case a DNS name and optional port number (i.e., server:port) for the

“master node” of the filesystem (i.e., where the NameNode service is running for

HDFS). If you are running in pseudo-distributed mode, then the master server will be

localhost. For local mode, the path will be a local path, file:///user/hive/warehouse/

financials.db.

If the authority is omitted, Hive uses the master-server name and port defined by

the property fs.default.name in the Hadoop configuration files, found in the

$HADOOP_HOME/conf directory.

To be clear, hdfs:///user/hive/warehouse/financials.db is equivalent to hdfs://master-

server/user/hive/warehouse/financials.db, where master-server is your master node’s

DNS name and optional port.

For completeness, when you specify a relative path (e.g., some/relative/path), Hive will

put this under your home directory in the distributed filesystem (e.g., hdfs:///user/<user-

name>) for HDFS. However, if you are running in local mode, your current working

directory is used as the parent of some/relative/path.

For script portability, it’s typical to omit the authority, only specifying it when referring

to another distributed filesystem instance (including S3 buckets).

Lastly, you can associate key-value properties with the database, although their only

function currently is to provide a way of adding information to the output of DESCRIBE

DATABASE EXTENDED <database>:

hive> CREATE DATABASE financials

> WITH DBPROPERTIES ('creator' = 'Mark Moneybags', 'date' = '2012-01-02');

hive> DESCRIBE DATABASE financials;

financials hdfs://master-server/user/hive/warehouse/financials.db

hive> DESCRIBE DATABASE EXTENDED financials;

financials hdfs://master-server/user/hive/warehouse/financials.db

{date=2012-01-02, creator=Mark Moneybags);

Databases in Hive | 51

The USE command sets a database as your working database, analogous to changing

working directories in a filesystem:

hive> USE financials;

Now, commands such as SHOW TABLES; will list the tables in this database.

Unfortunately, there is no command to show you which database is your current

working database! Fortunately, it’s always safe to repeat the USE … command; there is

no concept in Hive of nesting of databases.

Recall that we pointed out a useful trick in “Variables and Properties” on page 31 for

setting a property to print the current database as part of the prompt (Hive v0.8.0 and

later):

hive> set hive.cli.print.current.db=true;

hive (financials)> USE default;

hive (default)> set hive.cli.print.current.db=false;

hive> ...

Finally, you can drop a database:

hive> DROP DATABASE IF EXISTS financials;

The IF EXISTS is optional and suppresses warnings if financials doesn’t exist.

By default, Hive won’t permit you to drop a database if it contains tables. You can either

drop the tables first or append the CASCADE keyword to the command, which will cause

the Hive to drop the tables in the database first:

hive> DROP DATABASE IF EXISTS financials CASCADE;

Using the RESTRICT keyword instead of CASCADE is equivalent to the default behavior,

where existing tables must be dropped before dropping the database.

When a database is dropped, its directory is also deleted.

Alter Database

You can set key-value pairs in the DBPROPERTIES associated with a database using the

ALTER DATABASE command. No other metadata about the database can be changed,

including its name and directory location:

hive> ALTER DATABASE financials SET DBPROPERTIES ('edited-by' = 'Joe Dba');

There is no way to delete or “unset” a DBPROPERTY.

52 | Chapter 4: HiveQL: Data Definition

Creating Tables

The CREATE TABLE statement follows SQL conventions, but Hive’s version offers sig-

nificant extensions to support a wide range of flexibility where the data files for tables

are stored, the formats used, etc. We discussed many of these options in “Text File

Encoding of Data Values” on page 45 and we’ll return to more advanced options later

in Chapter 15. In this section, we describe the other options available for the CREATE

TABLE statement, adapting the employees table declaration we used previously in “Col-

lection Data Types” on page 43:

CREATE TABLE IF NOT EXISTS mydb.employees (

name STRING COMMENT 'Employee name',

salary FLOAT COMMENT 'Employee salary',

subordinates ARRAY<STRING> COMMENT 'Names of subordinates',

deductions MAP<STRING, FLOAT>

COMMENT 'Keys are deductions names, values are percentages',

address STRUCT<street:STRING, city:STRING, state:STRING, zip:INT>

COMMENT 'Home address')

COMMENT 'Description of the table'

TBLPROPERTIES ('creator'='me', 'created_at'='2012-01-02 10:00:00', ...)

LOCATION '/user/hive/warehouse/mydb.db/employees';

First, note that you can prefix a database name, mydb in this case, if you’re not currently

working in the target database.

If you add the option IF NOT EXISTS, Hive will silently ignore the statement if the table

already exists. This is useful in scripts that should create a table the first time they run.

However, the clause has a gotcha you should know. If the schema specified differs from

the schema in the table that already exists, Hive won’t warn you. If your intention is

for this table to have the new schema, you’ll have to drop the old table, losing your

data, and then re-create it. Consider if you should use one or more ALTER TABLE state-

ments to change the existing table schema instead. See “Alter Table” on page 66 for

details.

If you use IF NOT EXISTS and the existing table has a different schema

than the schema in the CREATE TABLE statement, Hive will ignore the

discrepancy.

You can add a comment to any column, after the type. Like databases, you can attach

a comment to the table itself and you can define one or more table properties. In most

cases, the primary benefit of TBLPROPERTIES is to add additional documentation in a

key-value format. However, when we examine Hive’s integration with databases such

as DynamoDB (see “DynamoDB” on page 225), we’ll see that the TBLPROPERTIES can

be used to express essential metadata about the database connection.

Creating Tables | 53

Hive automatically adds two table properties: last_modified_by holds the username of

the last user to modify the table, and last_modified_time holds the epoch time in sec-

onds of that modification.

A planned enhancement for Hive v0.10.0 is to add a SHOW TBLPROPERTIES

table_name command that will list just the TBLPROPERTIES for a table.

Finally, you can optionally specify a location for the table data (as opposed to meta-

data, which the metastore will always hold). In this example, we are showing the default

location that Hive would use, /user/hive/warehouse/mydb.db/employees, where /user/

hive/warehouse is the default “warehouse” location (as discussed previously),

mydb.db is the database directory, and employees is the table directory.

By default, Hive always creates the table’s directory under the directory for the enclos-

ing database. The exception is the default database. It doesn’t have a directory un-

der /user/hive/warehouse, so a table in the default database will have its directory created

directly in /user/hive/warehouse (unless explicitly overridden).

To avoid potential confusion, it’s usually better to use an external table

if you don’t want to use the default location table. See “External

Tables” on page 56 for details.

You can also copy the schema (but not the data) of an existing table:

CREATE TABLE IF NOT EXISTS mydb.employees2

LIKE mydb.employees;

This version also accepts the optional LOCATION clause, but note that no other properties,

including the schema, can be defined; they are determined from the original table.

The SHOW TABLES command lists the tables. With no additional arguments, it shows the

tables in the current working database. Let’s assume we have already created a few

other tables, table1 and table2, and we did so in the mydb database:

hive> USE mydb;

hive> SHOW TABLES;

employees

table1

table2

If we aren’t in the same database, we can still list the tables in that database:

hive> USE default;

hive> SHOW TABLES IN mydb;

employees

54 | Chapter 4: HiveQL: Data Definition

table1

table2

If we have a lot of tables, we can limit the ones listed using a regular expression, a

concept we’ll discuss in detail in “LIKE and RLIKE” on page 96:

hive> USE mydb;

hive> SHOW TABLES 'empl.*';

employees

Not all regular expression features are supported. If you know regular expressions, it’s

better to test a candidate regular expression to make sure it actually works!

The regular expression in the single quote looks for all tables with names starting with

empl and ending with any other characters (the .* part).

Using the IN database_name clause and a regular expression for the table

names together is not supported.

We can also use the DESCRIBE EXTENDED mydb.employees command to show details about

the table. (We can drop the mydb. prefix if we’re currently using the mydb database.) We

have reformatted the output for easier reading and we have suppressed many details

to focus on the items that interest us now:

hive> DESCRIBE EXTENDED mydb.employees;

name string Employee name

salary float Employee salary

subordinates array<string> Names of subordinates

deductions map<string,float> Keys are deductions names, values are percentages

address struct<street:string,city:string,state:string,zip:int> Home address

Detailed Table Information Table(tableName:employees, dbName:mydb, owner:me,

...

location:hdfs://master-server/user/hive/warehouse/mydb.db/employees,

parameters:{creator=me, created_at='2012-01-02 10:00:00',

last_modified_user=me, last_modified_time=1337544510,

comment:Description of the table, ...}, ...)

Replacing EXTENDED with FORMATTED provides more readable but also more verbose

output.

The first section shows the output of DESCRIBE without EXTENDED or FORMATTED (i.e., the

schema including the comments for each column).

If you only want to see the schema for a particular column, append the column to the

table name. Here, EXTENDED adds no additional output:

hive> DESCRIBE mydb.employees.salary;

salary float Employee salary

Creating Tables | 55

Returning to the extended output, note the line in the description that starts with

location:. It shows the full URI path in HDFS to the directory where Hive will keep

all the data for this table, as we discussed above.

We said that the last_modified_by and last_modified_time table prop-

erties are automatically created. However, they are only shown in the

Detailed Table Information if a user-specified table property has also

been defined!

Managed Tables

The tables we have created so far are called managed tables or sometimes called inter-

nal tables, because Hive controls the lifecycle of their data (more or less). As we’ve seen,

Hive stores the data for these tables in a subdirectory under the directory defined by

hive.metastore.warehouse.dir (e.g., /user/hive/warehouse), by default.

When we drop a managed table (see “Dropping Tables” on page 66), Hive deletes

the data in the table.

However, managed tables are less convenient for sharing with other tools. For example,

suppose we have data that is created and used primarily by Pig or other tools, but we

want to run some queries against it, but not give Hive ownership of the data. We can

define an external table that points to that data, but doesn’t take ownership of it.

External Tables

Suppose we are analyzing data from the stock markets. Periodically, we ingest the data

for NASDAQ and the NYSE from a source like Infochimps (http://infochimps.com/da

tasets) and we want to study this data with many tools. (See the data sets named

infochimps_dataset_4777_download_16185 and infochimps_dataset_4778_download_

16677, respectively, which are actually sourced from Yahoo! Finance.) The schema we’ll

use next matches the schemas of both these data sources. Let’s assume the data files

are in the distributed filesystem directory /data/stocks.

The following table declaration creates an external table that can read all the data files

for this comma-delimited data in /data/stocks:

CREATE EXTERNAL TABLE IF NOT EXISTS stocks (

exchange STRING,

symbol STRING,

ymd STRING,

price_open FLOAT,

price_high FLOAT,

price_low FLOAT,

price_close FLOAT,

volume INT,

price_adj_close FLOAT)

ROW FORMAT DELIMITED FIELDS TERMINATED BY ','

LOCATION '/data/stocks';

56 | Chapter 4: HiveQL: Data Definition

Downloa d f r o m W o w ! e B o o k < w w w.woweb o o k . c o m >

The EXTERNAL keyword tells Hive this table is external and the LOCATION … clause is

required to tell Hive where it’s located.

Because it’s external, Hive does not assume it owns the data. Therefore, dropping the

table does not delete the data, although the metadata for the table will be deleted.

There are a few other small differences between managed and external tables, where

some HiveQL constructs are not permitted for external tables. We’ll discuss those when

we come to them.

However, it’s important to note that the differences between managed and external

tables are smaller than they appear at first. Even for managed tables, you know where

they are located, so you can use other tools, hadoop dfs commands, etc., to modify and

even delete the files in the directories for managed tables. Hive may technically own

these directories and files, but it doesn’t have full control over them! Recall, in “Schema

on Read” on page 48, we said that Hive really has no control over the integrity of the

files used for storage and whether or not their contents are consistent with the table

schema. Even managed tables don’t give us this control.

Still, a general principle of good software design is to express intent. If the data is shared

between tools, then creating an external table makes this ownership explicit.

You can tell whether or not a table is managed or external using the output of DESCRIBE

EXTENDED tablename. Near the end of the Detailed Table Information output, you will

see the following for managed tables:

... tableType:MANAGED_TABLE)

For external tables, you will see the following:

... tableType:EXTERNAL_TABLE)

As for managed tables, you can also copy the schema (but not the data) of an existing

table:

CREATE EXTERNAL TABLE IF NOT EXISTS mydb.employees3

LIKE mydb.employees

LOCATION '/path/to/data';

If you omit the EXTERNAL keyword and the original table is external, the

new table will also be external. If you omit EXTERNAL and the original

table is managed, the new table will also be managed. However, if you

include the EXTERNAL keyword and the original table is managed, the new

table will be external. Even in this scenario, the LOCATION clause will

still be optional.

Creating Tables | 57

Partitioned, Managed Tables

The general notion of partitioning data is an old one. It can take many forms, but often

it’s used for distributing load horizontally, moving data physically closer to its most

frequent users, and other purposes.

Hive has the notion of partitioned tables. We’ll see that they have important

performance benefits, and they can help organize data in a logical fashion, such as

hierarchically.

We’ll discuss partitioned managed tables first. Let’s return to our employees table and

imagine that we work for a very large multinational corporation. Our HR people often

run queries with WHERE clauses that restrict the results to a particular country or to a

particular first-level subdivision (e.g., state in the United States or province in Canada).

(First-level subdivision is an actual term, used here, for example: http://www.common

datahub.com/state_source.jsp.) We’ll just use the word state for simplicity. We have

redundant state information in the address field. It is distinct from the state partition.

We could remove the state element from address. There is no ambiguity in queries,

since we have to use address.state to project the value inside the address. So, let’s

partition the data first by country and then by state:

CREATE TABLE employees (

name STRING,

salary FLOAT,

subordinates ARRAY<STRING>,

deductions MAP<STRING, FLOAT>,

address STRUCT<street:STRING, city:STRING, state:STRING, zip:INT>

)

PARTITIONED BY (country STRING, state STRING);

Partitioning tables changes how Hive structures the data storage. If we create this table

in the mydb database, there will still be an employees directory for the table:

hdfs://master_server/user/hive/warehouse/mydb.db/employees

However, Hive will now create subdirectories reflecting the partitioning structure. For

example:

...

.../employees/country=CA/state=AB

.../employees/country=CA/state=BC

...

.../employees/country=US/state=AL

.../employees/country=US/state=AK

...

Yes, those are the actual directory names. The state directories will contain zero or more

files for the employees in those states.

58 | Chapter 4: HiveQL: Data Definition

Once created, the partition keys (country and state, in this case) behave like regular

columns. There is one known exception, due to a bug (see “Aggregate func-

tions” on page 85). In fact, users of the table don’t need to care if these “columns”

are partitions or not, except when they want to optimize query performance.

For example, the following query selects all employees in the state of Illinois in the

United States:

SELECT * FROM employees

WHERE country = 'US' AND state = 'IL';

Note that because the country and state values are encoded in directory names, there

is no reason to have this data in the data files themselves. In fact, the data just gets in

the way in the files, since you have to account for it in the table schema, and this data

wastes space.

Perhaps the most important reason to partition data is for faster queries. In the previous

query, which limits the results to employees in Illinois, it is only necessary to scan the

contents of one directory. Even if we have thousands of country and state directories,

all but one can be ignored. For very large data sets, partitioning can dramatically im-

prove query performance, but only if the partitioning scheme reflects common range

filtering (e.g., by locations, timestamp ranges).

When we add predicates to WHERE clauses that filter on partition values, these predicates

are called partition filters.

Even if you do a query across the entire US, Hive only reads the 65 directories covering

the 50 states, 9 territories, and the District of Columbia, and 6 military “states” used

by the armed services. You can see the full list here: http://www.50states.com/abbrevia

tions.htm.

Of course, if you need to do a query for all employees around the globe, you can still

do it. Hive will have to read every directory, but hopefully these broader disk scans will

be relatively rare.

However, a query across all partitions could trigger an enormous MapReduce job if the

table data and number of partitions are large. A highly suggested safety measure is

putting Hive into “strict” mode, which prohibits queries of partitioned tables without

a WHERE clause that filters on partitions. You can set the mode to “nonstrict,” as in the

following session:

hive> set hive.mapred.mode=strict;

hive> SELECT e.name, e.salary FROM employees e LIMIT 100;

FAILED: Error in semantic analysis: No partition predicate found for

Alias "e" Table "employees"

hive> set hive.mapred.mode=nonstrict;

hive> SELECT e.name, e.salary FROM employees e LIMIT 100;

Partitioned, Managed Tables | 59

John Doe 100000.0

...

You can see the partitions that exist with the SHOW PARTITIONS command:

hive> SHOW PARTITIONS employees;

...

Country=CA/state=AB

country=CA/state=BC

...

country=US/state=AL

country=US/state=AK

...

If you have a lot of partitions and you want to see if partitions have been defined for

particular partition keys, you can further restrict the command with an optional PARTI

TION clause that specifies one or more of the partitions with specific values:

hive> SHOW PARTITIONS employees PARTITION(country='US');

country=US/state=AL

country=US/state=AK

...

hive> SHOW PARTITIONS employees PARTITION(country='US', state='AK');

country=US/state=AK

The DESCRIBE EXTENDED employees command shows the partition keys:

hive> DESCRIBE EXTENDED employees;

name string,

salary float,

...

address struct<...>,

country string,

state string

Detailed Table Information...

partitionKeys:[FieldSchema(name:country, type:string, comment:null),

FieldSchema(name:state, type:string, comment:null)],

...

The schema part of the output lists the country and state with the other columns,

because they are columns as far as queries are concerned. The Detailed Table Infor

mation includes the country and state as partition keys. The comments for both of these

keys are null; we could have added comments just as for regular columns.

You create partitions in managed tables by loading data into them. The following ex-

ample creates a US and CA (California) partition while loading data into it from a local

directory, $HOME/california-employees. You must specify a value for each partition

column. Notice how we reference the HOME environment variable in HiveQL:

LOAD DATA LOCAL INPATH '${env:HOME}/california-employees'

INTO TABLE employees

PARTITION (country = 'US', state = 'CA');

60 | Chapter 4: HiveQL: Data Definition

The directory for this partition, …/employees/country=US/state=CA, will be created by

Hive and all data files in $HOME/california-employees will be copied into it. See

“Loading Data into Managed Tables” on page 71 for more information on populating

tables.

External Partitioned Tables

You can use partitioning with external tables. In fact, you may find that this is your

most common scenario for managing large production data sets. The combi-

nation gives you a way to “share” data with other tools, while still optimizing query

performance.

You also have more flexibility in the directory structure used, as you define it yourself.

We’ll see a particularly useful example in a moment.

Let’s consider a new example that fits this scenario well: logfile analysis. Most organ-

izations use a standard format for log messages, recording a timestamp, severity (e.g.,

ERROR, WARNING, INFO), perhaps a server name and process ID, and then an arbitrary text

message. Suppose our Extract, Transform, and Load (ETL) process ingests and aggre-

gates logfiles in our environment, converting each log message to a tab-delimited record

and also decomposing the timestamp into separate year, month, and day fields, and a

combined hms field for the remaining hour, minute, and second parts of the timestamp,

for reasons that will become clear in a moment. You could do this parsing of log mes-

sages using the string parsing functions built into Hive or Pig, for example. Alterna-

tively, we could use smaller integer types for some of the timestamp-related fields to

conserve space. Here, we are ignoring subsequent resolution.

Here’s how we might define the corresponding Hive table:

CREATE EXTERNAL TABLE IF NOT EXISTS log_messages (

hms INT,

severity STRING,

server STRING,

process_id INT,

message STRING)

PARTITIONED BY (year INT, month INT, day INT)

ROW FORMAT DELIMITED FIELDS TERMINATED BY '\t';

We’re assuming that a day’s worth of log data is about the correct size for a useful

partition and finer grain queries over a day’s data will be fast enough.

Recall that when we created the nonpartitioned external stocks table, a LOCATION …

clause was required. It isn’t used for external partitioned tables. Instead, an ALTER

TABLE statement is used to add each partition separately. It must specify a value for each

partition key, the year, month, and day, in this case (see “Alter Table” on page 66 for

more details on this feature). Here is an example, where we add a partition for January

2nd, 2012:

ALTER TABLE log_messages ADD PARTITION(year = 2012, month = 1, day = 2)

LOCATION 'hdfs://master_server/data/log_messages/2012/01/02';

Partitioned, Managed Tables | 61

The directory convention we use is completely up to us. Here, we follow a hierarchical

directory structure, because it’s a logical way to organize our data, but there is no

requirement to do so. We could follow Hive’s directory naming convention (e.g., …/

exchange=NASDAQ/symbol=AAPL), but there is no requirement to do so.

An interesting benefit of this flexibility is that we can archive old data on inexpensive

storage, like Amazon’s S3, while keeping newer, more “interesting” data in HDFS. For

example, each day we might use the following procedure to move data older than a

month to S3:

• Copy the data for the partition being moved to S3. For example, you can use the

hadoop distcp command:

hadoop distcp /data/log_messages/2011/12/02 s3n://ourbucket/logs/2011/12/02

• Alter the table to point the partition to the S3 location:

ALTER TABLE log_messages PARTITION(year = 2011, month = 12, day = 2)

SET LOCATION 's3n://ourbucket/logs/2011/01/02';

• Remove the HDFS copy of the partition using the hadoop fs -rmr command:

hadoop fs -rmr /data/log_messages/2011/01/02

You don’t have to be an Amazon Elastic MapReduce user to use S3 this way. S3 support

is part of the Apache Hadoop distribution. You can still query this data, even queries

that cross the month-old “boundary,” where some data is read from HDFS and some

data is read from S3!

By the way, Hive doesn’t care if a partition directory doesn’t exist for a partition or if

it has no files. In both cases, you’ll just get no results for a query that filters for the

partition. This is convenient when you want to set up partitions before a separate pro-

cess starts writing data to them. As soon as data is there, queries will return results from

that data.

This feature illustrates another benefit: new data can be written to a dedicated directory

with a clear distinction from older data in other directories. Also, whether you move

old data to an “archive” location or delete it outright, the risk of tampering with newer

data is reduced since the data subsets are in separate directories.

As for nonpartitioned external tables, Hive does not own the data and it does not delete

the data if the table is dropped.

As for managed partitioned tables, you can see an external table’s partitions with SHOW

PARTITIONS:

hive> SHOW PARTITIONS log_messages;

...

year=2011/month=12/day=31

year=2012/month=1/day=1

year=2012/month=1/day=2

...

62 | Chapter 4: HiveQL: Data Definition

Similarly, the DESCRIBE EXTENDED log_messages shows the partition keys both as part

of the schema and in the list of partitionKeys:

hive> DESCRIBE EXTENDED log_messages;

...

message string,

year int,

month int,

day int

Detailed Table Information...

partitionKeys:[FieldSchema(name:year, type:int, comment:null),

FieldSchema(name:month, type:int, comment:null),

FieldSchema(name:day, type:int, comment:null)],

...

This output is missing a useful bit of information, the actual location of the partition

data. There is a location field, but it only shows Hive’s default directory that would be

used if the table were a managed table. However, we can get a partition’s location as

follows:

hive> DESCRIBE EXTENDED log_messages PARTITION (year=2012, month=1, day=2);

...

location:s3n://ourbucket/logs/2011/01/02,

...

We frequently use external partitioned tables because of the many benefits they pro-

vide, such as logical data management, performant queries, etc.

ALTER TABLE … ADD PARTITION is not limited to external tables. You can use it with

managed tables, too, when you have (or will have) data for partitions in directories

created outside of the LOAD and INSERT options we discussed above. You’ll need to

remember that not all of the table’s data will be under the usual Hive “warehouse”

directory, and this data won’t be deleted when you drop the managed table! Hence,

from a “sanity” perspective, it’s questionable whether you should dare to use this fea-

ture with managed tables.

Customizing Table Storage Formats

In “Text File Encoding of Data Values” on page 45, we discussed that Hive defaults to

a text file format, which is indicated by the optional clause STORED AS TEXTFILE, and

you can overload the default values for the various delimiters when creating the table.

Here we repeat the definition of the employees table we used in that discussion:

CREATE TABLE employees (

name STRING,

salary FLOAT,

subordinates ARRAY<STRING>,

deductions MAP<STRING, FLOAT>,

address STRUCT<street:STRING, city:STRING, state:STRING, zip:INT>

)

ROW FORMAT DELIMITED

Partitioned, Managed Tables | 63

FIELDS TERMINATED BY '\001'

COLLECTION ITEMS TERMINATED BY '\002'

MAP KEYS TERMINATED BY '\003'

LINES TERMINATED BY '\n'

STORED AS TEXTFILE;

TEXTFILE implies that all fields are encoded using alphanumeric characters, including

those from international character sets, although we observed that Hive uses non-

printing characters as “terminators” (delimiters), by default. When TEXTFILE is used,

each line is considered a separate record.

You can replace TEXTFILE with one of the other built-in file formats supported by Hive,

including SEQUENCEFILE and RCFILE, both of which optimize disk space usage and I/O

bandwidth performance using binary encoding and optional compression. These for-

mats are discussed in more detail in Chapter 11 and Chapter 15.

Hive draws a distinction between how records are encoded into files and how columns

are encoded into records. You customize these behaviors separately.

The record encoding is handled by an input format object (e.g., the Java code behind

TEXTFILE.) Hive uses a Java class (compiled module) named org.apache

.hadoop.mapred.TextInputFormat. If you are unfamiliar with Java, the dotted name syn-

tax indicates a hierarchical namespace tree of packages that actually corresponds to the

directory structure for the Java code. The last name, TextInputFormat, is a class in the

lowest-level package mapred.

The record parsing is handled by a serializer/deserializer or SerDe for short. For TEXT

FILE and the encoding we described in Chapter 3 and repeated in the example above,

the SerDe Hive uses is another Java class called org.apache.hadoop.hive.serde2.lazy.

LazySimpleSerDe.

For completeness, there is also an output format that Hive uses for writing the

output of queries to files and to the console. For TEXTFILE, the Java class

named org.apache.hadoop.hive.ql.io.HiveIgnoreKeyTextOutputFormat is used for

output.

Hive uses an input format to split input streams into records, an output

format to format records into output streams (i.e., the output of quer-

ies), and a SerDe to parse records into columns, when reading, and

encodes columns into records, when writing. We’ll explore these dis-

tinctions in greater depth in Chapter 15.

Third-party input and output formats and SerDes can be specified, a feature which

permits users to customize Hive for a wide range of file formats not supported natively.

Here is a complete example that uses a custom SerDe, input format, and output format

for files accessible through the Avro protocol, which we will discuss in detail in “Avro

Hive SerDe” on page 209:

64 | Chapter 4: HiveQL: Data Definition

CREATE TABLE kst

PARTITIONED BY (ds string)

ROW FORMAT SERDE 'com.linkedin.haivvreo.AvroSerDe'

WITH SERDEPROPERTIES ('schema.url'='http://schema_provider/kst.avsc')

STORED AS

INPUTFORMAT 'com.linkedin.haivvreo.AvroContainerInputFormat'

OUTPUTFORMAT 'com.linkedin.haivvreo.AvroContainerOutputFormat';

The ROW FORMAT SERDE … specifies the SerDe to use. Hive provides the WITH SERDEPRO

PERTIES feature that allows users to pass configuration information to the SerDe. Hive

knows nothing about the meaning of these properties. It’s up to the SerDe to decide

their meaning. Note that the name and value of each property must be a quoted string.

Finally, the STORED AS INPUTFORMAT … OUTPUTFORMAT … clause specifies the Java classes

to use for the input and output formats, respectively. If you specify one of these formats,

you are required to specify both of them.

Note that the DESCRIBE EXTENDED table command lists the input and output formats,

the SerDe, and any SerDe properties in the DETAILED TABLE INFORMATION. For our ex-

ample, we would see the following:

hive> DESCRIBE EXTENDED kst

...

inputFormat:com.linkedin.haivvreo.AvroContainerInputFormat,

outputFormat:com.linkedin.haivvreo.AvroContainerOutputFormat,

...

serdeInfo:SerDeInfo(name:null,

serializationLib:com.linkedin.haivvreo.AvroSerDe,

parameters:{schema.url=http://schema_provider/kst.avsc})

...

Finally, there are a few additional CREATE TABLE clauses that describe more details about

how the data is supposed to be stored. Let’s extend our previous stocks table example

from “External Tables” on page 56:

CREATE EXTERNAL TABLE IF NOT EXISTS stocks (

exchange STRING,

symbol STRING,

ymd STRING,

price_open FLOAT,

price_high FLOAT,

price_low FLOAT,

price_close FLOAT,

volume INT,

price_adj_close FLOAT)

CLUSTERED BY (exchange, symbol)

SORTED BY (ymd ASC)

INTO 96 BUCKETS

ROW FORMAT DELIMITED FIELDS TERMINATED BY ','

LOCATION '/data/stocks';

The CLUSTERED BY … INTO … BUCKETS clause, with an optional SORTED BY … clause is used

to optimize certain kinds of queries, which we discuss in detail in “Bucketing Table

Data Storage” on page 125.

Partitioned, Managed Tables | 65

Dropping Tables

The familiar DROP TABLE command from SQL is supported:

DROP TABLE IF EXISTS employees;

The IF EXISTS keywords are optional. If not used and the table doesn’t exist, Hive

returns an error.

For managed tables, the table metadata and data are deleted.

Actually, if you enable the Hadoop Trash feature, which is not on by

default, the data is moved to the .Trash directory in the distributed

filesystem for the user, which in HDFS is /user/$USER/.Trash. To enable

this feature, set the property fs.trash.interval to a reasonable positive

number. It’s the number of minutes between “trash checkpoints”; 1,440

would be 24 hours. While it’s not guaranteed to work for all versions of

all distributed filesystems, if you accidentally drop a managed table with

important data, you may be able to re-create the table, re-create any

partitions, and then move the files from .Trash to the correct directories

(using the filesystem commands) to restore the data.

For external tables, the metadata is deleted but the data is not.

Alter Table

Most table properties can be altered with ALTER TABLE statements, which change

metadata about the table but not the data itself. These statements can be used to fix

mistakes in schema, move partition locations (as we saw in “External Partitioned

Tables” on page 61), and do other operations.

ALTER TABLE modifies table metadata only. The data for the table is

untouched. It’s up to you to ensure that any modifications are consistent

with the actual data.

Renaming a Table

Use this statement to rename the table log_messages to logmsgs:

ALTER TABLE log_messages RENAME TO logmsgs;

Adding, Modifying, and Dropping a Table Partition

As we saw previously, ALTER TABLE table ADD PARTITION … is used to add a new partition

to a table (usually an external table). Here we repeat the same command shown pre-

viously with the additional options available:

66 | Chapter 4: HiveQL: Data Definition

ALTER TABLE log_messages ADD IF NOT EXISTS

PARTITION (year = 2011, month = 1, day = 1) LOCATION '/logs/2011/01/01'

PARTITION (year = 2011, month = 1, day = 2) LOCATION '/logs/2011/01/02'

PARTITION (year = 2011, month = 1, day = 3) LOCATION '/logs/2011/01/03'

...;

Multiple partitions can be added in the same query when using Hive v0.8.0 and later.

As always, IF NOT EXISTS is optional and has the usual meaning.

Hive v0.7.X allows you to use the syntax with multiple partition speci-

fications, but it actually uses just the first partition specification, silently

ignoring the others! Instead, use a separate ALTER STATEMENT statement

for each partition.

Similarly, you can change a partition location, effectively moving it:

ALTER TABLE log_messages PARTITION(year = 2011, month = 12, day = 2)

SET LOCATION 's3n://ourbucket/logs/2011/01/02';

This command does not move the data from the old location, nor does it delete the old

data.

Finally, you can drop a partition:

ALTER TABLE log_messages DROP IF EXISTS PARTITION(year = 2011, month = 12, day = 2);

The IF EXISTS clause is optional, as usual. For managed tables, the data for the partition

is deleted, along with the metadata, even if the partition was created using ALTER TABLE

… ADD PARTITION. For external tables, the data is not deleted.

There are a few more ALTER statements that affect partitions discussed later

in “Alter Storage Properties” on page 68 and “Miscellaneous Alter Table State-

ments” on page 69.

Changing Columns

You can rename a column, change its position, type, or comment:

ALTER TABLE log_messages

CHANGE COLUMN hms hours_minutes_seconds INT

COMMENT 'The hours, minutes, and seconds part of the timestamp'

AFTER severity;

You have to specify the old name, a new name, and the type, even if the name or type

is not changing. The keyword COLUMN is optional as is the COMMENT clause. If you aren’t

moving the column, the AFTER other_column clause is not necessary. In the example

shown, we move the column after the severity column. If you want to move the column

to the first position, use FIRST instead of AFTER other_column.

As always, this command changes metadata only. If you are moving columns, the data

must already match the new schema or you must change it to match by some other

means.

Alter Table | 67

Adding Columns

You can add new columns to the end of the existing columns, before any partition

columns.

ALTER TABLE log_messages ADD COLUMNS (

app_name STRING COMMENT 'Application name',

session_id LONG COMMENT 'The current session id');

The COMMENT clauses are optional, as usual. If any of the new columns are in the wrong

position, use an ALTER COLUMN table CHANGE COLUMN statement for each one to move it

to the correct position.

Deleting or Replacing Columns

The following example removes all the existing columns and replaces them with the

new columns specified:

ALTER TABLE log_messages REPLACE COLUMNS (

hours_mins_secs INT COMMENT 'hour, minute, seconds from timestamp',

severity STRING COMMENT 'The message severity'

message STRING COMMENT 'The rest of the message');

This statement effectively renames the original hms column and removes the server and

process_id columns from the original schema definition. As for all ALTER statements,

only the table metadata is changed.

The REPLACE statement can only be used with tables that use one of the native SerDe

modules: DynamicSerDe or MetadataTypedColumnsetSerDe. Recall that the SerDe deter-

mines how records are parsed into columns (deserialization) and how a record’s col-

umns are written to storage (serialization). See Chapter 15 for more details on SerDes.

Alter Table Properties

You can add additional table properties or modify existing properties, but not remove

them:

ALTER TABLE log_messages SET TBLPROPERTIES (

'notes' = 'The process id is no longer captured; this column is always NULL');

Alter Storage Properties

There are several ALTER TABLE statements for modifying format and SerDe properties.

The following statement changes the storage format for a partition to be SEQUENCE

FILE, as we discussed in “Creating Tables” on page 53 (see “Sequence

Files” on page 148 and Chapter 15 for more information):

ALTER TABLE log_messages

PARTITION(year = 2012, month = 1, day = 1)

SET FILEFORMAT SEQUENCEFILE;

68 | Chapter 4: HiveQL: Data Definition

The PARTITION clause is required if the table is partitioned.

You can specify a new SerDe along with SerDe properties or change the properties for

the existing SerDe. The following example specifies that a table will use a Java class

named com.example.JSONSerDe to process a file of JSON-encoded records:

ALTER TABLE table_using_JSON_storage

SET SERDE 'com.example.JSONSerDe'

WITH SERDEPROPERTIES (

'prop1' = 'value1',

'prop2' = 'value2');

The SERDEPROPERTIES are passed to the SerDe module (the Java class com.example.JSON

SerDe, in this case). Note that both the property names (e.g., prop1) and the values (e.g.,

value1) must be quoted strings.

The SERDEPROPERTIES feature is a convenient mechanism that SerDe implementations

can exploit to permit user customization. We’ll see a real-world example of a JSON

SerDe and how it uses SERDEPROPERTIES in “JSON SerDe” on page 208.

The following example demonstrates how to add new SERDEPROPERTIES for the current

SerDe:

ALTER TABLE table_using_JSON_storage

SET SERDEPROPERTIES (

'prop3' = 'value3',

'prop4' = 'value4');

You can alter the storage properties that we discussed in “Creating Tables”

on page 53:

ALTER TABLE stocks

CLUSTERED BY (exchange, symbol)

SORTED BY (symbol)

INTO 48 BUCKETS;

The SORTED BY clause is optional, but the CLUSTER BY and INTO … BUCKETS are required.

(See also “Bucketing Table Data Storage” on page 125 for information on the use of

data bucketing.)

Miscellaneous Alter Table Statements

In “Execution Hooks” on page 158, we’ll discuss a technique for adding execution

“hooks” for various operations. The ALTER TABLE … TOUCH statement is used to trigger

these hooks:

ALTER TABLE log_messages TOUCH

PARTITION(year = 2012, month = 1, day = 1);

The PARTITION clause is required for partitioned tables. A typical scenario for this state-

ment is to trigger execution of the hooks when table storage files have been modified

outside of Hive. For example, a script that has just written new files for the 2012/01/01

partition for log_message can make the following call to the Hive CLI:

Alter Table | 69

hive -e 'ALTER TABLE log_messages TOUCH PARTITION(year = 2012, month = 1, day = 1);'

This statement won’t create the table or partition if it doesn’t already exist. Use the

appropriate creation commands in that case.

The ALTER TABLE … ARCHIVE PARTITION statement captures the partition files into a Ha-

doop archive (HAR) file. This only reduces the number of files in the filesystem, re-

ducing the load on the NameNode, but doesn’t provide any space savings (e.g., through

compression):

ALTER TABLE log_messages ARCHIVE

PARTITION(year = 2012, month = 1, day = 1);

To reverse the operation, substitute UNARCHIVE for ARCHIVE. This feature is only available

for individual partitions of partitioned tables.

Finally, various protections are available. The following statements prevent the parti-

tion from being dropped and queried:

ALTER TABLE log_messages

PARTITION(year = 2012, month = 1, day = 1) ENABLE NO_DROP;

ALTER TABLE log_messages

PARTITION(year = 2012, month = 1, day = 1) ENABLE OFFLINE;

To reverse either operation, replace ENABLE with DISABLE. These operations also can’t

be used with nonpartitioned tables.

70 | Chapter 4: HiveQL: Data Definition

CHAPTER 5

HiveQL: Data Manipulation

This chapter continues our discussion of HiveQL, the Hive query language, focusing

on the data manipulation language parts that are used to put data into tables and to

extract data from tables to the filesystem.

This chapter uses SELECT ... WHERE clauses extensively when we discuss populating

tables with data queried from other tables. So, why aren’t we covering SELECT ...

WHERE clauses first, instead of waiting until the next chapter, Chapter 6?

Since we just finished discussing how to create tables, we wanted to cover the next

obvious topic: how to get data into these tables so you’ll have something to query! We

assume you already understand the basics of SQL, so these clauses won’t be new to

you. If they are, please refer to Chapter 6 for details.

Loading Data into Managed Tables

Since Hive has no row-level insert, update, and delete operations, the only way to put

data into an table is to use one of the “bulk” load operations. Or you can just write files

in the correct directories by other means.

We saw an example of how to load data into a managed table in “Partitioned, Managed

Tables” on page 58, which we repeat here with an addition, the use of the OVERWRITE

keyword:

LOAD DATA LOCAL INPATH '${env:HOME}/california-employees'

OVERWRITE INTO TABLE employees

PARTITION (country = 'US', state = 'CA');

This command will first create the directory for the partition, if it doesn’t already exist,

then copy the data to it.

If the target table is not partitioned, you omit the PARTITION clause.

It is conventional practice to specify a path that is a directory, rather than an individual

file. Hive will copy all the files in the directory, which give you the flexibility of organ-

izing the data into multiple files and changing the file naming convention, without

requiring a change to your Hive scripts. Either way, the files will be copied to the ap-

propriate location for the table and the names will be the same.

If the LOCAL keyword is used, the path is assumed to be in the local filesystem. The data

is copied into the final location. If LOCAL is omitted, the path is assumed to be in the

distributed filesystem. In this case, the data is moved from the path to the final location.

LOAD DATA LOCAL ... copies the local data to the final location in the

distributed filesystem, while LOAD DATA ... (i.e., without LOCAL) moves

the data to the final location.

The rationale for this inconsistency is the assumption that you usually don’t want

duplicate copies of your data files in the distributed filesystem.

Also, because files are moved in this case, Hive requires the source and target files and

directories to be in the same filesystem. For example, you can’t use LOAD DATA to load

(move) data from one HDFS cluster to another.

It is more robust to specify a full path, but relative paths can be used. When running

in local mode, the relative path is interpreted relative to the user’s working directory

when the Hive CLI was started. For distributed or pseudo-distributed mode, the path

is interpreted relative to the user’s home directory in the distributed filesystem, which

is /user/$USER by default in HDFS and MapRFS.

If you specify the OVERWRITE keyword, any data already present in the target directory

will be deleted first. Without the keyword, the new files are simply added to the target

directory. However, if files already exist in the target directory that match filenames

being loaded, the old files are overwritten.

Versions of Hive before v0.9.0 had the following bug: when the OVER

WRITE keyword was not used, an existing data file in the target directory

would be overwritten if its name matched the name of a data file being

written to the directory. Hence, data would be lost. This bug was fixed

in the v0.9.0 release.

The PARTITION clause is required if the table is partitioned and you must specify a value

for each partition key.

In the example, the data will now exist in the following directory:

hdfs://master_server/user/hive/warehouse/mydb.db/employees/country=US/state=CA

Another limit on the file path used, the INPATH clause, is that it cannot contain any

directories.

72 | Chapter 5: HiveQL: Data Manipulation

Hive does not verify that the data you are loading matches the schema for the table.

However, it will verify that the file format matches the table definition. For example,

if the table was created with SEQUENCEFILE storage, the loaded files must be sequence

files.

Inserting Data into Tables from Queries

The INSERT statement lets you load data into a table from a query. Reusing our employ

ees example from the previous chapter, here is an example for the state of Oregon,

where we presume the data is already in another table called staged_employees. For

reasons we’ll discuss shortly, let’s use different names for the country and state fields

in staged_employees, calling them cnty and st, respectively:

INSERT OVERWRITE TABLE employees

PARTITION (country = 'US', state = 'OR')

SELECT * FROM staged_employees se

WHERE se.cnty = 'US' AND se.st = 'OR';

With OVERWRITE, any previous contents of the partition (or whole table if not parti-

tioned) are replaced.

If you drop the keyword OVERWRITE or replace it with INTO, Hive appends the data rather

than replaces it. This feature is only available in Hive v0.8.0 or later.

This example suggests one common scenario where this feature is useful: data has been

staged in a directory, exposed to Hive as an external table, and now you want to put it

into the final, partitioned table. A workflow like this is also useful if you want the target

table to have a different record format than the source table (e.g., a different field de-

limiter).

However, if staged_employees is very large and you run 65 of these statements to cover

all states, then it means you are scanning staged_employees 65 times! Hive offers an

alternative INSERT syntax that allows you to scan the input data once and split it multiple

ways. The following example shows this feature for creating the employees partitions

for three states:

FROM staged_employees se

INSERT OVERWRITE TABLE employees

PARTITION (country = 'US', state = 'OR')

SELECT * WHERE se.cnty = 'US' AND se.st = 'OR'

INSERT OVERWRITE TABLE employees

PARTITION (country = 'US', state = 'CA')

SELECT * WHERE se.cnty = 'US' AND se.st = 'CA'

INSERT OVERWRITE TABLE employees

PARTITION (country = 'US', state = 'IL')

SELECT * WHERE se.cnty = 'US' AND se.st = 'IL';

We have used indentation to make it clearer how the clauses group together. Each

record read from staged_employees will be evaluated with each SELECT … WHERE … clause.

Those clauses are evaluated independently; this is not an IF … THEN … ELSE … construct!

Inserting Data into Tables from Queries | 73

In fact, by using this construct, some records from the source table can be written to

multiple partitions of the destination table or none of them.

If a record satisfied a given SELECT … WHERE … clause, it gets written to the specified table

and partition. To be clear, each INSERT clause can insert into a different table, when

desired, and some of those tables could be partitioned while others aren’t.

Hence, some records from the input might get written to multiple output locations and

others might get dropped!

You can mix INSERT OVERWRITE clauses and INSERT INTO clauses, as well.

Dynamic Partition Inserts

There’s still one problem with this syntax: if you have a lot of partitions to create, you

have to write a lot of SQL! Fortunately, Hive also supports a dynamic partition feature,

where it can infer the partitions to create based on query parameters. By comparison,

up until now we have considered only static partitions.

Consider this change to the previous example:

INSERT OVERWRITE TABLE employees

PARTITION (country, state)

SELECT ..., se.cnty, se.st

FROM staged_employees se;

Hive determines the values of the partition keys, country and state, from the last two

columns in the SELECT clause. This is why we used different names in staged_employ

ees, to emphasize that the relationship between the source column values and the out-

put partition values is by position only and not by matching on names.

Suppose that staged_employees has data for a total of 100 country and state pairs. After

running this query, employees will have 100 partitions!

You can also mix dynamic and static partitions. This variation of the previous query

specifies a static value for the country (US) and a dynamic value for the state:

INSERT OVERWRITE TABLE employees

PARTITION (country = 'US', state)

SELECT ..., se.cnty, se.st

FROM staged_employees se

WHERE se.cnty = 'US';

The static partition keys must come before the dynamic partition keys.

Dynamic partitioning is not enabled by default. When it is enabled, it works in “strict”

mode by default, where it expects at least some columns to be static. This helps protect

against a badly designed query that generates a gigantic number of partitions. For ex-

ample, you partition by timestamp and generate a separate partition for each second!

Perhaps you meant to partition by day or maybe hour instead. Several other properties

are also used to limit excess resource utilization. Table 5-1 describes these properties.

74 | Chapter 5: HiveQL: Data Manipulation

Table 5-1. Dynamic partitions properties

Name Default Description

hive.exec.dynamic.parti

tion

false Set to true to enable dynamic partitioning.

hive.exec.dynamic.parti

tion.mode

strict Set to nonstrict to enable all partitions to be determined

dynamically.

hive.exec.max.dynamic.par

titions.pernode

100 The maximum number of dynamic partitions that can be cre-

ated by each mapper or reducer. Raises a fatal error if one

mapper or reducer attempts to create more than the threshold.

hive.exec.max.dynamic.par

titions

+1000 The total number of dynamic partitions that can be created by

one statement with dynamic partitioning. Raises a fatal error

if the limit is exceeded.

hive.exec.max.cre

ated.files

100000 The maximum total number of files that can be created globally.

A Hadoop counter is used to track the number of files created.

Raises a fatal error if the limit is exceeded.

So, for example, our first example using dynamic partitioning for all partitions might

actually look this, where we set the desired properties just before use:

hive> set hive.exec.dynamic.partition=true;

hive> set hive.exec.dynamic.partition.mode=nonstrict;

hive> set hive.exec.max.dynamic.partitions.pernode=1000;

hive> INSERT OVERWRITE TABLE employees

> PARTITION (country, state)

> SELECT ..., se.cty, se.st

> FROM staged_employees se;

Creating Tables and Loading Them in One Query

You can also create a table and insert query results into it in one statement:

CREATE TABLE ca_employees

AS SELECT name, salary, address

FROM employees

WHERE se.state = 'CA';

This table contains just the name, salary, and address columns from the employee table

records for employees in California. The schema for the new table is taken from the

SELECT clause.

A common use for this feature is to extract a convenient subset of data from a larger,

more unwieldy table.

This feature can’t be used with external tables. Recall that “populating” a partition for

an external table is done with an ALTER TABLE statement, where we aren’t “loading”

data, per se, but pointing metadata to a location where the data can be found.

Creating Tables and Loading Them in One Query | 75

Downloa d f r o m W o w ! e B o o k < w w w.woweb o o k . c o m >

Exporting Data

How do we get data out of tables? If the data files are already formatted the way you

want, then it’s simple enough to copy the directories or files:

hadoop fs -cp source_path target_path

Otherwise, you can use INSERT … DIRECTORY …, as in this example:

INSERT OVERWRITE LOCAL DIRECTORY '/tmp/ca_employees'

SELECT name, salary, address

FROM employees

WHERE se.state = 'CA';

OVERWRITE and LOCAL have the same interpretations as before and paths are interpreted

following the usual rules. One or more files will be written to /tmp/ca_employees,

depending on the number of reducers invoked.

The specified path can also be a full URI (e.g., hdfs://master-server/tmp/ca_employees).

Independent of how the data is actually stored in the source table, it is written to files

with all fields serialized as strings. Hive uses the same encoding in the generated output

files as it uses for the tables internal storage.

As a reminder, we can look at the results from within the hive CLI:

hive> ! ls /tmp/ca_employees;

000000_0

hive> ! cat /tmp/payroll/000000_0

John Doe100000.0201 San Antonio CircleMountain ViewCA94040

Mary Smith80000.01 Infinity LoopCupertinoCA95014

...

Yes, the filename is 000000_0. If there were two or more reducers writing output, we

would have additional files with similar names (e.g., 000001_0).

The fields appear to be joined together without delimiters because the ^A and ^B

separators aren’t rendered.

Just like inserting data to tables, you can specify multiple inserts to directories:

FROM staged_employees se

INSERT OVERWRITE DIRECTORY '/tmp/or_employees'

SELECT * WHERE se.cty = 'US' and se.st = 'OR'

INSERT OVERWRITE DIRECTORY '/tmp/ca_employees'

SELECT * WHERE se.cty = 'US' and se.st = 'CA'

INSERT OVERWRITE DIRECTORY '/tmp/il_employees'

SELECT * WHERE se.cty = 'US' and se.st = 'IL';

There are some limited options for customizing the output of the data (other than

writing a custom OUTPUTFORMAT, as discussed in “Customizing Table Storage For-

mats” on page 63). To format columns, the built-in functions include those for

formatting strings, such as converting case, padding output, and more. See “Other

built-in functions” on page 88 for more details.

76 | Chapter 5: HiveQL: Data Manipulation

The field delimiter for the table can be problematic. For example, if it uses the default

^A delimiter. If you export table data frequently, it might be appropriate to use comma

or tab delimiters.

Another workaround is to define a “temporary” table with the storage configured to

match the desired output format (e.g., tab-delimited fields). Then write a query result

to that table and use INSERT OVERWRITE DIRECTORY, selecting from the temporary table.

Unlike many relational databases, there is no temporary table feature in Hive. You have

to manually drop any tables you create that aren’t intended to be permanent.

Exporting Data | 77

CHAPTER 6

HiveQL: Queries

After learning the many ways we can define and format tables, let’s learn how to run

queries. Of course, we have assumed all along that you have some prior knowledge of

SQL. We’ve used some queries already to illustrate several concepts, such as loading

query data into other tables in Chapter 5. Now we’ll fill in most of the details. Some

special topics will be covered in subsequent chapters.

We’ll move quickly through details that are familiar to users with prior SQL experience

and focus on what’s unique to HiveQL, including syntax and feature differences, as

well as performance implications.

SELECT … FROM Clauses

SELECT is the projection operator in SQL. The FROM clause identifies from which table,

view, or nested query we select records (see Chapter 7).

For a given record, SELECT specifies the columns to keep, as well as the outputs of

function calls on one or more columns (e.g., the aggregation functions like count(*)).

Recall again our partitioned employees table:

CREATE TABLE employees (

name STRING,

salary FLOAT,

subordinates ARRAY<STRING>,

deductions MAP<STRING, FLOAT>,

address STRUCT<street:STRING, city:STRING, state:STRING, zip:INT>

)

PARTITIONED BY (country STRING, state STRING);

Let’s assume we have the same contents we showed in “Text File Encoding of Data

Values” on page 45 for four employees in the US state of Illinois (abbreviated IL). Here

are queries of this table and the output they produce:

hive> SELECT name, salary FROM employees;

John Doe 100000.0

Mary Smith 80000.0

Todd Jones 70000.0

Bill King 60000.0

The following two queries are identical. The second version uses a table alias e, which

is not very useful in this query, but becomes necessary in queries with JOINs (see “JOIN

Statements” on page 98) where several different tables are used:

hive> SELECT name, salary FROM employees;

hive> SELECT e.name, e.salary FROM employees e;

When you select columns that are one of the collection types, Hive uses JSON (Java-

Script Object Notation) syntax for the output. First, let’s select the subordinates, an

ARRAY, where a comma-separated list surrounded with […] is used. Note that STRING

elements of the collection are quoted, while the primitive STRING name column is not:

hive> SELECT name, subordinates FROM employees;

John Doe ["Mary Smith","Todd Jones"]

Mary Smith ["Bill King"]

Todd Jones []

Bill King []

The deductions is a MAP, where the JSON representation for maps is used, namely a

comma-separated list of key:value pairs, surrounded with {...}:

hive> SELECT name, deductions FROM employees;

John Doe {"Federal Taxes":0.2,"State Taxes":0.05,"Insurance":0.1}

Mary Smith {"Federal Taxes":0.2,"State Taxes":0.05,"Insurance":0.1}

Todd Jones {"Federal Taxes":0.15,"State Taxes":0.03,"Insurance":0.1}

Bill King {"Federal Taxes":0.15,"State Taxes":0.03,"Insurance":0.1}

Finally, the address is a STRUCT, which is also written using the JSON map format:

hive> SELECT name, address FROM employees;

John Doe {"street":"1 Michigan Ave.","city":"Chicago","state":"IL","zip":60600}

Mary Smith {"street":"100 Ontario St.","city":"Chicago","state":"IL","zip":60601}

Todd Jones {"street":"200 Chicago Ave.","city":"Oak Park","state":"IL","zip":60700}

Bill King {"street":"300 Obscure Dr.","city":"Obscuria","state":"IL","zip":60100}

Next, let’s see how to reference elements of collections.

First, ARRAY indexing is 0-based, as in Java. Here is a query that selects the first element

of the subordinates array:

hive> SELECT name, subordinates[0] FROM employees;

John Doe Mary Smith

Mary Smith Bill King

Todd Jones NULL

Bill King NULL

Note that referencing a nonexistent element returns NULL. Also, the extracted STRING

values are no longer quoted!

To reference a MAP element, you also use ARRAY[...] syntax, but with key values instead

of integer indices:

hive> SELECT name, deductions["State Taxes"] FROM employees;

John Doe 0.05

80 | Chapter 6: HiveQL: Queries

Mary Smith 0.05

Todd Jones 0.03

Bill King 0.03

Finally, to reference an element in a STRUCT, you use “dot” notation, similar to the

table_alias.column mentioned above:

hive> SELECT name, address.city FROM employees;

John Doe Chicago

Mary Smith Chicago

Todd Jones Oak Park

Bill King Obscuria

These same referencing techniques are also used in WHERE clauses, which we discuss in

“WHERE Clauses” on page 92.

Specify Columns with Regular Expressions

We can even use regular expressions to select the columns we want. The following query

selects the symbol column and all columns from stocks whose names start with the

prefix price:1

hive> SELECT symbol, `price.*` FROM stocks;

AAPL 195.69 197.88 194.0 194.12 194.12

AAPL 192.63 196.0 190.85 195.46 195.46

AAPL 196.73 198.37 191.57 192.05 192.05

AAPL 195.17 200.2 194.42 199.23 199.23

AAPL 195.91 196.32 193.38 195.86 195.86

...

We’ll talk more about Hive’s use of regular expressions in the section “LIKE and

RLIKE” on page 96.

Computing with Column Values

Not only can you select columns in a table, but you can manipulate column values

using function calls and arithmetic expressions.

For example, let’s select the employees’ names converted to uppercase, their salaries,

federal taxes percentage, and the value that results if we subtract the federal taxes por-

tion from their salaries and round to the nearest integer. We could call a built-in func-

tion map_values to extract all the values from the deductions map and then add them

up with the built-in sum function.

The following query is long enough that we’ll split it over two lines. Note the secondary

prompt that Hive uses, an indented greater-than sign (>):

hive> SELECT upper(name), salary, deductions["Federal Taxes"],

> round(salary * (1 - deductions["Federal Taxes"])) FROM employees;

1. At the time of this writing, the Hive Wiki shows an incorrect syntax for specifying columns using regular

expressions.

SELECT … FROM Clauses | 81

JOHN DOE 100000.0 0.2 80000

MARY SMITH 80000.0 0.2 64000

TODD JONES 70000.0 0.15 59500

BILL KING 60000.0 0.15 51000

Let’s discuss arithmetic operators and then discuss the use of functions in expressions.

Arithmetic Operators

All the typical arithmetic operators are supported. Table 6-1 describes the specific

details.

Table 6-1. Arithmetic operators

Operator Types Description

A + B Numbers Add A and B.

A - B Numbers Subtract B from A.

A * B Numbers Multiply A and B.

A / B Numbers Divide A with B. If the operands are in-

teger types, the quotient of the division

is returned.

A % B Numbers The remainder of dividing A with B.

A & B Numbers Bitwise AND of A and B.

A | B Numbers Bitwise OR of A and B.

A ^ B Numbers Bitwise XOR of A and B.

~A Numbers Bitwise NOT of A.

Arithmetic operators take any numeric type. No type coercion is performed if the two

operands are of the same numeric type. Otherwise, if the types differ, then the value of

the smaller of the two types is promoted to wider type of the other value. (Wider in the

sense that a type with more bytes can hold a wider range of values.) For example, for

INT and BIGINT operands, the INT is promoted to BIGINT. For INT and FLOAT operands,

the INT is promoted to FLOAT. Note that our query contained (1 - deductions[…]). Since

the deductions are FLOATS, the 1 was promoted to FLOAT.

You have to be careful about data overflow or underflow when doing arithmetic. Hive

follows the rules for the underlying Java types, where no attempt is made to automat-

ically convert a result to a wider type if one exists, when overflow or underflow will

occur. Multiplication and division are most likely to trigger this problem.

It pays to be aware of the ranges of your numeric data values, whether or not those

values approach the upper or lower range limits of the types you are using in the cor-

responding schema, and what kinds of calculations people might do with the data.

If you are concerned about overflow or underflow, consider using wider types in the

schema. The drawback is the extra memory each data value will occupy.

82 | Chapter 6: HiveQL: Queries

You can also convert values to wider types in specific expressions, called casting. See

Table 6-2 below and “Casting” on page 109 for details.

Finally, it is sometimes useful to scale data values, such as dividing by powers of 10,

using log values, and so on. Scaling can also improve the accuracy and numerical sta-

bility of algorithms used in certain machine learning calculations, for example.

Using Functions

Our tax-deduction example also uses a built-in mathematical function, round(), for

finding the nearest integer for a DOUBLE value.

Mathematical functions

Table 6-2 describes the built-in mathematical functions, as of Hive v0.8.0, for working

with single columns of data.

Table 6-2. Mathematical functions

Return type Signature Description

BIGINT round(d) Return the BIGINT for the rounded value of DOUBLE d.

DOUBLE round(d, N) Return the DOUBLE for the value of d, a DOUBLE, rounded to

N decimal places.

BIGINT floor(d) Return the largest BIGINT that is <= d, a DOUBLE.

BIGINT ceil(d), ceiling(DOUBLE d) Return the smallest BIGINT that is >= d.

DOUBLE rand(), rand(seed) Return a pseudorandom DOUBLE that changes for each row.

Passing in an integer seed makes the return value

deterministic.

DOUBLE exp(d) Return e to the d, a DOUBLE.

DOUBLE ln(d) Return the natural logarithm of d, a DOUBLE.

DOUBLE log10(d) Return the base-10 logarithm of d, a DOUBLE.

DOUBLE log2(d) Return the base-2 logarithm of d, a DOUBLE.

DOUBLE log(base, d) Return the base-base logarithm of d, where base and d are

DOUBLEs.

DOUBLE pow(d, p), power(d, p) Return d raised to the power p, where d and p are DOUBLEs.

DOUBLE sqrt(d) Return the square root of d, a DOUBLE.

STRING bin(i) Return the STRING representing the binary value of i, a

BIGINT.

STRING hex(i) Return the STRING representing the hexadecimal value of i, a

BIGINT.

STRING hex(str) Return the STRING representing the hexadecimal value of s,

where each two characters in the STRING s is converted to its

hexadecimal representation.

SELECT … FROM Clauses | 83

Return type Signature Description

STRING unhex(i) The inverse of hex(str).

STRING conv(i, from_base, to_base) Return the STRING in base to_base, an INT, representing the

value of i, a BIGINT, in base from_base, an INT.

STRING conv(str, from_base,

to_base)

Return the STRING in base to_base, an INT, representing the

value of str, a STRING, in base from_base, an INT.

DOUBLE abs(d) Return the DOUBLE that is the absolute value of d, a DOUBLE.

INT pmod(i1, i2) Return the positive module INT for two INTs, i1 mod i2.

DOUBLE pmod(d1, d2) Return the positive module DOUBLE for two DOUBLEs, d1 mod

d2.

DOUBLE sin(d) Return the DOUBLE that is the sin of d, a DOUBLE, in radians.

DOUBLE asin(d) Return the DOUBLE that is the arcsin of d, a DOUBLE, in radians.

DOUBLE cos(d) Return the DOUBLE that is the cosine of d, a DOUBLE, in radians.

DOUBLE acos(d) Return the DOUBLE that is the arccosine of d, a DOUBLE, in

radians.

DOUBLE tan(d) Return the DOUBLE that is the tangent of d, a DOUBLE, in radians.

DOUBLE atan(d) Return the DOUBLE that is the arctangent of d, a DOUBLE, in

radians.

DOUBLE degrees(d) Return the DOUBLE that is the value of d, a DOUBLE, converted

from radians to degrees.

DOUBLE radians(d) Return the DOUBLE that is the value of d, a DOUBLE, converted

from degrees to radians.

INT positive(i) Return the INT value of i (i.e., it’s effectively the expression \

+i).

DOUBLE positive(d) Return the DOUBLE value of d (i.e., it’s effectively the expression

\+d).

INT negative(i) Return the negative of the INT value of i (i.e., it’s effectively the

expression -i).

DOUBLE negative(d) Return the negative of the DOUBLE value of d; effectively, the

expression -d.

FLOAT sign(d) Return the FLOAT value 1.0 if d, a DOUBLE, is positive; return

the FLOAT value -1.0 if d is negative; otherwise return 0.0.

DOUBLE e() Return the DOUBLE that is the value of the constant e,

2.718281828459045.

DOUBLE pi() Return the DOUBLE that is the value of the constant pi,

3.141592653589793.

Note the functions floor, round, and ceil (“ceiling”) for converting DOUBLE to BIGINT,

which is floating-point numbers to integer numbers. These functions are the preferred

technique, rather than using the cast operator we mentioned above.

84 | Chapter 6: HiveQL: Queries

Also, there are functions for converting integers to strings in different bases (e.g.,

hexadecimal).

Aggregate functions

A special kind of function is the aggregate function that returns a single value resulting

from some computation over many rows. More precisely, this is the User Defined Ag-

gregate Function, as we’ll see in “Aggregate Functions” on page 164. Perhaps the two

best known examples are count, which counts the number of rows (or values for a

specific column), and avg, which returns the average value of the specified column

values.

Here is a query that counts the number of our example employees and averages their

salaries:

hive> SELECT count(*), avg(salary) FROM employees;

4 77500.0

We’ll see other examples when we discuss GROUP BY in the section “GROUP BY Clau-

ses” on page 97.

Table 6-3 lists Hive’s built-in aggregate functions.

Table 6-3. Aggregate functions

Return type Signature Description

BIGINT count(*) Return the total number of retrieved rows, including rows

containing NULL values.

BIGINT count(expr) Return the number of rows for which the supplied

expression is not NULL.

BIGINT count(DISTINCT expr[, expr_.]) Return the number of rows for which the supplied

expression(s) are unique and not NULL.

DOUBLE sum(col) Return the sum of the values.

DOUBLE sum(DISTINCT col) Return the sum of the distinct values.

DOUBLE avg(col) Return the average of the values.

DOUBLE avg(DISTINCT col) Return the average of the distinct values.

DOUBLE min(col) Return the minimum value of the values.

DOUBLE max(col) Return the maximum value of the values.

DOUBLE variance(col), var_pop(col) Return the variance of a set of numbers in a collection:

col.

DOUBLE var_samp(col) Return the sample variance of a set of numbers.

DOUBLE stddev_pop(col) Return the standard deviation of a set of numbers.

DOUBLE stddev_samp(col) Return the sample standard deviation of a set of numbers.

DOUBLE covar_pop(col1, col2) Return the covariance of a set of numbers.

DOUBLE covar_samp(col1, col2) Return the sample covariance of a set of numbers.

SELECT … FROM Clauses | 85

Return type Signature Description

DOUBLE corr(col1, col2) Return the correlation of two sets of numbers.

DOUBLE percentile(int_expr, p) Return the percentile of int_expr at p (range: [0,1]),

where p is a DOUBLE.

ARRAY<DOU

BLE>

percentile(int_expr,

[p1, ...])

Return the percentiles of int_expr at p (range: [0,1]),

where p is a DOUBLE array.

DOUBLE percentile_approx(int_expr,

p , NB)

Return the approximate percentiles of int_expr at p

(range: [0,1]), where p is a DOUBLE and NB is the number

of histogram bins for estimating (default: 10,000 if not

specified).

DOUBLE percentile_approx(int_expr,

[p1, ...] , NB)

Return the approximate percentiles of int_expr at p

(range: [0,1]), where p is a DOUBLE array and NB is the

number of histogram bins for estimating (default: 10,000

if not specified).

ARRAY<STRUCT

{'x','y'}>

histogram_numeric(col, NB) Return an array of NB histogram bins, where the x value

is the center and the y value is the height of the bin.

ARRAY collect_set(col) Return a set with the duplicate elements from collection

col removed.

You can usually improve the performance of aggregation by setting the following prop-

erty to true, hive.map.aggr, as shown here:

hive> SET hive.map.aggr=true;

hive> SELECT count(*), avg(salary) FROM employees;

This setting will attempt to do “top-level” aggregation in the map phase, as in this

example. (An aggregation that isn’t top-level would be aggregation after performing a

GROUP BY.) However, this setting will require more memory.

As Table 6-3 shows, several functions accept DISTINCT … expressions. For example, we

could count the unique stock symbols this way:

hive> SELECT count(DISTINCT symbol) FROM stocks;

Wait, zero?? There is a bug when trying to use count(DISTINCT col)

when col is a partition column. The answer should be 743 for NASDAQ

and NYSE, at least as of early 2010 in the infochimps.org data set we

used.

Note that the Hive wiki currently claims that you can’t use more than one function(DIS

TINCT …) expression in a query. For example, the following is supposed to be disallowed,

but it actually works:

hive> SELECT count(DISTINCT ymd), count(DISTINCT volume) FROM stocks;

12110 26144

86 | Chapter 6: HiveQL: Queries

So, there are 12,110 trading days of data, over 40 years worth.

Table generating functions

The “inverse” of aggregate functions are so-called table generating functions, which take

single columns and expand them to multiple columns or rows. We will discuss them

extensively in “Table Generating Functions” on page 165, but to complete the contents

of this section, we will discuss them briefly now and list the few built-in table generating

functions available in Hive.

To explain by way of an example, the following query converts the subordinate array

in each employees record into zero or more new records. If an employee record has an

empty subordinates array, then no new records are generated. Otherwise, one new

record per subordinate is generated:

hive> SELECT explode(subordinates) AS sub FROM employees;

Mary Smith

Todd Jones

Bill King

We used a column alias, sub, defined using the AS sub clause. When using table gen-

erating functions, column aliases are required by Hive. There are many other particular

details that you must understand to use these functions correctly. We’ll wait until

“Table Generating Functions” on page 165 to discuss the details.

Table 6-4 lists the built-in table generating functions.

Table 6-4. Table generating functions

Return type Signature Description

N rows explode(array) Return 0 to many rows, one row for each element from

the input array.

N rows explode(map) (v0.8.0 and later) Return 0 to many rows, one row for each

map key-value pair, with a field for each map key and a

field for the map value.

tuple json_tuple(jsonStr, p1, p2, …,

pn)

Like get_json_object, but it takes multiple names

and returns a tuple. All the input parameters and output

column types are STRING.

tuple parse_url_tuple(url, part

name1, partname2, …, partna

meN) where N >= 1

Extract N parts from a URL. It takes a URL and the part

names to extract, returning a tuple. All the input pa-

rameters and output column types are STRING. The valid

partnames are case-sensitive and should only contain

a minimum of white space: HOST, PATH, QUERY, REF,

PROTOCOL, AUTHORITY, FILE, USERINFO,

QUERY:<KEY_NAME>.

N rows stack(n, col1, …, colM) Convert M columns into N rows of size M/N each.

SELECT … FROM Clauses | 87

Here is an example that uses parse_url_tuple where we assume a url_table exists that

contains a column of URLs called url:

SELECT parse_url_tuple(url, 'HOST', 'PATH', 'QUERY') as (host, path, query)

FROM url_table;

Compare parse_url_tuple with parse_url in Table 6-5 below.

Other built-in functions

Table 6-5 describes the rest of the built-in functions for working with strings, maps,

arrays, JSON, and timestamps, with or without the recently introduced TIMESTAMP type

(see “Primitive Data Types” on page 41).

Table 6-5. Other built-in functions

Return type Signature Description

BOOLEAN test in(val1, val2, …) Return true if test equals one of the values in the list.

INT length(s) Return the length of the string.

STRING reverse(s) Return a reverse copy of the string.

STRING concat(s1, s2, …) Return the string resulting from s1 joined with s2, etc.

For example, concat('ab', 'cd') results in

'abcd'. You can pass an arbitrary number of string ar-

guments and the result will contain all of them joined

together.

STRING concat_ws(separator, s1, s2,

…)

Like concat, but using the specified separator.

STRING substr(s, start_index) Return the substring of s starting from the

start_index position, where 1 is the index of the first

character, until the end of s. For example,

substr('abcd', 3) results in 'cd'.

STRING substr(s, int start, int

length)

Return the substring of s starting from the start posi-

tion with the given length, e.g., substr('abc

defgh', 3, 2) results in 'cd'.

STRING upper(s) Return the string that results from converting all charac-

ters of s to upper case, e.g., upper('hIvE') results in

'HIVE'.

STRING ucase(s) A synonym for upper().

STRING lower(s) Return the string that results from converting all charac-

ters of s to lower case, e.g., lower('hIvE') results in

'hive'.

STRING lcase(s) A synonym for lower().

STRING trim(s) Return the string that results from removing whitespace

from both ends of s, e.g., trim(' hive ') results in

'hive'.

88 | Chapter 6: HiveQL: Queries

Return type Signature Description

STRING ltrim(s) Return the string resulting from trimming spaces from

the beginning (lefthand side) of s, e.g., ltrim(' hive

') results in 'hive '.

STRING rtrim(s) Return the string resulting from trimming spaces from

the end (righthand side) of s, e.g., rtrim(' hive

') results in ' hive'.

STRING regexp_replace(s, regex,

replacement)

Return the string resulting from replacing all substrings

in s that match the Java regular expression re with

replacement.a If replacement is blank, the

matches are effectively deleted, e.g.,

regexp_replace('hive', '[ie]', 'z')

returns 'hzvz'.

STRING regexp_extract(subject,

regex_pattern, index)

Returns the substring for the index’s match using the

regex_pattern.

STRING parse_url(url, partname, key) Extracts the specified part from a URL. It takes a URL and

the partname to extract. The valid partnames are

case-sensitive: HOST, PATH, QUERY, REF, PROTOCOL,

AUTHORITY, FILE, USERINFO, QUERY:<key>. The

optional key is used for the last QUERY:<key> request.

Compare with parse_url_tuple described in Ta-

ble 6-4.

int size(map<K.V>) Return the number of elements in the map.

int size(array<T>) Return the number of elements in the array.

value of type cast(<expr> as <type>) Convert (“cast”) the result of the expression expr to

type, e.g., cast('1' as BIGINT) will convert the

string '1' to its integral representation. A NULL is re-

turned if the conversion does not succeed.

STRING from_unixtime(int unixtime) Convert the number of seconds from the Unix epoch

(1970-01-01 00:00:00 UTC) to a string representing the

timestamp of that moment in the current system time

zone in the format of '1970-01-01 00:00:00'.

STRING to_date(timestamp) Return the date part of a timestamp string, e.g.,

to_date("1970-01-01 00:00:00") returns

'1970-01-01'.

INT year(timestamp) Return the year part as an INT of a timestamp string, e.g.,

year("1970-11-01 00:00:00") returns 1970.

INT month(timestamp) Return the month part as an INT of a timestamp string,

e.g., month("1970-11-01 00:00:00") returns

11.

INT day(timestamp) Return the day part as an INT of a timestamp string, e.g.,

day("1970-11-01 00:00:00") returns 1.

STRING get_json_object(json_string,

path)

Extract the JSON object from a JSON string based on the

given JSON path, and return the JSON string of the

SELECT … FROM Clauses | 89

Return type Signature Description

extracted object. NULL is returned if the input JSON string

is invalid.

STRING space(n) Returns n spaces.

STRING repeat(s, n) Repeats s n times.

STRING ascii(s) Returns the integer value for the first ASCII character in

the string s.

STRING lpad(s, len, pad) Returns s exactly len length, prepending instances of

the string pad on its left, if necessary, to reach len char-

acters. If s is longer than len, it is truncated.

STRING rpad(s, len, pad) Returns s exactly len length, appending instances of the

string pad on its right, if necessary, to reach len char-

acters. If s is longer than len, it is truncated.

ARRAY<STRING

split(s, pattern) Returns an array of substrings of s, split on occurrences

of pattern.

INT find_in_set(s, commaSeparated

String)

Returns the index of the comma-separated string where

s is found, or NULL if it is not found.

INT locate(substr, str, pos]) Returns the index of str after pos where substr is

found.

INT instr(str, substr) Returns the index of str where substr is found.

MAP<STRING,S

TRING>

str_to_map(s, delim1, delim2) Creates a map by parsing s, using delim1 as the sepa-

rator between key-value pairs and delim2 as the key-

value separator.

ARRAY<ARRAY<

STRING>>

sentences(s, lang, locale) Splits s into arrays of sentences, where each sentence is

an array of words. The lang and country arguments

are optional; if omitted, the default locale is used.

ARRAY<STRUCT

<STRING,DOU

BLE>>

ngrams(array<array<string>>,

N, K, pf)

Estimates the top-K n-grams in the text. pf is the precision

factor.

ARRAY<STRUCT

<STRING,DOU

BLE>>

con

text_ngrams(array<array<strin

g>>,array<string>,int K, int

pf)

Like ngrams, but looks for n-grams that begin with the

second array of words in each outer array.

BOOLEAN in_file(s, filename) Returns true if s appears in the file named filename.

aSee http://docs.oracle.com/javase/tutorial/essential/regex/ for more on Java regular expression syntax.

Note that the time-related functions (near the end of the table) take integer or string

arguments. As of Hive v0.8.0, these functions also take TIMESTAMP arguments, but they

will continue to take integer or string arguments for backwards compatibility.

90 | Chapter 6: HiveQL: Queries

LIMIT Clause

The results of a typical query can return a large number of rows. The LIMIT clause puts

an upper limit on the number of rows returned:

hive> SELECT upper(name), salary, deductions["Federal Taxes"],

> round(salary * (1 - deductions["Federal Taxes"])) FROM employees

> LIMIT 2;

JOHN DOE 100000.0 0.2 80000

MARY SMITH 80000.0 0.2 64000

Column Aliases

You can think of the previous example query as returning a new relation with new

columns, some of which are anonymous results of manipulating columns in

employees. It’s sometimes useful to give those anonymous columns a name, called a

column alias. Here is the previous query with column aliases for the third and fourth

columns returned by the query, fed_taxes and salary_minus_fed_taxes, respectively:

hive> SELECT upper(name), salary, deductions["Federal Taxes"] as fed_taxes,

> round(salary * (1 - deductions["Federal Taxes"])) as salary_minus_fed_taxes

> FROM employees LIMIT 2;

JOHN DOE 100000.0 0.2 80000

MARY SMITH 80000.0 0.2 64000

Nested SELECT Statements

The column alias feature is especially useful in nested select statements. Let’s use the

previous example as a nested query:

hive> FROM (

> SELECT upper(name), salary, deductions["Federal Taxes"] as fed_taxes,

> round(salary * (1 - deductions["Federal Taxes"])) as salary_minus_fed_taxes

> FROM employees

> ) e

> SELECT e.name, e.salary_minus_fed_taxes

> WHERE e.salary_minus_fed_taxes > 70000;

JOHN DOE 100000.0 0.2 80000

The previous result set is aliased as e, from which we perform a second query to select

the name and the salary_minus_fed_taxes, where the latter is greater than 70,000. (We’ll

cover WHERE clauses in “WHERE Clauses” on page 92 below.)

CASE … WHEN … THEN Statements

The CASE … WHEN … THEN clauses are like if statements for individual columns in query

results. For example:

hive> SELECT name, salary,

> CASE

> WHEN salary < 50000.0 THEN 'low'

SELECT … FROM Clauses | 91

> WHEN salary >= 50000.0 AND salary < 70000.0 THEN 'middle'

> WHEN salary >= 70000.0 AND salary < 100000.0 THEN 'high'

> ELSE 'very high'

> END AS bracket FROM employees;

John Doe 100000.0 very high

Mary Smith 80000.0 high

Todd Jones 70000.0 high

Bill King 60000.0 middle

Boss Man 200000.0 very high

Fred Finance 150000.0 very high

Stacy Accountant 60000.0 middle

...

When Hive Can Avoid MapReduce

If you have been running the queries in this book so far, you have probably noticed

that a MapReduce job is started in most cases. Hive implements some kinds of queries

without using MapReduce, in so-called local mode, for example:

SELECT * FROM employees;

In this case, Hive can simply read the records from employees and dump the formatted

output to the console.

This even works for WHERE clauses that only filter on partition keys, with or without

LIMIT clauses:

SELECT * FROM employees

WHERE country = 'US' AND state = 'CA'

LIMIT 100;

Furthermore, Hive will attempt to run other operations in local mode if the

hive.exec.mode.local.auto property is set to true:

set hive.exec.mode.local.auto=true;

Otherwise, Hive uses MapReduce to run all other queries.

Trust us, you want to add set hive.exec.mode.local.auto=true; to your

$HOME/.hiverc file.

WHERE Clauses

While SELECT clauses select columns, WHERE clauses are filters; they select which records

to return. Like SELECT clauses, we have already used many simple examples of WHERE

clauses before defining the clause, on the assumption you have seen them before. Now

we’ll explore them in a bit more detail.

92 | Chapter 6: HiveQL: Queries

WHERE clauses use predicate expressions, applying predicate operators, which we’ll de-

scribe in a moment, to columns. Several predicate expressions can be joined with AND

and OR clauses. When the predicate expressions evaluate to true, the corresponding

rows are retained in the output.

We just used the following example that restricts the results to employees in the state

of California:

SELECT * FROM employees

WHERE country = 'US' AND state = 'CA';

The predicates can reference the same variety of computations over column values that

can be used in SELECT clauses. Here we adapt our previously used query involving

Federal Taxes, filtering for those rows where the salary minus the federal taxes is greater

than 70,000:

hive> SELECT name, salary, deductions["Federal Taxes"],

> salary * (1 - deductions["Federal Taxes"])

> FROM employees

> WHERE round(salary * (1 - deductions["Federal Taxes"])) > 70000;

John Doe 100000.0 0.2 80000.0

This query is a bit ugly, because the complex expression on the second line is duplicated

in the WHERE clause. The following variation eliminates the duplication, using a column

alias, but unfortunately it’s not valid:

hive> SELECT name, salary, deductions["Federal Taxes"],

> salary * (1 - deductions["Federal Taxes"]) as salary_minus_fed_taxes

> FROM employees

> WHERE round(salary_minus_fed_taxes) > 70000;

FAILED: Error in semantic analysis: Line 4:13 Invalid table alias or

column reference 'salary_minus_fed_taxes': (possible column names are:

name, salary, subordinates, deductions, address)

As the error message says, we can’t reference column aliases in the WHERE clause. How-

ever, we can use a nested SELECT statement:

hive> SELECT e.* FROM

> (SELECT name, salary, deductions["Federal Taxes"] as ded,

> salary * (1 - deductions["Federal Taxes"]) as salary_minus_fed_taxes

> FROM employees) e

> WHERE round(e.salary_minus_fed_taxes) > 70000;

John Doe 100000.0 0.2 80000.0

Boss Man 200000.0 0.3 140000.0

Fred Finance 150000.0 0.3 105000.0

Predicate Operators

Table 6-6 describes the predicate operators, which are also used in JOIN … ON and

HAVING clauses.

WHERE Clauses | 93

Table 6-6. Predicate operators

Operator Types Description

A = B Primitive types True if A equals B. False otherwise.

A <> B, A != B Primitive types NULL if A or B is NULL; true if A is not equal to B; false

otherwise.

A < B Primitive types NULL if A or B is NULL; true if A is less than B; false

otherwise.

A <= B Primitive types NULL if A or B is NULL; true if A is less than or equal to

B; false otherwise.

A > B Primitive types NULL if A or B is NULL; true if A is greater than B; false

otherwise.

A >= B Primitive types NULL if A or B is NULL; true if A is greater than or equal

to B; false otherwise.

A IS NULL All types True if A evaluates to NULL; false otherwise.

A IS NOT NULL All types False if A evaluates to NULL; true otherwise.

A LIKE B String True if A matches the SQL simplified regular expression

specification given by B; false otherwise. B is interpreted

as follows: 'x%' means A must begin with the prefix 'x',

'%x' means A must end with the suffix 'x', and '%x

%' means A must begin with, end with, or contain the

substring 'x'. Similarly, the underscore '_' matches a

single character. B must match the whole string A.

A RLIKE B, A REGEXP B String True if A matches the regular expression given by B; false

otherwise. Matching is done by the JDK regular expression

library and hence it follows the rules of that library. For

example, the regular expression must match the entire

string A, not just a subset. See below for more information

about regular expressions.

We’ll discuss LIKE and RLIKE in detail below (“LIKE and RLIKE” on page 96). First,

let’s point out an issue with comparing floating-point numbers that you should

understand.

Gotchas with Floating-Point Comparisons

A common gotcha arises when you compare floating-point numbers of different types

(i.e., FLOAT versus DOUBLE). Consider the following query of the employees table, which

is designed to return the employee’s name, salary, and federal taxes deduction, but only

if that tax deduction exceeds 0.2 (20%) of his or her salary:

hive> SELECT name, salary, deductions['Federal Taxes']

> FROM employees WHERE deductions['Federal Taxes'] > 0.2;

John Doe 100000.0 0.2

Mary Smith 80000.0 0.2

94 | Chapter 6: HiveQL: Queries

Boss Man 200000.0 0.3

Fred Finance 150000.0 0.3

Wait! Why are records with deductions['Federal Taxes'] = 0.2 being returned?

Is it a Hive bug? There is a bug filed against Hive for this issue, but it actually reflects

the behavior of the internal representation of floating-point numbers when they are

compared and it affects almost all software written in most languages on all modern

digital computers (see https://issues.apache.org/jira/browse/HIVE-2586).

When you write a floating-point literal value like 0.2, Hive uses a DOUBLE to hold the

value. We defined the deductions map values to be FLOAT, which means that Hive will

implicitly convert the tax deduction value to DOUBLE to do the comparison. This should

work, right?

Actually, it doesn’t work. Here’s why. The number 0.2 can’t be represented exactly in

a FLOAT or DOUBLE. (See http://docs.oracle.com/cd/E19957-01/806-3568/ncg_goldberg

.html for an in-depth discussion of floating-point number issues.) In this particular case,

the closest exact value is just slightly greater than 0.2, with a few nonzero bits at the

least significant end of the number.

To simplify things a bit, let’s say that 0.2 is actually 0.2000001 for FLOAT and

0.200000000001 for DOUBLE, because an 8-byte DOUBLE has more significant digits (after

the decimal point). When the FLOAT value from the table is converted to DOUBLE by Hive,

it produces the DOUBLE value 0.200000100000, which is greater than 0.200000000001.

That’s why the query results appear to use >= not >!

This issue is not unique to Hive nor Java, in which Hive is implemented. Rather, it’s a

general problem for all systems that use the IEEE standard for encoding floating-point

numbers!

However, there are two workarounds we can use in Hive.

First, if we read the data from a TEXTFILE (see Chapter 15), which is what we have been

assuming so far, then Hive reads the string “0.2” from the data file and converts it to a

real number. We could use DOUBLE instead of FLOAT in our schema. Then we would be

comparing a DOUBLE for the deductions['Federal Taxes'] with a double for the literal

0.2. However, this change will increase the memory footprint of our queries. Also, we

can’t simply change the schema like this if the data file is a binary file format like

SEQUENCEFILE (discussed in Chapter 15).

The second workaround is to explicitly cast the 0.2 literal value to FLOAT. Java has a

nice way of doing this: you append the letter f or F to the end of the number (e.g.,

0.2f). Unfortunately, Hive doesn’t support this syntax; we have to use the cast

operator.

Here is a modified query that casts the 0.2 literal value to FLOAT. With this change, the

expected results are returned by the query:

hive> SELECT name, salary, deductions['Federal Taxes'] FROM employees

> WHERE deductions['Federal Taxes'] > cast(0.2 AS FLOAT);

WHERE Clauses | 95

Boss Man 200000.0 0.3

Fred Finance 150000.0 0.3

Note the syntax inside the cast operator: number AS FLOAT.

Actually, there is also a third solution: avoid floating-point numbers for anything in-

volving money.

Use extreme caution when comparing floating-point numbers. Avoid

all implicit casts from smaller to wider types.

LIKE and RLIKE

Table 6-6 describes the LIKE and RLIKE predicate operators. You have probably seen

LIKE before, a standard SQL operator. It lets us match on strings that begin with or end

with a particular substring, or when the substring appears anywhere within the string.

For example, the following three queries select the employee names and addresses

where the street ends with Ave., the city begins with O, and the street contains Chicago:

hive> SELECT name, address.street FROM employees WHERE address.street LIKE '%Ave.';

John Doe 1 Michigan Ave.

Todd Jones 200 Chicago Ave.

hive> SELECT name, address.city FROM employees WHERE address.city LIKE 'O%';

Todd Jones Oak Park

Bill King Obscuria

hive> SELECT name, address.street FROM employees WHERE address.street LIKE '%Chi%';

Todd Jones 200 Chicago Ave.

A Hive extension is the RLIKE clause, which lets us use Java regular expressions, a more

powerful minilanguage for specifying matches. The rich details of regular expression

syntax and features are beyond the scope of this book. The entry for RLIKE in Ta-

ble 6-6 provides links to resources with more details on regular expressions. Here, we

demonstrate their use with an example, which finds all the employees whose street

contains the word Chicago or Ontario:

hive> SELECT name, address.street

> FROM employees WHERE address.street RLIKE '.*(Chicago|Ontario).*';

Mary Smith 100 Ontario St.

Todd Jones 200 Chicago Ave.

The string after the RLIKE keyword has the following interpretation. A period (.) matches

any character and a star (*) means repeat the “thing to the left” (period, in the two cases

shown) zero to many times. The expression (x|y) means match either x or y.

Hence, there might be no characters before “Chicago” or “Ontario” and there might

be no characters after them. Of course, we could have written this particular example

with two LIKE clauses:

96 | Chapter 6: HiveQL: Queries

Downloa d f r o m W o w ! e B o o k < w w w.woweb o o k . c o m >

SELECT name, address FROM employees

WHERE address.street LIKE '%Chicago%' OR address.street LIKE '%Ontario%';

General regular expression matches will let us express much richer matching criteria

that would become very unwieldy with joined LIKE clauses such as these.

For more details about regular expressions as implemented by Hive using Java, see the

documentation for the Java regular expression syntax at http://docs.oracle.com/javase/

6/docs/api/java/util/regex/Pattern.html or see Regular Expression Pocket Reference by

Tony Stubblebine (O’Reilly), Regular Expressions Cookbook by Jan Goyvaerts and Ste-

ven Levithan (O’Reilly), or Mastering Regular Expressions, 3rd Edition, by Jeffrey E.F.

Friedl (O’Reilly).

GROUP BY Clauses

The GROUP BY statement is often used in conjunction with aggregate functions to

group the result set by one or more columns and then perform an aggregation over each

group.

Let’s return to the stocks table we defined in “External Tables” on page 56. The fol-

lowing query groups stock records for Apple by year, then averages the closing price

for each year:

hive> SELECT year(ymd), avg(price_close) FROM stocks

> WHERE exchange = 'NASDAQ' AND symbol = 'AAPL'

> GROUP BY year(ymd);

1984 25.578625440597534

1985 20.193676221040867

1986 32.46102808021274

1987 53.88968399108163

1988 41.540079275138766

1989 41.65976212516664

1990 37.56268799823263

1991 52.49553383386182

1992 54.80338610251119

1993 41.02671956450572

1994 34.0813495847914

...

HAVING Clauses

The HAVING clause lets you constrain the groups produced by GROUP BY in a way that

could be expressed with a subquery, using a syntax that’s easier to express. Here’s the

previous query with an additional HAVING clause that limits the results to years where

the average closing price was greater than $50.0:

GROUP BY Clauses | 97

hive> SELECT year(ymd), avg(price_close) FROM stocks

> WHERE exchange = 'NASDAQ' AND symbol = 'AAPL'

> GROUP BY year(ymd)

> HAVING avg(price_close) > 50.0;

1987 53.88968399108163

1991 52.49553383386182

1992 54.80338610251119

1999 57.77071460844979

2000 71.74892876261757

2005 52.401745992993554

...

Without the HAVING clause, this query would require a nested SELECT statement:

hive> SELECT s2.year, s2.avg FROM

> (SELECT year(ymd) AS year, avg(price_close) AS avg FROM stocks

> WHERE exchange = 'NASDAQ' AND symbol = 'AAPL'

> GROUP BY year(ymd)) s2

> WHERE s2.avg > 50.0;

1987 53.88968399108163

...

JOIN Statements

Hive supports the classic SQL JOIN statement, but only equi-joins are supported.

Inner JOIN

In an inner JOIN, records are discarded unless join criteria finds matching records in

every table being joined. For example, the following query compares Apple (symbol

AAPL) and IBM (symbol IBM). The stocks table is joined against itself, a self-join, where

the dates, ymd (year-month-day) values must be equal in both tables. We say that the

ymd columns are the join keys in this query:

hive> SELECT a.ymd, a.price_close, b.price_close

> FROM stocks a JOIN stocks b ON a.ymd = b.ymd

> WHERE a.symbol = 'AAPL' AND b.symbol = 'IBM';

2010-01-04 214.01 132.45

2010-01-05 214.38 130.85

2010-01-06 210.97 130.0

2010-01-07 210.58 129.55

2010-01-08 211.98 130.85

2010-01-11 210.11 129.48

...

The ON clause specifies the conditions for joining records between the two tables. The

WHERE clause limits the lefthand table to AAPL records and the righthand table to IBM

records. You can also see that using table aliases for the two occurrences of stocks is

essential in this query.

As you may know, IBM is an older company than Apple. It has been a publicly traded

stock for much longer than Apple. However, since this is an inner JOIN, no IBM records

98 | Chapter 6: HiveQL: Queries

will be returned older than September 7, 1984, which was the first day that Apple was

publicly traded!

Standard SQL allows a non-equi-join on the join keys, such as the following example

that shows Apple versus IBM, but with all older records for Apple paired up with each

day of IBM data. It would be a lot of data (Example 6-1)!

Example 6-1. Query that will not work in Hive

SELECT a.ymd, a.price_close, b.price_close

FROM stocks a JOIN stocks b

ON a.ymd <= b.ymd

WHERE a.symbol = 'AAPL' AND b.symbol = 'IBM';

This is not valid in Hive, primarily because it is difficult to implement these kinds of

joins in MapReduce. It turns out that Pig offers a cross product feature that makes it

possible to implement this join, even though Pig’s native join feature doesn’t support

it, either.

Also, Hive does not currently support using OR between predicates in ON clauses.

To see a nonself join, let’s introduce the corresponding dividends data, also

available from infochimps.org, as described in “External Tables” on page 56:

CREATE EXTERNAL TABLE IF NOT EXISTS dividends (

ymd STRING,

dividend FLOAT

)

PARTITIONED BY (exchange STRING, symbol STRING)

ROW FORMAT DELIMITED FIELDS TERMINATED BY ',';

Here is an inner JOIN between stocks and dividends for Apple, where we use the ymd

and symbol columns as join keys:

hive> SELECT s.ymd, s.symbol, s.price_close, d.dividend

> FROM stocks s JOIN dividends d ON s.ymd = d.ymd AND s.symbol = d.symbol

> WHERE s.symbol = 'AAPL';

1987-05-11 AAPL 77.0 0.015

1987-08-10 AAPL 48.25 0.015

1987-11-17 AAPL 35.0 0.02

...

1995-02-13 AAPL 43.75 0.03

1995-05-26 AAPL 42.69 0.03

1995-08-16 AAPL 44.5 0.03

1995-11-21 AAPL 38.63 0.03

Yes, Apple paid a dividend years ago and only recently announced it would start doing

so again! Note that because we have an inner JOIN, we only see records approximately

every three months, the typical schedule of dividend payments, which are announced

when reporting quarterly results.

You can join more than two tables together. Let’s compare Apple, IBM, and GE side

by side:

JOIN Statements | 99

hive> SELECT a.ymd, a.price_close, b.price_close , c.price_close

> FROM stocks a JOIN stocks b ON a.ymd = b.ymd

> JOIN stocks c ON a.ymd = c.ymd

> WHERE a.symbol = 'AAPL' AND b.symbol = 'IBM' AND c.symbol = 'GE';

2010-01-04 214.01 132.45 15.45

2010-01-05 214.38 130.85 15.53

2010-01-06 210.97 130.0 15.45

2010-01-07 210.58 129.55 16.25

2010-01-08 211.98 130.85 16.6

2010-01-11 210.11 129.48 16.76

...

Most of the time, Hive will use a separate MapReduce job for each pair of things to

join. In this example, it would use one job for tables a and b, then a second job to join

the output of the first join with c.

Why not join b and c first? Hive goes from left to right.

However, this example actually benefits from an optimization we’ll discuss next.

Join Optimizations

In the previous example, every ON clause uses a.ymd as one of the join keys. In this case,

Hive can apply an optimization where it joins all three tables in a single MapReduce

job. The optimization would also be used if b.ymd were used in both ON clauses.

When joining three or more tables, if every ON clause uses the same join

key, a single MapReduce job will be used.

Hive also assumes that the last table in the query is the largest. It attempts to buffer the

other tables and then stream the last table through, while performing joins on individual

records. Therefore, you should structure your join queries so the largest table is last.

Recall our previous join between stocks and dividends. We actually made the mistake

of using the smaller dividends table last:

SELECT s.ymd, s.symbol, s.price_close, d.dividend

FROM stocks s JOIN dividends d ON s.ymd = d.ymd AND s.symbol = d.symbol

WHERE s.symbol = 'AAPL';

We should switch the positions of stocks and dividends:

SELECT s.ymd, s.symbol, s.price_close, d.dividend

FROM dividends d JOIN stocks s ON s.ymd = d.ymd AND s.symbol = d.symbol

WHERE s.symbol = 'AAPL';

100 | Chapter 6: HiveQL: Queries

It turns out that these data sets are too small to see a noticeable performance difference,

but for larger data sets, you’ll want to exploit this optimization.

Fortunately, you don’t have to put the largest table last in the query. Hive also provides

a “hint” mechanism to tell the query optimizer which table should be streamed:

SELECT /*+ STREAMTABLE(s) */ s.ymd, s.symbol, s.price_close, d.dividend

FROM stocks s JOIN dividends d ON s.ymd = d.ymd AND s.symbol = d.symbol

WHERE s.symbol = 'AAPL';

Now Hive will attempt to stream the stocks table, even though it’s not the last table in

the query.

There is another important optimization called map-side joins that we’ll return to in

“Map-side Joins” on page 105.

LEFT OUTER JOIN

The left-outer join is indicated by adding the LEFT OUTER keywords:

hive> SELECT s.ymd, s.symbol, s.price_close, d.dividend

> FROM stocks s LEFT OUTER JOIN dividends d ON s.ymd = d.ymd AND s.symbol = d.symbol

> WHERE s.symbol = 'AAPL';

...

1987-05-01 AAPL 80.0 NULL

1987-05-04 AAPL 79.75 NULL

1987-05-05 AAPL 80.25 NULL

1987-05-06 AAPL 80.0 NULL

1987-05-07 AAPL 80.25 NULL

1987-05-08 AAPL 79.0 NULL

1987-05-11 AAPL 77.0 0.015

1987-05-12 AAPL 75.5 NULL

1987-05-13 AAPL 78.5 NULL

1987-05-14 AAPL 79.25 NULL

1987-05-15 AAPL 78.25 NULL

1987-05-18 AAPL 75.75 NULL

1987-05-19 AAPL 73.25 NULL

1987-05-20 AAPL 74.5 NULL

...

In this join, all the records from the lefthand table that match the WHERE clause are

returned. If the righthand table doesn’t have a record that matches the ON criteria,

NULL is used for each column selected from the righthand table.

Hence, in this result set, we see that the every Apple stock record is returned and the

d.dividend value is usually NULL, except on days when a dividend was paid (May 11th,

1987, in this output).

OUTER JOIN Gotcha

Before we discuss the other outer joins, let’s discuss a gotcha you should understand.

JOIN Statements | 101

Recall what we said previously about speeding up queries by adding partition filters in

the WHERE clause. To speed up our previous query, we might choose to add predicates

that select on the exchange in both tables:

hive> SELECT s.ymd, s.symbol, s.price_close, d.dividend

> FROM stocks s LEFT OUTER JOIN dividends d ON s.ymd = d.ymd AND s.symbol = d.symbol

> WHERE s.symbol = 'AAPL'

> AND s.exchange = 'NASDAQ' AND d.exchange = 'NASDAQ';

1987-05-11 AAPL 77.0 0.015

1987-08-10 AAPL 48.25 0.015

1987-11-17 AAPL 35.0 0.02

1988-02-12 AAPL 41.0 0.02

1988-05-16 AAPL 41.25 0.02

...

However, the output has changed, even though we thought we were just adding an

optimization! We’re back to having approximately four stock records per year and we

have non-NULL entries for all the dividend values. In other words, we are back to the

original inner join!

This is actually common behavior for all outer joins in most SQL implementations. It

occurs because the JOIN clause is evaluated first, then the results are passed through

the WHERE clause. By the time the WHERE clause is reached, d.exchange is NULL most of the

time, so the “optimization” actually filters out all records except those on the day of

dividend payments.

One solution is straightforward; remove the clauses in the WHERE clause that reference

the dividends table:

hive> SELECT s.ymd, s.symbol, s.price_close, d.dividend

> FROM stocks s LEFT OUTER JOIN dividends d ON s.ymd = d.ymd AND s.symbol = d.symbol

> WHERE s.symbol = 'AAPL' AND s.exchange = 'NASDAQ';

...

1987-05-07 AAPL 80.25 NULL

1987-05-08 AAPL 79.0 NULL

1987-05-11 AAPL 77.0 0.015

1987-05-12 AAPL 75.5 NULL

1987-05-13 AAPL 78.5 NULL

...

This isn’t very satisfactory. You might wonder if you can move the predicates from the

WHERE clause into the ON clause, at least the partition filters. This does not work for outer

joins, despite documentation on the Hive Wiki that claims it should work (https://cwiki

.apache.org/confluence/display/Hive/LanguageManual+Joins).

hive> SELECT s.ymd, s.symbol, s.price_close, d.dividend

> FROM stocks s LEFT OUTER JOIN dividends d

> ON s.ymd = d.ymd AND s.symbol = d.symbol

> AND s.symbol = 'AAPL' AND s.exchange = 'NASDAQ' AND d.exchange = 'NASDAQ';

...

1962-01-02 GE 74.75 NULL

1962-01-02 IBM 572.0 NULL

1962-01-03 GE 74.0 NULL

1962-01-03 IBM 577.0 NULL

102 | Chapter 6: HiveQL: Queries

1962-01-04 GE 73.12 NULL

1962-01-04 IBM 571.25 NULL

1962-01-05 GE 71.25 NULL

1962-01-05 IBM 560.0 NULL

...

The partition filters are ignored for OUTER JOINTS. However, using such filter predicates

in ON clauses for inner joins does work!

Fortunately, there is solution that works for all joins; use nested SELECT statements:

hive> SELECT s.ymd, s.symbol, s.price_close, d.dividend FROM

> (SELECT * FROM stocks WHERE symbol = 'AAPL' AND exchange = 'NASDAQ') s

> LEFT OUTER JOIN

> (SELECT * FROM dividends WHERE symbol = 'AAPL' AND exchange = 'NASDAQ') d

> ON s.ymd = d.ymd;

...

1988-02-10 AAPL 41.0 NULL

1988-02-11 AAPL 40.63 NULL

1988-02-12 AAPL 41.0 0.02

1988-02-16 AAPL 41.25 NULL

1988-02-17 AAPL 41.88 NULL

...

The nested SELECT statement performs the required “push down” to apply the partition

filters before data is joined.

WHERE clauses are evaluated after joins are performed, so WHERE clauses

should use predicates that only filter on column values that won’t be

NULL. Also, contrary to Hive documentation, partition filters don’t work

in ON clauses for OUTER JOINS, although they do work for INNER JOINS!

RIGHT OUTER JOIN

Right-outer joins return all records in the righthand table that match the WHERE clause.

NULL is used for fields of missing records in the lefthand table.

Here we switch the places of stocks and dividends and perform a righthand join, but

leave the SELECT statement unchanged:

hive> SELECT s.ymd, s.symbol, s.price_close, d.dividend

> FROM dividends d RIGHT OUTER JOIN stocks s ON d.ymd = s.ymd AND d.symbol = s.symbol

> WHERE s.symbol = 'AAPL';

...

1987-05-07 AAPL 80.25 NULL

1987-05-08 AAPL 79.0 NULL

1987-05-11 AAPL 77.0 0.015

1987-05-12 AAPL 75.5 NULL

1987-05-13 AAPL 78.5 NULL

...

JOIN Statements | 103

FULL OUTER JOIN

Finally, a full-outer join returns all records from all tables that match the WHERE clause.

NULL is used for fields in missing records in either table.

If we convert the previous query to a full-outer join, we’ll actually get the same results,

since there is never a case where a dividend record exists without a matching stock

record:

hive> SELECT s.ymd, s.symbol, s.price_close, d.dividend

> FROM dividends d FULL OUTER JOIN stocks s ON d.ymd = s.ymd AND d.symbol = s.symbol

> WHERE s.symbol = 'AAPL';

...

1987-05-07 AAPL 80.25 NULL

1987-05-08 AAPL 79.0 NULL

1987-05-11 AAPL 77.0 0.015

1987-05-12 AAPL 75.5 NULL

1987-05-13 AAPL 78.5 NULL

...

LEFT SEMI-JOIN

A left semi-join returns records from the lefthand table if records are found in the right-

hand table that satisfy the ON predicates. It’s a special, optimized case of the more general

inner join. Most SQL dialects support an IN ... EXISTS construct to do the same thing.

For instance, the following query in Example 6-2 attempts to return stock records only

on the days of dividend payments, but it doesn’t work in Hive.

Example 6-2. Query that will not work in Hive

SELECT s.ymd, s.symbol, s.price_close FROM stocks s

WHERE s.ymd, s.symbol IN

(SELECT d.ymd, d.symbol FROM dividends d);

Instead, you use the following LEFT SEMI JOIN syntax:

hive> SELECT s.ymd, s.symbol, s.price_close

> FROM stocks s LEFT SEMI JOIN dividends d ON s.ymd = d.ymd AND s.symbol = d.symbol;

...

1962-11-05 IBM 361.5

1962-08-07 IBM 373.25

1962-05-08 IBM 459.5

1962-02-06 IBM 551.5

Note that the SELECT and WHERE clauses can’t reference columns from the righthand

table.

Right semi-joins are not supported in Hive.

104 | Chapter 6: HiveQL: Queries

The reason semi-joins are more efficient than the more general inner join is as follows.

For a given record in the lefthand table, Hive can stop looking for matching records in

the righthand table as soon as any match is found. At that point, the selected columns

from the lefthand table record can be projected.

Cartesian Product JOINs

A Cartesian product is a join where all the tuples in the left side of the join are paired

with all the tuples of the right table. If the left table has 5 rows and the right table has

6 rows, 30 rows of output will be produced:

SELECTS * FROM stocks JOIN dividends;

Using the table of stocks and dividends, it is hard to find a reason for a join of this type,

as the dividend of one stock is not usually paired with another. Additionally, Cartesian

products create a lot of data. Unlike other join types, Cartesian products are not exe-

cuted in parallel, and they are not optimized in any way using MapReduce.

It is critical to point out that using the wrong join syntax will cause a long, slow-running

Cartesian product query. For example, the following query will be optimized to an

inner join in many databases, but not in Hive:

hive > SELECT * FROM stocks JOIN dividends

> WHERE stock.symbol = dividends.symbol and stock.symbol='AAPL';

In Hive, this query computes the full Cartesian product before applying the WHERE

clause. It could take a very long time to finish. When the property hive.mapred.mode is

set to strict, Hive prevents users from inadvertently issuing a Cartesian product query.

We’ll discuss the features of strict mode more extensively in Chapter 10.

Cartesian product queries can be useful. For example, suppose there is

a table of user preferences, a table of news articles, and an algorithm

that predicts which articles a user would like to read. A Cartesian prod-

uct is required to generate the set of all users and all pages.

Map-side Joins

If all but one table is small, the largest table can be streamed through the mappers while

the small tables are cached in memory. Hive can do all the joining map-side, since it

can look up every possible match against the small tables in memory, thereby elimi-

nating the reduce step required in the more common join scenarios. Even on smaller

data sets, this optimization is noticeably faster than the normal join. Not only does it

eliminate reduce steps, it sometimes reduces the number of map steps, too.

The joins between stocks and dividends can exploit this optimization, as the dividends

data set is small enough to be cached.

JOIN Statements | 105

Before Hive v0.7, it was necessary to add a hint to the query to enable this optimization.

Returning to our inner join example:

SELECT /*+ MAPJOIN(d) */ s.ymd, s.symbol, s.price_close, d.dividend

FROM stocks s JOIN dividends d ON s.ymd = d.ymd AND s.symbol = d.symbol

WHERE s.symbol = 'AAPL';

Running this query versus the original on a fast MacBook Pro laptop yielded times of

approximately 23 seconds versus 33 seconds for the original unoptimized query, which

is roughly 30% faster using our sample stock data.

The hint still works, but it’s now deprecated as of Hive v0.7. However, you still have

to set a property, hive.auto.convert.join, to true before Hive will attempt the opti-

mization. It’s false by default:

hive> set hive.auto.convert.join=true;

hive> SELECT s.ymd, s.symbol, s.price_close, d.dividend

> FROM stocks s JOIN dividends d ON s.ymd = d.ymd AND s.symbol = d.symbol

> WHERE s.symbol = 'AAPL';

Note that you can also configure the threshold size for table files considered small

enough to use this optimization. Here is the default definition of the property (in bytes):

hive.mapjoin.smalltable.filesize=25000000

If you always want Hive to attempt this optimization, set one or both of these properties

in your $HOME/.hiverc file.

Hive does not support the optimization for right- and full-outer joins.

This optimization can also be used for larger tables under certain conditions when

the data for every table is bucketed, as discussed in “Bucketing Table Data Stor-

age” on page 125. Briefly, the data must be bucketed on the keys used in the ON clause

and the number of buckets for one table must be a multiple of the number of buckets

for the other table. When these conditions are met, Hive can join individual buckets

between tables in the map phase, because it does not need to fetch the entire contents

of one table to match against each bucket in the other table.

However, this optimization is not turned on by default. It must be enabled by setting

the property hive.optimize.bucketmapjoin:

set hive.optimize.bucketmapjoin=true;

If the bucketed tables actually have the same number of buckets and the data is sorted

by the join/bucket keys, then Hive can perform an even faster sort-merge join. Once

again, properties must be set to enable the optimization:

set hive.input.format=org.apache.hadoop.hive.ql.io.BucketizedHiveInputFormat;

set hive.optimize.bucketmapjoin=true;

set hive.optimize.bucketmapjoin.sortedmerge=true;

106 | Chapter 6: HiveQL: Queries

ORDER BY and SORT BY

The ORDER BY clause is familiar from other SQL dialects. It performs a total ordering of

the query result set. This means that all the data is passed through a single reducer,

which may take an unacceptably long time to execute for larger data sets.

Hive adds an alternative, SORT BY, that orders the data only within each reducer, thereby

performing a local ordering, where each reducer’s output will be sorted. Better perfor-

mance is traded for total ordering.

In both cases, the syntax differs only by the use of the ORDER or SORT keyword. You can

specify any columns you wish and specify whether or not the columns are ascending

using the ASC keyword (the default) or descending using the DESC keyword.

Here is an example using ORDER BY:

SELECT s.ymd, s.symbol, s.price_close

FROM stocks s

ORDER BY s.ymd ASC, s.symbol DESC;

Here is the same example using SORT BY instead:

SELECT s.ymd, s.symbol, s.price_close

FROM stocks s

SORT BY s.ymd ASC, s.symbol DESC;

The two queries look almost identical, but if more than one reducer is invoked, the

output will be sorted differently. While each reducer’s output files will be sorted, the

data will probably overlap with the output of other reducers.

Because ORDER BY can result in excessively long run times, Hive will require a LIMIT

clause with ORDER BY if the property hive.mapred.mode is set to strict. By default, it is

set to nonstrict.

DISTRIBUTE BY with SORT BY

DISTRIBUTE BY controls how map output is divided among reducers. All data that flows

through a MapReduce job is organized into key-value pairs. Hive must use this feature

internally when it converts your queries to MapReduce jobs.

Usually, you won’t need to worry about this feature. The exceptions are queries that

use the Streaming feature (see Chapter 14) and some stateful UDAFs (User-Defined

Aggregate Functions; see “Aggregate Functions” on page 164). There is one other sce-

nario where these clauses are useful.

By default, MapReduce computes a hash on the keys output by mappers and tries to

evenly distribute the key-value pairs among the available reducers using the hash values.

Unfortunately, this means that when we use SORT BY, the contents of one reducer’s

output will overlap significantly with the output of the other reducers, as far as sorted

order is concerned, even though the data is sorted within each reducer’s output.

DISTRIBUTE BY with SORT BY | 107

Say we want the data for each stock symbol to be captured together. We can use

DISTRIBUTE BY to ensure that the records for each stock symbol go to the same reducer,

then use SORT BY to order the data the way we want. The following query demonstrates

this technique:

hive> SELECT s.ymd, s.symbol, s.price_close

> FROM stocks s

> DISTRIBUTE BY s.symbol

> SORT BY s.symbol ASC, s.ymd ASC;

1984-09-07 AAPL 26.5

1984-09-10 AAPL 26.37

1984-09-11 AAPL 26.87

1984-09-12 AAPL 26.12

1984-09-13 AAPL 27.5

1984-09-14 AAPL 27.87

1984-09-17 AAPL 28.62

1984-09-18 AAPL 27.62

1984-09-19 AAPL 27.0

1984-09-20 AAPL 27.12

...

Of course, the ASC keywords could have been omitted as they are the defaults. The

ASC keyword is placed here for reasons that will be described shortly.

DISTRIBUTE BY works similar to GROUP BY in the sense that it controls how reducers

receive rows for processing, while SORT BY controls the sorting of data inside the reducer.

Note that Hive requires that the DISTRIBUTE BY clause come before the SORT BY clause.

CLUSTER BY

In the previous example, the s.symbol column was used in the DISTRIBUTE BY clause,

and the s.symbol and the s.ymd columns in the SORT BY clause. Suppose that the same

columns are used in both clauses and all columns are sorted by ascending order (the

default). In this case, the CLUSTER BY clause is a shor-hand way of expressing the same

query.

For example, let’s modify the previous query to drop sorting by s.ymd and use CLUSTER

BY on s.symbol:

hive> SELECT s.ymd, s.symbol, s.price_close

> FROM stocks s

> CLUSTER BY s.symbol;

2010-02-08 AAPL 194.12

2010-02-05 AAPL 195.46

2010-02-04 AAPL 192.05

2010-02-03 AAPL 199.23

2010-02-02 AAPL 195.86

2010-02-01 AAPL 194.73

2010-01-29 AAPL 192.06

2010-01-28 AAPL 199.29

108 | Chapter 6: HiveQL: Queries

2010-01-27 AAPL 207.88

...

Because the sort requirements are removed for the s.ymd, the output reflects the original

order of the stock data, which is sorted descending.

Using DISTRIBUTE BY ... SORT BY or the shorthand CLUSTER BY clauses is a way to exploit

the parallelism of SORT BY, yet achieve a total ordering across the output files.

Casting

We briefly mentioned in “Primitive Data Types” on page 41 that Hive will perform

some implicit conversions, called casts, of numeric data types, as needed. For example,

when doing comparisons between two numbers of different types. This topic is dis-

cussed more fully in “Predicate Operators” on page 93 and “Gotchas with Floating-

Point Comparisons” on page 94.

Here we discuss the cast() function that allows you to explicitly convert a value of one

type to another.

Recall our employees table uses a FLOAT for the salary column. Now, imagine for a

moment that STRING was used for that column instead. How could we work with the

values as FLOATS?

The following example casts the values to FLOAT before performing a comparison:

SELECT name, salary FROM employees

WHERE cast(salary AS FLOAT) < 100000.0;

The syntax of the cast function is cast(value AS TYPE). What would happen in the

example if a salary value was not a valid string for a floating-point number? In this

case, Hive returns NULL.

Note that the preferred way to convert floating-point numbers to integers is to use the

round() or floor() functions listed in Table 6-2, rather than to use the cast operator.

Casting BINARY Values

The new BINARY type introduced in Hive v0.8.0 only supports casting BINARY to

STRING. However, if you know the value is a number, you can nest cast() invocations,

as in this example where column b is a BINARY column:

SELECT (2.0*cast(cast(b as string) as double)) from src;

You can also cast STRING to BINARY.

Casting | 109

Queries that Sample Data

For very large data sets, sometimes you want to work with a representative sample of

a query result, not the whole thing. Hive supports this goal with queries that sample

tables organized into buckets.

In the following example, assume the numbers table has one number column with

values 1−10.

We can sample using the rand() function, which returns a random number. In the first

two queries, two distinct numbers are returned for each query. In the third query, no

results are returned:

hive> SELECT * from numbers TABLESAMPLE(BUCKET 3 OUT OF 10 ON rand()) s;

If we bucket on a column instead of rand(), then identical results are returned on mul-

tiple runs:

hive> SELECT * from numbers TABLESAMPLE(BUCKET 3 OUT OF 10 ON number) s;

hive> SELECT * from numbers TABLESAMPLE(BUCKET 5 OUT OF 10 ON number) s;

hive> SELECT * from numbers TABLESAMPLE(BUCKET 3 OUT OF 10 ON number) s;

The denominator in the bucket clause represents the number of buckets into which

data will be hashed. The numerator is the bucket number selected:

hive> SELECT * from numbers TABLESAMPLE(BUCKET 1 OUT OF 2 ON number) s;

hive> SELECT * from numbers TABLESAMPLE(BUCKET 2 OUT OF 2 ON number) s;

110 | Chapter 6: HiveQL: Queries

Block Sampling

Hive offers another syntax for sampling a percentage of blocks of an input path as an

alternative to sampling based on rows:

hive> SELECT * FROM numbersflat TABLESAMPLE(0.1 PERCENT) s;

This sampling is not known to work with all file formats. Also, the

smallest unit of sampling is a single HDFS block. Hence, for tables less

than the typical block size of 128 MB, all rows will be retuned.

Percentage-based sampling offers a variable to control the seed information for block-

based tuning. Different seeds produce different samples:

<name>hive.sample.seednumber</name>

<description>A number used for percentage sampling. By changing this

number, user will change the subsets of data sampled.</description>

</property>

Input Pruning for Bucket Tables

From a first look at the TABLESAMPLE syntax, an astute user might come to the conclusion

that the following query would be equivalent to the TABLESAMPLE operation:

hive> SELECT * FROM numbersflat WHERE number % 2 = 0;

It is true that for most table types, sampling scans through the entire table and selects

every Nth row. However, if the columns specified in the TABLESAMPLE clause match the

columns in the CLUSTERED BY clause, TABLESAMPLE queries only scan the required hash

partitions of the table:

hive> CREATE TABLE numbers_bucketed (number int) CLUSTERED BY (number) INTO 3 BUCKETS;

hive> SET hive.enforce.bucketing=true;

hive> INSERT OVERWRITE TABLE numbers_bucketed SELECT number FROM numbers;

hive> dfs -ls /user/hive/warehouse/mydb.db/numbers_bucketed;

/user/hive/warehouse/mydb.db/numbers_bucketed/000000_0

/user/hive/warehouse/mydb.db/numbers_bucketed/000001_0

/user/hive/warehouse/mydb.db/numbers_bucketed/000002_0

Queries that Sample Data | 111

hive> dfs -cat /user/hive/warehouse/mydb.db/numbers_bucketed/000001_0;

Because this table is clustered into three buckets, the following query can be used to

sample only one of the buckets efficiently:

hive> SELECT * FROM numbers_bucketed TABLESAMPLE (BUCKET 2 OUT OF 3 ON NUMBER) s;

UNION ALL

UNION ALL combines two or more tables. Each subquery of the union query must pro-

duce the same number of columns, and for each column, its type must match all the

column types in the same position. For example, if the second column is a FLOAT, then

the second column of all the other query results must be a FLOAT.

Here is an example the merges log data:

SELECT log.ymd, log.level, log.message

FROM (

SELECT l1.ymd, l1.level,

l1.message, 'Log1' AS source

FROM log1 l1

UNION ALL

SELECT l2.ymd, l2.level,

l2.message, 'Log2' AS source

FROM log1 l2

) log

SORT BY log.ymd ASC;

UNION may be used when a clause selects from the same source table. Logically, the same

results could be achieved with a single SELECT and WHERE clause. This technique increases

readability by breaking up a long complex WHERE clause into two or more UNION queries.

However, unless the source table is indexed, the query will have to make multiple passes

over the same source data. For example:

FROM (

FROM src SELECT src.key, src.value WHERE src.key < 100

UNION ALL

FROM src SELECT src.* WHERE src.key > 110

) unioninput

INSERT OVERWRITE DIRECTORY '/tmp/union.out' SELECT unioninput.*

112 | Chapter 6: HiveQL: Queries

CHAPTER 7

HiveQL: Views

A view allows a query to be saved and treated like a table. It is a logical construct, as it

does not store data like a table. In other words, materialized views are not currently

supported by Hive.

When a query references a view, the information in its definition is combined with the

rest of the query by Hive’s query planner. Logically, you can imagine that Hive executes

the view and then uses the results in the rest of the query.

Views to Reduce Query Complexity

When a query becomes long or complicated, a view may be used to hide the complexity

by dividing the query into smaller, more manageable pieces; similar to writing a func-

tion in a programming language or the concept of layered design in software. Encap-

sulating the complexity makes it easier for end users to construct complex queries from

reusable parts. For example, consider the following query with a nested subquery:

FROM (

SELECT * FROM people JOIN cart

ON (cart.people_id=people.id) WHERE firstname='john'

) a SELECT a.lastname WHERE a.id=3;

It is common for Hive queries to have many levels of nesting. In the following example,

the nested portion of the query is turned into a view:

CREATE VIEW shorter_join AS

SELECT * FROM people JOIN cart

ON (cart.people_id=people.id) WHERE firstname='john';

Now the view is used like any other table. In this query we added a WHERE clause to the

SELECT statement. This exactly emulates the original query:

SELECT lastname FROM shorter_join WHERE id=3;

113

Views that Restrict Data Based on Conditions

A common use case for views is restricting the result rows based on the value of one or

more columns. Some databases allow a view to be used as a security mechanism. Rather

than give the user access to the raw table with sensitive data, the user is given access to

a view with a WHERE clause that restricts the data. Hive does not currently support this

feature, as the user must have access to the entire underlying raw table for the view to

work. However, the concept of a view created to limit data access can be used to protect

information from the casual query:

hive> CREATE TABLE userinfo (

> firstname string, lastname string, ssn string, password string);

hive> CREATE VIEW safer_user_info AS

> SELECT firstname,lastname FROM userinfo;

Here is another example where a view is used to restrict data based on a WHERE clause.

In this case, we wish to provide a view on an employee table that only exposes employees

from a specific department:

hive> CREATE TABLE employee (firstname string, lastname string,

> ssn string, password string, department string);

hive> CREATE VIEW techops_employee AS

> SELECT firstname,lastname,ssn FROM userinfo WERE department='techops';

Views and Map Type for Dynamic Tables

Recall from Chapter 3 that Hive supports arrays, maps, and structs datatypes. These

datatypes are not common in traditional databases as they break first normal form.

Hive’s ability to treat a line of text as a map, rather than a fixed set of columns, combined

with the view feature, allows you to define multiple logical tables over one physical table.

For example, consider the following sample data file that treats an entire row as a map

rather than a list of fixed columns. Rather than using Hive’s default values for separa-

tors, this file uses ^A (Control-A) as the collection item separator (i.e., between key-

value pairs in this case, where the collection is a map) and ^B (Control-B) as the sepa-

rator between keys and values in the map. The long lines wrap in the following listing,

so we added a blank line between them for better clarity:

time^B1298598398404^Atype^Brequest^Astate^Bny^Acity^Bwhite

plains^Apart\^Bmuffler

time^B1298598398432^Atype^Bresponse^Astate^Bny^Acity^Btarry-

town^Apart\^Bmuffler

time^B1298598399404^Atype^Brequest^Astate^Btx^Acity^Baus-

tin^Apart^Bheadlight

Now we create our table:

114 | Chapter 7: HiveQL: Views

Downloa d f r o m W o w ! e B o o k < w w w.woweb o o k . c o m >

CREATE EXTERNAL TABLE dynamictable(cols map<string,string>)

ROW FORMAT DELIMITED

FIELDS TERMINATED BY '\004'

COLLECTION ITEMS TERMINATED BY '\001'

MAP KEYS TERMINATED BY '\002'

STORED AS TEXTFILE;

Because there is only one field per row, the FIELDS TERMINATED BY value actually has

no effect.

Now we can create a view that extracts only rows with type equal to requests and get

the city, state, and part into a view called orders:

CREATE VIEW orders(state, city, part) AS

SELECT cols["state"], cols["city"], cols["part"]

FROM dynamictable

WHERE cols["type"] = "request";

A second view is created named shipments. This view returns the time and part column

from rows where the type is response:

CREATE VIEW shipments(time, part) AS

SELECT cols["time"], cols["parts"]

FROM dynamictable

WHERE cols["type"] = "response";

For another example of this feature, see http://dev.bizo.com/2011/02/columns-in-hive

.html#!/2011/02/columns-in-hive.html.

View Odds and Ends

We said that Hive evaluates the view and then uses the results to evaluate the query.

However, as part of Hive’s query optimization, the clauses of both the query and view

may be combined together into a single actual query.

Nevertheless, the conceptual view still applies when the view and a query that uses it

both contain an ORDER BY clause or a LIMIT clause. The view’s clauses are evaluated

before the using query’s clauses.

For example, if the view has a LIMIT 100 clause and the query has a LIMIT 200 clause,

you’ll get at most 100 results.

While defining a view doesn’t “materialize” any data, the view is frozen to any subse-

quent changes to any tables and columns that the view uses. Hence, a query using a

view can fail if the referenced tables or columns no longer exist.

There are a few other clauses you can use when creating views. Modifying our last

example:

CREATE VIEW IF NOT EXISTS shipments(time, part)

COMMENT 'Time and parts for shipments.'

TBLPROPERTIES ('creator' = 'me')

AS SELECT ...;

View Odds and Ends | 115

As for tables, the IF NOT EXISTS and COMMENT … clauses are optional, and have the same

meaning they have for tables.

A view’s name must be unique compared to all other table and view names in the same

database.

You can also add a COMMENT for any or all of the new column names. The comments are

not “inherited” from the definition of the original table.

Also, if the AS SELECT contains an expression without an alias—e.g., size(cols) (the

number of items in cols)—then Hive will use _CN as the name, where N is a number

starting with 0. The view definition will fail if the AS SELECT clause is invalid.

Before the AS SELECT clause, you can also define TBLPROPERTIES, just like for tables. In

the example, we defined a property for the “creator” of the view.

The CREATE TABLE … LIKE … construct discussed in “Creating Tables” on page 53 can

also be used to copy a view, that is with a view as part of the LIKE expression:

CREATE TABLE shipments2

LIKE shipments;

You can also use the optional EXTERNAL keyword and LOCATION … clause, as before.

The behavior of this statement is different as of Hive v0.8.0 and previous

versions of Hive. For v0.8.0, the command creates a new table, not a

new view. It uses defaults for the SerDe and file formats. For earlier

versions, a new view is created.

A view is dropped in the same way as a table:

DROP VIEW IF EXISTS shipments;

As usual, IF EXISTS is optional.

A view will be shown using SHOW TABLES (there is no SHOW VIEWS), however DROP TABLE

cannot be used to delete a view.

As for tables, DESCRIBE shipments and DESCRIBE EXTENDED shipments displays the usual

data for the shipment view. With the latter, there will be a tableType value in the

Detailed Table Information indicating the “table” is a VIRTUAL_VIEW.

You cannot use a view as a target of an INSERT or LOAD command.

Finally, views are read-only. You can only alter the metadata TBLPROPERTIES for a view:

ALTER VIEW shipments SET TBLPROPERTIES ('created_at' = 'some_timestamp');

116 | Chapter 7: HiveQL: Views

CHAPTER 8

HiveQL: Indexes

Hive has limited indexing capabilities. There are no keys in the usual relational database

sense, but you can build an index on columns to speed some operations. The index

data for a table is stored in another table.

Also, the feature is relatively new, so it doesn’t have a lot of options yet. However, the

indexing process is designed to be customizable with plug-in Java code, so teams can

extend the feature to meet their needs.

Indexing is also a good alternative to partitioning when the logical partitions would

actually be too numerous and small to be useful. Indexing can aid in pruning some

blocks from a table as input for a MapReduce job. Not all queries can benefit from an

index—the EXPLAIN syntax and Hive can be used to determine if a given query is aided

by an index.

Indexes in Hive, like those in relational databases, need to be evaluated carefully.

Maintaining an index requires extra disk space and building an index has a processing

cost. The user must weigh these costs against the benefits they offer when querying a

table.

Creating an Index

Let’s create an index for our managed, partitioned employees table we described in

“Partitioned, Managed Tables” on page 58. Here is the table definition we used previ-

ously, for reference:

CREATE TABLE employees (

name STRING,

salary FLOAT,

subordinates ARRAY<STRING>,

deductions MAP<STRING, FLOAT>,

address STRUCT<street:STRING, city:STRING, state:STRING, zip:INT>

)

PARTITIONED BY (country STRING, state STRING);

Let’s index on the country partition only:

117

CREATE INDEX employees_index

ON TABLE employees (country)

AS 'org.apache.hadoop.hive.ql.index.compact.CompactIndexHandler'

WITH DEFERRED REBUILD

IDXPROPERTIES ('creator = 'me', 'created_at' = 'some_time')

IN TABLE employees_index_table

PARTITIONED BY (country, name)

COMMENT 'Employees indexed by country and name.';

In this case, we did not partition the index table to the same level of granularity as the

original table. We could choose to do so. If we omitted the PARTITIONED BY clause

completely, the index would span all partitions of the original table.

The AS ... clause specifies the index handler, a Java class that implements indexing.

Hive ships with a few representative implementations; the CompactIndexHandler shown

was in the first release of this feature. Third-party implementations can optimize certain

scenarios, support specific file formats, and more. We’ll provide more information on

implementing your own index handler in “Implementing a Custom Index Han-

dler” on page 119.

We’ll discuss the meaning of WITH DEFERRED REBUILD in the next section.

It’s not a requirement for the index handler to save its data in a new table, but if it does,

the IN TABLE ... clause is used. It supports many of the options available when creating

other tables. Specifically, the example doesn’t use the optional ROW FORMAT, STORED AS,

STORED BY, LOCATION, and TBLPROPERTIES clauses that we discussed in Chapter 4. All

would appear before the final COMMENT clause shown.

Currently, indexing external tables and views is supported except for data residing

in S3.

Bitmap Indexes

Hive v0.8.0 adds a built-in bitmap index handler. Bitmap indexes are commonly used

for columns with few distinct values. Here is our previous example rewritten to use the

bitmap index handler:

CREATE INDEX employees_index

ON TABLE employees (country)

AS 'BITMAP'

WITH DEFERRED REBUILD

IDXPROPERTIES ('creator = 'me', 'created_at' = 'some_time')

IN TABLE employees_index_table

PARTITIONED BY (country, name)

COMMENT 'Employees indexed by country and name.';

Rebuilding the Index

If you specified WITH DEFERRED REBUILD, the new index starts empty. At any time, the

index can be built the first time or rebuilt using the ALTER INDEX statement:

118 | Chapter 8: HiveQL: Indexes

ALTER INDEX employees_index

ON TABLE employees

PARTITION (country = 'US')

REBUILD;

If the PARTITION clause is omitted, the index is rebuilt for all partitions.

There is no built-in mechanism to trigger an automatic rebuild of the index if the un-

derlying table or a particular partition changes. However, if you have a workflow that

updates table partitions with data, one where you might already use the ALTER TABLE ...

TOUCH PARTITION(...) feature described in “Miscellaneous Alter Table State-

ments” on page 69, that same workflow could issue the ALTER INDEX ... REBUILD

command for a corresponding index.

The rebuild is atomic in the sense that if the rebuild fails, the index is left in the previous

state before the rebuild was started.

Showing an Index

The following command will show all the indexes defined for any column in the indexed

table:

SHOW FORMATTED INDEX ON employees;

FORMATTED is optional. It causes column titles to be added to the output. You can also

replace INDEX with INDEXES, as the output may list multiple indexes.

Dropping an Index

Dropping an index also drops the index table, if any:

DROP INDEX IF EXISTS employees_index ON TABLE employees;

Hive won’t let you attempt to drop the index table directly with DROP TABLE. As always,

IF EXISTS is optional and serves to suppress errors if the index doesn’t exist.

If the table that was indexed is dropped, the index itself and its table is dropped. Sim-

ilarly, if a partition of the original table is dropped, the corresponding partition index

is also dropped.

Implementing a Custom Index Handler

The full details for implementing a custom index handler are given on the Hive Wiki

page, https://cwiki.apache.org/confluence/display/Hive/IndexDev#CREATE_INDEX,

where the initial design of indexing is documented. Of course, you can use the

source code for org.apache.hadoop.hive.ql.index.compact.CompactIndexHandler as an

example.

Implementing a Custom Index Handler | 119

When the index is created, the Java code you implement for the index handler has to

do some initial validation and define the schema for the index table, if used. It also has

to implement the rebuilding process where it reads the table to be indexed and writes

to the index storage (e.g., the index table). The handler must clean up any nontable

storage it uses for the index when the index is dropped, relying on Hive to drop the

index table, as needed. Finally, the handler must participate in optimizing queries.

120 | Chapter 8: HiveQL: Indexes

CHAPTER 9

Schema Design

Hive looks and acts like a relational database. Users have a familiar nomenclature such

as tables and columns, as well as a query language that is remarkably similar to SQL

dialects they have used before. However, Hive is implemented and used in ways that

are very different from conventional relational databases. Often, users try to carry over

paradigms from the relational world that are actually Hive anti-patterns. This section

highlights some Hive patterns you should use and some anti-patterns you should avoid.

Table-by-Day

Table-by-day is a pattern where a table named supply is appended with a timestamp

such as supply_2011_01_01, supply_2011_01_02, etc. Table-by-day is an anti-pattern in

the database world, but due to common implementation challenges of ever-growing

data sets, it is still widely used:

hive> CREATE TABLE supply_2011_01_02 (id int, part string, quantity int);

hive> CREATE TABLE supply_2011_01_03 (id int, part string, quantity int);

hive> CREATE TABLE supply_2011_01_04 (id int, part string, quantity int);

hive> .... load data ...

hive> SELECT part,quantity supply_2011_01_02

> UNION ALL

> SELECT part,quantity from supply_2011_01_03

> WHERE quantity < 4;

With Hive, a partitioned table should be used instead. Hive uses expressions in the

WHERE clause to select input only from the partitions needed for the query. This query

will run efficiently, and it is clean and easy on the eyes:

hive> CREATE TABLE supply (id int, part string, quantity int)

> PARTITIONED BY (int day);

hive> ALTER TABLE supply add PARTITION (day=20110102);

121

hive> ALTER TABLE supply add PARTITION (day=20110103);

hive> ALTER TABLE supply add PARTITION (day=20110102);

hive> .... load data ...

hive> SELECT part,quantity FROM supply

> WHERE day>=20110102 AND day<20110103 AND quantity < 4;

Over Partitioning

The partitioning feature is very useful in Hive. This is because Hive typically performs

full scans over all input to satisfy a query (we’ll leave Hive’s indexing out for this

discussion). However, a design that creates too many partitions may optimize some

queries, but be detrimental for other important queries:

hive> CREATE TABLE weblogs (url string, time long )

> PARTITIONED BY (day int, state string, city string);

hive> SELECT * FROM weblogs WHERE day=20110102;

HDFS was designed for many millions of large files, not billions of small files. The first

drawback of having too many partitions is the large number of Hadoop files and di-

rectories that are created unnecessarily. Each partition corresponds to a directory that

usually contains multiple files. If a given table contains thousands of partitions, it may

have tens of thousands of files, possibly created every day. If the retention of this table

is multiplied over years, it will eventually exhaust the capacity of the NameNode to

manage the filesystem metadata. The NameNode must keep all metadata for the file-

system in memory. While each file requires a small number of bytes for its metadata

(approximately 150 bytes/file), the net effect is to impose an upper limit on the total

number of files that can be managed in an HDFS installation. Other filesystems, like

MapR and Amazon S3 don’t have this limitation.

MapReduce processing converts a job into multiple tasks. In the default case, each task

is a new JVM instance, requiring the overhead of start up and tear down. For small

files, a separate task will be used for each file. In pathological scenarios, the overhead

of JVM start up and tear down can exceed the actual processing time!

Hence, an ideal partition scheme should not result in too many partitions and their

directories, and the files in each directory should be large, some multiple of the file-

system block size.

A good strategy for time-range partitioning, for example, is to determine the approxi-

mate size of your data accumulation over different granularities of time, and start with

the granularity that results in “modest” growth in the number of partitions over time,

while each partition contains files at least on the order of the filesystem block size or

multiples thereof. This balancing keeps the partitions large, which optimizes

throughput for the general case query. Consider when the next level of granularity is

122 | Chapter 9: Schema Design

appropriate, especially if query WHERE clauses typically select ranges of smaller

granularities:

hive> CREATE TABLE weblogs (url string, time long, state string, city string )

> PARTITIONED BY (day int);

hive> SELECT * FROM weblogs WHERE day=20110102;

Another solution is to use two levels of partitions along different dimensions. For ex-

ample, the first partition might be by day and the second-level partition might be by

geographic region, like the state:

hive> CREATE TABLE weblogs (url string, time long, city string )

> PARTITIONED BY (day int, state string);

hive> SELECT * FROM weblogs WHERE day=20110102;

However, since some states will probably result in lots more data than others, you could

see imbalanced map tasks, as processing the larger states takes a lot longer than pro-

cessing the smaller states.

If you can’t find good, comparatively sized partition choices, consider using bucket-

ing as described in “Bucketing Table Data Storage” on page 125.

Unique Keys and Normalization

Relational databases typically use unique keys, indexes, and normalization to store data

sets that fit into memory or mostly into memory. Hive, however, does not have the

concept of primary keys or automatic, sequence-based key generation. Joins should be

avoided in favor of denormalized data, when feasible. The complex types, Array, Map,

and Struct, help by allowing the storage of one-to-many data inside a single row. This

is not to say normalization should never be utilized, but star-schema type designs are

nonoptimal.

The primary reason to avoid normalization is to minimize disk seeks, such as those

typically required to navigate foreign key relations. Denormalizing data permits it to

be scanned from or written to large, contiguous sections of disk drives, which optimizes

I/O performance. However, you pay the penalty of denormalization, data duplication

and the greater risk of inconsistent data.

For example, consider our running example, the employees table. Here it is again with

some changes for clarity:

CREATE TABLE employees (

name STRING,

salary FLOAT,

subordinates ARRAY<STRING>,

deductions MAP<STRING, FLOAT>

address STRUCT<street:STRING, city:STRING, state:STRING, zip:INT>);

The data model of this example breaks the traditional design rules in a few ways.

Unique Keys and Normalization | 123

First, we are informally using name as the primary key, although we all know that names

are often not unique! Ignoring that issue for now, a relational model would have a single

foreign key relation from an employee record to the manager record, using the name

key. We represented this relation the other way around: each employee has an ARRAY

of names of subordinates.

Second, the value for each deduction is unique to the employee, but the map keys are

duplicated data, even if you substitute “flags” (say, integers) for the actual key strings.

A normal relational model would have a separate, two-column table for the deduction

name (or flag) and value, with a one-to-many relationship between the employees and

this deductions table.

Finally, chances are that at least some employees live at the same address, but we are

duplicating the address for each employee, rather than using a one-to-one relationship

to an addresses table.

It’s up to us to manage referential integrity (or deal with the consequences), and to fix

the duplicates of a particular piece of data that has changed. Hive does not give us a

convenient way to UPDATE single records.

Still, when you have 10s of terabytes to many petabytes of data, optimizing speed makes

these limitations worth accepting.

Making Multiple Passes over the Same Data

Hive has a special syntax for producing multiple aggregations from a single pass

through a source of data, rather than rescanning it for each aggregation. This change

can save considerable processing time for large input data sets. We discussed the details

previously in Chapter 5.

For example, each of the following two queries creates a table from the same source

table, history:

hive> INSERT OVERWRITE TABLE sales

> SELECT * FROM history WHERE action='purchased';

hive> INSERT OVERWRITE TABLE credits

> SELECT * FROM history WHERE action='returned';

This syntax is correct, but inefficient. The following rewrite achieves the same thing,

but using a single pass through the source history table:

hive> FROM history

> INSERT OVERWRITE sales SELECT * WHERE action='purchased'

> INSERT OVERWRITE credits SELECT * WHERE action='returned';

The Case for Partitioning Every Table

Many ETL processes involve multiple processing steps. Each step may produce one or

more temporary tables that are only needed until the end of the next job. At first it may

124 | Chapter 9: Schema Design

appear that partitioning these temporary tables is unnecessary. However, imagine a

scenario where a mistake in step’s query or raw data forces a rerun of the ETL process

for several days of input. You will likely need to run the catch-up process a day at a

time in order to make sure that one job does not overwrite the temporary table before

other tasks have completed.

For example, this following design creates an intermediate table by the name

of distinct_ip_in_logs to be used by a subsequent processing step:

$ hive -hiveconf dt=2011-01-01

hive> INSERT OVERWRITE table distinct_ip_in_logs

> SELECT distinct(ip) as ip from weblogs

> WHERE hit_date='${hiveconf:dt}';

hive> CREATE TABLE state_city_for_day (state string,city string);

hive> INSERT OVERWRITE state_city_for_day

> SELECT distinct(state,city) FROM distinct_ip_in_logs

> JOIN geodata ON (distinct_ip_in_logs.ip=geodata.ip);

This approach works, however computing a single day causes the record of the previous

day to be removed via the INSERT OVERWRITE clause. If two instances of this process are

run at once for different days they could stomp on each others’ results.

A more robust approach is to carry the partition information all the way through the

process. This makes synchronization a nonissue. Also, as a side effect, this approach

allows you to compare the intermediate data day over day:

$ hive -hiveconf dt=2011-01-01

hive> INSERT OVERWRITE table distinct_ip_in_logs

> PARTITION (hit_date=${dt})

> SELECT distinct(ip) as ip from weblogs

> WHERE hit_date='${hiveconf:dt}';

hive> CREATE TABLE state_city_for_day (state string,city string)

> PARTITIONED BY (hit_date string);

hive> INSERT OVERWRITE table state_city_for_day PARTITION(${hiveconf:df})

> SELECT distinct(state,city) FROM distinct_ip_in_logs

> JOIN geodata ON (distinct_ip_in_logs.ip=geodata.ip)

> WHERE (hit_date='${hiveconf:dt}');

A drawback of this approach is that you will need to manage the intermediate table

and delete older partitions, but these tasks are easy to automate.

Bucketing Table Data Storage

Partitions offer a convenient way to segregate data and to optimize queries. However,

not all data sets lead to sensible partitioning, especially given the concerns raised earlier

about appropriate sizing.

Bucketing is another technique for decomposing data sets into more manageable parts.

Bucketing Table Data Storage | 125

For example, suppose a table using the date dt as the top-level partition and the

user_id as the second-level partition leads to too many small partitions. Recall that if

you use dynamic partitioning to create these partitions, by default Hive limits the max-

imum number of dynamic partitions that may be created to prevent the extreme case

where so many partitions are created they overwhelm the filesystem’s ability to manage

them and other problems. So, the following commands might fail:

hive> CREATE TABLE weblog (url STRING, source_ip STRING)

> PARTITIONED BY (dt STRING, user_id INT);

hive> FROM raw_weblog

> INSERT OVERWRITE TABLE page_view PARTITION(dt='2012-06-08', user_id)

> SELECT server_name, url, source_ip, dt, user_id;

Instead, if we bucket the weblog table and use user_id as the bucketing column, the

value of this column will be hashed by a user-defined number into buckets. Records

with the same user_id will always be stored in the same bucket. Assuming the number

of users is much greater than the number of buckets, each bucket will have many users:

hive> CREATE TABLE weblog (user_id INT, url STRING, source_ip STRING)

> PARTITIONED BY (dt STRING)

> CLUSTERED BY (user_id) INTO 96 BUCKETS;

However, it is up to you to insert data correctly into the table! The specification in

CREATE TABLE only defines metadata, but has no effect on commands that actually pop-

ulate the table.

This is how to populate the table correctly, when using an INSERT … TABLE statement.

First, we set a property that forces Hive to choose the correct number of reducers cor-

responding to the target table’s bucketing setup. Then we run a query to populate the

partitions. For example:

hive> SET hive.enforce.bucketing = true;

hive> FROM raw_logs

> INSERT OVERWRITE TABLE weblog

> PARTITION (dt='2009-02-25')

> SELECT user_id, url, source_ip WHERE dt='2009-02-25';

If we didn’t use the hive.enforce.bucketing property, we would have to set the number

of reducers to match the number of buckets, using set mapred.reduce.tasks=96. Then

the INSERT query would require a CLUSTER BY clause after the SELECT clause.

As for all table metadata, specifying bucketing doesn’t ensure that the

table is properly populated. Follow the previous example to ensure that

you correctly populate bucketed tables.

Bucketing has several advantages. The number of buckets is fixed so it does not fluc-

tuate with data. Buckets are ideal for sampling. If two tables are bucketed by user_id,

126 | Chapter 9: Schema Design

Hive can create a logically correct sampling. Bucketing also aids in doing efficient map-

side joins, as we discussed in “Map-side Joins” on page 105.

Adding Columns to a Table

Hive allows the definition of a schema over raw data files, unlike many databases that

force the conversion and importation of data following a specific format. A benefit of

this separation of concerns is the ability to adapt a table definition easily when new

columns are added to the data files.

Hive offers the SerDe abstraction, which enables the extraction of data from input. The

SerDe also enables the output of data, though the output feature is not used as fre-

quently because Hive is used primarily as a query mechanism. A SerDe usually parses

from left to right, splitting rows by specified delimiters into columns. The SerDes tend

to be very forgiving. For example, if a row has fewer columns than expected, the missing

columns will be returned as null. If the row has more columns than expected, they will

be ignored. Adding new columns to the schema involves a single ALTER TABLE ADD COL

UMN command. This is very useful as log formats tend to only add more information to

a message:

hive> CREATE TABLE weblogs (version LONG, url STRING)

> PARTITIONED BY (hit_date int)

> ROW FORMAT DELIMITED FIELDS TERMINATED BY '\t';

hive> ! cat log1.txt

1 /mystuff

1 /toys

hive> LOAD DATA LOCAL INPATH 'log1.txt' int weblogs partition(20110101);

hive> SELECT * FROM weblogs;

1 /mystuff 20110101

1 /toys 20110101

Over time a new column may be added to the underlying data. In the following example

the column user_id is added to the data. Note that some older raw data files may not

have this column:

hive> ! cat log2.txt

2 /cars bob

2 /stuff terry

hive> ALTER TABLE weblogs ADD COLUMNS (user_id string);

hive> LOAD DATA LOCAL INPATH 'log2.txt' int weblogs partition(20110102);

hive> SELECT * from weblogs

1 /mystuff 20110101 NULL

1 /toys 20110101 NULL

2 /cars 20110102 bob

2 /stuff 20110102 terry

Adding Columns to a Table | 127

Note that with this approach, columns cannot be added in the beginning or the middle.

Using Columnar Tables

Hive typically uses row-oriented storage, however Hive also has a columnar SerDe that

stores information in a hybrid row-column orientated form. While this format can be

used for any type of data there are some data sets that it is optimal for.

Repeated Data

Given enough rows, fields like state and age will have the same data repeated many

times. This type of data benefits from column-based storage.

state uid age

NY Bob 40

NJ Sara 32

NY Peter 14

NY Sandra 4

Many Columns

The table below has a large number of columns.

state uid age server tz many_more …

NY Bob 40 web1 est stuff

NJ Sara 32 web1 est stuff

NY Peter 14 web3 pst stuff

NY Sandra 4 web45 pst stuff

Queries typically only use a single column or a small set of columns. Column-based

storage will make analyzing the table data faster:

hive> SELECT distinct(state) from weblogs;

You can reference the section “RCFile” on page 202 to see how to use this format.

(Almost) Always Use Compression!

In almost all cases, compression makes data smaller on disk, which usually makes

queries faster by reducing I/O overhead. Hive works seamlessly with many compression

types. The only compelling reason to not use compression is when the data produced

128 | Chapter 9: Schema Design

is intended for use by an external system, and an uncompressed format, such as text,

is the most compatible.

But compression and decompression consumes CPU resources. MapReduce jobs tend

to be I/O bound, so the extra CPU overhead is usually not a problem. However, for

workflows that are CPU intensive, such as some machine-learning algorithms, com-

pression may actually reduce performance by stealing valuable CPU resources from

more essential operations.

See Chapter 11 for more on how to use compression.

(Almost) Always Use Compression! | 129

CHAPTER 10

Tuning

HiveQL is a declarative language where users issue declarative queries and Hive figures

out how to translate them into MapReduce jobs. Most of the time, you don’t need to

understand how Hive works, freeing you to focus on the problem at hand. While the

sophisticated process of query parsing, planning, optimization, and execution is the

result of many years of hard engineering work by the Hive team, most of the time you

can remain oblivious to it.

However, as you become more experienced with Hive, learning about the theory behind

Hive, and the low-level implementation details, will let you use Hive more effectively,

especially where performance optimizations are concerned.

This chapter covers several different topics related to tuning Hive performance. Some

tuning involves adjusting numeric configuration parameters (“turning the knobs”),

while other tuning steps involve enabling or disabling specific features.

Using EXPLAIN

The first step to learning how Hive works (after reading this book…) is to use the

EXPLAIN feature to learn how Hive translates queries into MapReduce jobs.

Consider the following example:

hive> DESCRIBE onecol;

number int

hive> SELECT * FROM onecol;

hive> SELECT SUM(number) FROM onecol;

Now, put the EXPLAIN keyword in front of the last query to see the query plan and other

information. The query will not be executed.

131

hive> EXPLAIN SELECT SUM(number) FROM onecol;

The output requires some explaining and practice to understand.

First, the abstract syntax tree is printed. This shows how Hive parsed the query into

tokens and literals, as part of the first step in turning the query into the ultimate result:

ABSTRACT SYNTAX TREE:

(TOK_QUERY

(TOK_FROM (TOK_TABREF (TOK_TABNAME onecol)))

(TOK_INSERT (TOK_DESTINATION (TOK_DIR TOK_TMP_FILE))

(TOK_SELECT

(TOK_SELEXPR

(TOK_FUNCTION sum (TOK_TABLE_OR_COL number))))))

(The indentation of the actual output was changed to fit the page.)

For those not familiar with parsers and tokenizers, this can look overwhelming. How-

ever, even if you are a novice in this area, you can study the output to get a sense for

what Hive is doing with the SQL statement. (As a first step, ignore the TOK_ prefixes.)

Even though our query will write its output to the console, Hive will actually write the

output to a temporary file first, as shown by this part of the output:

'(TOK_INSERT (TOK_DESTINATION (TOK_DIR TOK_TMP_FILE))'

Next, we can see references to our column name number, our table name onecol, and

the sum function.

A Hive job consists of one or more stages, with dependencies between different stages.

As you might expect, more complex queries will usually involve more stages and more

stages usually requires more processing time to complete.

A stage could be a MapReduce job, a sampling stage, a merge stage, a limit stage, or a

stage for some other task Hive needs to do. By default, Hive executes these stages one

at a time, although later we’ll discuss parallel execution in “Parallel Execu-

tion” on page 136.

Some stages will be short, like those that move files around. Other stages may also

finish quickly if they have little data to process, even though they require a map or

reduce task:

STAGE DEPENDENCIES:

Stage-1 is a root stage

Stage-0 is a root stage

The STAGE PLAN section is verbose and complex. Stage-1 is the bulk of the processing

for this job and happens via a MapReduce job. A TableScan takes the input of the table

and produces a single output column number. The Group By Operator applies the

sum(number) and produces an output column _col0 (a synthesized name for an anony-

mous result). All this is happening on the map side of the job, under the Map Operator

Tree:

132 | Chapter 10: Tuning

Downloa d f r o m W o w ! e B o o k < w w w.woweb o o k . c o m >

STAGE PLANS:

Stage: Stage-1

Map Reduce

Alias -> Map Operator Tree:

onecol

TableScan

alias: onecol

Select Operator

expressions:

expr: number

type: int

outputColumnNames: number

Group By Operator

aggregations:

expr: sum(number)

bucketGroup: false

mode: hash

outputColumnNames: _col0

Reduce Output Operator

sort order:

tag: -1

value expressions:

expr: _col0

type: bigint

On the reduce side, under the Reduce Operator Tree, we see the same Group by Opera

tor but this time it is applying sum on _col0. Finally, in the reducer we see the File

Output Operator, which shows that the output will be text, based on the string output

format: HiveIgnoreKeyTextOutputFormat:

Reduce Operator Tree:

Group By Operator

aggregations:

expr: sum(VALUE._col0)

bucketGroup: false

mode: mergepartial

outputColumnNames: _col0

Select Operator

expressions:

expr: _col0

type: bigint

outputColumnNames: _col0

File Output Operator

compressed: false

GlobalTableId: 0

table:

input format: org.apache.hadoop.mapred.TextInputFormat

output format:

org.apache.hadoop.hive.ql.io.HiveIgnoreKeyTextOutputFormat

Because this job has no LIMIT clause, Stage-0 is a no-op stage:

Stage: Stage-0

Fetch Operator

limit: -1

Using EXPLAIN | 133

Understanding the intricate details of how Hive parses and plans every query is not

useful all of the time. However, it is a nice to have for analyzing complex or poorly

performing queries, especially as we try various tuning steps. We can observe what

effect these changes have at the “logical” level, in tandem with performance measure-

ments.

EXPLAIN EXTENDED

Using EXPLAIN EXTENDED produces even more output. In an effort to “go green,” we

won’t show the entire output, but we will show you the Reduce Operator Tree to

demonstrate the different output:

Reduce Operator Tree:

Group By Operator

aggregations:

expr: sum(VALUE._col0)

bucketGroup: false

mode: mergepartial

outputColumnNames: _col0

Select Operator

expressions:

expr: _col0

type: bigint

outputColumnNames: _col0

File Output Operator

compressed: false

GlobalTableId: 0

directory: file:/tmp/edward/hive_2012-[long number]/-ext-10001

NumFilesPerFileSink: 1

Stats Publishing Key Prefix:

file:/tmp/edward/hive_2012-[long number]/-ext-10001/

table:

input format: org.apache.hadoop.mapred.TextInputFormat

output format:

org.apache.hadoop.hive.ql.io.HiveIgnoreKeyTextOutputFormat

properties:

columns _col0

columns.types bigint

escape.delim \

serialization.format 1

TotalFiles: 1

GatherStats: false

MultiFileSpray: false

We encourage you to compare the two outputs for the Reduce Operator Tree.

Limit Tuning

The LIMIT clause is commonly used, often by people working with the CLI. However,

in many cases a LIMIT clause still executes the entire query, then only returns a handful

134 | Chapter 10: Tuning

of results. Because this behavior is generally wasteful, it should be avoided when

possible. Hive has a configuration property to enable sampling of source data for use

with LIMIT:

<name>hive.limit.optimize.enable</name>

<description>Whether to enable to optimization to

try a smaller subset of data for simple LIMIT first.</description>

</property>

Once the hive.limit.optimize.enable is set to true, two variables control its operation,

hive.limit.row.max.size and hive.limit.optimize.limit.file:

<name>hive.limit.row.max.size</name>

<description>When trying a smaller subset of data for simple LIMIT,

how much size we need to guarantee each row to have at least.

</description>

</property>

<name>hive.limit.optimize.limit.file</name>

<description>When trying a smaller subset of data for simple LIMIT,

maximum number of files we can sample.</description>

</property>

A drawback of this feature is the risk that useful input data will never get processed.

For example, any query that requires a reduce step, such as most JOIN and GROUP BY

operations, most calls to aggregate functions, etc., will have very different results. Per-

haps this difference is okay in many cases, but it’s important to understand.

Optimized Joins

We discussed optimizing join performance in “Join Optimizations” on page 100 and

“Map-side Joins” on page 105. We won’t reproduce the details here, but just remind

yourself that it’s important to know which table is the largest and put it last in the

JOIN clause, or use the /* streamtable(table_name) */ directive.

If all but one table is small enough, typically to fit in memory, then Hive can perform

a map-side join, eliminating the need for reduce tasks and even some map tasks. Some-

times even tables that do not fit in memory are good candidates because removing the

reduce phase outweighs the cost of bringing semi-large tables into each map tasks.

Local Mode

Many Hadoop jobs need the full scalability benefits of Hadoop to process large data

sets. However, there are times when the input to Hive is very small. In these cases, the

Local Mode | 135

overhead of launching tasks for queries consumes a significant percentage of the overall

job execution time. In many of these cases, Hive can leverage the lighter weight of the

local mode to perform all the tasks for the job on a single machine and sometimes in

the same process. The reduction in execution times can be dramatic for small data sets.

You can explicitly enable local mode temporarily, as in this example:

hive> set oldjobtracker=${hiveconf:mapred.job.tracker};

hive> set mapred.job.tracker=local;

hive> set mapred.tmp.dir=/home/edward/tmp;

hive> SELECT * from people WHERE firstname=bob;

...

hive> set mapred.job.tracker=${oldjobtracker};

You can also tell Hive to automatically apply this optimization by setting

hive.exec.mode.local.auto to true, perhaps in your $HOME/.hiverc.

To set this property permanently for all users, change the value in your $HIVE_HOME/

conf/hive-site.xml:

<name>hive.exec.mode.local.auto</name>

Let hive determine whether to run in local mode automatically

</description>

</property>

Parallel Execution

Hive converts a query into one or more stages. Stages could be a MapReduce stage, a

sampling stage, a merge stage, a limit stage, or other possible tasks Hive needs to do.

By default, Hive executes these stages one at a time. However, a particular job may

consist of some stages that are not dependent on each other and could be executed in

parallel, possibly allowing the overall job to complete more quickly. However, if more

stages are run simultaneously, the job may complete much faster.

Setting hive.exec.parallel to true enables parallel execution. Be careful in a shared

cluster, however. If a job is running more stages in parallel, it will increase its cluster

utilization:

<name>hive.exec.parallel</name>

<description>Whether to execute jobs in parallel</description>

</property>

136 | Chapter 10: Tuning

Strict Mode

Strict mode is a setting in Hive that prevents users from issuing queries that could have

unintended and undesirable effects.

Setting the property hive.mapred.mode to strict disables three types of queries.

First, queries on partitioned tables are not permitted unless they include a partition

filter in the WHERE clause, limiting their scope. In other words, you’re prevented from

queries that will scan all partitions. The rationale for this limitation is that partitioned

tables often hold very large data sets that may be growing rapidly. An unrestricted

partition could consume unacceptably large resources over such a large table:

hive> SELECT DISTINCT(planner_id) FROM fracture_ins WHERE planner_id=5;

FAILED: Error in semantic analysis: No Partition Predicate Found for

Alias "fracture_ins" Table "fracture_ins"

The following enhancement adds a partition filter—the table partitions—to the

WHERE clause:

hive> SELECT DISTINCT(planner_id) FROM fracture_ins

> WHERE planner_id=5 AND hit_date=20120101;

... normal results ...

The second type of restricted query are those with ORDER BY clauses, but no LIMIT clause.

Because ORDER BY sends all results to a single reducer to perform the ordering, forcing

the user to specify a LIMIT clause prevents the reducer from executing for an extended

period of time:

hive> SELECT * FROM fracture_ins WHERE hit_date>2012 ORDER BY planner_id;

FAILED: Error in semantic analysis: line 1:56 In strict mode,

limit must be specified if ORDER BY is present planner_id

To issue this query, add a LIMIT clause:

hive> SELECT * FROM fracture_ins WHERE hit_date>2012 ORDER BY planner_id

> LIMIT 100000;

... normal results ...

The third and final type of query prevented is a Cartesian product. Users coming from

the relational database world may expect that queries that perform a JOIN not with an

ON clause but with a WHERE clause will have the query optimized by the query planner,

effectively converting the WHERE clause into an ON clause. Unfortunately, Hive does not

perform this optimization, so a runaway query will occur if the tables are large:

hive> SELECT * FROM fracture_act JOIN fracture_ads

> WHERE fracture_act.planner_id = fracture_ads.planner_id;

FAILED: Error in semantic analysis: In strict mode, cartesian product

is not allowed. If you really want to perform the operation,

+set hive.mapred.mode=nonstrict+

Here is a properly constructed query with JOIN and ON clauses:

Strict Mode | 137

hive> SELECT * FROM fracture_act JOIN fracture_ads

> ON (fracture_act.planner_id = fracture_ads.planner_id);

... normal results ...

Tuning the Number of Mappers and Reducers

Hive is able to parallelize queries by breaking the query into one or more MapReduce

jobs. Each of which might have multiple mapper and reducer tasks, at least some of

which can run in parallel. Determining the optimal number of mappers and reducers

depends on many variables, such as the size of the input and the operation being per-

formed on the data.

A balance is required. Having too many mapper or reducer tasks causes excessive over-

head in starting, scheduling, and running the job, while too few tasks means the

inherent parallelism of the cluster is underutilized.

When running a Hive query that has a reduce phase, the CLI prints information about

how the number of reducers can be tuned. Let’s see an example that uses a GROUP BY

query, because they always require a reduce phase. In contrast, many other queries are

converted into map-only jobs:

hive> SELECT pixel_id, count FROM fracture_ins WHERE hit_date=20120119

> GROUP BY pixel_id;

Total MapReduce jobs = 1

Launching Job 1 out of 1

Number of reduce tasks not specified. Estimated from input data size: 3

In order to change the average load for a reducer (in bytes):

set hive.exec.reducers.bytes.per.reducer=<number>

In order to limit the maximum number of reducers:

set hive.exec.reducers.max=<number>

In order to set a constant number of reducers:

set mapred.reduce.tasks=<number>

...

Hive is determining the number of reducers from the input size. This can be confirmed

using the dfs -count command, which works something like the Linux du -s command;

it computes a total size for all the data under a given directory:

[edward@etl02 ~]$ hadoop dfs -count /user/media6/fracture/ins/* | tail -4

1 8 2614608737 hdfs://.../user/media6/fracture/ins/hit_date=20120118

1 7 2742992546 hdfs://.../user/media6/fracture/ins/hit_date=20120119

1 17 2656878252 hdfs://.../user/media6/fracture/ins/hit_date=20120120

1 2 362657644 hdfs://.../user/media6/fracture/ins/hit_date=20120121

(We’ve reformatted the output and elided some details for space.)

The default value of hive.exec.reducers.bytes.per.reducer is 1 GB. Changing this

value to 750 MB causes Hive to estimate four reducers for this job:

hive> set hive.exec.reducers.bytes.per.reducer=750000000;

hive> SELECT pixel_id,count(1) FROM fracture_ins WHERE hit_date=20120119

> GROUP BY pixel_id;

138 | Chapter 10: Tuning

Total MapReduce jobs = 1

Launching Job 1 out of 1

Number of reduce tasks not specified. Estimated from input data size: 4

...

This default typically yields good results. However, there are cases where a query’s map

phase will create significantly more data than the input size. In the case of excessive

map phase data, the input size of the default might be selecting too few reducers. Like-

wise the map function might filter a large portion of the data from the data set and then

fewer reducers may be justified.

A quick way to experiment is by setting the number of reducers to a fixed size, rather

than allowing Hive to calculate the value. If you remember, the Hive default estimate

is three reducers. Set mapred.reduce.tasks to different numbers and determine if more

or fewer reducers results in faster run times. Remember that benchmarking like this is

complicated by external factors such as other users running jobs simultaneously. Ha-

doop has a few seconds overhead to start up and schedule map and reduce tasks. When

executing performance tests, it’s important to keep these factors in mind, especially if

the jobs are small.

The hive.exec.reducers.max property is useful for controlling resource utilization on

shared clusters when dealing with large jobs. A Hadoop cluster has a fixed number of

map and reduce “slots” to allocate to tasks. One large job could reserve all of the slots

and block other jobs from starting. Setting hive.exec.reducers.max can stop a query

from taking too many reducer resources. It is a good idea to set this value in your

$HIVE_HOME/conf/hive-site.xml. A suggested formula is to set the value to the result

of this calculation:

(Total Cluster Reduce Slots * 1.5) / (avg number of queries running)

The 1.5 multiplier is a fudge factor to prevent underutilization of the cluster.

JVM Reuse

JVM reuse is a Hadoop tuning parameter that is very relevant to Hive performance,

especially scenarios where it’s hard to avoid small files and scenarios with lots of tasks,

most which have short execution times.

The default configuration of Hadoop will typically launch map or reduce tasks in a

forked JVM. The JVM start-up may create significant overhead, especially when

launching jobs with hundreds or thousands of tasks. Reuse allows a JVM instance to

be reused up to N times for the same job. This value is set in Hadoop’s mapred-

site.xml (in $HADOOP_HOME/conf):

<name>mapred.job.reuse.jvm.num.tasks</name>

<description>How many tasks to run per jvm. If set to -1, there is no limit.

JVM Reuse | 139

</description>

</property>

A drawback of this feature is that JVM reuse will keep reserved task slots open until

the job completes, in case they are needed for reuse. If an “unbalanced” job has some

reduce tasks that run considerably longer than the others, the reserved slots will sit idle,

unavailable for other jobs, until the last task completes.

Indexes

Indexes may be used to accelerate the calculation speed of a GROUP BY query.

Hive contains an implementation of bitmap indexes since v0.8.0. The main use case

for bitmap indexes is when there are comparatively few values for a given column. See

“Bitmap Indexes” on page 118 for more information.

Dynamic Partition Tuning

As explained in “Dynamic Partition Inserts” on page 74, dynamic partition INSERT

statements enable a succinct SELECT statement to create many new partitions for inser-

tion into a partitioned table.

This is a very powerful feature, however if the number of partitions is high, a large

number of output handles must be created on the system. This is a somewhat uncom-

mon use case for Hadoop, which typically creates a few files at once and streams large

amounts of data to them.

Out of the box, Hive is configured to prevent dynamic partition inserts from creating

more than 1,000 or so partitions. While it can be bad for a table to have too many

partitions, it is generally better to tune this setting to the larger value and allow these

queries to work.

First, it is always good to set the dynamic partition mode to strict in your hive-

site.xml, as discussed in “Strict Mode” on page 137. When strict mode is on, at

least one partition has to be static, as demonstrated in “Dynamic Partition In-

serts” on page 74:

<name>hive.exec.dynamic.partition.mode</name>

<value>strict</value>

<description>In strict mode, the user must specify at least one

static partition in case the user accidentally overwrites all

partitions.</description>

</property>

Then, increase the other relevant properties to allow queries that will create a large

number of dynamic partitions, for example:

<name>hive.exec.max.dynamic.partitions</name>

140 | Chapter 10: Tuning

<description>Maximum number of dynamic partitions allowed to be

created in total.</description>

</property>

<name>hive.exec.max.dynamic.partitions.pernode</name>

<description>Maximum number of dynamic partitions allowed to be

created in each mapper/reducer node.</description>

</property>

Another setting controls how many files a DataNode will allow to be open at once. It

must be set in the DataNode’s $HADOOP_HOME/conf/hdfs-site.xml.

In Hadoop v0.20.2, the default value is 256, which is too low. The value affects the

number of maximum threads and resources, so setting it to a very high number is not

recommended. Note also that in Hadoop v0.20.2, changing this variable requires re-

starting the DataNode to take effect:

<name>dfs.datanode.max.xcievers</name>

</property>

Speculative Execution

Speculative execution is a feature of Hadoop that launches a certain number of dupli-

cate tasks. While this consumes more resources computing duplicate copies of data

that may be discarded, the goal of this feature is to improve overall job progress by

getting individual task results faster, and detecting then black-listing slow-running

TaskTrackers.

Hadoop speculative execution is controlled in the $HADOOP_HOME/conf/mapred-

site.xml file by the following two variables:

<name>mapred.map.tasks.speculative.execution</name>

<description>If true, then multiple instances of some map tasks

may be executed in parallel.</description>

</property>

<name>mapred.reduce.tasks.speculative.execution</name>

<description>If true, then multiple instances of some reduce tasks

may be executed in parallel.</description>

</property>

However, Hive provides its own variable to control reduce-side speculative execution:

Speculative Execution | 141

<name>hive.mapred.reduce.tasks.speculative.execution</name>

<description>Whether speculative execution for

reducers should be turned on. </description>

</property>

It is hard to give a concrete recommendation about tuning these speculative execution

variables. If you are very sensitive to deviations in runtime, you may wish to turn these

features on. However, if you have long-running map or reduce tasks due to large

amounts of input, the waste could be significant.

Single MapReduce MultiGROUP BY

Another special optimization attempts to combine multiple GROUP BY operations in a

query into a single MapReduce job. For this optimization to work, a common set of

GROUP BY keys is required:

<name>hive.multigroupby.singlemr</name>

<value>false</value>

<description>Whether to optimize multi group by query to generate single M/R

job plan. If the multi group by query has common group by keys, it will be

optimized to generate single M/R job.</description>

</property>

Virtual Columns

Hive provides two virtual columns: one for the input filename for split and the other

for the block offset in the file. These are helpful when diagnosing queries where Hive

is producing unexpected or null results. By projecting these “columns,” you can see

which file and row is causing problems:

hive> set hive.exec.rowoffset=true;

hive> SELECT INPUT__FILE__NAME, BLOCK__OFFSET__INSIDE__FILE, line

> FROM hive_text WHERE line LIKE '%hive%' LIMIT 2;

har://file/user/hive/warehouse/hive_text/folder=docs/

data.har/user/hive/warehouse/hive_text/folder=docs/README.txt 2243

http://hive.apache.org/

har://file/user/hive/warehouse/hive_text/folder=docs/

data.har/user/hive/warehouse/hive_text/folder=docs/README.txt 3646

- Hive 0.8.0 ignores the hive-default.xml file, though we continue

(We wrapped the long output and put a blank line between the two output rows.)

A third virtual column provides the row offset of the file. It must be enabled explicitly:

<name>hive.exec.rowoffset</name>

142 | Chapter 10: Tuning

<description>Whether to provide the row offset virtual column</description>

</property>

Now it can be used in queries:

hive> SELECT INPUT__FILE__NAME, BLOCK__OFFSET__INSIDE__FILE,

> ROW__OFFSET__INSIDE__BLOCK

> FROM hive_text WHERE line LIKE '%hive%' limit 2;

file:/user/hive/warehouse/hive_text/folder=docs/README.txt 2243 0

file:/user/hive/warehouse/hive_text/folder=docs/README.txt 3646 0

Virtual Columns | 143

CHAPTER 11

Other File Formats and Compression

One of Hive’s unique features is that Hive does not force data to be converted to a

specific format. Hive leverages Hadoop’s InputFormat APIs to read data from a variety

of sources, such as text files, sequence files, or even custom formats. Likewise, the

OutputFormat API is used to write data to various formats.

While Hadoop offers linear scalability in file storage for uncompressed data, storing

data in compressed form has many benefits. Compression typically saves significant

disk storage; for example, text-based files may compress 40% or more. Compression

also can increase throughput and performance. This may seem counterintuitive be-

cause compressing and decompressing data incurs extra CPU overhead, however, the

I/O savings resulting from moving fewer bytes into memory can result in a net perfor-

mance gain.

Hadoop jobs tend to be I/O bound, rather than CPU bound. If so, compression will

improve performance. However, if your jobs are CPU bound, then compression will

probably lower your performance. The only way to really know is to experiment with

different options and measure the results.

Determining Installed Codecs

Based on your Hadoop version, different codecs will be available to you. The set feature

in Hive can be used to display the value of hiveconf or Hadoop configuration values.

The codecs available are in a comma-separated list named io.compression.codec:

# hive -e "set io.compression.codecs"

io.compression.codecs=org.apache.hadoop.io.compress.GzipCodec,

org.apache.hadoop.io.compress.DefaultCodec,

org.apache.hadoop.io.compress.BZip2Codec,

org.apache.hadoop.io.compress.SnappyCodec

145

Choosing a Compression Codec

Using compression has the advantage of minimizing the disk space required for files

and the overhead of disk and network I/O. However, compressing and decompressing

files increases the CPU overhead. Therefore, compression is best used for I/O-bound

jobs, where there is extra CPU capacity, or when disk space is at a premium.

All recent versions of Hadoop have built-in support for the GZip and BZip2 compres-

sion schemes, including native Linux libraries that accelerate compression and decom-

pression for these formats. Bundled support for Snappy compression was recently

added, but if your version of Hadoop doesn’t support it, you can add the appropriate

libraries yourself.1 Finally, LZO compression is often used.2

So, why do we need different compression schemes? Each scheme makes a trade-off

between speed and minimizing the size of the compressed output. BZip2 creates the

smallest compressed output, but with the highest CPU overhead. GZip is next in terms

of compressed size versus speed. Hence, if disk space utilization and I/O overhead are

concerns, both are attractive choices.

LZO and Snappy create larger files but are much faster, especially for decompression.

They are good choices if disk space and I/O overhead are less important than rapid

decompression of frequently read data.

Another important consideration is whether or not the compression format is splitta-

ble. MapReduce wants to split very large input files into splits (often one split per file-

system block, i.e., a multiple of 64 MB), where each split is sent to a separate map

process. This can only work if Hadoop knows the record boundaries in the file. In text

files, each line is a record, but these boundaries are obscured by GZip and Snappy.

However, BZip2 and LZO provide block-level compression, where each block has

complete records, so Hadoop can split these files on block boundaries.

The desire for splittable files doesn’t rule out GZip and Snappy. When you create your

data files, you could partition them so that they are approximately the desired size.

Typically the number of output files is equal to the number of reducers. If you are using

N reducers you typically get N output files. Be careful, if you have a large nonsplittable

file, a single task will have to read the entire file beginning to end.

There’s much more we could say about compression, but instead we’ll refer you to

Hadoop: The Definitive Guide by Tom White (O’Reilly) for more details, and we’ll focus

now on how to tell Hive what format you’re using.

From Hive’s point of view, there are two aspects to the file format. One aspect is how

the file is delimited into rows (records). Text files use \n (linefeed) as the default row

delimiter. When you aren’t using the default text file format, you tell Hive the name of

1. See http://code.google.com/p/hadoop-snappy/.

2. See http://wiki.apache.org/hadoop/UsingLzoCompression.

146 | Chapter 11: Other File Formats and Compression

an InputFormat and an OutputFormat to use. Actually, you will specify the names of Java

classes that implement these formats. The InputFormat knows how to read splits and

partition them into records, and the OutputFormat knows how to write these splits back

to files or console output.

The second aspect is how records are partitioned into fields (or columns). Hive uses

^A by default to separate fields in text files. Hive uses the name SerDe, which is short

for serializer/deserializer for the “module” that partitions incoming records (the deser-

ializer) and also knows how to write records in this format (the serializer). This time

you will specify a single Java class that performs both jobs.

All this information is specified as part of the table definition when you create the table.

After creation, you query the table as you normally would, agnostic to the underlying

format. Hence, if you’re a user of Hive, but not a Java developer, don’t worry about

the Java aspects. The developers on your team will help you specify this information

when needed, after which you’ll work as you normally do.

Enabling Intermediate Compression

Intermediate compression shrinks the data shuffled between the map and reduce tasks

for a job. For intermediate compression, choosing a codec that has lower CPU cost is

typically more important than choosing a codec that results in the most compression.

The property hive.exec.compress.intermediate defaults to false and should be set to

true by default:

<name>hive.exec.compress.intermediate</name>

<description> This controls whether intermediate files produced by Hive between

multiple map-reduce jobs are compressed. The compression codec and other options

are determined from hadoop config variables mapred.output.compress* </description>

</property>

The property that controls intermediate compression for other Hadoop

jobs is mapred.compress.map.output.

Hadoop compression has a DefaultCodec. Changing the codec involves setting the

mapred.map.output.compression.codec property. This is a Hadoop variable and can be

set in the $HADOOP_HOME/conf/mapred-site.xml or the $HADOOP_HOME/conf/

hive-site.xml. SnappyCodec is a good choice for intermediate compression because it

combines good compression performance with low CPU cost:

<name>mapred.map.output.compression.codec</name>

<value>org.apache.hadoop.io.compress.SnappyCodec</value>

<description> This controls whether intermediate files produced by Hive

Enabling Intermediate Compression | 147

between multiple map-reduce jobs are compressed. The compression codec

and other options are determined from hadoop config variables

mapred.output.compress* </description>

</property>

Final Output Compression

When Hive writes output to a table, that content can also be compressed. The property

hive.exec.compress.output controls this feature. You may wish to leave this value set

to false in the global configuration file, so that the default output is uncompressed

clear text. Users can turn on final compression by setting the property to true on a

query-by-query basis or in their scripts:

<name>hive.exec.compress.output</name>

<value>false</value>

<description> This controls whether the final outputs of a query

(to a local/hdfs file or a Hive table) is compressed. The compression

codec and other options are determined from hadoop config variables

mapred.output.compress* </description>

</property>

The property that controls final compression for other Hadoop jobs is

mapred.output.compress.

If hive.exec.compress.output is set true, a codec can be chosen. GZip compression is

a good choice for output compression because it typically reduces the size of files sig-

nificantly, but remember that GZipped files aren’t splittable by subsequent MapReduce

jobs:

<name>mapred.output.compression.codec</name>

<value>org.apache.hadoop.io.compress.GzipCodec</value>

<description>If the job outputs are compressed, how should they be compressed?

</description>

</property>

Sequence Files

Compressing files results in space savings but one of the downsides of storing raw

compressed files in Hadoop is that often these files are not splittable. Splittable files

can be broken up and processed in parts by multiple mappers in parallel. Most com-

pressed files are not splittable because you can only start reading from the beginning.

The sequence file format supported by Hadoop breaks a file into blocks and then op-

tionally compresses the blocks in a splittable way.

148 | Chapter 11: Other File Formats and Compression

To use sequence files from Hive, add the STORED AS SEQUENCEFILE clause to a CREATE

TABLE statement:

CREATE TABLE a_sequence_file_table STORED AS SEQUENCEFILE;

Sequence files have three different compression options: NONE, RECORD, and BLOCK.

RECORD is the default. However, BLOCK compression is usually more efficient and it still

provides the desired splittability. Like many other compression properties, this one is

not Hive-specific. It can be defined in Hadoop’s mapred-site.xml file, in Hive’s hive-

site.xml, or as needed in scripts or before individual queries:

<name>mapred.output.compression.type</name>

<value>BLOCK</value>

<description>If the job outputs are to compressed as SequenceFiles,

how should they be compressed? Should be one of NONE, RECORD or BLOCK.

</description>

</property>

Compression in Action

We have introduced a number of compression-related properties in Hive, and different

permutations of these options result in different output. Let’s use these properties in

some examples and show what they produce. Remember that variables set by the CLI

persist across the rest of the queries in the session, so between examples you should

revert the settings or simply restart the Hive session:

hive> SELECT * FROM a;

4 5

3 2

hive> DESCRIBE a;

a int

b int

First, let’s enable intermediate compression. This won’t affect the final output, however

the job counters will show less physical data transferred for the job, since the shuffle

sort data was compressed:

hive> set hive.exec.compress.intermediate=true;

hive> CREATE TABLE intermediate_comp_on

> ROW FORMAT DELIMITED FIELDS TERMINATED BY '\t'

> AS SELECT * FROM a;

Moving data to: file:/user/hive/warehouse/intermediate_comp_on

Table default.intermediate_comp_on stats: [num_partitions: 0, num_files: 1,

num_rows: 2, total_size: 8, raw_data_size: 6]

...

As expected, intermediate compression did not affect the final output, which remains

uncompressed:

hive> dfs -ls /user/hive/warehouse/intermediate_comp_on;

Found 1 items

Compression in Action | 149

/user/hive/warehouse/intermediate_comp_on/000000_0

hive> dfs -cat /user/hive/warehouse/intermediate_comp_on/000000_0;

4 5

3 2

We can also chose an intermediate compression codec other then the default codec. In

this case we chose GZIP, although Snappy is normally a better option. The first line is

wrapped for space:

hive> set mapred.map.output.compression.codec

=org.apache.hadoop.io.compress.GZipCodec;

hive> set hive.exec.compress.intermediate=true;

hive> CREATE TABLE intermediate_comp_on_gz

> ROW FORMAT DELIMITED FIELDS TERMINATED BY '\t'

> AS SELECT * FROM a;

Moving data to: file:/user/hive/warehouse/intermediate_comp_on_gz

Table default.intermediate_comp_on_gz stats:

[num_partitions: 0, num_files: 1, num_rows: 2, total_size: 8, raw_data_size: 6]

hive> dfs -cat /user/hive/warehouse/intermediate_comp_on_gz/000000_0;

4 5

3 2

Next, we can enable output compression:

hive> set hive.exec.compress.output=true;

hive> CREATE TABLE final_comp_on

> ROW FORMAT DELIMITED FIELDS TERMINATED BY '\t'

> AS SELECT * FROM a;

Moving data to: file:/tmp/hive-edward/hive_2012-01-15_11-11-01_884_.../-ext-10001

Moving data to: file:/user/hive/warehouse/final_comp_on

Table default.final_comp_on stats:

[num_partitions: 0, num_files: 1, num_rows: 2, total_size: 16, raw_data_size: 6]

hive> dfs -ls /user/hive/warehouse/final_comp_on;

Found 1 items

/user/hive/warehouse/final_comp_on/000000_0.deflate

The output table statistics show that the total_size is 16, but the raw_data_size is 6.

The extra space is overhead for the deflate algorithm. We can also see the output file

is named .deflate.

Trying to cat the file is not suggested, as you get binary output. However, Hive can

query this data normally:

hive> dfs -cat /user/hive/warehouse/final_comp_on/000000_0.deflate;

... UGLYBINARYHERE ...

hive> SELECT * FROM final_comp_on;

4 5

3 2

150 | Chapter 11: Other File Formats and Compression

This ability to seamlessly work with compressed files is not Hive-specific; Hadoop’s

TextInputFormat is at work here. While the name is confusing in this case, TextInput

Format understands file extensions such as .deflate or .gz and decompresses these files

on the fly. Hive is unaware if the underlying files are uncompressed or compressed

using any of the supported compression schemes.

Let’s change the codec used by output compression to see the results (another line wrap

for space):

hive> set hive.exec.compress.output=true;

hive> set mapred.output.compression.codec

=org.apache.hadoop.io.compress.GzipCodec;

hive> CREATE TABLE final_comp_on_gz

> ROW FORMAT DELIMITED FIELDS TERMINATED BY '\t'

> AS SELECT * FROM a;

Moving data to: file:/user/hive/warehouse/final_comp_on_gz

Table default.final_comp_on_gz stats:

[num_partitions: 0, num_files: 1, num_rows: 2, total_size: 28, raw_data_size: 6]

hive> dfs -ls /user/hive/warehouse/final_comp_on_gz;

Found 1 items

/user/hive/warehouse/final_comp_on_gz/000000_0.gz

As you can see, the output folder now contains zero or more .gz files. Hive has a quick

hack to execute local commands like zcat from inside the Hive shell. The ! tells Hive

to fork and run the external command and block until the system returns a result.

zcat is a command-line utility that decompresses and displays output:

hive> ! /bin/zcat /user/hive/warehouse/final_comp_on_gz/000000_0.gz;

4 5

3 2

hive> SELECT * FROM final_comp_on_gz;

4 5

3 2

Time taken: 0.159 seconds

Using output compression like this results in binary compressed files that are small

and, as a result, operations on them are very fast. However, recall that the number of

output files is a side effect of how many mappers or reducers processed the data. In the

worst case scenario, you can end up with one large binary file in a directory that is not

splittable. This means that subsequent steps that have to read this data cannot work

in parallel. The answer to this problem is to use sequence files:

hive> set mapred.output.compression.type=BLOCK;

hive> set hive.exec.compress.output=true;

hive> set mapred.output.compression.codec=org.apache.hadoop.io.compress.GzipCodec;

hive> CREATE TABLE final_comp_on_gz_seq

> ROW FORMAT DELIMITED FIELDS TERMINATED BY '\t'

> STORED AS SEQUENCEFILE

> AS SELECT * FROM a;

Compression in Action | 151

Moving data to: file:/user/hive/warehouse/final_comp_on_gz_seq

Table default.final_comp_on_gz_seq stats:

[num_partitions: 0, num_files: 1, num_rows: 2, total_size: 199, raw_data_size: 6]

hive> dfs -ls /user/hive/warehouse/final_comp_on_gz_seq;

Found 1 items

/user/hive/warehouse/final_comp_on_gz_seq/000000_0

Sequence files are binary. But it is a nice exercise to see the header. To confirm the

results are what was intended (output wrapped):

hive> dfs -cat /user/hive/warehouse/final_comp_on_gz_seq/000000_0;

SEQ[]org.apache.hadoop.io.BytesWritable[]org.apache.hadoop.io.BytesWritable[]

org.apache.hadoop.io.compress.GzipCodec[]

Because of the meta-information embedded in the sequence file and in the Hive met-

astore, Hive can query the table without any specific settings. Hadoop also offers the

dfs -text command to strip the header and compression away from sequence files and

return the raw result:

hive> dfs -text /user/hive/warehouse/final_comp_on_gz_seq/000000_0;

4 5

3 2

hive> select * from final_comp_on_gz_seq;

4 5

3 2

Finally, let’s use intermediate and output compression at the same time and set different

compression codecs for each while saving the final output to sequence files! These

settings are commonly done for production environments where data sets are large and

such settings improve performance:

hive> set mapred.map.output.compression.codec

=org.apache.hadoop.io.compress.SnappyCodec;

hive> set hive.exec.compress.intermediate=true;

hive> set mapred.output.compression.type=BLOCK;

hive> set hive.exec.compress.output=true;

hive> set mapred.output.compression.codec

=org.apache.hadoop.io.compress.GzipCodec;

hive> CREATE TABLE final_comp_on_gz_int_compress_snappy_seq

> ROW FORMAT DELIMITED FIELDS TERMINATED BY '\t'

> STORED AS SEQUENCEFILE AS SELECT * FROM a;

Archive Partition

Hadoop has a format for storage known as HAR, which stands for Hadoop ARchive. A

HAR file is like a TAR file that lives in the HDFS filesystem as a single file. However,

internally it can contain multiple files and directories. In some use cases, older direc-

tories and files are less commonly accessed than newer files. If a particular partition

contains thousands of files it will require significant overhead to manage it in the HDFS

152 | Chapter 11: Other File Formats and Compression

NameNode. By archiving the partition it is stored as a single, large file, but it can still

be accessed by hive. The trade-off is that HAR files will be less efficient to query. Also,

HAR files are not compressed, so they don’t save any space.

In the following example, we’ll use Hive’s own documentation as data.

First, create a partitioned table and load it with the text data from the Hive package:

hive> CREATE TABLE hive_text (line STRING) PARTITIONED BY (folder STRING);

hive> ! ls $HIVE_HOME;

LICENSE

README.txt

RELEASE_NOTES.txt

hive> ALTER TABLE hive_text ADD PARTITION (folder='docs');

hive> LOAD DATA INPATH '${env:HIVE_HOME}/README.txt'

> INTO TABLE hive_text PARTITION (folder='docs');

Loading data to table default.hive_text partition (folder=docs)

hive> LOAD DATA INPATH '${env:HIVE_HOME}/RELEASE_NOTES.txt'

> INTO TABLE hive_text PARTITION (folder='docs');

Loading data to table default.hive_text partition (folder=docs)

hive> SELECT * FROM hive_text WHERE line LIKE '%hive%' LIMIT 2;

http://hive.apache.org/ docs

- Hive 0.8.0 ignores the hive-default.xml file, though we continue docs

Some versions of Hadoop, such as Hadoop v0.20.2, will require the JAR containing

the Hadoop archive tools to be placed on the Hive auxlib:

$ mkdir $HIVE_HOME/auxlib

$ cp $HADOOP_HOME/hadoop-0.20.2-tools.jar $HIVE_HOME/auxlib/

Take a look at the underlying structure of the table, before we archive it. Note the

location of the table’s data partition, since it’s a managed, partitioned table:

hive> dfs -ls /user/hive/warehouse/hive_text/folder=docs;

Found 2 items

/user/hive/warehouse/hive_text/folder=docs/README.txt

/user/hive/warehouse/hive_text/folder=docs/RELEASE_NOTES.txt

The ALTER TABLE ... ARCHIVE PARTITION statement converts the table into an archived

table:

hive> SET hive.archive.enabled=true;

hive> ALTER TABLE hive_text ARCHIVE PARTITION (folder='docs');

intermediate.archived is

file:/user/hive/warehouse/hive_text/folder=docs_INTERMEDIATE_ARCHIVED

intermediate.original is

file:/user/hive/warehouse/hive_text/folder=docs_INTERMEDIATE_ORIGINAL

Creating data.har for file:/user/hive/warehouse/hive_text/folder=docs

in file:/tmp/hive-edward/hive_..._3862901820512961909/-ext-10000/partlevel

Please wait... (this may take a while)

Moving file:/tmp/hive-edward/hive_..._3862901820512961909/-ext-10000/partlevel

Archive Partition | 153

Downloa d f r o m W o w ! e B o o k < w w w.woweb o o k . c o m >

to file:/user/hive/warehouse/hive_text/folder=docs_INTERMEDIATE_ARCHIVED

Moving file:/user/hive/warehouse/hive_text/folder=docs

to file:/user/hive/warehouse/hive_text/folder=docs_INTERMEDIATE_ORIGINAL

Moving file:/user/hive/warehouse/hive_text/folder=docs_INTERMEDIATE_ARCHIVED

to file:/user/hive/warehouse/hive_text/folder=docs

(We reformatted the output slightly so it would fit, and used ... to replace two time-

stamp strings in the original output.)

The underlying table has gone from two files to one Hadoop archive (HAR file):

hive> dfs -ls /user/hive/warehouse/hive_text/folder=docs;

Found 1 items

/user/hive/warehouse/hive_text/folder=docs/data.har

The ALTER TABLE ... UNARCHIVE PARTITION command extracts the files from the HAR

and puts them back into HDFS:

ALTER TABLE hive_text UNARCHIVE PARTITION (folder='docs');

Compression: Wrapping Up

Hive’s ability to read and write different types of compressed files is a big performance

win as it saves disk space and processing overhead. This flexibility also aids in integra-

tion with other tools, as Hive can query many native file types without the need to write

custom “adapters” in Java.

154 | Chapter 11: Other File Formats and Compression

CHAPTER 12

Developing

Hive won’t provide everything you could possibly need. Sometimes a third-party library

will fill a gap. At other times, you or someone else who is a Java developer will need to

write user-defined functions (UDFs; see Chapter 13), SerDes (see “Record Formats:

SerDes” on page 205), input and/or output formats (see Chapter 15), or other

enhancements.

This chapter explores working with the Hive source code itself, including the new

Plugin Developer Kit introduced in Hive v0.8.0.

Changing Log4J Properties

Hive can be configured with two separate Log4J configuration files found in

$HIVE_HOME/conf. The hive-log4j.properties file controls the logging of the CLI or

other locally launched components. The hive-exec-log4j.properties file controls the log-

ging inside the MapReduce tasks. These files do not need to be present inside the Hive

installation because the default properties come built inside the Hive JARs. In fact, the

actual files in the conf directory have the .template extension, so they are ignored by

default. To use either of them, copy it with a name that removes the .template extension

and edit it to taste:

$ cp conf/hive-log4j.properties.template conf/hive-log4j.properties

$ ... edit file ...

It is also possible to change the logging configuration of Hive temporarily without

copying and editing the Log4J files. The hiveconf switch can be specified on start-up

with definitions of any properties in the log4.properties file. For example, here we set

the default logger to the DEBUG level and send output to the console appender:

$ bin/hive -hiveconf hive.root.logger=DEBUG,console

12/03/27 08:46:01 WARN conf.HiveConf: hive-site.xml not found on CLASSPATH

12/03/27 08:46:01 DEBUG conf.Configuration: java.io.IOException: config()

155

Connecting a Java Debugger to Hive

When enabling more verbose output does not help find the solution to the problem

you are troubleshooting, attaching a Java debugger will give you the ability to step

through the Hive code and hopefully find the problem.

Remote debugging is a feature of Java that is manually enabled by setting specific com-

mand-line properties for the JVM. The Hive shell script provides a switch and help

screen that makes it easy to set these properties (some output truncated for space):

$ bin/hive --help --debug

Allows to debug Hive by connecting to it via JDI API

Usage: hive --debug[:comma-separated parameters list]

Parameters:

recursive=<y|n> Should child JVMs also be started in debug mode. Default: y

port=<port_number> Port on which main JVM listens for debug connection. Defaul...

mainSuspend=<y|n> Should main JVM wait with execution for the debugger to con...

childSuspend=<y|n> Should child JVMs wait with execution for the debugger to c...

swapSuspend Swaps suspend options between main and child JVMs

Building Hive from Source

Running Apache releases is usually a good idea, however you may wish to use features

that are not part of a release, or have an internal branch with nonpublic customizations.

Hence, you’ll need to build Hive from source. The minimum requirements for building

Hive are a recent Java JDK, Subversion, and ANT. Hive also contains components such

as Thrift-generated classes that are not built by default. Rebuilding Hive requires a

Thrift compiler, too.

The following commands check out a Hive release and builds it, produces output in

the hive-trunk/build/dist directory:

$ svn co http://svn.apache.org/repos/asf/hive/trunk hive-trunk

$ cd hive-trunk

$ ant package

$ ls build/dist/

bin examples LICENSE README.txt scripts

conf lib NOTICE RELEASE_NOTES.txt

Running Hive Test Cases

Hive has a unique built-in infrastructure for testing. Hive does have traditional JUnit

tests, however the majority of the testing happens by running queries saved in .q files,

then comparing the results with a previous run saved in Hive source.1 There are multiple

1. That is, they are more like feature or acceptance tests.

156 | Chapter 12: Developing

directories inside the Hive source folder. “Positive” tests are those that should pass,

while “negative” tests should fail.

An example of a positive test is a well-formed query. An example of a negative test is a

query that is malformed or tries doing something that is not allowed by HiveQL:

$ ls -lah ql/src/test/queries/

total 76K

drwxrwxr-x. 7 edward edward 4.0K May 28 2011 .

drwxrwxr-x. 8 edward edward 4.0K May 28 2011 ..

drwxrwxr-x. 3 edward edward 20K Feb 21 20:08 clientnegative

drwxrwxr-x. 3 edward edward 36K Mar 8 09:17 clientpositive

drwxrwxr-x. 3 edward edward 4.0K May 28 2011 negative

drwxrwxr-x. 3 edward edward 4.0K Mar 12 09:25 positive

Take a look at ql/src/test/queries/clientpositive/cast1.q. The first thing you should know

is that a src table is the first table automatically created in the test process. It is a table

with two columns, key and value, where key is an INT and value is a STRING. Because

Hive does not currently have the ability to do a SELECT without a FROM clause, selecting

a single row from the src table is the trick used to test out functions that don’t really

need to retrieve table data; inputs can be “hard-coded” instead.

As you can see in the following example queries, the src table is never referenced in the

SELECT clauses:

hive> CREATE TABLE dest1(c1 INT, c2 DOUBLE, c3 DOUBLE,

> c4 DOUBLE, c5 INT, c6 STRING, c7 INT) STORED AS TEXTFILE;

hive> EXPLAIN

> FROM src INSERT OVERWRITE TABLE dest1

> SELECT 3 + 2, 3.0 + 2, 3 + 2.0, 3.0 + 2.0,

> 3 + CAST(2.0 AS INT) + CAST(CAST(0 AS SMALLINT) AS INT),

> CAST(1 AS BOOLEAN), CAST(TRUE AS INT) WHERE src.key = 86;

hive> FROM src INSERT OVERWRITE TABLE dest1

> SELECT 3 + 2, 3.0 + 2, 3 + 2.0, 3.0 + 2.0,

> 3 + CAST(2.0 AS INT) + CAST(CAST(0 AS SMALLINT) AS INT),

> CAST(1 AS BOOLEAN), CAST(TRUE AS INT) WHERE src.key = 86;

hive> SELECT dest1.* FROM dest1;

The results of the script are found here: ql/src/test/results/clientpositive/cast1.q.out. The

result file is large and printing the complete results inline will kill too many trees. How-

ever, portions of the file are worth noting.

This command invokes a positive and a negative test case for the Hive client:

ant test -Dtestcase=TestCliDriver -Dqfile=mapreduce1.q

ant test -Dtestcase=TestNegativeCliDriver -Dqfile=script_broken_pipe1.q

The two particular tests only parse queries. They do not actually run the client. They

are now deprecated in favor of clientpositive and clientnegative.

Building Hive from Source | 157

You can also run multiple tests in one ant invocation to save time (the last -Dqfile=…

string was wrapped for space; it’s all one string):

ant test -Dtestcase=TestCliDriver -Dqfile=avro_change_schema.q,avro_joins.q,

avro_schema_error_message.q,avro_evolved_schemas.q,avro_sanity_test.q,

avro_schema_literal.q

Execution Hooks

PreHooks and PostHooks are utilities that allow user code to hook into parts of Hive

and execute custom code. Hive’s testing framework uses hooks to echo commands that

produce no output, so that the results show up inside tests:

PREHOOK: query: CREATE TABLE dest1(c1 INT, c2 DOUBLE, c3 DOUBLE,

c4 DOUBLE, c5 INT, c6 STRING, c7 INT) STORED AS TEXTFILE

PREHOOK: type: CREATETABLE

POSTHOOK: query: CREATE TABLE dest1(c1 INT, c2 DOUBLE, c3 DOUBLE,

c4 DOUBLE, c5 INT, c6 STRING, c7 INT) STORED AS TEXTFILE

Setting Up Hive and Eclipse

Eclipse is an open source IDE (Integrated Development Environment). The following

steps allow you to use Eclipse to work with the Hive source code:

$ ant clean package eclipse-files

$ cd metastore

$ ant model-jar

$ cd ../ql

$ ant gen-test

Once built, you can import the project into Eclipse and use it as you normally would.

Create a workspace in Eclipse, as normal. Then use the File → Import command and

then select General → Existing Projects into Workspace. Select the directory where Hive

is installed.

When the list of available projects is shown in the wizard, you’ll see one named hive-

trunk, which you should select and click Finish.

Figure 12-1 shows how to start the Hive Command CLI Driver from within Eclipse.

Hive in a Maven Project

You can set up Hive as a dependency in Maven builds. The Maven repository http://

mvnrepository.com/artifact/org.apache.hive/hive-service contains the most recent relea-

ses. This page also lists the dependencies hive-service requires.

Here is the top-level dependency definition for Hive v0.9.0, not including the tree of

transitive dependencies, which is quite deep:

158 | Chapter 12: Developing

<groupId>org.apache.hive</groupId>

<artifactId>hive-service</artifactId>

</dependency>

The pom.xml file for hive_test, which we discuss next, provides a complete example

of the transitive dependencies for Hive v0.9.0. You can find that file at https://github

.com/edwardcapriolo/hive_test/blob/master/pom.xml.

Unit Testing in Hive with hive_test

The optimal way to write applications to work with Hive is to access Hive with Thrift

through the HiveService. However, the Thrift service was traditionally difficult to bring

up in an embedded environment due to Hive’s many JAR dependencies and the met-

astore component.

Figure 12-1. Starting the Hive Command CLI Driver from within Eclipse

Unit Testing in Hive with hive_test | 159

Hive_test fetches all the Hive dependencies from Maven, sets up the metastore and

Thrift service locally, and provides test classes to make unit testing easier. Also, because

it is very lightweight and unit tests run quickly, this is in contrast to the elaborate test

targets inside Hive, which have to rebuild the entire project to execute any unit test.

Hive_test is ideal for testing code such as UDFs, input formats, SerDes, or any com-

ponent that only adds a pluggable feature for the language. It is not useful for internal

Hive development because all the Hive components are pulled from Maven and are

external to the project.

In your Maven project, create a pom.xml and include hive_test as a dependency, as

shown here:

<groupId>com.jointhegrid</groupId>

<version>3.0.1-SNAPSHOT</version>

</dependency>

Then create a version of hive-site.xml:

$ cp $HIVE_HOME/conf/* src/test/resources/

$ vi src/test/resources/hive-site.xml

Unlike a normal hive-site.xml, this version should not save any data to a

permanent place. This is because unit tests are not supposed to create or preserve any

permanent state. javax.jdo.option.ConnectionURL is set to use a feature in Derby that

only stores the database in main memory. The warehouse directory hive

.metastore.warehouse.dir is set to a location inside /tmp that will be deleted on each

run of the unit test:

<name>javax.jdo.option.ConnectionURL</name>

<value>jdbc:derby:memory:metastore_db;create=true</value>

<description>JDBC connect string for a JDBC metastore</description>

</property>

<name>hive.metastore.warehouse.dir</name>

<value>/tmp/warehouse</value>

<description>location of default database for the warehouse</description>

</property>

</configuration>

Hive_test provides several classes that extend JUnit test cases. HiveTestService set up

the environment, cleared out the warehouse directory, and launched a metastore and

HiveService in-process. This is typically the component to extend for testing. However,

other components, such as HiveTestEmbedded are also available:

package com.jointhegrid.hive_test;

160 | Chapter 12: Developing

import java.io.BufferedWriter;

import java.io.IOException;

import java.io.OutputStreamWriter;

import org.apache.hadoop.fs.FSDataOutputStream;

import org.apache.hadoop.fs.Path;

/* Extending HiveTestService creates and initializes

the metastore and thrift service in an embedded mode */

public class ServiceHiveTest extends HiveTestService {

public ServiceHiveTest() throws IOException {

super();

}

public void testExecute() throws Exception {

/* Use the Hadoop filesystem API to create a

data file */

Path p = new Path(this.ROOT_DIR, "afile");

FSDataOutputStream o = this.getFileSystem().create(p);

BufferedWriter bw = new BufferedWriter(new OutputStreamWriter(o));

bw.write("1\n");

bw.write("2\n");

bw.close();

/* ServiceHive is a component that connections

to an embedded or network HiveService based

on the constructor used */

ServiceHive sh = new ServiceHive();

/* We can now interact through the HiveService

and assert on results */

sh.client.execute("create table atest (num int)");

sh.client.execute("load data local inpath '"

+ p.toString() + "' into table atest");

sh.client.execute("select count(1) as cnt from atest");

String row = sh.client.fetchOne();

assertEquals("2", row);

sh.client.execute("drop table atest");

}

The New Plugin Developer Kit

Hive v0.8.0 introduced a Plugin Developer Kit (PDK). Its intent is to allow developers

to build and test plug-ins without the Hive source. Only Hive binary code is required.

The PDK is relatively new and has some subtle bugs of its own that can make it difficult

to use. If you want to try using the PDK anyway, consult the wiki page, https://cwiki

.apache.org/Hive/plugindeveloperkit.html, but note that this page has a few errors, at

least at the time of this writing.

The New Plugin Developer Kit | 161

CHAPTER 13

Functions

User-Defined Functions (UDFs) are a powerful feature that allow users to extend

HiveQL. As we’ll see, you implement them in Java and once you add them to your

session (interactive or driven by a script), they work just like built-in functions, even

the online help. Hive has several types of user-defined functions, each of which per-

forms a particular “class” of transformations on input data.

In an ETL workload, a process might have several processing steps. The Hive language

has multiple ways to pipeline the output from one step to the next and produce multiple

outputs during a single query. Users also have the ability to create their own functions

for custom processing. Without this feature a process might have to include a custom

MapReduce step or move the data into another system to apply the changes. Intercon-

necting systems add complexity and increase the chance of misconfigurations or other

errors. Moving data between systems is time consuming when dealing with gigabyte-

or terabyte-sized data sets. In contrast, UDFs run in the same processes as the tasks for

your Hive queries, so they work efficiently and eliminate the complexity of integration

with other systems. This chapter covers best practices associated with creating and

using UDFs.

Discovering and Describing Functions

Before writing custom UDFs, let’s familiarize ourselves with the ones that are already

part of Hive. Note that it’s common in the Hive community to use “UDF” to refer to

any function, user-defined or built-in.

The SHOW FUNCTIONS command lists the functions currently loaded in the Hive session,

both built-in and any user-defined functions that have been loaded using the techniques

we will discuss shortly:

hive> SHOW FUNCTIONS;

abs

acos

and

array

163

array_contains

...

Functions usually have their own documentation. Use DESCRIBE FUNCTION to display a

short description:

hive> DESCRIBE FUNCTION concat;

concat(str1, str2, ... strN) - returns the concatenation of str1, str2, ... strN

Functions may also contain extended documentation that can be accessed by adding

the EXTENDED keyword:

hive> DESCRIBE FUNCTION EXTENDED concat;

concat(str1, str2, ... strN) - returns the concatenation of str1, str2, ... strN

Returns NULL if any argument is NULL.

Example:

> SELECT concat('abc', 'def') FROM src LIMIT 1;

'abcdef'

Calling Functions

To use a function, simply call it by name in a query, passing in any required arguments.

Some functions take a specific number of arguments and argument types, while other

functions accept a variable number of arguments with variable types. Just like key-

words, the case of function names is ignored:

SELECT concat(column1,column2) AS x FROM table;

Standard Functions

The term user-defined function (UDF) is also used in a narrower sense to refer to any

function that takes a row argument or one or more columns from a row and returns a

single value. Most functions fall into this category.

Examples include many of the mathematical functions, like round() and floor(), for

converting DOUBLES to BIGINTS, and abs(), for taking the absolute value of a number.

Other examples include string manipulation functions, like ucase(), which converts

the string to upper case; reverse(), which reverses a string; and concat(), which joins

multiple input strings into one output string.

Note that these UDFs can return a complex object, such as an array, map, or struct.

Aggregate Functions

Another type of function is an aggregate function. All aggregate functions, user-defined

and built-in, are referred to generically as user-defined aggregate functions (UDAFs).

An aggregate function takes one or more columns from zero to many rows and returns

a single result. Examples include the math functions: sum(), which returns a sum of all

164 | Chapter 13: Functions

inputs; avg(), which computes the average of the values; min() and max(), which return

the lowest and highest values, respectively:

hive> SELECT avg(price_close)

> FROM stocks

> WHERE exchange = 'NASDAQ' AND symbol = 'AAPL';

Aggregate methods are often combined with GROUP BY clauses. We saw this example in

“GROUP BY Clauses” on page 97:

hive> SELECT year(ymd), avg(price_close) FROM stocks

> WHERE exchange = 'NASDAQ' AND symbol = 'AAPL'

> GROUP BY year(ymd);

1984 25.578625440597534

1985 20.193676221040867

1986 32.46102808021274

...

Table 6-3 in Chapter 6 lists the built-in aggregate functions in HiveQL.

Table Generating Functions

A third type of function supported by Hive is a table generating function. As for the

other function kinds, all table generating functions, user-defined and built-in, are often

referred to generically as user-defined table generating functions (UDTFs).

Table generating functions take zero or more inputs and produce multiple columns or

rows of output. The array function takes a list of arguments and returns the list as a

single array type. Suppose we start with this query using an array:

hive> SELECT array(1,2,3) FROM dual;

[1,2,3]

The explode() function is a UDTF that takes an array of input and iterates through the

list, returning each element from the list in a separate row.

hive> SELECT explode(array(1,2,3)) AS element FROM src;

However, Hive only allows table generating functions to be used in limited ways. For

example, we can’t project out any other columns from the table, a significant limitation.

Here is a query we would like to write with the employees table we have used before.

We want to list each manager-subordinate pair.

Example 13-1. Invalid use of explode

hive> SELECT name, explode(subordinates) FROM employees;

FAILED: Error in semantic analysis: UDTF's are not supported outside

the SELECT clause, nor nested in expressions

However, Hive offers a LATERAL VIEW feature to allow this kind of query:

Table Generating Functions | 165

hive> SELECT name, sub

> FROM employees

> LATERAL VIEW explode(subordinates) subView AS sub;

John Doe Mary Smith

John Doe Todd Jones

Mary Smith Bill King

Note that there are no output rows for employees who aren’t managers (i.e., who have

no subordinates), namely Bill King and Todd Jones. Hence, explode outputs zero to

many new records.

The LATERAL VIEW wraps the output of the explode call. A view alias and column alias

are required, subView and sub, respectively, in this case.

The list of built-in, table generating functions can be found in Table 6-4 in Chapter 6.

A UDF for Finding a Zodiac Sign from a Day

Let’s tackle writing our own UDF. Imagine we have a table with each user’s birth date

stored as a column of a table. With that information, we would like to determine the

user’s Zodiac sign. This process can be implemented with a standard function (UDF

in the most restrictive sense). Specifically, we assume we have a discrete input either as

a date formatted as a string or as a month and a day. The function must return a discrete

single column of output.

Here is a sample data set, which we’ll put in a file called littlebigdata.txt in our home

directory:

edward capriolo,edward@media6degrees.com,2-12-1981,209.191.139.200,M,10

bob,bob@test.net,10-10-2004,10.10.10.1,M,50

sara connor,sara@sky.net,4-5-1974,64.64.5.1,F,2

Load this data set into a table called littlebigdata:

hive > CREATE TABLE IF NOT EXISTS littlebigdata(

> name STRING,

> email STRING,

> bday STRING,

> ip STRING,

> gender STRING,

> anum INT)

> ROW FORMAT DELIMITED FIELDS TERMINATED BY ',';

hive> LOAD DATA LOCAL INPATH '${env:HOME}/littlebigdata.txt'

> INTO TABLE littlebigdata;

The input for the function will be a date and the output will be a string representing

the user’s Zodiac sign.

Here is a Java implementation of the UDF we need:

package org.apache.hadoop.hive.contrib.udf.example;

import java.util.Date;

166 | Chapter 13: Functions

import java.text.SimpleDateFormat;

import org.apache.hadoop.hive.ql.exec.UDF;

@Description(name = "zodiac",

value = "_FUNC_(date) - from the input date string "+

"or separate month and day arguments, returns the sign of the Zodiac.",

extended = "Example:\n"

+ " > SELECT _FUNC_(date_string) FROM src;\n"

+ " > SELECT _FUNC_(month, day) FROM src;")

public class UDFZodiacSign extends UDF{

private SimpleDateFormat df;

public UDFZodiacSign(){

df = new SimpleDateFormat("MM-dd-yyyy");

}

public String evaluate( Date bday ){

return this.evaluate( bday.getMonth(), bday.getDay() );

}

public String evaluate(String bday){

Date date = null;

try {

date = df.parse(bday);

} catch (Exception ex) {

return null;

}

return this.evaluate( date.getMonth()+1, date.getDay() );

}

public String evaluate( Integer month, Integer day ){

if (month==1) {

if (day < 20 ){

return "Capricorn";

} else {

return "Aquarius";

}

if (month==2){

if (day < 19 ){

return "Aquarius";

} else {

return "Pisces";

}

/* ...other months here */

return null;

}

To write a UDF, start by extending the UDF class and implements and the evaluate()

function. During query processing, an instance of the class is instantiated for each usage

of the function in a query. The evaluate() is called for each input row. The result of

A UDF for Finding a Zodiac Sign from a Day | 167

evaluate() is returned to Hive. It is legal to overload the evaluate method. Hive will

pick the method that matches in a similar way to Java method overloading.

The @Description(...) is an optional Java annotation. This is how function documen-

tation is defined and you should use these annotations to document your own UDFs.

When a user invokes DESCRIBE FUNCTION ..., the _FUNC_ strings will be replaced with

the function name the user picks when defining a “temporary” function, as discussed

below.

The arguments and return types of the UDF’s evaluate() function can

only be types that Hive can serialize. For example, if you are working

with whole numbers, a UDF can take as input a primitive int, an Inte

ger wrapper object, or an IntWritable, which is the Hadoop wrapper

for integers. You do not have to worry specifically about what the caller

is sending because Hive will convert the types for you if they do not

match. Remember that null is valid for any type in Hive, but in Java

primitives are not objects and cannot be null.

To use the UDF inside Hive, compile the Java code and package the UDF bytecode

class file into a JAR file. Then, in your Hive session, add the JAR to the classpath and

use a CREATE FUNCTION statement to define a function that uses the Java class:

hive> ADD JAR /full/path/to/zodiac.jar;

hive> CREATE TEMPORARY FUNCTION zodiac

> AS 'org.apache.hadoop.hive.contrib.udf.example.UDFZodiacSign';

Note that quotes are not required around the JAR file path and currently it needs to be

a full path to the file on a local filesystem. Hive not only adds this JAR to the classpath,

it puts the JAR file in the distributed cache so it’s available around the cluster.

Now the Zodiac UDF can be used like any other function. Notice the word TEMPO

RARY found inside the CREATE FUNCTION statement. Functions declared will only be avail-

able in the current session. You will have to add the JAR and create the function in each

session. However, if you use the same JAR files and functions frequently, you can add

these statements to your $HOME/.hiverc file:

hive> DESCRIBE FUNCTION zodiac;

zodiac(date) - from the input date string or separate month and day

arguments, returns the sign of the Zodiac.

hive> DESCRIBE FUNCTION EXTENDED zodiac;

zodiac(date) - from the input date string or separate month and day

arguments, returns the sign of the Zodiac.

Example:

> SELECT zodiac(date_string) FROM src;

> SELECT zodiac(month, day) FROM src;

hive> SELECT name, bday, zodiac(bday) FROM littlebigdata;

edward capriolo 2-12-1981 Aquarius

168 | Chapter 13: Functions

bob 10-10-2004 Libra

sara connor 4-5-1974 Aries

To recap, our UDF allows us to do custom transformations inside the Hive language.

Hive can now convert the user’s birthday to the corresponding Zodiac sign while it is

doing any other aggregations and transformations.

If we’re finished with the function, we can drop it:

hive> DROP TEMPORARY FUNCTION IF EXISTS zodiac;

As usual, the IF EXISTS is optional. It suppresses errors if the function doesn’t exist.

UDF Versus GenericUDF

In our Zodiac example we extended the UDF class. Hive offers a counterpart called

GenericUDF. GenericUDF is a more complex abstraction, but it offers support for better

null handling and makes it possible to handle some types of operations programmati-

cally that a standard UDF cannot support. An example of a generic UDF is the Hive

CASE ... WHEN statement, which has complex logic depending on the arguments to the

statement. We will demonstrate how to use the GenericUDF class to write a user-defined

function, called nvl(), which returns a default value if null is passed in.

The nvl() function takes two arguments. If the first argument is non-null, it is returned.

If the first argument is null, the second argument is returned. The GenericUDF frame-

work is a good fit for this problem. A standard UDF could be used as a solution but it

would be cumbersome because it requires overloading the evaluate method to handle

many different input types. GenericUDF will detect the type of input to the function

programmatically and provide an appropriate response.

We begin with the usual laundry list of import statements:

package org.apache.hadoop.hive.ql.udf.generic;

import org.apache.hadoop.hive.ql.exec.Description;

import org.apache.hadoop.hive.ql.exec.UDFArgumentException;

import org.apache.hadoop.hive.ql.exec.UDFArgumentLengthException;

import org.apache.hadoop.hive.ql.exec.UDFArgumentTypeException;

import org.apache.hadoop.hive.ql.metadata.HiveException;

import org.apache.hadoop.hive.ql.udf.generic.GenericUDF;

import org.apache.hadoop.hive.ql.udf.generic.GenericUDFUtils;

import org.apache.hadoop.hive.serde2.objectinspector.ObjectInspector;

Next, we use the @Description annotation to document the UDF:

@Description(name = "nvl",

value = "_FUNC_(value,default_value) - Returns default value if value"

+" is null else returns value",

extended = "Example:\n"

+ " > SELECT _FUNC_(null,'bla') FROM src LIMIT 1;\n")

UDF Versus GenericUDF | 169

Now the class extends GenericUDF, a requirement to exploit the generic handling we

want.

The initialize() method is called and passed an ObjectInspector for each argument.

The goal of this method is to determine the return type from the arguments. The user

can also throw an Exception to signal that bad types are being sent to the method. The

returnOIResolver is a built-in class that determines the return type by finding the type

of non-null variables and using that type:

public class GenericUDFNvl extends GenericUDF {

private GenericUDFUtils.ReturnObjectInspectorResolver returnOIResolver;

private ObjectInspector[] argumentOIs;

@Override

public ObjectInspector initialize(ObjectInspector[] arguments)

throws UDFArgumentException {

argumentOIs = arguments;

if (arguments.length != 2) {

throw new UDFArgumentLengthException(

"The operator 'NVL' accepts 2 arguments.");

}

returnOIResolver = new GenericUDFUtils.ReturnObjectInspectorResolver(true);

if (!(returnOIResolver.update(arguments[0]) && returnOIResolver

.update(arguments[1]))) {

throw new UDFArgumentTypeException(2,

"The 1st and 2nd args of function NLV should have the same type, "

+ "but they are different: \"" + arguments[0].getTypeName()

+ "\" and \"" + arguments[1].getTypeName() + "\"");

}

return returnOIResolver.get();

}

...

The evaluate method has access to the values passed to the method stored in an array

of DeferredObject values. The returnOIResolver created in the initialize method is

used to get values from the DeferredObjects. In this case, the function returns the first

non-null value:

...

@Override

public Object evaluate(DeferredObject[] arguments) throws HiveException {

Object retVal = returnOIResolver.convertIfNecessary(arguments[0].get(),

argumentOIs[0]);

if (retVal == null ){

retVal = returnOIResolver.convertIfNecessary(arguments[1].get(),

argumentOIs[1]);

}

return retVal;

}

...

The final method to override is getDisplayString(), which is used inside the Hadoop

tasks to display debugging information when the function is being used:

170 | Chapter 13: Functions

...

@Override

public String getDisplayString(String[] children) {

StringBuilder sb = new StringBuilder();

sb.append("if ");

sb.append(children[0]);

sb.append(" is null ");

sb.append("returns");

sb.append(children[1]);

return sb.toString() ;

}

To test the generic nature of the UDF, it is called several times, each time passing values

of different types, as shown the following example:

hive> ADD JAR /path/to/jar.jar;

hive> CREATE TEMPORARY FUNCTION nvl

> AS 'org.apache.hadoop.hive.ql.udf.generic.GenericUDFNvl';

hive> SELECT nvl( 1 , 2 ) AS COL1,

> nvl( NULL, 5 ) AS COL2,

> nvl( NULL, "STUFF" ) AS COL3

> FROM src LIMIT 1;

1 5 STUFF

Permanent Functions

Until this point we have bundled our code into JAR files, then used ADD JAR and CREATE

TEMPORARY FUNCTION to make use of them.

Your function may also be added permanently to Hive, however this requires a small

modification to a Hive Java file and then rebuilding Hive.

Inside the Hive source code, a one-line change is required to the FunctionRegistry class

found at ql/src/java/org/apache/hadoop/hive/ql/exec/FunctionRegistry.java. Then you

rebuild Hive following the instructions that come with the source distribution.

While it is recommended that you redeploy the entire new build, only the hive-exec-

*.jar, where \* is the version number, needs to be replaced.

Here is an example change to FunctionRegistry where the new nvl() function is added

to Hive’s list of built-in functions:

...

registerUDF("parse_url", UDFParseUrl.class, false);

registerGenericUDF("nvl", GenericUDFNvl.class);

registerGenericUDF("split", GenericUDFSplit.class);

...

Permanent Functions | 171

User-Defined Aggregate Functions

Users are able to define aggregate functions, too. However, the interface is more com-

plex to implement. Aggregate functions are processed in several phases. Depending on

the transformation the UDAF performs, the types returned by each phase could be

different. For example, a sum() UDAF could accept primitive integer input, create in-

teger PARTIAL data, and produce a final integer result. However, an aggregate like

median() could take primitive integer input, have an intermediate list of integers as

PARTIAL data, and then produce a final integer as the result.

For an example of a generic user-defined aggregate function, see the source code for

GenericUDAFAverage available at http://svn.apache.org/repos/asf/hive/branches/branch-0

.8/ql/src/java/org/apache/hadoop/hive/ql/udf/generic/GenericUDAFAverage.java.

Aggregations execute inside the context of a map or reduce task, which

is a Java process with memory limitations. Therefore, storing large

structures inside an aggregate may exceed available heap space. The

min() UDAF only requires a single element be stored in memory for

comparison. The collectset() UDAF uses a set internally to de-

duplicate data in order to limit memory usage. percentile_approx()

uses approximations to achieve a near correct result while limiting

memory usage. It is important to keep memory usage in mind when

writing a UDAF. You can increase your available memory to some extent

by adjusting mapred.child.java.opts, but that solution does not scale:

<name>mapred.child.java.opts</name>

</property>

Creating a COLLECT UDAF to Emulate GROUP_CONCAT

MySQL has a useful function known as GROUP_CONCAT, which combines all the

elements of a group into a single string using a user-specified delimiter. Below is an

example MySQL query that shows how to use its version of this function:

mysql > CREATE TABLE people (

name STRING,

friendname STRING );

mysql > SELECT * FROM people;

bob sara

bob john

bob ted

john sara

ted bob

ted sara

mysql > SELECT name, GROUP_CONCAT(friendname SEPARATOR ',')

FROM people

172 | Chapter 13: Functions

Downloa d f r o m W o w ! e B o o k < w w w.woweb o o k . c o m >

GROUP BY name;

bob sara,john,ted

john sara

ted bob,sara

We can do the same transformation in Hive without the need for additional grammar

in the language. First, we need an aggregate function that builds a list of all input to

the aggregate. Hive already has a UDAF called collect_set that adds all input into a

java.util.Set collection. Sets automatically de-duplicate entries on insertion, which

is undesirable for GROUP CONCAT. To build collect, we will take the code in col

lect_set and replace instances of Set with instances of ArrayList. This will stop the

de-duplication. The result of the aggregate will be a single array of all values.

It is important to remember that the computation of your aggregation must be arbi-

trarily divisible over the data. Think of it as writing a divide-and-conquer algorithm

where the partitioning of the data is completely out of your control and handled by

Hive. More formally, given any subset of the input rows, you should be able to compute

a partial result, and also be able to merge any pair of partial results into another partial

result.

The following code is available on Github. All the input to the aggregation must be

primitive types. Rather than returning an ObjectInspector, like GenericUDFs, aggregates

return a subclass of GenericUDAFEvaluator:

@Description(name = "collect", value = "_FUNC_(x) - Returns a list of objects. "+

"CAUTION will easily OOM on large data sets" )

public class GenericUDAFCollect extends AbstractGenericUDAFResolver {

static final Log LOG = LogFactory.getLog(GenericUDAFCollect.class.getName());

public GenericUDAFCollect() {

}

@Override

public GenericUDAFEvaluator getEvaluator(TypeInfo[] parameters)

throws SemanticException {

if (parameters.length != 1) {

throw new UDFArgumentTypeException(parameters.length - 1,

"Exactly one argument is expected.");

}

if (parameters[0].getCategory() != ObjectInspector.Category.PRIMITIVE) {

throw new UDFArgumentTypeException(0,

"Only primitive type arguments are accepted but "

+ parameters[0].getTypeName() + " was passed as parameter 1.");

}

return new GenericUDAFMkListEvaluator();

}

Table 13-1 describes the methods that are part of the base class.

User-Defined Aggregate Functions | 173

Table 13-1. Methods in AbstractGenericUDAFResolver

Method Description

init Called by Hive to initialize an instance of the UDAF evaluator

class.

getNewAggregationBuffer Return an object that will be used to store temporary aggre-

gation results.

iterate Process a new row of data into the aggregation buffer.

terminatePartial Return the contents of the current aggregation in a persistable

way. Here, persistable means the return value can only be built

up in terms of Java primitives, arrays, primitive wrappers

(e.g., Double), Hadoop Writables, Lists, and Maps. Do NOT

use your own classes (even if they implement java.io

.Serializable).

merge Merge a partial aggregation returned by

terminatePartial into the current aggregation.

terminate Return the final result of the aggregation to Hive.

In the init method, the object inspectors for the result type are set, after determining

what mode the evaluator is in.

The iterate() and terminatePartial() methods are used on the map side, while ter

minate() and merge() are used on the reduce side to produce the final result. In all cases

the merges are building larger lists:

public static class GenericUDAFMkListEvaluator extends GenericUDAFEvaluator {

private PrimitiveObjectInspector inputOI;

private StandardListObjectInspector loi;

private StandardListObjectInspector internalMergeOI;

@Override

public ObjectInspector init(Mode m, ObjectInspector[] parameters)

throws HiveException {

super.init(m, parameters);

if (m == Mode.PARTIAL1) {

inputOI = (PrimitiveObjectInspector) parameters[0];

return ObjectInspectorFactory

.getStandardListObjectInspector(

(PrimitiveObjectInspector) ObjectInspectorUtils

.getStandardObjectInspector(inputOI));

} else {

if (!(parameters[0] instanceof StandardListObjectInspector)) {

inputOI = (PrimitiveObjectInspector) ObjectInspectorUtils

.getStandardObjectInspector(parameters[0]);

return (StandardListObjectInspector) ObjectInspectorFactory

.getStandardListObjectInspector(inputOI);

} else {

internalMergeOI = (StandardListObjectInspector) parameters[0];

inputOI = (PrimitiveObjectInspector)

internalMergeOI.getListElementObjectInspector();

174 | Chapter 13: Functions

loi = (StandardListObjectInspector) ObjectInspectorUtils

.getStandardObjectInspector(internalMergeOI);

return loi;

}

...

The remaining methods and class definition define MkArrayAggregationBuffer as well

as top-level methods that modify the contents of the buffer:

You may have noticed that Hive tends to avoid allocating objects with

new whenever possible. Hadoop and Hive use this pattern to create fewer

temporary objects and thus less work for the JVM’s Garbage Collec

tion algorithms. Keep this in mind when writing UDFs, because refer-

ences are typically reused. Assuming immutable objects will lead to

bugs!

...

static class MkArrayAggregationBuffer implements AggregationBuffer {

List<Object> container;

}

@Override

public void reset(AggregationBuffer agg) throws HiveException {

((MkArrayAggregationBuffer) agg).container =

new ArrayList<Object>();

}

@Override

public AggregationBuffer getNewAggregationBuffer()

throws HiveException {

MkArrayAggregationBuffer ret = new MkArrayAggregationBuffer();

reset(ret);

return ret;

}

// Mapside

@Override

public void iterate(AggregationBuffer agg, Object[] parameters)

throws HiveException {

assert (parameters.length == 1);

Object p = parameters[0];

if (p != null) {

MkArrayAggregationBuffer myagg = (MkArrayAggregationBuffer) agg;

putIntoList(p, myagg);

}

// Mapside

@Override

public Object terminatePartial(AggregationBuffer agg)

User-Defined Aggregate Functions | 175

throws HiveException {

MkArrayAggregationBuffer myagg = (MkArrayAggregationBuffer) agg;

ArrayList<Object> ret = new ArrayList<Object>(myagg.container.size());

ret.addAll(myagg.container);

return ret;

}

@Override

public void merge(AggregationBuffer agg, Object partial)

throws HiveException {

MkArrayAggregationBuffer myagg = (MkArrayAggregationBuffer) agg;

ArrayList<Object> partialResult =

(ArrayList<Object>) internalMergeOI.getList(partial);

for(Object i : partialResult) {

putIntoList(i, myagg);

}

@Override

public Object terminate(AggregationBuffer agg) throws HiveException {

MkArrayAggregationBuffer myagg = (MkArrayAggregationBuffer) agg;

ArrayList<Object> ret = new ArrayList<Object>(myagg.container.size());

ret.addAll(myagg.container);

return ret;

}

private void putIntoList(Object p, MkArrayAggregationBuffer myagg) {

Object pCopy =

ObjectInspectorUtils.copyToStandardObject(p,this.inputOI);

myagg.container.add(pCopy);

}

Using collect will return a single row with a single array of all of the aggregated values:

hive> dfs -cat $HOME/afile.txt;

twelve 12

twelve 1

eleven 11

eleven 10

hive> CREATE TABLE collecttest (str STRING, countVal INT)

> ROW FORMAT DELIMITED FIELDS TERMINATED BY '09' LINES TERMINATED BY '10';

hive> LOAD DATA LOCAL INPATH '${env:HOME}/afile.txt' INTO TABLE collecttest;

hive> SELECT collect(str) FROM collecttest;

[twelve,twelve,eleven,eleven]

The concat_ws() takes a delimiter as its first argument. The remaining arguments can

be string types or arrays of strings. The returned result contains the argument joined

together by the delimiter. Hence, we have converted the array into a single comma-

separated string:

hive> SELECT concat_ws( ',' , collect(str)) FROM collecttest;

twelve,twleve,eleven,eleven

176 | Chapter 13: Functions

GROUP_CONCAT can be done by combining GROUP BY, COLLECT and concat_ws() as shown

here:

hive> SELECT str, concat_ws( ',' , collect(cast(countVal AS STRING)))

> FROM collecttest GROUP BY str;

eleven 11,10

twelve 12,1

User-Defined Table Generating Functions

While UDFs can be used be return arrays or structures, they cannot return multiple

columns or multiple rows. User-Defined Table Generating Functions, or UDTFs, ad-

dress this need by providing a programmatic interface to return multiple columns and

even multiple rows.

UDTFs that Produce Multiple Rows

We have already used the explode method in several examples. Explode takes an array

as input and outputs one row for each element in the array. An alternative way to do

this would have the UDTF generate the rows based on some input. We will demonstrate

this with a UDTF that works like a for loop. The function receives user inputs of the

start and stop values and then outputs N rows:

hive> SELECT forx(1,5) AS i FROM collecttest;

Our class extends the GenericUDTF interface. We declare three integer variables for the

start, end, and increment. The forwardObj array will be used to return result rows:

package com.jointhegrid.udf.collect;

import java.util.ArrayList;

import org.apache.hadoop.hive.ql.exec.UDFArgumentException;

import org.apache.hadoop.hive.ql.metadata.HiveException;

import org.apache.hadoop.hive.ql.udf.generic.GenericUDFUtils.*;

import org.apache.hadoop.hive.ql.udf.generic.GenericUDTF;

import org.apache.hadoop.hive.serde2.objectinspector.*;

import org.apache.hadoop.hive.serde2.objectinspector.primitive.*;

import org.apache.hadoop.io.IntWritable;

public class GenericUDTFFor extends GenericUDTF {

IntWritable start;

IntWritable end;

IntWritable inc;

Object[] forwardObj = null;

...

User-Defined Table Generating Functions | 177

Because the arguments to this function are constant, the value can be determined in

the initialize method. Nonconstant values are typically not available until the evaluate

method. The third argument for increment is optional, as it defaults to 1:

...

@Override

public StructObjectInspector initialize(ObjectInspector[] args)

throws UDFArgumentException {

start=((WritableConstantIntObjectInspector) args[0])

.getWritableConstantValue();

end=((WritableConstantIntObjectInspector) args[1])

.getWritableConstantValue();

if (args.length == 3) {

inc =((WritableConstantIntObjectInspector) args[2])

.getWritableConstantValue();

} else {

inc = new IntWritable(1);

}

...

This function returns only a single column and its type is always an integer. We need

to give it a name, but the user can always override this later:

...

this.forwardObj = new Object[1];

ArrayList<String> fieldNames = new ArrayList<String>();

ArrayList<ObjectInspector> fieldOIs = new ArrayList<ObjectInspector>();

fieldNames.add("col0");

fieldOIs.add(

PrimitiveObjectInspectorFactory.getPrimitiveJavaObjectInspector(

PrimitiveCategory.INT));

return ObjectInspectorFactory.getStandardStructObjectInspector(

fieldNames, fieldOIs);

}

...

The process method is where the interesting work happens. Notice that the return type

is void. This is because UDTF can forward zero or more rows, unlike a UDF, which

has a single return. In this case the call to the forward method is nested inside a for

loop, which causes it to forward a row for each iteration:

...

@Override

public void process(Object[] args)

throws HiveException, UDFArgumentException {

for (int i = start.get(); i < end.get(); i = i + inc.get()) {

this.forwardObj[0] = new Integer(i);

forward(forwardObj);

}

@Override

178 | Chapter 13: Functions

public void close() throws HiveException {

}

UDTFs that Produce a Single Row with Multiple Columns

An example of a UDTF that returns multiple columns but only one row is the

parse_url_tuple function, which is a built-in Hive function. It takes as input a param-

eter that is a URL and one or more constants that specify the parts of the URL the user

wants returned:

hive> SELECT parse_url_tuple(weblogs.url, 'HOST', 'PATH')

> AS (host, path) FROM weblogs;

google.com /index.html

hotmail.com /a/links.html

The benefit of this type of UDFT is the URL only needs to be parsed once, then returns

multiple columns—a clear performance win. The alternative, using UDFs, involves

writing several UDFs to extract specific parts of the URL. Using UDFs requires writing

more code as well as more processing time because the URL is parsed multiple times.

For example, something like the following:

SELECT PARSE_HOST(a.url) as host, PARSE_PORT(url) FROM weblogs;

UDTFs that Simulate Complex Types

A UDTF can be used as a technique for adding more complex types to Hive. For ex-

ample, a complex type can be serialized as an encoded string and a UDTF will deseri-

alize the complex type when needed. Suppose we have a Java class named Book. Hive

cannot work with this datatype directly, however a Book could be encoded to and

decoded from a string format:

public class Book {

public Book () { }

public String isbn;

public String title;

public String [] authors;

/* note: this system will not work if your table is

using '|' or ',' as the field delimiter! */

public void fromString(String parts){

String [] part = part.split("\|");

isbn = Integer.parseInt( part[0] );

title = part[1] ;

authors = part[2].split(",");

}

public String toString(){

return isbn+"\t"+title+"\t"+StringUtils.join(authors, ",");

}

User-Defined Table Generating Functions | 179

Imagine we have a flat text file with books in this format. For now lets assume we could

not use a delimited SerDe to split on | and ,:

hive> SELECT * FROM books;

5555555|Programming Hive|Edward,Dean,Jason

In the pipe-delimited raw form it is possible to do some parsing of the data:

hive> SELECT cast(split(book_info,"\|")[0] AS INTEGER) AS isbn FROM books

> WHERE split(book_info,"\|")[1] = "Programming Hive";

5555555

This HiveQL works correctly, however it could be made easier for the end user. For

example, writing this type of query may require consulting documentation regarding

which fields and types are used, remembering casting conversion rules, and so forth.

By contrast, a UDTF makes this HiveQL simpler and more readable. In the following

example, the parse_book() UDTF is introduced:

hive> FROM (

> parse_book(book_info) AS (isbn, title, authors) FROM Book ) a

> SELECT a.isbn

> WHERE a.title="Programming Hive"

> AND array_contains (authors, 'Edward');

5555555

The function parse_book() allows Hive to return multiple columns of different types

representing the fields of a book:

package com.jointhegrid.udf.collect;

import java.util.ArrayList;

import org.apache.hadoop.hive.ql.exec.UDFArgumentException;

import org.apache.hadoop.hive.ql.metadata.HiveException;

import org.apache.hadoop.hive.ql.udf.generic.GenericUDTF;

import org.apache.hadoop.hive.serde2.objectinspector.ObjectInspector;

import org.apache.hadoop.hive.serde2.objectinspector.ObjectInspectorFactory;

import org.apache.hadoop.hive.serde2.objectinspector.PrimitiveObjectInspector

.PrimitiveCategory;

import org.apache.hadoop.hive.serde2.objectinspector.StructObjectInspector;

import org.apache.hadoop.hive.serde2.objectinspector.primitive

.PrimitiveObjectInspectorFactory;

import org.apache.hadoop.hive.serde2.objectinspector.primitive

.WritableConstantStringObjectInspector;

import org.apache.hadoop.hive.serde2.objectinspector.primitive

.WritableStringObjectInspector;

import org.apache.hadoop.io.Text;

public class UDTFBook extends GenericUDTF{

private Text sent;

Object[] forwardObj = null;

...

180 | Chapter 13: Functions

The function will return three properties and ISBN as an integer, a title as a string, and

authors as an array of strings. Notice that we can return nested types with all UDFs,

for example we can return an array of array of strings:

...

@Override

public StructObjectInspector initialize(ObjectInspector[] args)

throws UDFArgumentException {

ArrayList<String> fieldNames = new ArrayList<String>();

ArrayList<ObjectInspector> fieldOIs = new ArrayList<ObjectInspector>();

fieldNames.add("isbn");

fieldOIs.add(PrimitiveObjectInspectorFactory.getPrimitiveJavaObjectInspector(

PrimitiveCategory.INT));

fieldNames.add("title");

fieldOIs.add(PrimitiveObjectInspectorFactory.getPrimitiveJavaObjectInspector(

PrimitiveCategory.STRING));

fieldNames.add("authors");

fieldOIs.add( ObjectInspectorFactory.getStandardListObjectInspector(

PrimitiveObjectInspectorFactory.getPrimitiveJavaObjectInspector(

PrimitiveCategory.STRING)

)

);

forwardObj= new Object[3];

return ObjectInspectorFactory.getStandardStructObjectInspector(

fieldNames, fieldOIs);

}

...

The process method only returns a single row. However, each element in the object

array will be bound to a specific variable:

...

@Override

public void process(Object[] os) throws HiveException {

sent = new Text(((StringObjectInspector)args[0])

.getPrimitiveJavaObject(os[0]));

String parts = new String(this.sent.getBytes());

String [] part = parts.split("\\|");

forwardObj[0]=Integer.parseInt( part[0] );

forwardObj[1]=part[1] ;

forwardObj[2]=part[2].split(",");

this.forward(forwardObj);

}

@Override

public void close() throws HiveException {

}

User-Defined Table Generating Functions | 181

We have followed the call to the book UDTF with AS, which allows the result columns

to be named by the user. They can then be used in other parts of the query without

having to parse information from the book again:

client.execute(

"create temporary function book as 'com.jointhegrid.udf.collect.UDTFBook'");

client.execute("create table booktest (str string) ");

client.execute(

"load data local inpath '" + p.toString() + "' into table booktest");

client.execute("select book(str) AS (book, title, authors) from booktest");

[555 Programming Hive "Dean","Jason","Edward"]

Accessing the Distributed Cache from a UDF

UDFs may access files inside the distributed cache, the local filesystem, or even the

distributed filesystem. This access should be used cautiously as the overhead is

significant.

A common usage of Hive is the analyzing of web logs. A popular operation is deter-

mining the geolocation of web traffic based on the IP address. Maxmind makes a GeoIP

database available and a Java API to search this database. By wrapping a UDF around

this API, location information may be looked up about an IP address from within a

Hive query.

The GeoIP API uses a small data file. This is ideal for showing the functionality of

accessing a distributed cache file from a UDF. The complete code for this example is

found at https://github.com/edwardcapriolo/hive-geoip/.

ADD FILE is used to cache the necessary data files with Hive. ADD JAR is used to add the

required Java JAR files to the cache and the classpath. Finally, the temporary function

must be defined as the final step before performing queries:

hive> ADD FILE GeoIP.dat;

hive> ADD JAR geo-ip-java.jar;

hive> ADD JAR hive-udf-geo-ip-jtg.jar;

hive> CREATE TEMPORARY FUNCTION geoip

> AS 'com.jointhegrid.hive.udf.GenericUDFGeoIP';

hive> SELECT ip, geoip(source_ip, 'COUNTRY_NAME', './GeoIP.dat') FROM weblogs;

209.191.139.200 United States

10.10.0.1 Unknown

The two examples returned include an IP address in the United States and a private IP

address that has no fixed address.

The geoip() function takes three arguments: the IP address in either string or long

format, a string that must match one of the constants COUNTRY_NAME or DMA_CODE, and a

final argument that is the name of the data file that has already been placed in the

distributed cache.

182 | Chapter 13: Functions

The first call to the UDF (which triggers the first call to the evaluate Java function in

the implementation) will instantiate a LookupService object that uses the file located in

the distributed cache. The lookup service is saved in a reference so it only needs to be

initialized once in the lifetime of a map or reduce task that initializes it. Note that the

LookupService has its own internal caching, LookupService.GEOIP\_MEMORY_CACHE, so

that optimization should avoid frequent disk access when looking up IPs.

Here is the source code for evaluate():

@Override

public Object evaluate(DeferredObject[] arguments) throws HiveException {

if (argumentOIs[0] instanceof LongObjectInspector) {

this.ipLong = ((LongObjectInspector)argumentOIs[0]).get(arguments[0].get());

} else {

this.ipString = ((StringObjectInspector)argumentOIs[0])

.getPrimitiveJavaObject(arguments[0].get());

}

this.property = ((StringObjectInspector)argumentOIs[1])

.getPrimitiveJavaObject(arguments[1].get());

if (this.property != null) {

this.property = this.property.toUpperCase();

}

if (ls ==null){

if (argumentOIs.length == 3){

this.database = ((StringObjectInspector)argumentOIs[1])

.getPrimitiveJavaObject(arguments[2].get());

File f = new File(database);

if (!f.exists())

throw new HiveException(database+" does not exist");

try {

ls = new LookupService ( f , LookupService.GEOIP_MEMORY_CACHE );

} catch (IOException ex){

throw new HiveException (ex);

}

...

An if statement in evaluate determines which data the method should return. In our

example, the country name is requested:

...

if (COUNTRY_PROPERTIES.contains(this.property)) {

Country country = ipString != null ?

ls.getCountry(ipString) : ls.getCountry(ipLong);

if (country == null) {

return null;

} else if (this.property.equals(COUNTRY_NAME)) {

return country.getName();

} else if (this.property.equals(COUNTRY_CODE)) {

return country.getCode();

}

assert(false);

} else if (LOCATION_PROPERTIES.contains(this.property)) {

...

Accessing the Distributed Cache from a UDF | 183

}

Annotations for Use with Functions

In this chapter we mentioned the Description annotation and how it is used to provide

documentation for Hive methods at runtime. Other annotations exist for UDFs that

can make functions easier to use and even increase the performance of some Hive

queries:

public @interface UDFType {

boolean deterministic() default true;

boolean stateful() default false;

boolean distinctLike() default false;

}

Deterministic

By default, deterministic is automatically turned on for most queries because they are

inherently deterministic by nature. An exception is the function rand().

If a UDF is not deterministic, it is not included in the partition pruner.

An example of a nondeterministic query using rand() is the following:

SELECT * FROM t WHERE rand() < 0.01;

If rand() were deterministic, the result would only be calculated a single time in the

computation state. Because a query with rand() is nondeterministic, the result of

rand() is recomputed for each row.

Stateful

Almost all the UDFs are stateful by default; a UDF that is not stateful is rand() because

it returns a different value for each invocation. The Stateful annotation may be used

under the following conditions:

• A stateful UDF can only be used in the SELECT list, not in other clauses such as

WHERE/ON/ORDER/GROUP.

• When a stateful UDF is present in a query, the implication is the SELECT will be

treated similarly to TRANSFORM (i.e., a DISTRIBUTE/CLUSTER/SORT clause), then run

inside the corresponding reducer to ensure the results are as expected.

• If stateful is set to true, the UDF should also be treated as nondeterministic (even

if the deterministic annotation explicitly returns true).

See https://issues.apache.org/jira/browse/HIVE-1994 for more details.

184 | Chapter 13: Functions

DistinctLike

Used for cases where the function behaves like DISTINCT even when applied to a non-

distinct column of values. Examples include min and max functions that return a distinct

value even though the underlying numeric data can have repeating values.

Macros

Macros provide the ability to define functions in HiveQL that call other functions and

operators. When appropriate for the particular situation, macros are a convenient

alternative to writing UDFs in Java or using Hive streaming, because they require no

external code or scripts.

To define a macro, use the CREATE TEMPORARY MACRO syntax. Here is an example that

creates a SIGMOID function calculator:

hive> CREATE TEMPORARY MACRO SIGMOID (x DOUBLE) 1.0 / (1.0 + EXP(-x));

hive> SELECT SIGMOID(2) FROM src LIMIT 1;

Macros | 185

CHAPTER 14

Streaming

Hive works by leveraging and extending the components of Hadoop, common

abstractions such as InputFormat, OutputFormat, Mapper, and Reducer, plus its own

abstractions, like SerializerDeserializer (SerDe), User-Defined Functions (UDFs),

and StorageHandlers.

These components are all Java components, but Hive hides the complexity of imple-

menting and using these components by letting the user work with SQL abstractions,

rather than Java code.

Streaming offers an alternative way to transform data. During a streaming job, the Ha-

doop Streaming API opens an I/O pipe to an external process. Data is then passed to

the process, which operates on the data it reads from the standard input and writes the

results out through the standard output, and back to the Streaming API job. While Hive

does not leverage the Hadoop streaming API directly, it works in a very similar way.

This pipeline computing model is familiar to users of Unix operating systems and their

descendants, like Linux and Mac OS X.

Streaming is usually less efficient than coding the comparable UDFs or

InputFormat objects. Serializing and deserializing data to pass it in and

out of the pipe is relatively inefficient. It is also harder to debug the whole

program in a unified manner. However, it is useful for fast prototyping

and for leveraging existing code that is not written in Java. For Hive

users who don’t want to write Java code, it can be a very effective

approach.

Hive provides several clauses to use streaming: MAP(), REDUCE(), and TRANSFORM(). An

important point to note is that MAP() does not actually force streaming during the map

phase nor does reduce force streaming to happen in the reduce phase. For this reason,

the functionally equivalent yet more generic TRANSFORM() clause is suggested to avoid

misleading the reader of the query.

187

For our streaming examples we will use a small table named a, with columns named

col1 and col2, both of type INT, and two rows:

hive> CREATE TABLE a (col1 INT, col2 INT)

> ROW FORMAT DELIMITED FIELDS TERMINATED BY '\t';

hive> SELECT * FROM a;

4 5

3 2

hive> DESCRIBE a;

a int

b int

Identity Transformation

The most basic streaming job is an identity operation. The /bin/cat command echoes

the data sent to it and meets the requirements. In this example, /bin/cat is assumed to

be installed on all TaskTracker nodes. Any Linux system should have it! Later, we will

show how Hive can “ship” applications with the job when they aren’t already installed

around the cluster:

hive> SELECT TRANSFORM (a, b)

> USING '/bin/cat' AS newA, newB

> FROM default.a;

4 5

3 2

Changing Types

The return columns from TRANSFORM are typed as strings, by default. There is an alter-

native syntax that casts the results to different types.

hive> SELECT TRANSFORM (col1, col2)

> USING '/bin/cat' AS (newA INT , newB DOUBLE) FROM a;

4 5.0

3 2.0

Projecting Transformation

The cut command can be used with streaming to extract or project specific fields. In

other words, this behaves like the SELECT statement:

hive> SELECT TRANSFORM (a, b)

> USING '/bin/cut -f1'

> AS newA, newB FROM a;

4 NULL

3 NULL

188 | Chapter 14: Streaming

Note that the query attempts to read more columns than are actually returned from the

external process, so newB is always NULL. By default, TRANSFORM assumes two columns

but there can be any number of them:

hive> SELECT TRANSFORM (a, b)

> USING '/bin/cut -f1'

> AS newA FROM a;

Manipulative Transformations

The /bin/sed program (or /usr/bin/sed on Mac OS X systems) is a stream editor. It

takes the input stream, edits it according to the user’s specification, and then writes the

results to the output stream. The example below replaces the string 4 with the string 10:

hive> SELECT TRANSFORM (a, b)

> USING '/bin/sed s/4/10/'

> AS newA, newB FROM a;

10 5

3 2

Using the Distributed Cache

All of the streaming examples thus far have used applications such as cat and sed that

are core parts of Unix operating systems and their derivatives. When a query requires

files that are not already installed on every TaskTracker, users can use the distributed

cache to transmit data or program files across the cluster that will be cleaned up when

the job is complete.

This is helpful, because installing (and sometimes removing) lots of little components

across large clusters can be a burden. Also, the cache keeps one job’s cached files sep-

arate from those files belonging to other jobs.

The following example is a bash shell script that converts degrees in Celsius to degrees

in Fahrenheit:

while read LINE

res=$(echo "scale=2;((9/5) * $LINE) + 32" | bc)

echo $res

done

To test this script, launch it locally. It will not prompt for input. Type 100 and then

strike Enter. The process prints 212.00 to the standard output. Then enter another

number and the program returns another result. You can continue entering numbers

or use Control-D to end the input.

Using the Distributed Cache | 189

Downloa d f r o m W o w ! e B o o k < w w w.woweb o o k . c o m >

#!/bin/bash

$ sh ctof.sh

100

212.00

32.00

Hive’s ADD FILE feature adds files to the distributed cache. The added file is put in the

current working directory of each task. This allows the transform task to use the script

without needing to know where to find it:

hive> ADD FILE ${env:HOME}/prog_hive/ctof.sh;

Added resource: /home/edward/prog_hive/ctof.sh

hive> SELECT TRANSFORM(col1) USING 'ctof.sh' AS convert FROM a;

39.20

37.40

Producing Multiple Rows from a Single Row

The examples shown thus far have taken one row of input and produced one row of

output. Streaming can be used to produce multiple rows of output for each input row.

This functionality produces output similar to the EXPLODE() UDF and the LATERAL

VIEW syntax1.

Given an input file $HOME/kv_data.txt that looks like:

k1=v1,k2=v2

k4=v4,k5=v5,k6=v6

k7=v7,k7=v7,k3=v7

We would like the data in a tabular form. This will allow the rows to be processed by

familiar HiveQL operators:

k1 v1

k2 v2

k4 k4

Create this Perl script and save it as $HOME/split_kv.pl:

#!/usr/bin/perl

while (<STDIN>) {

my $line = $_;

chomp($line);

my @kvs = split(/,/, $line);

foreach my $p (@kvs) {

my @kv = split(/=/, $p);

print $kv[0] . "\t" . $kv[1] . "\n";

}

1. The source code and concept for this example comes from Larry Ogrodnek, “Custom Map Scripts and

Hive”, Bizo development (blog), July 14, 2009.

190 | Chapter 14: Streaming

Create a kv_data table. The entire table is defined as a single string column. The row

format does not need to be configured because the streaming script will do all the

tokenization of the fields:

hive> CREATE TABLE kv_data ( line STRING );

hive> LOAD DATA LOCAL INPATH '${env:HOME}/kv_data.txt' INTO TABLE kv_data;

Use the transform script on the source table. The ragged, multiple-entry-per-row format

is converted into a two-column result set of key-value pairs:

hive> SELECT TRANSFORM (line)

> USING 'perl split_kv.pl'

> AS (key, value) FROM kv_data;

k1 v1

k2 v2

k4 v4

k5 v5

k6 v6

k7 v7

k3 v7

Calculating Aggregates with Streaming

Streaming can also be used to do aggregating operations like Hive’s built-in SUM func-

tion. This is possible because streaming processes can return zero or more rows of

output for every given input.

To accomplish aggregation in an external application, declare an accumulator before

the loop that reads from the input stream and output the sum after the completion of

the input:

#!/usr/bin/perl

my $sum=0;

while (<STDIN>) {

my $line = $_;

chomp($line);

$sum=${sum}+${line};

}

print $sum;

Create a table and populate it with integer data, one integer per line, for testing:

hive> CREATE TABLE sum (number INT);

hive> LOAD DATA LOCAL INPATH '${env:HOME}/data_to_sum.txt' INTO TABLE sum;

Calculating Aggregates with Streaming | 191

hive> SELECT * FROM sum;

Add the streaming program to the distributed cache and use it in a TRANSFORM query.

The process returns a single row, which is the sum of the input:

hive> ADD FILE ${env:HOME}/aggregate.pl;

Added resource: /home/edward/aggregate.pl

hive> SELECT TRANSFORM (number)

> USING 'perl aggregate.pl' AS total FROM sum;

Unfortunately, it is not possible to do multiple TRANSFORMs in a single query like the

UDAF SUM() can do. For example:

hive> SELECT sum(number) AS one, sum(number) AS two FROM sum;

14 14

Also, without using CLUSTER BY or DISTRIBUTE BY for the intermediate data, this job may

run single, very long map and reduce tasks. While not all operations can be done in

parallel, many can. The next section discusses how to do streaming in parallel, when

possible.

CLUSTER BY, DISTRIBUTE BY, SORT BY

Hive offers syntax to control how data is distributed and sorted. These features can be

used on most queries, but are particularly useful when doing streaming processes. For

example, data for the same key may need to be sent to the same processing node, or

data may need to be sorted by a specific column, or by a function. Hive provides several

ways to control this behavior.

The first way to control this behavior is the CLUSTER BY clause, which ensures like data

is routed to the same reduce task and sorted.

To demonstrate the use of CLUSTER BY, let’s see a nontrivial example: another way to

perform the Word Count algorithm that we introduced in Chapter 1. Now, we will use

the TRANSFORM feature and two Python scripts, one to tokenize lines of text into words,

and the second to accept a stream of word occurrences and an intermediate count of

the words (mostly the number “1”) and then sum up the counts for each word.

Here is the first Python script that tokenizes lines of text on whitespace (which doesn’t

properly handle punctuation, etc.):

192 | Chapter 14: Streaming

import sys

for line in sys.stdin:

words = line.strip().split()

for word in words:

print "%s\t1" % (word.lower())

Without explaining all the Python syntax, this script imports common functions from

a sys module, then it loops over each line on the “standard input,” stdin, splits each

line on whitespace into a collection of words, then iterates over the word and writes

each word, followed by a tab, \t, and the “count” of one.2

Before we show the second Python script, let’s discuss the data that’s passed to it. We’ll

use CLUSTER BY for the words output from the first Python script in our TRANSFORM Hive

query. This will have the effect of causing all occurrences of the word\t1 “pairs” for a

give, word to be grouped together, one pair per line:

word1 1

word2 1

word3 1

...

Hence, the second Python script will be more complex, because it needs to cache the

word it’s currently processing and the count of occurrences seen so far. When the word

changes, the script must output the count for the previous word and reset its caches.

So, here it is:

import sys

(last_key, last_count) = (None, 0)

for line in sys.stdin:

(key, count) = line.strip().split("\t")

if last_key and last_key != key:

print "%s\t%d" % (last_key, last_count)

(last_key, last_count) = (key, int(count))

else:

last_key = key

last_count += int(count)

if last_key:

print "%s\t%d" % (last_key, last_count)

We’ll assume that both Python scripts are in your home directory.

Finally, here is the Hive query that glues it all together. We’ll start by repeating a CREATE

TABLE statement for an input table of lines of text, one that we used in Chapter 1. Any

text file could serve as the data for this table. Next we’ll show the TABLE for the output

2. This is the most naive approach. We could cache the counts of words seen and then write the final count.

That would be faster, by minimizing I/O overhead, but it would also be more complex to implement.

CLUSTER BY, DISTRIBUTE BY, SORT BY | 193

of word count. It will have two columns, the word and count, and data will be tab-

delimited. Finally, we show the TRANSFORM query that glues it all together:

hive> CREATE TABLE docs (line STRING);

hive> CREATE TABLE word_count (word STRING, count INT)

> ROW FORMAT DELIMITED FIELDS TERMINATED BY '\t';

hive> FROM (

> FROM docs

> SELECT TRANSFORM (line) USING '${env:HOME}/mapper.py'

> AS word, count

> CLUSTER BY word) wc

> INSERT OVERWRITE TABLE word_count

> SELECT TRANSFORM (wc.word, wc.count) USING '${env:HOME}/reducer.py'

> AS word, count;

The USING clauses specify an absolute path to the Python scripts.

A more flexible alternative to CLUSTER BY is to use DISTRIBUTE BY and SORT BY. This is

used in the general case when you wish to partition the data by one column and sort it

by another. In fact, CLUSTER BY word is equivalent to DISTRIBUTE BY word SORT BY word

ASC.

The following version of the TRANSFORM query outputs the word count results in reverse

order:

FROM (

FROM docs

SELECT TRANSFORM (line) USING '/.../mapper.py'

AS word, count

DISTRIBUTE BY word SORT BY word DESC) wc

INSERT OVERWRITE TABLE word_count

SELECT TRANSFORM (wc.word, wc.count) USING '/.../reducer.py'

AS word, count;

Using either CLUSTER BY or DISTRIBUTE BY with SORT BY is important. Without these

directives, Hive may not be able to parallelize the job properly. All the data might be

sent to a single reducer, which would extend the job processing time.

GenericMR Tools for Streaming to Java

Typically, streaming is used to integrate non-Java code into Hive. Streaming works with

applications written in essentially any language, as we saw. It is possible to use Java for

streaming, and Hive includes a GenericMR API that attempts to give the feel of the

Hadoop MapReduce API to streaming:

FROM (

FROM src

MAP value, key

USING 'java -cp hive-contrib-0.9.0.jar

org.apache.hadoop.hive.contrib.mr.example.IdentityMapper'

AS k, v

194 | Chapter 14: Streaming

CLUSTER BY k) map_output

REDUCE k, v

USING 'java -cp hive-contrib-0.9.0.jar

org.apache.hadoop.hive.contrib.mr.example.WordCountReduce'

AS k, v;

To understand how the IdentityMapper is written, we can take a look at the interfaces

GenericMR provides. The Mapper interface is implemented to build custom Mapper

implementations. It provides a map method where the column data is sent as a string

array, String []:

package org.apache.hadoop.hive.contrib.mr;

public interface Mapper {

void map(String[] record, Output output) throws Exception;

}

The IdentityMapper makes no changes to the input data and passes it to the collector.

This is functionally equivalent to the identity streaming done with /bin/cat earlier in

the chapter:

package org.apache.hadoop.hive.contrib.mr.example;

import org.apache.hadoop.hive.contrib.mr.GenericMR;

import org.apache.hadoop.hive.contrib.mr.Mapper;

import org.apache.hadoop.hive.contrib.mr.Output;

public final class IdentityMapper {

public static void main(final String[] args) throws Exception {

new GenericMR().map(System.in, System.out, new Mapper() {

@Override

public void map(final String[] record, final Output output) throws Exception {

output.collect(record);

}

});

}

private IdentityMapper() {

}

The Reducer interface provides the first column as a String, and the remaining columns

are available through the record Iterator. Each iteration returns a pair of Strings,

where the 0th element is the key repeated and the next element is the value. The output

object is the same one used to emit results:

package org.apache.hadoop.hive.contrib.mr;

import java.util.Iterator;

public interface Reducer {

void reduce(String key, Iterator<String[]> records, Output output)

throws Exception;

}

GenericMR Tools for Streaming to Java | 195

WordCountReduce has an accumulator that is added by each element taken from the

records Iterator. When all the records have been counted, a single two-element array

of the key and the count is emitted:

package org.apache.hadoop.hive.contrib.mr.example;

import java.util.Iterator;

import org.apache.hadoop.hive.contrib.mr.GenericMR;

import org.apache.hadoop.hive.contrib.mr.Output;

import org.apache.hadoop.hive.contrib.mr.Reducer;

public final class WordCountReduce {

private WordCountReduce() {

}

public static void main(final String[] args) throws Exception {

new GenericMR().reduce(System.in, System.out, new Reducer() {

public void reduce(String key, Iterator<String[]> records, Output output)

throws Exception {

int count = 0;

while (records.hasNext()) {

// note we use col[1] -- the key is provided again as col[0]

count += Integer.parseInt(records.next()[1]);

}

output.collect(new String[] {key, String.valueOf(count)});

}

});

}

Calculating Cogroups

It’s common in MapReduce applications to join together records from multiple data

sets and then stream them through a final TRANSFORM step. Using UNION ALL and CLUSTER

BY, we can perform this generalization of a GROUP BY operation

Pig provides a native COGROUP BY operation.

Suppose we have several sources of logfiles, with similar schema, that we wish to bring

together and analyze with a reduce_script:

196 | Chapter 14: Streaming

FROM (

FROM order_log ol

-- User Id, order Id, and timestamp:

SELECT ol.userid AS uid, ol.orderid AS id, av.ts AS ts

UNION ALL

FROM clicks_log cl

SELECT cl.userid AS uid, cl.id AS id, ac.ts AS ts

) union_msgs

SELECT union_msgs.uid, union_msgs.id, union_msgs.ts

CLUSTER BY union_msgs.uid, union_msgs.ts) map

INSERT OVERWRITE TABLE log_analysis

SELECT TRANSFORM(map.uid, map.id, map.ts) USING 'reduce_script'

AS (uid, id, ...);

Calculating Cogroups | 197

CHAPTER 15

Customizing Hive File and

Record Formats

Hive functionality can be customized in several ways. First, there are the variables and

properties that we discussed in “Variables and Properties” on page 31. Second, you

may extend Hive using custom UDFs, or user-defined functions, which was discussed

in Chapter 13. Finally, you can customize the file and record formats, which we discuss

now.

File Versus Record Formats

Hive draws a clear distinction between the file format, how records are encoded in a

file, the record format, and how the stream of bytes for a given record are encoded in

the record.

In this book we have been using text files, with the default STORED AS TEXTFILE in CREATE

TABLE statements (see “Text File Encoding of Data Values” on page 45), where each

line in the file is a record. Most of the time those records have used the default sepa-

rators, with occasional examples of data that use commas or tabs as field separators.

However, a text file could contain JSON or XML “documents.”

For Hive, the file format choice is orthogonal to the record format. We’ll first discuss

options for file formats, then we’ll discuss different record formats and how to use them

in Hive.

Demystifying CREATE TABLE Statements

Throughout the book we have shown examples of creating tables. You may have no-

ticed that CREATE TABLE has a variety of syntax. Examples of this syntax are STORED AS

SEQUENCEFILE, ROW FORMAT DELIMITED , SERDE, INPUTFORMAT, OUTPUTFORMAT. This chapter

will cover much of this syntax and give examples, but as a preface note that some syntax

199

is sugar for other syntax, that is, syntax used to make concepts easier (sweeter) to

understand. For example, specifying STORED AS SEQUENCEFILE is an alternative to spec-

ifying an INPUTFORMAT of org.apache.hadoop.mapred.SequenceFileInputFormat and an

OUTPUTFORMAT of org.apache.hadoop.hive.ql.io.HiveSequenceFileOutputFormat.

Let’s create some tables and use DESCRIBE TABLE EXTENDED to peel away the sugar and

expose the internals. First, we will create and then describe a simple table (we have

formatted the output here, as Hive otherwise would not have indented the output):

hive> create table text (x int) ;

hive> describe extended text;

x int

Detailed Table Information

Table(tableName:text, dbName:default, owner:edward, createTime:1337814583,

lastAccessTime:0, retention:0,

sd:StorageDescriptor(

cols:[FieldSchema(name:x, type:int, comment:null)],

location:file:/user/hive/warehouse/text,

inputFormat:org.apache.hadoop.mapred.TextInputFormat,

outputFormat:org.apache.hadoop.hive.ql.io.HiveIgnoreKeyTextOutputFormat,

compressed:false,

numBuckets:-1,

serdeInfo:SerDeInfo(

name:null,

serializationLib:org.apache.hadoop.hive.serde2.lazy.LazySimpleSerDe,

parameters:{serialization.format=1}

bucketCols:[], sortCols:[], parameters:{}), partitionKeys:[],

parameters:{transient_lastDdlTime=1337814583},

viewOriginalText:null, viewExpandedText:null, tableType:MANAGED_TABLE

)

Now let’s create a table using STORED AS SEQUENCEFILE for comparison:

hive> CREATE TABLE seq (x int) STORED AS SEQUENCEFILE;

hive> DESCRIBE EXTENDED seq;

x int

Detailed Table Information

Table(tableName:seq, dbName:default, owner:edward, createTime:1337814571,

lastAccessTime:0, retention:0,

sd:StorageDescriptor(

cols:[FieldSchema(name:x, type:int, comment:null)],

location:file:/user/hive/warehouse/seq,

inputFormat:org.apache.hadoop.mapred.SequenceFileInputFormat,

outputFormat:org.apache.hadoop.hive.ql.io.HiveSequenceFileOutputFormat,

compressed:false, numBuckets:-1,

serdeInfo:SerDeInfo(

name:null,

serializationLib:org.apache.hadoop.hive.serde2.lazy.LazySimpleSerDe,

parameters:{serialization.format=1}

200 | Chapter 15: Customizing Hive File and Record Formats

bucketCols:[], sortCols:[], parameters:{}), partitionKeys:[],

parameters:{transient_lastDdlTime=1337814571},

viewOriginalText:null, viewExpandedText:null, tableType:MANAGED_TABLE

)

Time taken: 0.107 seconds

Unless you have been blinded by Hive’s awesomeness, you would have picked up on

the difference between these two tables. That STORED AS SEQUENCEFILE has changed the

InputFormat and the OutputFormat:

inputFormat:org.apache.hadoop.mapred.TextInputFormat,

outputFormat:org.apache.hadoop.hive.ql.io.HiveIgnoreKeyTextOutputFormat,

inputFormat:org.apache.hadoop.mapred.SequenceFileInputFormat,

outputFormat:org.apache.hadoop.hive.ql.io.HiveSequenceFileOutputFormat,

Hive uses the InputFormat when reading data from the table, and it uses the OutputFor

mat when writing data to the table.

InputFormat reads key-value pairs from files; Hive currently ignores the

key and works only with the data found in the value by default. The

reason for this is that the key, which comes from TextInputFormat, is a

long integer that represents the byte offset in the block (which is not

user data).

The rest of the chapter describes other aspects of the table metadata.

File Formats

We discussed in “Text File Encoding of Data Values” on page 45 that the simplest data

format to use is the text format, with whatever delimiters you prefer. It is also the default

format, equivalent to creating a table with the clause STORED AS TEXTFILE.

The text file format is convenient for sharing data with other tools, such as Pig, Unix

text tools like grep, sed, and awk, etc. It’s also convenient for viewing or editing files

manually. However, the text format is not space efficient compared to binary formats.

We can use compression, as we discussed in Chapter 11, but we can also gain more

efficient usage of disk space and better disk I/O performance by using binary file

formats.

SequenceFile

The first alternative is the SequenceFile format, which we can specify using the STORED

AS SEQUENCEFILE clause during table creation.

Sequence files are flat files consisting of binary key-value pairs. When Hive converts

queries to MapReduce jobs, it decides on the appropriate key-value pairs to use for a

given record.

File Formats | 201

The sequence file is a standard format supported by Hadoop itself, so it is an acceptable

choice when sharing files between Hive and other Hadoop-related tools. It’s less suit-

able for use with tools outside the Hadoop ecosystem. As we discussed in Chap-

ter 11, sequence files can be compressed at the block and record level, which is very

useful for optimizing disk space utilization and I/O, while still supporting the ability

to split files on block boundaries for parallel processing.

Another efficient binary format that is supported natively by Hive is RCFile.

RCFile

Most Hadoop and Hive storage is row oriented, which is efficient in most cases. The

efficiency can be attributed to several factors: most tables have a smaller number

(1−20) of columns. Compression on blocks of a file is efficient for dealing with repeating

data, and many processing and debugging tools (more, head, awk) work well with row-

oriented data.

Not all tools and data stores take a row-oriented approach; column-oriented organi-

zation is a good storage option for certain types of data and applications. For example,

if a given table has hundreds of columns but most queries use only a few of the columns,

it is wasteful to scan entire rows then discard most of the data. However, if the data is

stored by column instead of by row, then only the data for the desired columns has to

be read, improving performance.

It also turns out that compression on columns is typically very efficient, especially

when the column has low cardinality (only a few distinct entries). Also, some column-

oriented stores do not physically need to store null columns.

Hive’s RCFile is designed for these scenarios.

While books like Programming Hive are invaluable sources of information, sometimes

the best place to find information is inside the source code itself. A good description

of how Hive’s column storage known as RCFile works is found in the source code:

cd hive-trunk

find . -name "RCFile*"

vi ./ql/src/java/org/apache/hadoop/hive/ql/io/RCFile.java

* <p>

* RCFile stores columns of a table in a record columnar way. It first

* partitions rows horizontally into row splits. and then it vertically

* partitions each row split in a columnar way. RCFile first stores the meta

* data of a row split, as the key part of a record, and all the data of a row

* split as the value part.

* </p>

A powerful aspect of Hive is that converting data between different formats is simple.

Storage information is stored in the tables metadata. When a query SELECTs from one

table and INSERTs into another, Hive uses the metadata about the tables and handles

the conversion automatically. This makes for easy evaluation of the different options

without writing one-off programs to convert data between the different formats.

202 | Chapter 15: Customizing Hive File and Record Formats

Creating a table using the ColumnarSerDe, RCFileInputFormat, and RCFileOutputFormat:

hive> select * from a;

4 5

3 2

Time taken: 0.336 seconds

hive> create table columnTable (key int , value int)

> ROW FORMAT SERDE

> 'org.apache.hadoop.hive.serde2.columnar.ColumnarSerDe'

> STORED AS

> INPUTFORMAT 'org.apache.hadoop.hive.ql.io.RCFileInputFormat'

> OUTPUTFORMAT 'org.apache.hadoop.hive.ql.io.RCFileOutputFormat';

hive> FROM a INSERT OVERWRITE TABLE columnTable SELECT a.col1, a.col2;

RCFile’s cannot be opened with the tools that open typical sequence files. However,

Hive provides an rcfilecat tool to display the contents of RCFiles:

$ bin/hadoop dfs -text /user/hive/warehouse/columntable/000000_0

text: java.io.IOException: WritableName can't load class:

org.apache.hadoop.hive.ql.io.RCFile$KeyBuffer

$ bin/hive --service rcfilecat /user/hive/warehouse/columntable/000000_0

4 5

3 2

Example of a Custom Input Format: DualInputFormat

Many databases allow users to SELECT without FROM. This can be used to perform simple

calculations, such as SELECT 1+2. If Hive did not allow this type of query, then a user

would instead select from an existing table and limit the results to a single row. Or the

user may create a table with a single row. Some databases provide a table named

dual, which is a single row table to be used in this manner.

By default, a standard Hive table uses the TextInputFormat. The TextInputFormat cal-

culates zero or more splits for the input. Splits are opened by the framework and a

RecordReader is used to read the data. Each row of text becomes an input record. To

create an input format that works with a dual table, we need to create an input format

that returns one split with one row, regardless of the input path specified.1

In the example below, DualInputFormat returns a single split:

public class DualInputFormat implements InputFormat{

public InputSplit[] getSplits(JobConf jc, int i) throws IOException {

InputSplit [] splits = new DualInputSplit[1];

splits[0]= new DualInputSplit();

return splits;

}

public RecordReader<Text,Text> getRecordReader(InputSplit split, JobConf jc,

Reporter rprtr) throws IOException {

1. The source code for the DualInputFormat is available at: https://github.com/edwardcapriolo/

DualInputFormat.

File Formats | 203

return new DualRecordReader(jc, split);

}

In the example below the split is a single row. There is nothing to serialize or deserialize:

public class DualInputSplit implements InputSplit {

public long getLength() throws IOException {

return 1;

}

public String[] getLocations() throws IOException {

return new String [] { "localhost" };

}

public void write(DataOutput d) throws IOException {

}

public void readFields(DataInput di) throws IOException {

}

The DualRecordReader has a Boolean variable hasNext. After the first invocation of

next(), its value is set to false. Thus, this record reader returns a single row and then

is finished with virtual input:

public class DualRecordReader implements RecordReader<Text,Text>{

boolean hasNext=true;

public DualRecordReader(JobConf jc, InputSplit s) {

}

public DualRecordReader(){

}

public long getPos() throws IOException {

return 0;

}

public void close() throws IOException {

}

public float getProgress() throws IOException {

if (hasNext)

return 0.0f;

else

return 1.0f;

}

public Text createKey() {

return new Text("");

}

public Text createValue() {

return new Text("");

}

public boolean next(Text k, Text v) throws IOException {

if (hasNext){

hasNext=false;

return true;

} else {

return hasNext;

}

204 | Chapter 15: Customizing Hive File and Record Formats

Downloa d f r o m W o w ! e B o o k < w w w.woweb o o k . c o m >

We can create a table using our DualInputFormat and the default HiveIgnoreKeyTextOut

putFormat. Selecting from the table confirms that it returns a single empty row. Input

Formats should be placed inside the Hadoop lib directory or preferably inside the Hive

auxlib directory.

client.execute("add jar dual.jar");

client.execute("create table dual (fake string) "+

"STORED AS INPUTFORMAT 'com.m6d.dualinputformat.DualInputFormat'"+

"OUTPUTFORMAT 'org.apache.hadoop.hive.ql.io.HiveIgnoreKeyTextOutputFormat'");

client.execute("select count(1) as cnt from dual");

String row = client.fetchOne();

assertEquals("1", row);

client.execute("select * from dual");

row = client.fetchOne();

assertEquals( "", row);

Record Formats: SerDes

SerDe is short for serializer/deserializer. A SerDe encapsulates the logic for converting

the unstructured bytes in a record, which is stored as part of a file, into a record that

Hive can use. SerDes are implemented using Java. Hive comes with several built-in

SerDes and many other third-party SerDes are available.

Internally, the Hive engine uses the defined InputFormat to read a record of data. That

record is then passed to the SerDe.deserialize() method.

A lazy SerDe does not fully materialize an object until individual attributes are

necessary.

The following example uses a RegexSerDe to parse a standard formatted Apache web

log. The RegexSerDe is included as a standard feature as a part of the Hive distribution:

CREATE TABLE serde_regex(

host STRING,

identity STRING,

user STRING,

time STRING,

request STRING,

status STRING,

size STRING,

referer STRING,

agent STRING)

ROW FORMAT SERDE 'org.apache.hadoop.hive.contrib.serde2.RegexSerDe'

WITH SERDEPROPERTIES (

"input.regex" = "([^ ]*) ([^ ]*) ([^ ]*) (-|\\[[^\\]]*\\])

([^ \"]*|\"[^\"]*\") (-|[0-9]*) (-|[0-9]*)(?: ([^ \"]*|\"[^\"]*\")

([^ \"]*|\"[^\"]*\"))?",

"output.format.string" = "%1$s %2$s %3$s %4$s %5$s %6$s %7$s %8$s %9$s"

)

STORED AS TEXTFILE;

Now we can load data and write queries:

Record Formats: SerDes | 205

hive> LOAD DATA LOCAL INPATH "../data/files/apache.access.log" INTO TABLE serde_regex;

hive> LOAD DATA LOCAL INPATH "../data/files/apache.access.2.log" INTO TABLE serde_regex;

hive> SELECT * FROM serde_regex ORDER BY time;

(The long regular expression was wrapped to fit.)

CSV and TSV SerDes

What about CSV (comma-separated values) and TSV (tab-separated values) files? Of

course, for simple data such as numerical data, you can just use the default test file

format and specify the field delimiter, as we saw previously. However, this simplistic

approach doesn’t handle strings with embedded commas or tabs, nor does it handle

other common conventions, like whether or not to quote all or no strings, or the op-

tional presence of a “column header” row as the first line in each file.

First, it’s generally safer to remove the column header row, if present. Then one of

several third-party SerDes are available for properly parsing CSV or TSV files. For CSV

files, consider CSVSerde:

ADD JAR /path/to/csv-serde.jar;

CREATE TABLE stocks(ymd STRING, ...)

ROW FORMAT SERDE 'com.bizo.hive.serde.csv.CSVSerde'

STORED AS TEXTFILE

...;

While TSV support should be similar, there are no comparable third-party TSV SerDes

available at the time of this writing.

ObjectInspector

Underneath the covers, Hive uses what is known as an ObjectInspector to transform

raw records into objects that Hive can access.

Think Big Hive Reflection ObjectInspector

Think Big Analytics has created an ObjectInspector based on Java reflection called

BeansStructObjectInspector. Using the JavaBeans model for introspection, any

“property” on objects that are exposed through get methods or as public member

variables may be referenced in queries.

An example of how to use the BeansStructObjectInspector is as follows:

public class SampleDeserializer implements Deserializer {

@Override

public ObjectInspector getObjectInspector() throws SerDeException {

return BeansStructObjectInspector.getBeansObjectInspector(YourObject.class);

206 | Chapter 15: Customizing Hive File and Record Formats

}

XML UDF

XML is inherently unstructured, which makes Hive a powerful database platform for

XML. One of the reasons Hadoop is ideal as an XML database platform is the com-

plexity and resource consumption to parse and process potentially large XML docu-

ments. Because Hadoop parallelizes processing of XML documents, Hive becomes a

perfect tool for accelerating XML-related data solutions. Additionally, HiveQL natively

enables access to XML’s nested elements and values, then goes further by allowing joins

on any of the nested fields, values, and attributes.

XPath (XML Path Language) is a global standard created by the W3C for addressing

parts of an XML document. Using XPath as an expressive XML query language, Hive

becomes extremely useful for extracting data from XML documents and into the Hive

subsystem.

XPath models an XML document as a tree of nodes. Basic facilities are provided for

access to primitive types, such as string, numeric, and Boolean types.

While commercial solutions such as Oracle XML DB and MarkLogic provide native

XML database solutions, open source Hive leverages the advantages provided by the

parallel petabyte processing of the Hadoop infrastructure to enable widely effective

XML database vivification.

XPath-Related Functions

Hive contains a number of XPath-related UDFs since the 0.6.0 release (Table 15-1).

Table 15-1. XPath UDFs

Name Description

xpath Returns a Hive array of strings

xpath_string Returns a string

xpath_boolean Returns a Boolean

xpath_short Returns a short integer

xpath_int Returns an integer

xpath_long Returns a long integer

xpath_float Returns a floating-point number

xpath_double, xpath_number Returns a double-precision floating-point number

Here are some examples where these functions are run on string literals:

XPath-Related Functions | 207

hive> SELECT xpath(\'<a><b id="foo">b1</b><b id="bar">b2</b></a>\',\'//@id\')

> FROM src LIMIT 1;

[foo","bar]

hive> SELECT xpath (\'<a><b class="bb">b1</b><b>b2</b><b>b3</b><c class="bb">c1</c>

<c>c2</c></a>\', \'a/*[@class="bb"]/text()\')

> FROM src LIMIT 1;

[b1","c1]

(The long XML string was wrapped for space.)

hive> SELECT xpath_double (\'<a><b>2</b><c>4</c></a>\', \'a/b + a/c\')

> FROM src LIMIT 1;

6.0

JSON SerDe

What if you want to query JSON (JavaScript Object Notation) data with Hive? If each

JSON “document” is on a separate line, you can use TEXTFILE as the input and output

format, then use a JSON SerDe to parse each JSON document as a record.

There is a third-party JSON SerDe that started as a Google “Summer of Code”

project and was subsequently cloned and forked by other contributors. Think Big An-

alytics created its own fork and added an enhancement we’ll go over in the discussion

that follows.

In the following example, this SerDe is used to extract a few fields from JSON data for

a fictitious messaging system. Not all the available fields are exposed. Those that are

exposed become available as columns in the table:

CREATE EXTERNAL TABLE messages (

msg_id BIGINT,

tstamp STRING,

text STRING,

user_id BIGINT,

user_name STRING

)

ROW FORMAT SERDE "org.apache.hadoop.hive.contrib.serde2.JsonSerde"

WITH SERDEPROPERTIES (

"msg_id"="$.id",

"tstamp"="$.created_at",

"text"="$.text",

"user_id"="$.user.id",

"user_name"="$.user.name"

)

LOCATION '/data/messages';

The WITH SERDEPROPERTIES is a Hive feature that allows the user to define properties

that will be passed to the SerDe. The SerDe interprets those properties as it sees fit.

Hive doesn’t know or care what they mean.

In this case, the properties are used to map fields in the JSON documents to columns

in the table. A string like $.user.id means to take each record, represented by $, find

208 | Chapter 15: Customizing Hive File and Record Formats

the user key, which is assumed to be a JSON map in this case, and finally extract the

value for the id key inside the user. This value for the id is used as the value for the

user_id column.

Once defined, the user runs queries as always, blissfully unaware that the queries are

actually getting data from JSON!

Avro Hive SerDe

Avro is a serialization systemit’s main feature is an evolvable schema-driven binary data

format. Initially, Avro’s goals appeared to be in conflict with Hive since both wish to

provide schema or metadata information. However Hive and the Hive metastore have

pluggable design and can defer to the Avro support to infer the schema.

The Hive Avro SerDe system was created by LinkedIn and has the following features:

• Infers the schema of the Hive table from the Avro schema

• Reads all Avro files within a table against a specified schema, taking advantage of

Avro’s backwards compatibility

• Supports arbitrarily nested schemas

• Translates all Avro data types into equivalent Hive types. Most types map exactly,

but some Avro types do not exist in Hive and are automatically converted by Hive

with Avro

• Understands compressed Avro files

• Transparently converts the Avro idiom of handling nullable types as Union[T,

null] into just T and returns null when appropriate

• Writes any Hive table to Avro files

Defining Avro Schema Using Table Properties

Create an Avro table by specifying the AvroSerDe, AvroContainerInputFormat, and Avro

ContainerOutputFormat. Avro has its own schema definition language. This schema

definition language can be stored in the table properties as a string literal using the

property avro.schema.literal. The schema specifies three columns: number as int,

firstname as string, and lastname as string.

CREATE TABLE doctors

ROW FORMAT

SERDE 'org.apache.hadoop.hive.serde2.avro.AvroSerDe'

STORED AS

INPUTFORMAT 'org.apache.hadoop.hive.ql.io.avro.AvroContainerInputFormat'

OUTPUTFORMAT 'org.apache.hadoop.hive.ql.io.avro.AvroContainerOutputFormat'

TBLPROPERTIES ('avro.schema.literal'='{

"namespace": "testing.hive.avro.serde",

"name": "doctors",

"type": "record",

Avro Hive SerDe | 209

"fields": [

{

"name":"number",

"type":"int",

"doc":"Order of playing the role"

{

"name":"first_name",

"type":"string",

"doc":"first name of actor playing role"

{

"name":"last_name",

"type":"string",

"doc":"last name of actor playing role"

}

]

}');

When the DESCRIBE command is run, Hive shows the name and types of the columns.

In the output below you will notice that the third column of output states from deser

ializer. This shows that the SerDe itself returned the information from the column

rather than static values stored in the metastore:

hive> DESCRIBE doctors;

number int from deserializer

first_name string from deserializer

last_name string from deserializer

Defining a Schema from a URI

It is also possible to provide the schema as a URI. This can be a path to a file in HDFS

or a URL to an HTTP server. To do this, specify avro.schema.url in table properties

and do not specify avro.schema.literal.

The schema can be a file in HDFS:

TBLPROPERTIES ('avro.schema.url'='hdfs://hadoop:9020/path/to.schema')

The schema can also be stored on an HTTP server:

TBLPROPERTIES ('avro.schema.url'='http://site.com/path/to.schema')

Evolving Schema

Over time fields may be added or deprecated from data sets. Avro is designed with this

in mind. An evolving schema is one that changes over time. Avro allows fields to be

null. It also allows for default values to be returned if the column is not defined in the

data file.

For example, if the Avro schema is changed and a field added, the default field supplies

a value if the column is not found:

210 | Chapter 15: Customizing Hive File and Record Formats

{

"name":"extra_field",

"type":"string",

"doc:":"an extra field not in the original file",

"default":"fishfingers and custard"

}

Binary Output

There are several kinds of binary output. We have already seen compression of files,

sequence files (compressed or not), and related file types.

Sometimes, it’s also useful to read and write streams of bytes. For example, you may

have tools that expect a stream of bytes, without field separators of any kind, and you

either use Hive to generate suitable files for those tools or you want to query such files

with Hive. You may also want the benefits of storing numbers in compact binary forms

instead of strings like “5034223,” which consume more space. A common example is

to query the output of the tcpdump command to analyze network behavior.

The following table expects its own files to be in text format, but it writes query results

as binary streams:

CREATE TABLE binary_table (num1 INT, num2 INT)

ROW FORMAT SERDE 'org.apache.hadoop.hive.serde2.lazy.LazySimpleSerDe'

WITH SERDEPROPERTIES ('serialization.last.column.takes.rest'='true')

STORED AS

INPUTFORMAT 'org.apache.hadoop.mapred.TextInputFormat'

OUTPUTFORMAT 'org.apache.hadoop.hive.ql.io.HiveBinaryOutputFormat';

Here’s a SELECT TRANSFORM query that reads binary data from a src table, streams it

through the shell cat command and overwrites the contents of a destination1 table:

INSERT OVERWRITE TABLE destination1

SELECT TRANSFORM(*)

USING 'cat' AS mydata STRING

ROW FORMAT SERDE 'org.apache.hadoop.hive.serde2.lazy.LazySimpleSerDe'

WITH SERDEPROPERTIES ('serialization.last.column.takes.rest'='true')

RECORDREADER 'org.apache.hadoop.hive.ql.exec.BinaryRecordReader'

FROM src;

Binary Output | 211

CHAPTER 16

Hive Thrift Service

Hive has an optional component known as HiveServer or HiveThrift that allows access

to Hive over a single port. Thrift is a software framework for scalable cross-language

services development. See http://thrift.apache.org/ for more details. Thrift allows clients

using languages including Java, C++, Ruby, and many others, to programmatically

access Hive remotely.

The CLI is the most common way to access Hive. However, the design of the CLI can

make it difficult to use programmatically. The CLI is a fat client; it requires a local copy

of all the Hive components and configuration as well as a copy of a Hadoop client and

its configuration. Additionally, it works as an HDFS client, a MapReduce client, and a

JDBC client (to access the metastore). Even with the proper client installation, having

all of the correct network access can be difficult, especially across subnets or

datacenters.

Starting the Thrift Server

To Get started with the HiveServer, start it in the background using the service knob

for hive:

$ cd $HIVE_HOME

$ bin/hive --service hiveserver &

Starting Hive Thrift Server

A quick way to ensure the HiveServer is running is to use the netstat command to

determine if port 10,000 is open and listening for connections:

$ netstat -nl | grep 10000

tcp 0 0 :::10000 :::* LISTEN

(Some whitespace removed.) As mentioned, the HiveService uses Thrift. Thrift provides

an interface language. With the interface, the Thrift compiler generates code that cre-

ates network RPC clients for many languages. Because Hive is written in Java, and Java

bytecode is cross-platform, the clients for the Thrift server are included in the Hive

213

release. One way to use these clients is by starting a Java project with an IDE and

including these libraries or fetching them through Maven.

Setting Up Groovy to Connect to HiveService

For this example we will use Groovy. Groovy is an agile and dynamic language for the

Java Virtual Machine. Groovy is ideal for prototyping because it integrates with Java

and provides a read-eval-print-loop (REPL) for writing code on the fly:

$ curl -o http://dist.groovy.codehaus.org/distributions/groovy-binary-1.8.6.zip

$ unzip groovy-binary-1.8.6.zip

Next, add all Hive JARs to Groovy’s classpath by editing the groovy-starter.conf.

This will allow Groovy to communicate with Hive without having to manually load

JAR files each session:

# load required libraries

load !{groovy.home}/lib/*.jar

# load user specific libraries

load !{user.home}/.groovy/lib/*.jar

# tools.jar for ant tasks

load ${tools.jar}

load /home/edward/hadoop/hadoop-0.20.2_local/*.jar

load /home/edward/hadoop/hadoop-0.20.2_local/lib/*.jar

load /home/edward/hive-0.9.0/lib/*.jar

Groovy has an @grab annotation that can fetch JAR files from Maven

web repositories, but currently some packaging issues with Hive prevent

this from working correctly.

Groovy provides a shell found inside the distribution at bin/groovysh. Groovysh provides

a REPL for interactive programming. Groovy code is similar to Java code, although it

does have other forms including closures. For the most part, you can write Groovy as

you would write Java.

Connecting to HiveServer

From the REPL, import Hive- and Thrift-related classes. These classes are used to con-

nect to Hive and create an instance of HiveClient. HiveClient has the methods users

will typically use to interact with Hive:

$ $HOME/groovy/groovy-1.8.0/bin/groovysh

Groovy Shell (1.8.0, JVM: 1.6.0_23)

Type 'help' or '\h' for help.

groovy:000> import org.apache.hadoop.hive.service.*;

214 | Chapter 16: Hive Thrift Service

groovy:000> import org.apache.thrift.protocol.*;

groovy:000> import org.apache.thrift.transport.*;

groovy:000> transport = new TSocket("localhost" , 10000);

groovy:000> protocol = new TBinaryProtocol(transport);

groovy:000> client = new HiveClient(protocol);

groovy:000> transport.open();

groovy:000> client.execute("show tables");

Getting Cluster Status

The getClusterStatus method retrieves information from the Hadoop JobTracker. This

can be used to collect performance metrics and can also be used to wait for a lull to

launch a job:

groovy:000> client.getClusterStatus()

===> HiveClusterStatus(taskTrackers:50, mapTasks:52, reduceTasks:40,

maxMapTasks:480, maxReduceTasks:240, state:RUNNING)

Result Set Schema

After executing a query, you can get the schema of the result set using the get

Schema() method. If you call this method before a query, it may return a null schema:

groovy:000> client.getSchema()

===> Schema(fieldSchemas:null, properties:null)

groovy:000> client.execute("show tables");

===> null

groovy:000> client.getSchema()

===> Schema(fieldSchemas:[FieldSchema(name:tab_name, type:string,

comment:from deserializer)], properties:null)

Fetching Results

After a query is run, you can fetch results with the fetchOne() method. Retrieving large

result sets with the Thrift interface is not suggested. However, it does offer several

methods to retrieve data using a one-way cursor. The fetchOne() method retrieves an

entire row:

groovy:000> client.fetchOne()

===> cookjar_small

Instead of retrieving rows one at a time, the entire result set can be retrieved as a string

array using the fetchAll() method:

groovy:000> client.fetchAll()

===> [macetest, missing_final, one, time_to_serve, two]

Also available is fetchN, which fetches N rows at a time.

Fetching Results | 215

Retrieving Query Plan

After a query is started, the getQueryPlan() method is used to retrieve status information

about the query. The information includes information on counters and the state of

the job:

groovy:000> client.execute("SELECT * FROM time_to_serve");

===> null

groovy:000> client.getQueryPlan()

===> QueryPlan(queries:[Query(queryId:hadoop_20120218180808_...-aedf367ea2f3,

queryType:null, queryAttributes:{queryString=SELECT * FROM time_to_serve},

queryCounters:null, stageGraph:Graph(nodeType:STAGE, roots:null,

adjacencyList:null), stageList:null, done:true, started:true)],

done:false, started:false)

(A long number was elided.)

Metastore Methods

The Hive service also connects to the Hive metastore via Thrift. Generally, users should

not call metastore methods that modify directly and should only interact with Hive via

the HiveQL language. Users should utilize the read-only methods that

provide meta-information about tables. For example, the get_partition_names

(String,String,short) method can be used to determine which partitions are available

to a query:

groovy:000> client.get_partition_names("default", "fracture_act", (short)0)

[ hit_date=20120218/mid=001839,hit_date=20120218/mid=001842,

hit_date=20120218/mid=001846 ]

It is important to remember that while the metastore API is relatively stable in terms

of changes, the methods inside, including their signatures and purpose, can change

between releases. Hive tries to maintain compatibility in the HiveQL language, which

masks changes at these levels.

Example Table Checker

The ability to access the metastore programmatically provides the capacity to monitor

and enforce conditions across your deployment. For example, a check can be written

to ensure that all tables use compression, or that tables with names that start with zz

should not exist longer than 10 days. These small “Hive-lets” can be written quickly

and executed remotely, if necessary.

Finding tables not marked as external

By default, managed tables store their data inside the warehouse directory, which is /

user/hive/warehouse by default. Usually, external tables do not use this directory, but

there is nothing that prevents you from putting them there. Enforcing a rule that

216 | Chapter 16: Hive Thrift Service

managed tables should only be inside the warehouse directory will keep the environ-

ment sane.

In the following application, the outer loop iterates through the list returned

from get_all_databases(). The inner loop iterates through the list returned

from get_all_tables(database). The Table object returned from get_table(data

base,table) has all the information about the table in the metastore. We determine the

location of the table and check that the type matches the string MANAGED_TABLE. External

tables have a type EXTERNAL. A list of “bad” table names is returned:

public List<String> check(){

List<String> bad = new ArrayList<String>();

for (String database: client.get_all_databases() ){

for (String table: client.get_all_tables(database) ){

try {

Table t = client.get_table(database,table);

URI u = new URI(t.getSd().getLocation());

if (t.getTableType().equals("MANAGED_TABLE") &&

! u.getPath().contains("/user/hive/warehouse") ){

System.out.println(t.getTableName()

+ " is a non external table mounted inside /user/hive/warehouse" );

bad.add(t.getTableName());

}

} catch (Exception ex){

System.err.println("Had exception but will continue " +ex);

}

return bad;

}

Administrating HiveServer

The Hive CLI creates local artifacts like the .hivehistory file along with entries in /tmp

and hadoop.tmp.dir. Because the HiveService becomes the place where Hadoop jobs

launch from, there are some considerations when deploying it.

Productionizing HiveService

HiveService is a good alternative to having the entire Hive client install local to the

machine that launches the job. Using it in production does bring up some added issues

that need to be addressed. The work that used to be done on the client machine, in

planning and managing the tasks, now happens on the server. If you are launching

many clients simultaneously, this could cause too much load for a single HiveService.

A simple solution is to use a TCP load balancer or proxy to alternate connections be-

tween a pool of backend servers.

Administrating HiveServer | 217

There are several ways to do TCP load balancing and you should consult your network

administrator for the best solution. We suggest a simple solution that uses the hap

roxy tool to balance connections between backend ThriftServers.

First, inventory your physical ThriftServers and document the virtual server that will

be your proxy (Tables 16-1 and 16-2).

Table 16-1. Physical server inventory

Short name Hostname and port

HiveService1 hiveservice1.example.pvt:10000

HiveService2 hiveservice2.example.pvt:10000

Table 16-2. Proxy Configuration

Hostname IP

hiveprimary.example.pvt 10.10.10.100

Install ha-proxy (HAP). Depending on your operating system and distribution these

steps may be different. Example assumes a RHEL/CENTOS distribution:

$sudo yum install haproxy

Use the inventory prepared above to build the configuration file:

$ more /etc/haproxy/haproxy.cfg

listen hiveprimary 10.10.10.100:10000

balance leastconn

mode tcp

server hivethrift1 hiveservice1.example.pvt:10000 check

server hivethrift2 hiveservice1.example.pvt:10000 check

Start HAP via the system init script. After you have confirmed it is working, add it to

the default system start-up with chkconfig:

$ sudo /etc/init.d/haproxy start

$ sudo chkconfig haproxy on

Cleanup

Hive offers the configuration variable hive.start.cleanup.scratchdir, which is set to

false. Setting it to true will cause the service to clean up its scratch directory on restart:

<name>hive.start.cleanup.scratchdir</name>

<description>To clean up the Hive scratchdir while

starting the Hive server</description>

</property>

218 | Chapter 16: Hive Thrift Service

Hive ThriftMetastore

Typically, a Hive session connects directly to a JDBC database, which it uses as a met-

astore. Hive provides an optional component known as the ThriftMetastore. In this

setup, the Hive client connects to the ThriftMetastore, which in turn communicates to

the JDBC Metastore. Most deployments will not require this component. It is useful

for deployments that have non-Java clients that need access to information in the met-

astore. Using the metastore will require two separate configurations.

ThriftMetastore Configuration

The ThriftMetastore should be set up to communicate with the actual metastore using

JDBC. Then start up the metastore in the following manner:

$ cd ~

$ bin/hive --service metastore &

[1] 17096

Starting Hive Metastore Server

Confirm the metastore is running using the netstat command:

$ netstat -an | grep 9083

tcp 0 0 :::9083 :::* LISTEN

Client Configuration

Clients like the CLI should communicate with the metastore directory:

<name>hive.metastore.local</name>

<value>false</value>

<description>controls whether to connect to remove metastore server

or open a new metastore server in Hive Client JVM</description>

</property>

<name>hive.metastore.uris</name>

<value>thrift://metastore_server:9083</value>

<description>controls whether to connect to remove metastore server

or open a new metastore server in Hive Client JVM</description>

</property>

This change should be seamless from the user experience. Although, there are some

nuances with Hadoop Security and the metastore having to do work as the user.

Hive ThriftMetastore | 219

CHAPTER 17

Storage Handlers and NoSQL

Storage Handlers are a combination of InputFormat, OutputFormat, SerDe, and specific

code that Hive uses to treat an external entity as a standard Hive table. This allows the

user to issue queries seamlessly whether the table represents a text file stored in Hadoop

or a column family stored in a NoSQL database such as Apache HBase, Apache Cas-

sandra, and Amazon DynamoDB. Storage handlers are not only limited to NoSQL da-

tabases, a storage handler could be designed for many different kinds of data stores.

A specific storage handler may only implement some of the capabilities.

For example, a given storage handler may allow read-only access or im-

pose some other restriction.

Storage handlers offer a streamlined system for ETL. For example, a Hive query could

be run that selects a data table that is backed by sequence files, however it could output

Storage Handler Background

Hadoop has an abstraction known as InputFormat that allows data from different sour-

ces and formats to be used as input for a job. The TextInputFormat is a concrete im-

plementation of InputFormat. It works by providing Hadoop with information on how

to split a given path into multiple tasks, and it provides a RecordReader that provides

methods for reading data from each split.

Hadoop also has an abstraction known as OutputFormat, which takes the output from

a job and outputs it to an entity. The TextOutputFormat is a concrete implementation

of OutputFormat. It works by persisting output to a file which could be stored on HDFS

or locally.

Input and output that represent physical files are common in Hadoop, however Input

Format and OutputFormat abstractions can be used to load and persist data from other

221

sources including relational databases, NoSQL stores like Cassandra or HBase, or any-

thing that InputFormat or OutputFormat can be designed around!

In the HiveQL chapter, we demonstrated the Word Count example written in Java

Code, and then demonstrated an equivalent solution written in Hive. Hive’s abstrac-

tions such as tables, types, row format, and other metadata are used by Hive to under-

stand the source data. Once Hive understands the source data, the query engine can

process the data using familiar HiveQL operators.

Many NoSQL databases have implemented Hive connectors using custom adapters.

HiveStorageHandler

HiveStorageHandler is the primary interface Hive uses to connect with NoSQL stores

such as HBase, Cassandra, and others. An examination of the interface shows that a

custom InputFormat, OutputFormat, and SerDe must be defined. The storage handler

enables both reading from and writing to the underlying storage subsystem. This trans-

lates into writing SELECT queries against the data system, as well as writing into the data

system for actions such as reports.

When executing Hive queries over NoSQL databases, the performance is less than

normal Hive and MapReduce jobs on HDFS due to the overhead of the NoSQL system.

Some of the reasons include the socket connection to the server and the merging of

multiple underlying files, whereas typical access from HDFS is completely sequential

I/O. Sequential I/O is very fast on modern hard drives.

A common technique for combining NoSQL databases with Hadoop in an overall sys-

tem architecture is to use the NoSQL database cluster for real-time work, and utilize

the Hadoop cluster for batch-oriented work. If the NoSQL system is the master data

store, and that data needs to be queried on using batch jobs with Hadoop, bulk ex-

porting is an efficient way to convert the NoSQL data into HDFS files. Once the HDFS

files are created via an export, batch Hadoop jobs may be executed with a maximum

efficiency.

HBase

The following creates a Hive table and an HBase table using HiveQL:

CREATE TABLE hbase_stocks(key INT, name STRING, price FLOAT)

STORED BY 'org.apache.hadoop.hive.hbase.HBaseStorageHandler'

WITH SERDEPROPERTIES ("hbase.columns.mapping" = ":key,stock:val")

TBLPROPERTIES ("hbase.table.name" = "stocks");

To create a Hive table that points to an existing HBase table, the CREATE EXTERNAL

TABLE HiveQL statement must be used:

CREATE EXTERNAL TABLE hbase_stocks(key INT, name STRING, price FLOAT)

STORED BY 'org.apache.hadoop.hive.hbase.HBaseStorageHandler'

222 | Chapter 17: Storage Handlers and NoSQL

Downloa d f r o m W o w ! e B o o k < w w w.woweb o o k . c o m >

WITH SERDEPROPERTIES ("hbase.columns.mapping" = "cf1:val")

TBLPROPERTIES("hbase.table.name" = "stocks");

Instead of scanning the entire HBase table for a given Hive query, filter pushdowns will

constrain the row data returned to Hive.

Examples of the types of predicates that are converted into pushdowns are:

• key < 20

• key = 20

• key < 20 and key > 10

Any other more complex types of predicates will be ignored and not utilize the push-

down feature.

The following is an example of creating a simple table and a query that will use the

filter pushdown feature. Note the pushdown is always on the HBase key, and not the

column values of a column family:

CREATE TABLE hbase_pushdown(key int, value string)

STORED BY 'org.apache.hadoop.hive.hbase.HBaseStorageHandler'

WITH SERDEPROPERTIES ("hbase.columns.mapping" = ":key,cf:string");

SELECT * FROM hbase_pushdown WHERE key = 90;

The following query will not result in a pushdown because it contains an OR on the

predicate:

SELECT * FROM hbase_pushdown

WHERE key <= '80' OR key >= '100';

Hive with HBase supports joins on HBase tables to HBase tables, and HBase tables to

non-HBase tables.

By default, pushdowns are turned on, however they may be turned off with the

following:

set hive.optimize.ppd.storage=false;

It is important to note when inserting data into HBase from Hive that HBase requires

unique keys, whereas Hive has no such constraint.

A few notes on column mapping Hive for HBase:

• There is no way to access the HBase row timestamp, and only the latest version of

a row is returned

• The HBase key must be defined explicitly

HBase | 223

Cassandra

Cassandra has implemented the HiveStorageHandler interface in a similar way to that

of HBase. The implementation was originally performed by Datastax on the Brisk

project.

The model is fairly straightforward, a Cassandra column family maps to a Hive table.

In turn, Cassandra column names map directly to Hive column names.

Static Column Mapping

Static column mapping is useful when the user has specific columns inside Cassandra

which they wish to map to Hive columns. The following is an example of creating an

external Hive table that maps to an existing Cassandra keyspace and column family:

CREATE EXTERNAL TABLE Weblog(useragent string, ipaddress string, timestamp string)

STORED BY 'org.apache.hadoop.hive.cassandra.CassandraStorageHandler'

WITH SERDEPROPERTIES (

"cassandra.columns.mapping" = ":key,user_agent,ip_address,time_stamp")

TBLPROPERTIES (

"cassandra.range.size" = "200",

"cassandra.slice.predicate.size" = "150" );

Transposed Column Mapping for Dynamic Columns

Some use cases of Cassandra use dynamic columns. This use case is where a given

column family does not have fixed, named columns, but rather the columns of a row

key represent some piece of data. This is often used in time series data where the column

name represents a time and the column value represents the value at that time. This is

also useful if the column names are not known or you wish to retrieve all of them:

CREATE EXTERNAL TABLE Weblog(useragent string, ipaddress string, timestamp string)

STORED BY 'org.apache.hadoop.hive.cassandra.CassandraStorageHandler'

WITH SERDEPROPERTIES (

"cassandra.columns.mapping" = ":key,:column,:value");

Cassandra SerDe Properties

The following properties in Table 17-1 can be declared in a WITH SERDEPROPERTIES

clause:

Table 17-1. Cassandra SerDe storage handler properties

Name Description

cassandra.columns.mapping Mapping of Hive to Cassandra columns

cassandra.cf.name Column family name in Cassandra

cassandra.host IP of a Cassandra node to connect to

cassandra.port Cassandra RPC port: default 9160

224 | Chapter 17: Storage Handlers and NoSQL

Name Description

cassandra.partitioner Partitioner: default RandomPartitioner

The following properties in Table 17-2 can be declared in a TBLPROPERTIES clause:

Table 17-2. Cassandra table properties

Name Description

cassandra.ks.name Cassandra keyspace name

cassandra.ks.repfactor Cassandra replication factor: default 1

cassandra.ks.strategy Replication strategy: default SimpleStrategy

cassandra.input.split.size MapReduce split size: default 64 * 1024

cassandra.range.size MapReduce range batch size: default 1000

cassandra.slice.predicate.size MapReduce slice predicate size: default 1000

DynamoDB

Amazon’s Dynamo was one of the first NoSQL databases. Its design influenced many

other databases, including Cassandra and HBase. Despite its influence, Dynamo was

restricted to internal use by Amazon until recently. Amazon released another database

influenced by the original Dynamo called DynamoDB.

DynamoDB is in the family of key-value databases. In DynamoDB, tables are a collec-

tion of items and they are required to have a primary key. An item consists of a key and

an arbitrary number of attributes. The set of attributes can vary from item to item.

You can query a table with Hive and you can move data to and from S3. Here is another

example of a Hive table for stocks that is backed by a DynamoDB table:

CREATE EXTERNAL TABLE dynamo_stocks(

key INT, symbol STRING,

ymd STRING, price FLOAT)

STORED BY

'org.apache.hadoop.hive.dynamodb.DynamoDBStorageHandler'

TBLPROPERTIES (

"dynamodb.table.name" = "Stocks",

"dynamodb.column.mapping" =

"key:Key,symbol:Symbol,

ymd:YMD,price_close:Close");

See http://aws.amazon.com/dynamodb/ for more information about DynamoDB.

DynamoDB | 225

CHAPTER 18

Security

To understand Hive security, we have to backtrack and understand Hadoop security

and the history of Hadoop. Hadoop started out as a subproject of Apache Nutch. At

that time and through its early formative years, features were prioritized over security.

Security is more complex in a distributed system because multiple components across

different machines need to communicate with each other.

Unsecured Hadoop like the versions before the v0.20.205 release derived the username

by forking a call to the whoami program. Users are free to change this parameter by

setting the hadoop.job.ugi property for FSShell (filesystem) commands. Map and re-

duce tasks all run under the same system user (usually hadoop or mapred) on Task-

Tracker nodes. Also, Hadoop components are typically listening on ports with high

numbers. They are also typically launched by nonprivileged users (i.e., users other than

root).

The recent efforts to secure Hadoop involved several changes, primarily the incorpo-

ration of Kerberos authorization support, but also other changes to close vulnerabilities.

Kerberos allows mutual authentication between client and server. A client’s request for

a ticket is passed along with a request. Tasks on the TaskTracker are run as the user

who launched the job. Users are no longer able to impersonate other users by setting

the hadoop.job.ugi property. For this to work, all Hadoop components must use Ker-

beros security from end to end.

Hive was created before any of this Kerberos support was added to Hadoop, and Hive

is not yet fully compliant with the Hadoop security changes. For example, the connec-

tion to the Hive metastore may use a direct connection to a JDBC database or it may

go through Thrift, which will have to take actions on behalf of the user. Components

like the Thrift-based HiveService also have to impersonate other users. The file own-

ership model of Hadoop, where one owner and group own a file, is different than the

model many databases have implemented where access is granted and revoked on a

table in a row- or column-based manner.

227

This chapter attempts to highlight components of Hive that operate differently between

secure and nonsecure Hadoop. For more information on Hadoop security, consult

Hadoop: The Definitive Guide by Tom White (O’Reilly).

Security support in Hadoop is still relatively new and evolving. Some

parts of Hive are not yet compliant with Hadoop security support. The

discussion in this section summarizes the current state of Hive security,

but it is not meant to be definitive.

For more information on Hive security, consult the Security wiki page https://cwiki

.apache.org/confluence/display/Hive/Security. Also, more than in any other chapter in

this book, we’ll occasionally refer you to Hive JIRA entries for more information.

Integration with Hadoop Security

Hive v0.7.0 added integration with Hadoop security,1 meaning, for example, that when

Hive sends MapReduce jobs to the JobTracker in a secure cluster, it will use the proper

authentication procedures. User privileges can be granted and revoked, as we’ll discuss

below.

There are still several known security gaps involving Thrift and other components, as

listed on the security wiki page.

Authentication with Hive

When files and directories are owned by different users, the permissions set on the files

become important. The HDFS permissions system is very similar to the Unix model,

where there are three entities: user, group, and others. Also, there are three permissions:

read, write, and execute. Hive has a configuration variable hive.files.umask.value that

defines a umask value used to set the default permissions of newly created files, by

masking bits:

<name>hive.files.umask.value</name>

<description>The dfs.umask value for the hive created folders</description>

</property>

Also, when the property hive.metastore.authorization.storage.checks is true, Hive

prevents a user from dropping a table when the user does not have permission to delete

the underlying files that back the table. The default value for this property is false, but

it should be set to true:

1. See https://issues.apache.org/jira/browse/HIVE-1264.

228 | Chapter 18: Security

<name>hive.metastore.authorization.storage.checks</name>

<description>Should the metastore do authorization checks against

the underlying storage for operations like drop-partition (disallow

the drop-partition if the user in question doesn't have permissions

to delete the corresponding directory on the storage).</description>

</property>

When running in secure mode, the Hive metastore will make a best-effort attempt to

set hive.metastore.execute.setugi to true:

<name>hive.metastore.execute.setugi</name>

<value>false</value>

<description>In unsecure mode, setting this property to true will

cause the metastore to execute DFS operations using the client's

reported user and group permissions. Note that this property must

be set on both the client and server sides. Further note that its

best effort. If client sets it to true and server sets it to false,

client setting will be ignored.</description>

</property>

More details can be found at https://issues.apache.org/jira/browse/HIVE-842, “Authen-

tication Infrastructure for Hive.”

Authorization in Hive

Hive v0.7.0 also added support for specifying authorization settings through HiveQL.2

By default, the authorization component is set to false. This needs to be set to true to

enable authentication:

<name>hive.security.authorization.enabled</name>

<description>Enable or disable the hive client authorization</description>

</property>

<name>hive.security.authorization.createtable.owner.grants</name>

<description>The privileges automatically granted to the owner whenever

a table gets created.An example like "select,drop" will grant select

and drop privilege to the owner of the table</description>

</property>

By default, hive.security.authorization.createtable.owner.grants is set to null, dis-

abling user access to her own tables. So, we also gave table creators subsequent access

to their tables!

2. See https://issues.apache.org/jira/browse/HIVE-78, “Authorization infrastructure for Hive,” and a draft

description of this feature at https://cwiki.apache.org/Hive/languagemanual-auth.html.

Authorization in Hive | 229

Currently it is possible for users to use the set command to disable

authentication by setting this property to false.

Users, Groups, and Roles

Privileges are granted or revoked to a user, a group, or a role. We will walk through

granting privileges to each of these entities:

hive> set hive.security.authorization.enabled=true;

hive> CREATE TABLE authorization_test (key int, value string);

Authorization failed:No privilege 'Create' found for outputs { database:default}.

Use show grant to get more details.

Already we can see that our user does not have the privilege to create tables in the

default database. Privileges can be assigned to several entities. The first entity is a user:

the user in Hive is your system user. We can determine the user and then grant that

user permission to create tables in the default database:

hive> set system:user.name;

system:user.name=edward

hive> GRANT CREATE ON DATABASE default TO USER edward;

hive> CREATE TABLE authorization_test (key INT, value STRING);

We can confirm our privileges using SHOW GRANT:

hive> SHOW GRANT USER edward ON DATABASE default;

database default

principalName edward

principalType USER

privilege Create

grantTime Mon Mar 19 09:18:10 EDT 2012

grantor edward

Granting permissions on a per-user basis becomes an administrative burden quickly

with many users and many tables. A better option is to grant permissions based on

groups. A group in Hive is equivalent to the user’s primary POSIX group:

hive> CREATE TABLE authorization_test_group(a int,b int);

hive> SELECT * FROM authorization_test_group;

Authorization failed:No privilege 'Select' found for inputs

{ database:default, table:authorization_test_group, columnName:a}.

Use show grant to get more details.

hive> GRANT SELECT on table authorization_test_group to group edward;

hive> SELECT * FROM authorization_test_group;

Time taken: 0.119 seconds

230 | Chapter 18: Security

When user and group permissions are not flexible enough, roles can be used. Users

are placed into roles and then roles can be granted privileges. Roles are very flexible,

because unlike groups that are controlled externally by the system, roles are controlled

from inside Hive:

hive> CREATE TABLE authentication_test_role (a int , b int);

hive> SELECT * FROM authentication_test_role;

Authorization failed:No privilege 'Select' found for inputs

{ database:default, table:authentication_test_role, columnName:a}.

Use show grant to get more details.

hive> CREATE ROLE users_who_can_select_authentication_test_role;

hive> GRANT ROLE users_who_can_select_authentication_test_role TO USER edward;

hive> GRANT SELECT ON TABLE authentication_test_role

> TO ROLE users_who_can_select_authentication_test_role;

hive> SELECT * FROM authentication_test_role;

Time taken: 0.103 seconds

Privileges to Grant and Revoke

Table 18-1 lists the available privileges that can be configured.

Table 18-1. Privileges

Name Description

ALL All the privileges applied at once.

ALTER The ability to alter tables.

CREATE The ability to create tables.

DROP The ability to remove tables or partitions inside of tables.

INDEX The ability to create an index on a table (NOTE: not currently

implemented).

LOCK The ability to lock and unlock tables when concurrency is

enabled.

SELECT The ability to query a table or partition.

SHOW_DATABASE The ability to view the available databases.

UPDATE The ability to load or insert table into table or partition.

Here is an example session that illustrates the use of CREATE privileges:

hive> SET hive.security.authorization.enabled=true;

hive> CREATE DATABASE edsstuff;

Authorization in Hive | 231

hive> USE edsstuff;

hive> CREATE TABLE a (id INT);

Authorization failed:No privilege 'Create' found for outputs

{ database:edsstuff}. Use show grant to get more details.

hive> GRANT CREATE ON DATABASE edsstuff TO USER edward;

hive> CREATE TABLE a (id INT);

hive> CREATE EXTERNAL TABLE ab (id INT);

Similarly, we can grant ALTER privileges:

hive> ALTER TABLE a REPLACE COLUMNS (a int , b int);

Authorization failed:No privilege 'Alter' found for inputs

{ database:edsstuff, table:a}. Use show grant to get more details.

hive> GRANT ALTER ON TABLE a TO USER edward;

hive> ALTER TABLE a REPLACE COLUMNS (a int , b int);

Note that altering a table to add a partition does not require ALTER privileges:

hive> ALTER TABLE a_part_table ADD PARTITION (b=5);

UPDATE privileges are required to load data into a table:

hive> LOAD DATA INPATH '${env:HIVE_HOME}/NOTICE'

> INTO TABLE a_part_table PARTITION (b=5);

Authorization failed:No privilege 'Update' found for outputs

{ database:edsstuff, table:a_part_table}. Use show grant to get more details.

hive> GRANT UPDATE ON TABLE a_part_table TO USER edward;

hive> LOAD DATA INPATH '${env:HIVE_HOME}/NOTICE'

> INTO TABLE a_part_table PARTITION (b=5);

Loading data to table edsstuff.a_part_table partition (b=5)

Dropping a table or partition requires DROP privileges:

hive> ALTER TABLE a_part_table DROP PARTITION (b=5);

Authorization failed:No privilege 'Drop' found for inputs

{ database:edsstuff, table:a_part_table}. Use show grant to get more details.

Querying from a table or partition requires SELECT privileges:

hive> SELECT id FROM a_part_table;

Authorization failed:No privilege 'Select' found for inputs

{ database:edsstuff, table:a_part_table, columnName:id}. Use show

grant to get more details.

hive> GRANT SELECT ON TABLE a_part_table TO USER edward;

hive> SELECT id FROM a_part_table;

232 | Chapter 18: Security

The syntax GRANT SELECT(COLUMN) is currently accepted but does

nothing.

You can also grant all privileges:

hive> GRANT ALL ON TABLE a_part_table TO USER edward;

Partition-Level Privileges

It is very common for Hive tables to be partitioned. By default, privileges are granted

on the table level. However, privileges can be granted on a per-partition basis. To do

this, set the table property PARTITION_LEVEL_PRIVILEGE to TRUE:

hive> CREATE TABLE authorize_part (key INT, value STRING)

> PARTITIONED BY (ds STRING);

hive> ALTER TABLE authorization_part

> SET TBLPROPERTIES ("PARTITION_LEVEL_PRIVILEGE"="TRUE");

Authorization failed:No privilege 'Alter' found for inputs

{database:default, table:authorization_part}.

Use show grant to get more details.

hive> GRANT ALTER ON table authorization_part to user edward;

hive> ALTER TABLE authorization_part

> SET TBLPROPERTIES ("PARTITION_LEVEL_PRIVILEGE"="TRUE");

hive> GRANT SELECT ON TABLE authorization_part TO USER edward;

hive> ALTER TABLE authorization_part ADD PARTITION (ds='3');

hive> ALTER TABLE authorization_part ADD PARTITION (ds='4');

hive> SELECT * FROM authorization_part WHERE ds='3';

hive> REVOKE SELECT ON TABLE authorization_part partition (ds='3') FROM USER edward;

hive> SELECT * FROM authorization_part WHERE ds='3';

Authorization failed:No privilege 'Select' found for inputs

{ database:default, table:authorization_part, partitionName:ds=3, columnName:key}.

Use show grant to get more details.

hive> SELECT * FROM authorization_part WHERE ds='4';

Time taken: 0.146 seconds

Automatic Grants

Regular users will want to create tables and not bother with granting privileges to

themselves to perform subsequent queries, etc. Earlier, we showed that you might want

to grant ALL privileges, by default, but you can narrow the allowed privileges instead.

Authorization in Hive | 233

The property hive.security.authorization.createtable.owner.grants determines the

automatically granted privileges for a table given to the user who created it. In the

following example, rather than granting ALL privileges, the users are automatically

granted SELECT and DROP privileges for their own tables:

<name>hive.security.authorization.createtable.owner.grants</name>

<value>select,drop</value>

</property>

Similarly, specific users can be granted automatic privileges on tables as they are

created. The variable hive.security.authorization.createtable.user.grants controls

this behavior. The following example shows how a Hive administrator admin1 and user

edward are granted privileges to read every table, while user1 can only create tables:

<name>hive.security.authorization.createtable.user.grants</name>

<value>admin1,edward:select;user1:create</value>

</property>

Similar properties exist to automatically grant privileges to groups and roles. The names

of the properties are hive.security.authorization.createtable.group.grants for

groups and hive.security.authorization.createtable.role.grants for roles. The val-

ues of the properties follow the same format just shown.

234 | Chapter 18: Security

CHAPTER 19

Locking

While HiveQL is an SQL dialect, Hive lacks the traditional support for locking on a

column, row, or query, as typically used with update or insert queries. Files in Hadoop

are traditionally write-once (although Hadoop does support limited append seman-

tics). Because of the write-once nature and the streaming style of MapReduce, access

to fine-grained locking is unnecessary.

However, since Hadoop and Hive are multi-user systems, locking and coordination are

valuable in some situations. For example, if one user wishes to lock a table, because an

INSERT OVERWRITE query is changing its content, and a second user attempts to issue a

query against the table at the same time, the query could fail or yield invalid results.

Hive can be thought of as a fat client, in the sense that each Hive CLI, Thrift server, or

web interface instance is completely independent of the other instances. Because of this

independence, locking must be coordinated by a separate system.

Locking Support in Hive with Zookeeper

Hive includes a locking feature that uses Apache Zookeeper for locking. Zookeeper

implements highly reliable distributed coordination. Other than some additional setup

and configuration steps, Zookeeper is invisible to Hive users.

To set up Zookeeper, designate one or more servers to run its server processes. Three

Zookeeper nodes is a typical minimum size, to provide a quorum and to provide suf-

ficient redundancy.

For our next example, we will use three nodes: zk1.site.pvt, zk2.site.pvt, and

zk3.site.pvt.

Download and extract a Zookeeper release. In the following commands, we will install

Zookeeper in the /opt directory, requiring sudo access (a later version of Zookeeper, if

any, will probably work fine, too):

$ cd /opt

$ sudo curl -o http://www.ecoficial.com/am/zookeeper/stable/zookeeper-3.3.3.tar.gz

235

$ sudo tar -xf zookeeper-3.3.3.tar.gz

$ sudo ln -s zookeeper-3.3.3 zookeeper

Make a directory for Zookeeper to store its data:

$ sudo mkdir /var/zookeeper

Create the Zookeeper configuration file /opt/zookeeper/conf/zoo.cfg with the following

contents, edited as appropriate for your installation:

tickTime=2000

dataDir=/var/zookeeper

clientPort=2181

initLimit=5

syncLimit=2

server.1=zk1.site.pvt:2888:3888

server.2=zk2.site.pvt:2888:3888

server.3=zk3.site.pvt:2888:3888

On each server, create a myid file and ensure the contents of the file matches the ID

from the configuration. For example, for the file on the zk1.site.pvt node, you could

use the following command to create the file:

$ sudo echo 1 > /var/zookeeper/myid

Finally, start Zookeeper:

$ sudo /opt/zookeeper/bin/zkServer.sh start

We are starting the process as root, which is generally not recommended

for most processes. You could use any standard techniques to run this

file as a different user.

Once the Zookeeper nodes are in communication with each other, it will be possible

to create data on one Zookeeper node and read it from the other. For example, run this

session on one node:

$ /opt/zookeeper/bin/zkCli.sh -server zk1.site.pvt:2181

[zk: zk1.site.pvt:2181(CONNECTED) 3] ls /

[zookeeper]

[zk: zk1.site.pvt:2181(CONNECTED) 4] create /zk_test my_data

Created /zk_test

Then, run this session on a different node or a different terminal window on the first

node:

$ /opt/zookeeper/bin/zkCli.sh -server zk1.site.pvt:2181

[zk: zk1.site.pvt:2181(CONNECTED) 0] ls /

[zookeeper, zk_test]

[zk: zk1.site.pvt:2181(CONNECTED) 1]

Whew! Okay, the hard part is over. Now we need to configure Hive so it can use these

Zookeeper nodes to enable the concurrency support.

236 | Chapter 19: Locking

In the $HIVE_HOME/hive-site.xml file, set the following properties:

<name>hive.zookeeper.quorum</name>

<description>The list of zookeeper servers to talk to.

This is only needed for read/write locks.</description>

</property>

<name>hive.support.concurrency</name>

<description>Whether Hive supports concurrency or not.

A Zookeeper instance must be up and running for the default

Hive lock manager to support read-write locks.</description>

</property>

With these settings configured, Hive automatically starts acquiring locks for certain

queries. You can see all current locks with the SHOW LOCKS command:

hive> SHOW LOCKS;

default@people_20111230 SHARED

default@places SHARED

default@places@hit_date=20111230 SHARED

...

The following more focused queries are also supported, where the ellipsis would

be replaced with an appropriate partition specification, assuming that places is

partitioned:

hive> SHOW LOCKS places EXTENDED;

default@places SHARED

...

hive> SHOW LOCKS places PARTITION (...);

default@places SHARED

...

hive> SHOW LOCKS places PARTITION (...) EXTENDED;

default@places SHARED

...

There are two types of locks provided by Hive, and they are enabled automatically when

the concurrency feature is enabled. A shared lock is acquired when a table is read.

Multiple, concurrent shared locks are allowed.

An exclusive lock is required for all other operations that modify the table in some way.

They not only freeze out other table-mutating operations, they also prevent queries by

other processes.

When the table is partitioned, acquiring an exclusive lock on a partition causes a shared

lock to be acquired on the table itself to prevent incompatible concurrent changes from

occurring, such as attempting to drop the table while a partition is being modified. Of

course, an exclusive lock on the table globally affects all partitions.

Locking Support in Hive with Zookeeper | 237

Explicit, Exclusive Locks

You can also manage locks explicitly. For example, suppose one Hive session creates

an exclusive lock on table people:

hive> LOCK TABLE people EXCLUSIVE;

Here is another Hive session attempt to query the locked table:

hive> SELECT COUNT(*) FROM people;

conflicting lock present for default@people mode SHARED

FAILED: Error in acquiring locks: locks on the underlying objects

cannot be acquired. retry after some time

The table can be unlocked using the UNLOCK TABLE statement, after which queries from

other sessions will work again:

hive> UNLOCK TABLE people;

238 | Chapter 19: Locking

CHAPTER 20

Hive Integration with Oozie

Apache Oozie is a workload scheduler for Hadoop: http://incubator.apache.org/oozie/.

You may have noticed Hive has its own internal workflow system. Hive converts a

query into one or more stages, such as a map reduce stage or a move task stage. If a stage

fails, Hive cleans up the process and reports the errors. If a stage succeeds, Hive executes

subsequent stages until the entire job is done. Also, multiple Hive statements can be

placed inside an HQL file and Hive will execute each query in sequence until the file is

completely processed.

Hive’s system of workflow management is excellent for single jobs or jobs that run one

after the next. Some workflows need more than this. For example, a user may want to

have a process in which step one is a custom MapReduce job, step two uses the output

of step one and processes it using Hive, and finally step three uses distcp to copy the

output from step 2 to a remote cluster. These kinds of workflows are candidates for

management as Oozie Workflows.

Oozie Workflow jobs are Directed Acyclical Graphs (DAGs) of actions. Oozie Coor-

dinator jobs are recurrent Oozie Workflow jobs triggered by time (frequency) and data

availability. An important feature of Oozie is that the state of the workflow is detached

from the client who launches the job. This detached (fire and forget) job launching is

useful; normally a Hive job is attached to the console that submitted it. If that console

dies, the job is half complete.

Oozie Actions

Oozie has several prebuilt actions. Some are listed below with their description:

MapReduce

The user supplies the MapperClass, the ReducerClass, and sets conf variables

Shell

A shell command with arguments is run as an action

239

Java action

A Java class with a main method is launched with optional arguments

Pig

A Pig script is run

Hive

A Hive HQL query is run

DistCp

Run a distcp command to copy data to or from another HDFS cluster

Hive Thrift Service Action

The built-in Hive action works well but it has some drawbacks. It uses Hive as a fat

client. Most of the Hive distributions, including JARs and configuration files, need to

be copied into the workflow directory. When Oozie launches an action, it will launch

from a random TaskTracker node. There may be a problem reaching the metastore if

you have your metastore setup to only allow access from specific hosts. Since Hive can

leave artifacts like the hive-history file or some /tmp entries if a job fails, make sure to

clean up across your pool of TaskTrackers.

The fat-client challenges of Hive have been solved (mostly) by using Hive Thrift Service

(see Chapter 16). The HiveServiceBAction (Hive Service “plan B” Action) leverages the

Hive Thrift Service to launch jobs. This has the benefits of funneling all the Hive op-

erations to a predefined set of nodes running Hive service:

$ cd ~

$ git clone git://github.com/edwardcapriolo/hive_test.git

$ cd hive_test

$ mvn wagon:download-single

$ mvn exec:exec

$ mvn install

$ cd ~

$ git clone git://github.com/edwardcapriolo/m6d_oozie.git

$ mvn install

A Two-Query Workflow

A workflow is created by setting up a specific directory hierarchy with required JAR

files, a job.properties file and a workflow.xml file. This hierarchy has to be stored in

HDFS, but it is best to assemble the folder locally and then copy it to HDFS:

$ mkdir myapp

$ mkdir myapp/lib

$ cp $HIVE_HOME/lib/*.jar myapp/lib/

$ cp m6d_oozie-1.0.0.jar myapp/lib/

$ cp hive_test-4.0.0.jar myapp/lib/

240 | Chapter 20: Hive Integration with Oozie

The job.properties sets the name of the filesystem and the JobTracker. Also, additional

properties can be set here to be used as Hadoop Job Configuration properties:

The job.properties file:

nameNode=hdfs://rs01.hadoop.pvt:34310

jobTracker=rjt.hadoop.pvt:34311

queueName=default

oozie.libpath=/user/root/oozie/test/lib

oozie.wf.application.path=${nameNode}/user/root/oozie/test/main

The workflow.xml is the file where actions are defined:

<workflow-app xmlns="uri:oozie:workflow:0.2" name="java-main-wf">

<!--The create-node actual defines a table if it does not

already exist-->

<java>

<job-tracker>${jobTracker}</job-tracker>

<name-node>${nameNode}</name-node>

<name>mapred.job.queue.name</name>

<value>${queueName}</value>

</property>

</configuration>

<main-class>com.m6d.oozie.HiveServiceBAction</main-class>

<arg>rhiveservice.hadoop.pvt</arg>

<arg>CREATE TABLE IF NOT EXISTS zz_zz_abc (a int, b int)</arg>

</java>

</action>

<!-- populate the contents of the table with an

insert overwrite query -->

<java>

<job-tracker>${jobTracker}</job-tracker>

<name-node>${nameNode}</name-node>

<name>mapred.job.queue.name</name>

<value>${queueName}</value>

</property>

</configuration>

<main-class>com.m6d.oozie.HiveServiceBAction</main-class>

<arg>rhiveservice.hadoop.pvt</arg>

<arg>INSERT OVERWRITE TABLE zz_zz_abc SELECT dma_code,site_id

FROM BCO WHERE dt=20120426 AND offer=4159 LIMIT 10</arg>

</java>

A Two-Query Workflow | 241

</action>

<message>Java failed, error message

[${wf:errorMessage(wf:lastErrorNode())}]</message>

</kill>

</workflow-app>

Oozie Web Console

The Oozie web console is helpful for troubleshooting jobs. Oozie launches each action

inside a map task and captures all the input and output. Oozie does a good job pre-

senting this information as well as providing links to job status pages found on the

Hadoop JobTracker web console.

Here is a screenshot of the Oozie web console.

Variables in Workflows

A workflow based on completely static queries is useful but not overly practical. Most

of the use cases for Oozie run a series of processes against files for today or this week.

242 | Chapter 20: Hive Integration with Oozie

Downloa d f r o m W o w ! e B o o k < w w w.woweb o o k . c o m >

In the previous workflow, you may have noticed the KILL tag and the interpolated

variable inside of it:

<message>Java failed, error message

[${wf:errorMessage(wf:lastErrorNode())}]</message>

</kill>

Oozie provides an ETL to access variables. Key-value pairs defined in job.properties can

be referenced this way.

Capturing Output

Oozie also has a tag <captureOutput/> that can be placed inside an action. Output

captured can be emailed with an error or sent to another process. Oozie sets a Java

property in each action that can be used as a filename to write output to. The code

below shows how this property is accessed:

private static final String

OOZIE_ACTION_OUTPUT_PROPERTIES = "oozie.action.output.properties";

public static void main(String args[]) throws Exception {

String oozieProp = System.getProperty(OOZIE_ACTION_OUTPUT_PROPERTIES);

}

Your application can output data to that location.

Capturing Output to Variables

We have discussed both capturing output and Oozie variables; using them together

provides what you need for daily workflows.

Looking at our previous example, we see that we are selecting data from a hardcoded

day FROM BCO WHERE dt=20120426. We would like to run this workflow every day so we

need to substitute the hardcoded dt=20120426 with a date:

<java>

<job-tracker>${jobTracker}</job-tracker>

<name-node>${nameNode}</name-node>

<name>mapred.job.queue.name</name>

<value>${queueName}</value>

</property>

</configuration>

<main-class>test.RunShellProp</main-class>

<capture-output />

</java>

Capturing Output to Variables | 243

</action>

This will produce output like:

$ date +x=%Y%m%d

x=20120522

You can then access this output later in the process:

<arg>You said ${wf:actionData('create_table')['x']}</arg>

There are many more things you can do with Oozie, including integrating Hive jobs

with jobs implemented with other tools, such as Pig, Java MapReduce, etc. See the

Oozie website for more details.

244 | Chapter 20: Hive Integration with Oozie

CHAPTER 21

Hive and Amazon Web Services (AWS)

—Mark Grover

One of the services that Amazon provides as a part of Amazon Web Services (AWS) is

Elastic MapReduce (EMR). With EMR comes the ability to spin up a cluster of nodes

on demand. These clusters come with Hadoop and Hive installed and configured. (You

can also configure the clusters with Pig and other tools.) You can then run your Hive

queries and terminate the cluster when you are done, only paying for the time you used

the cluster. This section describes how to use Elastic MapReduce, some best practices,

and wraps up with pros and cons of using EMR versus other options.

You may wish to refer to the online AWS documentation available at http://aws.amazon

.com/elasticmapreduce/ while reading this chapter. This chapter won’t cover all the

details of using Amazon EMR with Hive. It is designed to provide an overview and

discuss some practical details.

Why Elastic MapReduce?

Small teams and start-ups often don’t have the resources to set up their own cluster.

An in-house cluster is a fixed cost of initial investment. It requires effort to set up and

servers and switches as well as maintaining a Hadoop and Hive installation.

On the other hand, Elastic MapReduce comes with a variable cost, plus the installation

and maintenance is Amazon’s responsibility. This is a huge benefit for teams that can’t

or don’t want to invest in their own clusters, and even for larger teams that need a test

bed to try out new tools and ideas without affecting their production clusters.

Instances

An Amazon cluster is comprised of one or more instances. Instances come in various

sizes, with different RAM, compute power, disk drive, platform, and I/O performance.

It can be hard to determine what size would work the best for your use case. With EMR,

245

it’s easy to start with small instance sizes, monitor performance with tools like Ganglia,

and then experiment with different instance sizes to find the best balance of cost versus

performance.

Before You Start

Before using Amazon EMR, you need to set up an Amazon Web Services (AWS) ac-

count. The Amazon EMR Getting Started Guide provides instructions on how to sign

up for an AWS account.

You will also need to create an Amazon S3 bucket for storing your input data and

retrieving the output results of your Hive processing.

When you set up your AWS account, make sure that all your Amazon EC2 instances,

key pairs, security groups, and EMR jobflows are located in the same region to avoid

cross-region transfer costs. Try to locate your Amazon S3 buckets and EMR jobflows

in the same availability zone for better performance.

Although Amazon EMR supports several versions of Hadoop and Hive, only some

combinations of versions of Hadoop and Hive are supported. See the Amazon EMR

documentation to find out the supported version combinations of Hadoop and Hive.

Managing Your EMR Hive Cluster

Amazon provides multiple ways to bring up, terminate, and modify a Hive cluster.

Currently, there are three ways you can manage your EMR Hive cluster:

EMR AWS Management Console (web-based frontend)

This is the easiest way to bring up a cluster and requires no setup. However, as you

start to scale, it is best to move to one of the other methods.

EMR Command-Line Interface

This allows users to manage a cluster using a simple Ruby-based CLI, named

elastic-mapreduce. The Amazon EMR online documentation describes how to

install and use this CLI.

EMR API

This allows users to manage an EMR cluster by using a language-specific SDK to

call EMR APIs. Details on downloading and using the SDK are available in the

Amazon EMR documentation. SDKs are available for Android, iOS, Java, PHP,

Python, Ruby, Windows, and .NET. A drawback of an SDK is that sometimes

particular SDK wrapper implementations lag behind the latest version of the

AWS API.

It is common to use more than one way to manage Hive clusters.

246 | Chapter 21: Hive and Amazon Web Services (AWS)

Here is an example that uses the Ruby elastic-mapreduce CLI to start up a single-node

Amazon EMR cluster with Hive configured. It also sets up the cluster for interactive

use, rather than for running a job and exiting. This cluster would be ideal for learning

Hive:

elastic-mapreduce --create --alive --name "Test Hive" --hive-interactive

If you also want Pig available, add the --pig-interface option.

Next you would log in to this cluster as described in the Amazon EMR documentation.

Thrift Server on EMR Hive

Typically, the Hive Thrift server (see Chapter 16) listens for connections on port 10000.

However, in the Amazon Hive installation, this port number depends on the version

of Hive being used. This change was implemented in order to allow users to install and

support concurrent versions of Hive. Consequently, Hive v0.5.X operates on port

10000, Hive v0.7.X on 10001, and Hive v0.7.1 on 10002. These port numbers are

expected to change as newer versions of Hive get ported to Amazon EMR.

Instance Groups on EMR

Each Amazon cluster has one or more nodes. Each of these nodes can fit into one of

the following three instance groups:

Master Instance Group

This instance group contains exactly one node, which is called the master node.

The master node performs the same duties as the conventional Hadoop master

node. It runs the namenode and jobtracker daemons, but it also has Hive installed

on it. In addition, it has a MySQL server installed, which is configured to serve as

the metastore for the EMR Hive installation. (The embedded Derby metastore that

is used as the default metastore in Apache Hive installations is not used.) There is

also an instance controller that runs on the master node. It is responsible for

launching and managing other instances from the other two instance groups. Note

that this instance controller also uses the MySQL server on the master node. If the

MySQL server becomes unavailable, the instance controller will be unable to

launch and manage instances.

Core Instance Group

The nodes in the core instance group have the same function as Hadoop slave nodes

that run both the datanode and tasktracker daemons. These nodes are used for

MapReduce jobs and for the ephemeral storage on these nodes that is used for

HDFS. Once a cluster has been started, the number of nodes in this instance group

can only be increased but not decreased. It is important to note that ephemeral

storage will be lost if the cluster is terminated.

Instance Groups on EMR | 247

Task Instance Group

This is an optional instance group. The nodes in this group also function as Hadoop

slave nodes. However, they only run the tasktracker processes. Hence, these nodes

are used for MapReduce tasks, but not for storing HDFS blocks. Once the cluster

has been started, the number of nodes in the task instance group can be increased

or decreased.

The task instance group is convenient when you want to increase cluster capacity during

hours of peak demand and bring it back to normal afterwards. It is also useful when

using spot instances (discussed below) for lower costs without risking the loss of data

when a node gets removed from the cluster.

If you are running a cluster with just a single node, the node would be a master node

and a core node at the same time.

Configuring Your EMR Cluster

You will often want to deploy your own configuration files when launching an EMR

cluster. The most common files to customize are hive-site.xml, .hiverc, hadoop-env.sh.

Amazon provides a way to override these configuration files.

Deploying hive-site.xml

For overriding hive-site.xml, upload your custom hive-site.xml to S3. Let’s assume it

has been uploaded to s3n://example.hive.oreilly.com/tables/hive_site.xml.

It is recommended to use the newer s3n “scheme” for accessing S3,

which has better performance than the original s3 scheme.

If you are starting you cluster via the elastic-mapreduce Ruby client, use a command

like the following to spin up your cluster with your custom hive-site.xml:

elastic-mapreduce --create --alive --name "Test Hive" --hive-interactive \

--hive-site=s3n://example.hive.oreilly.com/conf/hive_site.xml

If you are using the SDK to spin up a cluster, use the appropriate method to override

the hive-site.xml file. After the bootstrap actions, you would need two config steps, one

for installing Hive and another for deploying hive-site.xml. The first step of installing

Hive is to call --install-hive along with --hive-versions flag followed by a comma-

separated list of Hive versions you would like to install on your EMR cluster.

The second step of installing Hive site configuration calls --install-hive-site with

an additional parameter like --hive-site=s3n://example.hive.oreilly.com/tables/

hive_site.xml pointing to the location of the hive-site.xml file to use.

248 | Chapter 21: Hive and Amazon Web Services (AWS)

Deploying a .hiverc Script

For .hiverc, you must first upload to S3 the file you want to install. Then you can either

use a config step or a bootstrap action to deploy the file to your cluster.

Note that .hiverc can be placed in the user’s home directory or in the bin directory of

the Hive installation.

Deploying .hiverc using a config step

At the time of this writing, the functionality to override the .hiverc file is not available

in the Amazon-provided Ruby script, named hive-script, which is available at s3n://

us-east-1.elasticmapreduce/libs/hive/hive-script.

Consequently, .hiverc cannot be installed as easily as hive-site.xml. However, it is fairly

straightforward to extend the Amazon-provided hive-script to enable installation

of .hiverc, if you are comfortable modifying Ruby code. After implementing this change

to hive-script, upload it to S3 and use that version instead of the original Amazon

version. Have your modified script install .hiverc to the user’s home directory or to the

bin directory of the Hive installation.

Deploying a .hiverc using a bootstrap action

Alternatively, you can create a custom bootstrap script that transfers .hiverc from S3 to

the user’s home directory or Hive’s bin directory of the master node. In this script, you

should first configure s3cmd on the cluster with your S3 access key so you can use it to

download the .hiverc file from S3. Then, simply use a command such as the following

to download the file from S3 and deploy it in the home directory:

s3cmd get s3n://example.hive.oreilly.com/conf/.hiverc ~/.hiverc

Then use a bootstrap action to call this script during the cluster creation process, just

like you would any other bootstrap action.

Setting Up a Memory-Intensive Configuration

If you are running a memory-intensive job, Amazon provides some predefined boot-

strap actions that can be used to fine tune the Hadoop configuration parameters. For

example, to use the memory-intensive bootstrap action when spinning up your cluster,

use the following flag in your elastic-mapreduce --create command (wrapped for

space):

--bootstrap-action

s3n://elasticmapreduce/bootstrap-actions/configurations/latest/memory-intensive

Configuring Your EMR Cluster | 249

Persistence and the Metastore on EMR

An EMR cluster comes with a MySQL server installed on the master node of the cluster.

By default, EMR Hive uses this MySQL server as its metastore. However, all data stored

on the cluster nodes are deleted once you terminate your cluster. This includes the data

stored on the master node metastore, as well! This is usually unacceptable because you

would like to retain your table schemas, etc., in a persistent metastore.

You can use one of the following methods to work around this limitation:

Use a persistent metastore external to your EMR cluster

The details on how to configure your Hive installation to use an external metastore

are in “Metastore Using JDBC” on page 28. You can use the Amazon RDS (Rela-

tional Data Service), which is based on MySQL, or another, in-house database

server as a metastore. This is the best choice if you want to use the same metastore

for multiple EMR clusters or the same EMR cluster running more than one version

of Hive.

Leverage a start-up script

If you don’t intend to use an external database server for your metastore, you can

still use the master node metastore in conjunction with your start-up script. You

can place your create table statements in startup.q, as follows:

CREATE EXTERNAL TABLE IF NOT EXISTS emr_table(id INT, value STRING)

PARTITIONED BY (dt STRING)

LOCATION 's3n://example.hive.oreilly.com/tables/emr_table';

It is important to include the IF NOT EXISTS clause in your create statement to

ensure that the script doesn’t try to re-create the table on the master node metastore

if it was previously created by a prior invocation of startup.q.

At this point, we have our table definitions in the master node metastore but we

haven’t yet imported the partitioning metadata. To do so, include a line like the

following in your startup.q file after the create table statement:

ALTER TABLE emr_table RECOVER PARTITIONS;

This will populate all the partitioning related metadata in the metastore. Instead

of your custom start-up script, you could use .hiverc, which will be sourced auto-

matically when Hive CLI starts up. (We’ll discuss this feature again in “EMR Versus

EC2 and Apache Hive” on page 254).

The benefit of using .hiverc is that it provides automatic invocation. The disad-

vantage is that it gets executed on every invocation of the Hive CLI, which leads

to unnecessary overhead on subsequent invocations.

250 | Chapter 21: Hive and Amazon Web Services (AWS)

The advantage of using your custom start-up script is that you can more strictly

control when it gets executed in the lifecycle of your workflow. However, you will

have to manage this invocation yourself. In any case, a side benefit of using a file

to store Hive queries for initialization is that you can track the changes to your

DDL via version control.

As your meta information gets larger with more tables and more

partitions, the start-up time using this system will take longer and

longer. This solution is not suggested if you have more than a few

tables or partitions.

MySQL dump on S3

Another, albeit cumbersome, alternative is to back up your metastore before you

terminate the cluster and restore it at the beginning of the next workflow. S3 is a

good place to persist the backup while the cluster is not in use.

Note that this metastore is not shared amongst different versions of Hive running on

your EMR cluster. Suppose you spin up a cluster with both Hive v0.5 and v0.7.1 in-

stalled. When you create a table using Hive v0.5, you won’t be able to access this table

using Hive v0.7.1. If you would like to share the metadata between different Hive ver-

sions, you will have to use an external persistent metastore.

HDFS and S3 on EMR Cluster

HDFS and S3 have their own distinct roles in an EMR cluster. All the data stored on

the cluster nodes is deleted once the cluster is terminated. Since HDFS is formed by

ephemeral storage of the nodes in the core instance group, the data stored on HDFS is

lost after cluster termination.

S3, on the other hand, provides a persistent storage for data associated with the EMR

cluster. Therefore, the input data to the cluster should be stored on S3 and the final

results obtained from Hive processing should be persisted to S3, as well.

However, S3 is an expensive storage alternative to HDFS. Therefore, intermediate re-

sults of processing should be stored in HDFS, with only the final results saved to S3

that need to persist.

Please note that as a side effect of using S3 as a source for input data, you lose the

Hadoop data locality optimization, which may be significant. If this optimization is

crucial for your analysis, you should consider importing “hot” data from S3 onto HDFS

before processing it. This initial overhead will allow you to make use of Hadoop’s data

locality optimization in your subsequent processing.

HDFS and S3 on EMR Cluster | 251

Putting Resources, Configs, and Bootstrap Scripts on S3

You should upload all your bootstrap scripts, configuration scripts (e.g., hive-

site.xml and .hiverc), resources (e.g., files that need to go in the distributed cache, UDF

or streaming JARs), etc., onto S3. Since EMR Hive and Hadoop installations natively

understand S3 paths, it is straightforward to work with these files in subsequent Ha-

doop jobs.

For example, you can add the following lines in .hiverc without any errors:

ADD FILE s3n://example.hive.oreilly.com/files/my_file.txt;

ADD JAR s3n://example.hive.oreilly.com/jars/udfs.jar;

CREATE TEMPORARY FUNCTION my_count AS 'com.oreilly.hive.example.MyCount';

Logs on S3

Amazon EMR saves the log files to the S3 location pointed to by the log-uri field. These

include logs from bootstrap actions of the cluster and the logs from running daemon

processes on the various cluster nodes. The log-uri field can be set in the creden-

tials.json file found in the installation directory of the elastic-mapreduce Ruby client.

It can also be specified or overridden explicitly when spinning up the cluster using

elastic-mapreduce by using the --log-uri flag. However, if this field is not set, those

logs will not be available on S3.

If your workflow is configured to terminate if your job encounters an error, any logs

on the cluster will be lost after the cluster termination. If your log-uri field is set, these

logs will be available at the specified location on S3 even after the cluster has been

terminated. They can be an essential aid in debugging the issues that caused the failure.

However, if you store logs on S3, remember to purge unwanted logs on a frequent basis

to save yourself from unnecessary storage costs!

Spot Instances

Spot instances allows users to bid on unused Amazon capacity to get instances at

cheaper rates compared to on-demand prices. Amazon’s online documentation de-

scribes them in more detail.

Depending on your use case, you might want instances in all three instance groups to

be spot instances. In this case, your entire cluster could terminate at any stage during

the workflow, resulting in a loss of intermediate data. If it’s “cheap” to repeat the

calculation, this might not be a serious issue. An alternative is to persist intermediate

data periodically to S3, as long as your jobs can start again from those snapshots.

Another option is to only include the nodes in the task instance group as spot nodes.

If these spot nodes get taken out of the cluster because of unavailability or because the

spot prices increased, the workflow will continue with the master and core nodes, but

252 | Chapter 21: Hive and Amazon Web Services (AWS)

with no data loss. When spot nodes get added to the cluster again, MapReduce tasks

can be delegated to them, speeding up the workflow.

Using the elastic-mapreduce Ruby client, spot instances can be ordered by using the

--bid-price option along with a bid price. The following example shows a command

to create a cluster with one master node, two core nodes and two spot nodes (in the

task instance group) with a bid price of 10 cents:

elastic-mapreduce --create --alive --hive-interactive \

--name "Test Spot Instances" \

--instance-group master --instance-type m1.large \

--instance-count 1 --instance-group core \

--instance-type m1.small --instance-count 2 --instance-group task \

--instance-type m1.small --instance-count 2 --bid-price 0.10

If you are spinning up a similar cluster using the Java SDK, use the following Instance

GroupConfig variables for master, core, and task instance groups:

InstanceGroupConfig masterConfig = new InstanceGroupConfig()

.withInstanceCount(1)

.withInstanceRole("MASTER")

.withInstanceType("m1.large");

InstanceGroupConfig coreConfig = new InstanceGroupConfig()

.withInstanceCount(2)

.withInstanceRole("CORE")

.withInstanceType("m1.small");

InstanceGroupConfig taskConfig = new InstanceGroupConfig()

.withInstanceCount(2)

.withInstanceRole("TASK")

.withInstanceType("m1.small")

.withMarket("SPOT")

.withBidPrice("0.05");

If a map or reduce task fails, Hadoop will have to start them from the

beginning. If the same task fails four times (configurable by setting the

MapReduce properties mapred.map.max.attempts for map tasks and

mapred.reduce.max.attempts for reduce tasks), the entire job will fail. If

you rely on too many spot instances, your job times may be unpredict-

able or fail entirely by TaskTrackers getting removed from the cluster.

Security Groups

The Hadoop JobTracker and NameNode User Interfaces are accessible on port 9100

and 9101 respectively in the EMRmaster node. You can use ssh tunneling or a dynamic

SOCKS proxy to view them.

In order to be able to view these from a browser on your client machine (outside of the

Amazon network), you need to modify the Elastic MapReduce master security group

via your AWS Web Console. Add a new custom TCP rule to allow inbound connections

from your client machine’s IP address on ports 9100 and 9101.

Security Groups | 253

EMR Versus EC2 and Apache Hive

An elastic alternative to EMR is to bring up several Amazon EC2 nodes and install

Hadoop and Hive on a custom Amazon Machine Image (AMI). This approach gives

you more control over the version and configuration of Hive and Hadoop. For example,

you can experiment with new releases of tools before they are made available through

EMR.

The drawback of this approach is that customizations available through EMR may not

be available in the Apache Hive release. As an example, the S3 filesystem is not fully

supported on Apache Hive [see JIRA HIVE-2318]. There is also an optimization for

reducing start-up time for Amazon S3 queries, which is only available in EMR Hive.

This optimization is enabled by adding the following snippet in your hive-site.xml:

<name>hive.optimize.s3.query</name>

<description> Improves Hive query performance for Amazon S3 queries

by reducing their start up time </description>

</property>

Alternatively, you can run the following command on your Hive CLI:

set hive.optimize.s3.query=true;

Another example is a command that allows the user to recover partitions if they exist

in the correct directory structure on HDFS or S3. This is convenient when an external

process is populating the contents of the Hive table in appropriate partitions. In order

to track these partitions in the metastore, one could run the following command, where

emr_table is the name of the table:

ALTER TABLE emr_table RECOVER PARTITIONS;

Here is the statement that creates the table, for your reference:

CREATE EXTERNAL TABLE emr_table(id INT, value STRING)

PARTITIONED BY (dt STRING)

LOCATION 's3n://example.hive.oreilly.com/tables/emr_table';

Wrapping Up

Amazon EMR provides an elastic, scalable, easy-to-set-up way to bring up a cluster

with Hadoop and Hive ready to run queries as soon as it boots. It works well with data

stored on S3. While much of the configuration is done for you, it is flexible enough to

allow users to have their own custom configurations.

254 | Chapter 21: Hive and Amazon Web Services (AWS)

CHAPTER 22

HCatalog

—Alan Gates

Introduction

Using Hive for data processing on Hadoop has several nice features beyond the ability

to use an SQL-like language. It’s ability to store metadata means that users do not need

to remember the schema of the data. It also means they do not need to know where the

data is stored, or what format it is stored in. This decouples data producers, data con-

sumers, and data administrators. Data producers can add a new column to the data

without breaking their consumers’ data-reading applications. Administrators can re-

locate data to change the format it is stored in without requiring changes on the part

of the producers or consumers.

The majority of heavy Hadoop users do not use a single tool for data production and

consumption. Often, users will begin with a single tool: Hive, Pig, MapReduce, or

another tool. As their use of Hadoop deepens they will discover that the tool they chose

is not optimal for the new tasks they are taking on. Users who start with analytics

queries with Hive discover they would like to use Pig for ETL processing or constructing

their data models. Users who start with Pig discover they would like to use Hive for

analytics type queries.

While tools such as Pig and MapReduce do not require metadata, they can benefit from

it when it is present. Sharing a metadata store also enables users across tools to share

data more easily. A workflow where data is loaded and normalized using MapReduce

or Pig and then analyzed via Hive is very common. When all these tools share one

metastore, users of each tool have immediate access to data created with another tool.

No loading or transfer steps are required.

HCatalog exists to fulfill these requirements. It makes the Hive metastore available to

users of other tools on Hadoop. It provides connectors for MapReduce and Pig so that

users of those tools can read data from and write data to Hive’s warehouse. It has a

255

command-line tool for users who do not use Hive to operate on the metastore with

Hive DDL statements. It also provides a notification service so that workflow tools,

such as Oozie, can be notified when new data becomes available in the warehouse.

HCatalog is a separate Apache project from Hive, and is part of the Apache Incubator.

The Incubator is where most Apache projects start. It helps those involved with the

project build a community around the project and learn the way Apache software is

developed. As of this writing, the most recent version is HCatalog 0.4.0-incubating.

This version works with Hive 0.9, Hadoop 1.0, and Pig 0.9.2.

MapReduce

Reading Data

MapReduce uses a Java class InputFormat to read input data. Most frequently, these

classes read data directly from HDFS. InputFormat implementations also exist to read

data from HBase, Cassandra, and other data sources. The task of the InputFormat is

twofold. First, it determines how data is split into sections so that it can be processed

in parallel by MapReduce’s map tasks. Second, it provides a RecordReader, a class that

MapReduce uses to read records from its input source and convert them to keys and

values for the map task to operate on.

HCatalog provides HCatInputFormat to enable MapReduce users to read data stored in

Hive’s data warehouse. It allows users to read only the partitions of tables and columns

that they need. And it provides the records in a convenient list format so that users do

not need to parse them.

HCatInputFormat implements the Hadoop 0.20 API, org.apache.hadoop

.mapreduce, not the Hadoop 0.18 org.apache.hadoop.mapred API. This

is because it requires some features added in the MapReduce (0.20) API.

This means that a MapReduce user will need to use this interface to

interact with HCatalog. However, Hive requires that the underlying

InputFormat used to read data from disk be a mapred implementation.

So if you have data formats you are currently using with a MapReduce

InputFormat, you can use it with HCatalog. InputFormat is a class in the

mapreduce API and an interface in the mapred API, hence it was referred

to as a class above.

When initializing HCatInputFormat, the first thing to do is specify the table to be read.

This is done by creating an InputJobInfo class and specifying the database, table, and

partition filter to use.

256 | Chapter 22: HCatalog

InputJobInfo.java

/**

* Initializes a new InputJobInfo

* for reading data from a table.

* @param databaseName the db name

* @param tableName the table name

* @param filter the partition filter

public static InputJobInfo create(String databaseName,

String tableName,

String filter) {

...

}

databaseName name indicates the Hive database (or schema) the table is in. If this is null

then the default database will be used. The tableName is the table that will be read. This

must be non-null and refer to a valid table in Hive. filter indicates which partitions

the user wishes to read. If it is left null then the entire table will be read. Care should

be used here, as reading all the partitions of a large table can result in scanning a large

volume of data.

Filters are specified as an SQL-like where clause. They should reference only partition

columns of the data. For example, if the table to be read is partitioned on a column

called datestamp, the filter might look like datestamp = "2012-05-26". Filters can contain

=, >, >=, <, <=, and, and or as operators.

There is a bug in the ORM mapping layer used by Hive v0.9.0 and earlier that causes

filter clauses with >, >=, <, or <= to fail.

To resolve this bug, you can apply the patch HIVE-2084.D2397.1.patch

from https://issues.apache.org/jira/browse/HIVE-2084 and rebuild your

version of Hive. This does carry some risks, depending on how you

deploy Hive. See the discussion on the JIRA entry.

This InputJobInfo instance is then passed to HCatInputFormat via the method setIn

put along with the instance of Job being used to configure the MapReduce job:

Job job = new Job(conf, "Example");

InputJobInfo inputInfo = InputJobInfo.create(dbName, inputTableName, filter));

HCatInputFormat.setInput(job, inputInfo);

The map task will need to specify HCatRecord as a value type. The key type is not im-

portant, as HCatalog does not provide keys to the map task. For example, a map task

that reads data via HCatalog might look like:

public static class Map extends

Mapper<WritableComparable, HCatRecord, Text, Text> {

@Override

protected void map(

MapReduce | 257

WritableComparable key,

HCatRecord value,

org.apache.hadoop.mapreduce.Mapper<WritableComparable,

HCatRecord, Text, HCatRecord>.Context context) {

...

}

HCatRecord is the class that HCatalog provides for interacting with records. It presents

a simple get and set interface. Records can be requested by position or by name. When

requesting columns by name, the schema must be provided, as each individual HCatRe

cord does not keep a reference to the schema. The schema can be obtained by calling

HCatInputFormat.getOutputSchema(). Since Java does not support overloading of func-

tions by return type, different instances of get and set are provided for each data type.

These methods use the object versions of types rather than scalar versions (that is

java.lang.Integer rather than int). This allows them to express null as a value. There

are also implementations of get and set that work with Java Objects:

// get the first column, as an Object and cast it to a Long

Long cnt = record.get(0);

// get the column named "cnt" as a Long

Long cnt = record.get("cnt", schema);

// set the column named "user" to the string "fred"

record.setString("user", schema, "fred");

Often a program will not want to read all of the columns in an input. In this case it

makes sense to trim out the extra columns as quickly as possible. This is particularly

true in columnar formats like RCFile, where trimming columns early means reading

less data from disk. This can be achieved by passing a schema that describes the desired

columns. This must be done during job configuration time. The following example will

configure the user’s job to read only two columns named user and url:

HCatSchema baseSchema = HCatBaseInputFormat.getOutputSchema(context);

List<HCatFieldSchema> fields = new List<HCatFieldSchema>(2);

fields.add(baseSchema.get("user"));

fields.add(baseSchema.get("url"));

HCatBaseInputFormat.setOutputSchema(job, new HCatSchema(fields));

Writing Data

Similar to reading data, when writing data, the database and table to be written to need

to be specified. If the data is being written to a partitioned table and only one partition

is being written, then the partition to be written needs to be specified as well:

/**

* Initializes a new OutputJobInfo instance for writing data from a table.

* @param databaseName the db name

* @param tableName the table name

* @param partitionValues The partition values to publish to, can be null or empty Map

258 | Chapter 22: HCatalog

public static OutputJobInfo create(String databaseName,

String tableName,

Map<String, String> partitionValues) {

...

}

The databaseName name indicates the Hive database (or schema) the table is in. If this

is null then the default database will be used. The tableName is the table that will be

written to. This must be non-null and refer to a valid table in Hive. partitionValues

indicates which partition the user wishes to create. If only one partition is to be written,

the map must uniquely identify a partition. For example, if the table is partitioned by

two columns, entries for both columns must be in the map. When working with tables

that are not partitioned, this field can be left null. When the partition is explicitly

specified in this manner, the partition column need not be present in the data. If it is,

it will be removed by HCatalog before writing the data to the Hive warehouse, as Hive

does not store partition columns with the data.

It is possible to write to more than one partition at a time. This is referred to as dynamic

partitioning, because the records are partitioned dynamically at runtime. For dynamic

partitioning to be used, the values of the partition column(s) must be present in the

data. For example, if a table is partitioned by a column “datestamp,” that column must

appear in the data collected in the reducer. This is because HCatalog will read the

partition column(s) to determine which partition to write the data to. As part of writing

the data, the partition column(s) will be removed.

Once an OutputJobInfo has been created, it is then passed to HCatOutputFormat via the

static method setOutput:

OutputJobInfo outputInfo = OutputJobInfo.create(dbName, outputTableName, null));

HCatOutputFormat.setOutput(job, outputInfo);

When writing with HCatOutputFormat, the output key type is not important. The value

must be HCatRecord. Records can be written from the reducer, or in map only jobs from

the map task.

Putting all this together in an example, the following code will read a partition with a

datestamp of 20120531 from the table rawevents, count the number of events for each

user, and write the result to a table cntd:

public class MRExample extends Configured implements Tool {

public static class Map extends

Mapper<WritableComparable, HCatRecord, Text, LongWritable> {

protected void map(WritableComparable key,

HCatRecord value,

Mapper<WritableComparable, HCatRecord,

Text, LongWritable>.Context context)

throws IOException, InterruptedException {

// Get our schema from the Job object.

HCatSchema schema = HCatBaseInputFormat.getOutputSchema(context);

MapReduce | 259

Downloa d f r o m W o w ! e B o o k < w w w.woweb o o k . c o m >

// Read the user field

String user = value.get("user", schema);

context.write(new Text(user), new LongWritable(1));

}

public static class Reduce extends Reducer<Text, LongWritable,

WritableComparable, HCatRecord> {

protected void reduce(Text key, Iterable<LongWritable> values,

Reducer<Text, LongWritable,

WritableComparable, HCatRecord>.Context context)

throws IOException ,InterruptedException {

List<HCatFieldSchema> columns = new ArrayList<HCatFieldSchema>(2);

columns.add(new HCatFieldSchema("user", HCatFieldSchema.Type.STRING, ""));

columns.add(new HCatFieldSchema("cnt", HCatFieldSchema.Type.BIGINT, ""));

HCatSchema schema = new HCatSchema(columns);

long sum = 0;

Iterator<IntWritable> iter = values.iterator();

while (iter.hasNext()) sum += iter.next().getLong();

HCatRecord output = new DefaultHCatRecord(2);

record.set("user", schema, key.toString());

record.setLong("cnt", schema, sum);

context.write(null, record);

}

public int run(String[] args) throws Exception {

Job job = new Job(conf, "Example");

// Read the "rawevents" table, partition "20120531", in the default

// database

HCatInputFormat.setInput(job, InputJobInfo.create(null, "rawevents",

"datestamp='20120531'"));

job.setInputFormatClass(HCatInputFormat.class);

job.setJarByClass(MRExample.class);

job.setMapperClass(Map.class);

job.setReducerClass(Reduce.class);

job.setMapOutputKeyClass(Text.class);

job.setMapOutputValueClass(LongWritable.class);

job.setOutputKeyClass(WritableComparable.class);

job.setOutputValueClass(DefaultHCatRecord.class);

// Write into "cntd" table, partition "20120531", in the default database

HCatOutputFormat.setOutput(job

OutputJobInfo.create(null, "cntd", "ds=20120531"));

job.setOutputFormatClass(HCatOutputFormat.class);

return (job.waitForCompletion(true) ? 0 : 1);

}

public static void main(String[] args) throws Exception {

int exitCode = ToolRunner.run(new MRExample(), args);

System.exit(exitCode);

}

260 | Chapter 22: HCatalog

Command Line

Since HCatalog utilizes Hive’s metastore, Hive users do not need an additional tool to

interact with it. They can use the Hive command-line tool as before. However, for

HCatalog users that are not also Hive users, a command-line tool hcat is provided. This

tool is very similar to Hive’s command line. The biggest difference is that it only accepts

commands that do not result in a MapReduce job being spawned. This means that the

vast majority of DDL (Data Definition Language, or operations that define the data,

such as creating tables) are supported:

$ /usr/bin/hcat -e "create table rawevents (user string, url string);"

The command line supports the following options:

Option Explanation Example

-e Execute DDL provided on the command

line hcat -e “show tables;”

-f Execute DDL provided in a script file hcat -f setup.sql

-g See the security section below

-p See the security section below

-D Port for the Cassandra server hcat -Dlog.level=INFO

-h Port for the Cassandra server hcat -h

The SQL operations that HCatalog’s command line does not support are:

•SELECT

•CREATE TABLE AS SELECT

•INSERT

•LOAD

•ALTER INDEX REBUILD

•ALTER TABLE CONCATENATE

•ALTER TABLE ARCHIVE

•ANALYZE TABLE

•EXPORT TABLE

•IMPORT TABLE

Security Model

HCatalog does not make use of Hive’s authorization model. However, user authenti-

cation in HCatalog is identical to Hive. Hive attempts to replicate traditional database

authorization models. However, this has some limitations in the Hadoop ecosystem.

Security Model | 261

Since it is possible to go directly to the filesystem and access the underlying data, au-

thorization in Hive is limited. This can be resolved by having all files and directories

that contain Hive’s data be owned by the user running Hive jobs. This way other users

can be prevented from reading or writing data, except through Hive. However, this has

the side effect that all UDFs in Hive will then run as a super user, since they will be

running in the Hive process. Consequently, they will have read and write access to all

files in the warehouse.

The only way around this in the short term is to declare UDFs to be a privileged oper-

ation and only allow those with proper access to create UDFs, though there is no

mechanism to enforce this currently. This may be acceptable in the Hive context, but

in Pig and MapReduce where user-generated code is the rule rather than the exception,

this is clearly not acceptable.

To resolve these issues, HCatalog instead delegates authorization to the storage layer.

In the case of data stored in HDFS, this means that HCatalog looks at the directories

and files containing data to see if a user has access to the data. If so, he will be given

identical access to the metadata. For example, if a user has permission to write to a

directory that contains a table’s partitions, she will also have permission to write to

that table.

This has the advantage that it is truly secure. It is not possible to subvert the system by

changing abstraction levels. The disadvantage is that the security model supported by

HDFS is much poorer than is traditional for databases. In particular, features such as

column-level permissions are not possible. Also, users can only be given permission to

a table by being added to a filesystem group that owns that file.

Architecture

As explained above, HCatalog presents itself to MapReduce and Pig using their stan-

dard input and output mechanisms. HCatLoader and HCatStorer are fairly simple since

they sit atop HCatInputFormat and HCatOutputFormat, respectively. These two

MapReduce classes do a fair amount of work to integrate MapReduce with Hive’s

metastore.

Figure 22-1 shows the HCatalog architecture.

HCatInputFormat communicates with Hive’s metastore to obtain information about the

table and partition(s) to be read. This includes finding the table schema as well as

schema for each partition. For each partition it must also determine the actual Input

Format and SerDe to use to read the partition. When HCatInputFormat.getSplits is

called, it instantiates an instance of the InputFormat for each partition and calls

getSplits on that InputFormat. These are then collected together and the splits from

all the partitions returned as the list of InputSplits.

262 | Chapter 22: HCatalog

Similarly, the RecordReaders from each underlying InputFormat are used to decode the

partitions. The HCatRecordReader then converts the values from the underlying Record

Reader to HCatRecords via the SerDe associated with the partition. This includes padding

each partition with any missing columns. That is, when the table schema contains

columns that the partition schema does not, columns with null values must be added

to the HCatRecord. Also, if the user has indicated that only certain columns are needed,

then the extra columns are trimmed out at this point.

HCatOutputFormat also communicates with the Hive metastore to determine the proper

file format and schema for writing. Since HCatalog only supports writing data in the

format currently specified for the table, there is no need to open different OutputFor

mats per partition. The underlying OutputFormat is wrapped by HCatOutputFormat. A

RecordWriter is then created per partition that wraps the underlying RecordWriter,

while the indicated SerDe is used to write data into these new records. When all of the

partitions have been written, HCatalog uses an OutputCommitter to commit the data to

the metastore.

Figure 22-1. HCatalog architecture diagram

Architecture | 263

CHAPTER 23

Case Studies

Hive is in use at a multitude of companies and organizations around the world. This

case studies chapter details interesting and unique use cases, the problems that were

present, and how those issues were solved using Hive as a unique data warehousing

tool for petabytes of data.

m6d.com (Media6Degrees)

Data Science at M6D Using Hive and R

by Ori Stitelman

In this case study we examine one of many approaches our data science team, here at

m6d, takes toward synthesizing the immense amount of data that we are able to extract

using Hive. m6d is a display advertising prospecting company. Our role is to create

machine learning algorithms that are specifically tailored toward finding the best new

prospects for an advertising campaign. These algorithms are layered on top of a delivery

engine that is tied directly into a myriad of real time bidding exchanges that provide a

means to purchase locations on websites to display banner advertisements on behalf

of our clients. The m6d display advertising engine is involved in billions of auctions a

day and tens of millions of advertisements daily. Naturally, such a system produces an

immense amount of data. A large portion of the records that are generated by our

company’s display advertising delivery system are housed in m6d’s Hadoop cluster

and, as a result, Hive is the primary tool our data science team uses to interact with the

these logs.

Hive gives our data science team a way to extract and manipulate large amounts of

data. In fact, it allows us to extract samples and summarize data that prior to using

Hive could not be analyzed as efficiently, or at all, because of the immense size. Despite

the fact that Hive allows us access to huge amounts of data at rates many times faster

than before, it does not change the fact that most of the tools that we were previously

familiar with as data scientists are not always able to analyze data samples of the size

265

we can now produce. In summary, Hive provides us a great tool to extract huge amounts

of data; however, the toolbox of data science, or statistical learning, methods that we

as data scientists are used to using cannot easily accommodate the new larger data sets

without substantial changes.

Many different software packages have been developed or are under development for

both supervised and unsupervised learning on large data sets. Some of these software

packages are stand alone software implementations, such as Vowpal Wabbit and BBR,

while others are implementations within a larger infrastructure such as Mahout for

Hadoop or the multitude of “large data” packages for R. A portion of these algorithms

take advantage of parallel programing approaches while others rely on different meth-

ods to achieve scalability.

The primary tool for statistical learning for several of the data scientists in our team is

R. It provides a large array of packages that are able to perform many statistical learning

methods. More importantly, we have a lot of experience with it, know how its packages

perform, understand their features, and are very familiar with its documentation. How-

ever, one major drawback of R is that by default it loads the entire data set into memory.

This is a major limitation considering that the majority of the data sets that we extract

from Hive and are able to analyze today are much larger than what can fit in memory.

Moreover, once the data in R is larger than what is able to fit in memory, the system

will start swapping, which leads to the system thrashing and massive decreases in pro-

cessing speed.1

In no way are we advocating ignoring the new tools that are available. Obviously, it is

important to take advantage of the best of these scalable technologies, but only so much

time can be spent investigating and testing new technology. So now we are left with a

choice of either using the new tools that are available for large data sets or downsam-

pling our data to fit into the tools that we are more familiar with. If we decide to use

the new tools, we can gain signal by letting our data learn off of more data, and as a

result the variance in our estimates will decrease. This is particularly appealing in sit-

uations where the outcome is very rare. However, learning these new tools takes time

and there is an opportunity cost of using that time to learn new tools rather than an-

swering other questions that have particular value to the company.

Alternatively, we can downsample the data to obtain something that can fit in the old

tools we have at our disposal, but must deal with a loss of signal and increased variance

in our estimates. However, this allows us to deal with tools with which we are familiar

and the features that they provide. Thus, we are able to retain the functionality of our

current toolbox at the price of losing some signal. However, these are not the only two

possible approaches. In this case study, we highlight a way that we can both retain the

functionality of the current toolbox as well as gain signal, or decrease variance, by using

a larger sample, or all, of the data available to us.

1. http://www.r-bloggers.com/taking-r-to-the-limit-large-datasets-predictive-modeling-with-pmml-and

-adapa/

266 | Chapter 23: Case Studies

Figure 23-1 shows the probability of converting versus the score from an algorithm

designed to rank prospects for an advertising campaign. Higher scores should indicate

a higher probability of conversion. This plot clearly shows that the top individuals are

converting at a lower rate than some of the lower scoring browsers. That is, browsers

with scores greater than 1 convert at a lower rate than individuals with scores between

0.5 and 1.0. Considering that some campaigns only target a very small percentage of

the overall population, it is important the best prospects are among the top scorers.

Figure 23-1. Probability of conversion versus score

The line that expresses the relationship between the score and the conversion proba-

bility seen in Figure 23-1 is generated using a generalized additive model (GAM) in the

statistical programming package R.2 The details about GAMs will not be presented

here. For the purpose of this case study it can be thought of as a black box that produces

for each score a prediction of the conversion rate. The browsers can then be re-ranked

2. For more details on the use of generalized additive models (GAM), see Hastie et al. 2001. The R package

used to implement the GAM for the purpose of the analysis presented here is the mgcv package available

at http://cran.r-project.org/.

m6d.com (Media6Degrees) | 267

based on the predicted conversion rate; thus, the predicted conversion rate becomes

the new score.

The new ranking can be generated in the following way. First, extract the scores for

each individual browser and then follow them for some designated period of time, say

five days, and record if they took the desired action, and thus converted. Consider a

Hive table called scoretable that has the following information and is partitioned on

date and subpartitioned by offer.

Name Type Description

score double The score is the score generated by the

initial algorithm that does not neces-

sarily rank order appropriately.

convert int The variable convert is a binary variable

that is equal to one if the individual

browser takes the desired action in the

following five days and equal to zero if

not.

date int The day that the browser was given the

particular score.

offer int An ID of an offer.

The following query can then be used to extract a set of data from scoretable for use in

R to estimate the GAM line that predicts conversion for different levels of score like in

the preceding table:

SELECT score,convert

FROM scoretable

WHERE date >= (…) AND date <= (…)

AND offer = (…);

1.2347 0

3.2322 1

0.0013 0

0.3441 0

This data is then loaded into R and the following code is used to create the predicted

conversion probability versus score, as in the preceding table:

library(mgcv)

g1=gam(convert~s(score),family=binomial,data=[data frame name])

The issue with this approach is that it only can be used for a limited number of days of

data because the data set gets too large and R begins thrashing for any more than three

days of data. Moreover, it takes approximately 10 minutes of time for each campaign

to do this for about three days of data. So, running this analysis for about 300 campaigns

for a single scoring method took about 50 hours for three days of data.

By simply extracting the data from Hive in a slightly different way and making use of

the feature of the gam function in mgcv that allows for frequency weights, the same

268 | Chapter 23: Case Studies

analysis may be done using more data, and thus gaining more signal, at a much faster

rate. This is done by selecting the data from Hive by rounding the score to the nearest

hundredth and getting frequency weights for each rounded score, convert combination

by using a GROUP BY query. This is a very common approach for dealing with large data

sets and in the case of these scores there should be no loss of signal due to rounding

because there is no reason to believe that individuals with scores that differ by less than

0.001 are any different from each other. The following query would select such a data

set:

SELECT round(score,2) as score,convert,count(1) AS freq

FROM scoretable

WHERE date >= [start.date] and date <= [end.date] and offer = [chosen.offer]

GROUP BY round(score,2),convert;

1.23 0 500

3.23 1 22

0.00 0 127

0.34 0 36

The resulting data set is significantly smaller than the original approach presented that

does not use frequency weights. In fact, the initial data set for each offer consisted of

millions of records, and this new data set consists of approximately 6,500 rows per

offer. The new data is then loaded into R and the following code may be used to generate

the new GAM results:

library(mgcv)

g2=gam(convert~s(score),family=binomial,weights=freq,

data=[frequency weight data frame name])

(We wrapped the line.)

The previously presented approach took 10 minutes per offer to create the GAM for

only three days of data, compared to the frequency-weighted approach which was able

to create the GAM based on seven days of data in approximately 10 seconds. Thus, by

using frequency weights, the analysis for the 300 campaigns was able to be done in 50

minutes compared to 50 hours using the originally presented approach. This increase

in speed was also realized while using more than twice the amount of data resulting in

more precise estimates of the predicted conversion probabilities. In summary, the fre-

quency weights allowed for a more precise estimate of the GAM in significantly less

time.

In the presented case study, we showed how by rounding the continuous variables and

grouping like variables with frequency weights, we were both able to get more precise

estimates by using more data and fewer computational resources, resulting in quicker

estimates. The example shown was for a model with a single feature, score. In general,

this is an approach that will work well for a low number of features or a larger number

of sparse features. The above approach may be extended to higher dimensional prob-

lems as well using some other small tricks. One way this can be done for a larger number

of variables is by bucketing the variables, or features, into binary variables and then

using GROUP BY queries and frequency weights for those features. However, as the

m6d.com (Media6Degrees) | 269

number of features increases, and those features are not sparse, there is little value

gained by such an approach and other alternative methods must be explored, or soft-

ware designed for larger data sets must be embraced.

M6D UDF Pseudorank

by David Ha and Rumit Patel

Sorting data and identifying the top N elements is straightforward. You order the whole

data set by some criteria and limit the result set to N. But there are times when you

need to group like elements together and find the top N elements within that group

only. For example, identifying the top 10 requested songs for each recording artist or

the top 100 best-selling items per product category and country. Several database plat-

forms define a rank() function that can support these scenarios, but until Hive provides

an implementation, we can create a user-defined function to produce the results we

want. We will call this function p_rank() for psuedorank, leaving the name rank() for

the Hive implementation.

Say we have the following product sales data and we want to see the top three items

per category and country:

Category Country Product Sales

movies us chewblanca 100

movies us war stars iv 150

movies us war stars iii 200

movies us star wreck 300

movies gb titanus 100

movies gb spiderella 150

movies gb war stars iii 200

movies gb war stars iv 300

office us red pens 30

office us blue pens 50

office us black pens 60

office us pencils 70

office gb rulers 30

office gb blue pens 40

office gb black pens 50

office gb binder clips 60

In most SQL systems:

270 | Chapter 23: Case Studies

SELECT

category,country,product,sales,rank

FROM (

SELECT

category,country,product, sales,

rank() over (PARTITION BY category, country ORDER BY sales DESC) rank

FROM p_rank_demo) t

WHERE rank <= 3

To achieve the same result using HiveQL, the first step is partitioning the data into

groups, which we can achieve using the DISTRIBUTE BY clause. We must ensure that all

rows with the same category and country are sent to the same reducer:

DISTRIBUTE BY

category,

country

The next step is ordering the data in each group by descending sales using the SORT

BY clause. While ORDER BY effects a total ordering across all data, SORT BY affects the

ordering of data on a specific reducer. You must repeat the partition columns named

in the DISTRIBUTE BY clause:

SORT BY

category,

country,

sales DESC

Putting everything together, we have:

ADD JAR p-rank-demo.jar;

CREATE TEMPORARY FUNCTION p_rank AS 'demo.PsuedoRank';

SELECT

category,country,product,sales,rank

FROM (

SELECT

category,country,product,sales,

p_rank(category, country) rank

FROM (

SELECT

category,country,product,

sales

FROM p_rank_demo

DISTRIBUTE BY

category,country

SORT BY

category,country,sales desc) t1) t2

WHERE rank <= 3

The subquery t1 organizes the data so that all data belonging to the same category and

country are sorted by descending sales count. The next query t2 then uses p_rank() to

assign a rank to each row within the group. The outermost query filters the rank to be

in the top three:

m6d.com (Media6Degrees) | 271

Category Country Product Sales Rank

movies gb war stars iv 300 1

movies gb war stars iii 200 2

movies gb spiderella 150 3

movies us star wreck 300 1

movies us war stars iii 200 2

movies us war stars iv 150 3

office gb binder clips 60 1

office gb black pens 50 2

office gb blue pens 40 3

office us pencils 70 1

office us black pens 60 2

office us blue pens 50 3

p_rank() is implemented as a generic UDF whose parameters are all the identifying

attributes of the group, which, in this case, are category and country. The function

remembers the previous arguments, and so long as the successive arguments match, it

increments and returns the rank. Whenever the arguments do not match, the function

resets the rank back to 1 and starts over.

This is just one simple example of how p_rank() can be used. You can also find the

10th to 15th bestsellers by category and country. Or, if you precalculate the counts of

products in each category and country, you can use p_rank() to calculate percentiles

using a join. For example, if there were 1,000 products in the “movies” and “us” group,

the 50th, 70th, and 95th quantiles would have rank 500, 700, and 950, respectively.

Please know that p_rank() is not a direct substitute for rank() because there will be

differences in some circumstances. rank() returns the same value when there are ties,

but p_rank() will keep incrementing, so plan accordingly and test with your data.

Lastly, here is the implementation. It is public domain so feel free to use, improve, and

modify it to suit your needs:

package demo;

import org.apache.hadoop.hive.ql.exec.UDFArgumentException;

import org.apache.hadoop.hive.ql.metadata.HiveException;

import org.apache.hadoop.hive.ql.udf.generic.GenericUDF;

import org.apache.hadoop.hive.serde2.objectinspector.ObjectInspector;

import org.apache.hadoop.hive.serde2.objectinspector.primitive.

PrimitiveObjectInspectorFactory;

public class PsuedoRank extends GenericUDF {

/**

* The rank within the group. Resets whenever the group changes.

272 | Chapter 23: Case Studies

private long rank;

/**

* Key of the group that we are ranking. Use the string form

* of the objects since deferred object and equals do not work

* as expected even for equivalent values.

private String[] groupKey;

@Override

public ObjectInspector initialize(ObjectInspector[] oi)

throws UDFArgumentException {

return PrimitiveObjectInspectorFactory.javaLongObjectInspector;

}

@Override

public Object evaluate(DeferredObject[] currentKey) throws HiveException {

if (!sameAsPreviousKey(currentKey)) {

rank = 1;

}

return new Long(rank++);

}

/**

* Returns true if the current key and the previous keys are the same.

* If the keys are not the same, then sets {@link #groupKey} to the

* current key.

private boolean sameAsPreviousKey(DeferredObject[] currentKey)

throws HiveException {

if (null == currentKey && null == groupKey) {

return true;

}

String[] previousKey = groupKey;

copy(currentKey);

if (null == groupKey && null != previousKey) {

return false;

}

if (null != groupKey && null == previousKey) {

return false;

}

if (groupKey.length != previousKey.length) {

return false;

}

for (int index = 0; index < previousKey.length; index++) {

if (!groupKey[index].equals(previousKey[index])) {

return false;

}

return true;

}

/**

* Copies the given key to {@link #groupKey} for future

* comparisons.

m6d.com (Media6Degrees) | 273

private void copy(DeferredObject[] currentKey)

throws HiveException {

if (null == currentKey) {

groupKey = null;

} else {

groupKey = new String[currentKey.length];

for (int index = 0; index < currentKey.length; index++) {

groupKey[index] = String.valueOf(currentKey[index].get());

}

@Override

public String getDisplayString(String[] children) {

StringBuilder sb = new StringBuilder();

sb.append("PsuedoRank (");

for (int i = 0; i < children.length; i++) {

if (i > 0) {

sb.append(", ");

}

sb.append(children[i]);

}

sb.append(")");

return sb.toString();

}

M6D Managing Hive Data Across Multiple MapReduce Clusters

Although Hadoop clusters are designed to scale from 10 to 10,000 nodes, sometimes

deployment-specific requirements involve running more than one filesystem or Job-

Tracker. At M6D, we have such requirements, for example we have hourly and daily

process reports using Hadoop and Hive that are business critical and must complete

in a timely manner. However our systems also support data science and sales engineers

that periodically run ad hoc reporting. While using the fair share scheduler and capacity

scheduler meets many of our requirements, we need more isolation than schedulers

can provide. Also, because HDFS has no snapshot or incremental backup type features,

we require a solution that will prevent an accidental delete or drop table operations

from destroying data.

Our solution is to run two distinct Hadoop deployments. Data can have a replication

factor of two or three on the primary deployment and additionally be replicated to a

second deployment. This decision allows us to have guaranteed resources dedicated to

our time-sensitive production process as well as our ad hoc users. Additionally, we

protected against any accidental drop tables or data deletes. This design does incur

some overhead in having to administer two deployments and setup and administer the

replication processes, but this overhead is justified in our case.

274 | Chapter 23: Case Studies

Downloa d f r o m W o w ! e B o o k < w w w.woweb o o k . c o m >

Our two deployments are known as production and research. They each have their own

dedicated Data Nodes and Task Trackers. Each NameNode and JobTracker is a failover

setup using DRBD and Linux-HA. Both deployments are on the same switching net-

work (Tables 23-1 and 23-2).

Table 23-1. Production

NameNode hdfs.hadoop.pvt:54310

JobTracker jt.hadoop.pvt:54311

Table 23-2. Research

NameNode rs01.hadoop.pvt:34310

JobTracker rjt.hadoop.pvt:34311

Cross deployment queries with Hive

A given table zz_mid_set exists on Production and we wish to be able to query it from

Research without having to transfer the data between clusters using distcp. Generally,

we try to avoid this because it breaks our isolation design but it is nice to know that

this can be done.

Use the describe extended command to determine the columns of a table as well as its

location:

hive> set fs.default.name;

fs.default.name=hdfs://hdfs.hadoop.pvt:54310

hive> set mapred.job.tracker;

mapred.job.tracker=jt.hadoop.pvt:54311

hive> describe extended zz_mid_set;

adv_spend_id int

transaction_id bigint

time string

client_id bigint

visit_info string

event_type tinyint

level int

location:hdfs://hdfs.hadoop.pvt:54310/user/hive/warehouse/zz_mid_set

Time taken: 0.063 seconds

hive> select count(1) from zz_mid_set;

1795928

On the second cluster, craft a second CREATE TABLE statement with the same columns.

Create the second table as EXTERNAL, in this way if the table is dropped on the second

cluster the files are not deleted on the first cluster. Notice that for the location we

specified a full URI. In fact, when you specify a location as a relative URI, Hive stores

it as a full URI:

m6d.com (Media6Degrees) | 275

hive> set fs.default.name;

fs.default.name=hdfs://rs01.hadoop.pvt:34310

hive> set mapred.job.tracker;

mapred.job.tracker=rjt.hadoop.pvt:34311

hive> CREATE TABLE EXTERNAL table_in_another_cluster

( adv_spend_id int, transaction_id bigint, time string, client_id bigint,

visit_info string, event_type tinyint, level int)

LOCATION 'hdfs://hdfs.hadoop.pvt:54310/user/hive/warehouse/zz_mid_set';

hive> select count(*) FROM table_in_another_cluster;

1795928

It is important to note that this cross-deployment access works because both clusters

have network access to each other. The TaskTrackers of the deployment we submit the

job to will have to be able to access the NameNode and DataNodes of the other de-

ployment. Hadoop was designed to move processing closer to data. This is done by

scheduling tasks to run on nodes where the data is located. In this scenario TaskTrack-

ers connect to another cluster’s DataNodes. Which means a general performance de-

crease and network usage increase.

Replicating Hive data between deployments

Replicating Hadoop and Hive data is easier than replicating a traditional database.

Unlike a database running multiple transactions that change the underlying data fre-

quently, Hadoop and Hive data is typically “write once.” Adding new partitions does

not change the existing ones, and typically new partitions are added on time-based

intervals.

Early iterations of replication systems were standalone systems that used distcp and

generated Hive statements to add partitions on an interval. When we wanted to repli-

cate a new table, we could copy an existing program and make changes for different

tables and partitions. Over time we worked out a system that could do this in a more

automated manner without having to design a new process for each table to replicate.

The process that creates the partition also creates an empty HDFS file named:

/replication/default.fracture_act/hit_date=20110304,mid=3000

The replication daemon constantly scans the replication hierarchy. If it finds a file, it

looks up the table and partition in Hive’s metadata. It then uses the results to replicate

the partition. On a successful replication the file is then deleted.

Below is the main loop of the program. First, we do some checking to make sure the

table is defined in the source and destination metastores:

public void run(){

while (goOn){

Path base = new Path(pathToConsume);

FileStatus [] children = srcFs.listStatus(base);

for (FileStatus child: children){

try {

openHiveService();

String db = child.getPath().getName().split("\\.")[0];

276 | Chapter 23: Case Studies

String hiveTable = child.getPath().getName().split("\\.")[1];

Table table = srcHive.client.get_table(db, hiveTable);

if (table == null){

throw new RuntimeException(db+" "+hiveTable+

" not found in source metastore");

}

Table tableR = destHive.client.get_table(db,hiveTable);

if (tableR == null){

throw new RuntimeException(db+" "+hiveTable+

" not found in dest metastore");

}

Using the database and table name we can look up the location information inside the

metastore. We then do a sanity check to ensure the information does not already exist:

URI localTable = new URI(tableR.getSd().getLocation());

FileStatus [] partitions = srcFs.listStatus(child.getPath());

for (FileStatus partition : partitions){

try {

String replaced = partition.getPath().getName()

.replace(",", "/").replace("'","");

Partition p = srcHive.client.get_partition_by_name(

db, hiveTable, replaced);

URI partUri = new URI(p.getSd().getLocation());

String path = partUri.getPath();

DistCp distCp = new DistCp(destConf.conf);

String thdfile = "/tmp/replicator_distcp";

Path tmpPath = new Path(thdfile);

destFs.delete(tmpPath,true);

if (destFs.exists( new Path(localTable.getScheme()+

"://"+localTable.getHost()+":"+localTable.getPort()+path) ) ){

throw new RuntimeException("Target path already exists "

+localTable.getScheme()+"://"+localTable.getHost()+

":"+localTable.getPort()+path );

}

Hadoop DistCP is not necessarily made to be run programmatically. However, we can

pass a string array identical to command-line arguments to its main function. After, we

check to confirm the returned result was a 0:

String [] dargs = new String [4];

dargs[0]="-log";

dargs[1]=localTable.getScheme()+"://"+localTable.getHost()+":"+

localTable.getPort()+thdfile;

dargs[2]=p.getSd().getLocation();

dargs[3]=localTable.getScheme()+"://"+localTable.getHost()+":"+

localTable.getPort()+path;

int result =ToolRunner.run(distCp,dargs);

if (result != 0){

throw new RuntimeException("DistCP failed "+ dargs[2] +" "+dargs[3]);

}

Finally, we re-create the ALTER TABLE statement that adds the partition:

String HQL = "ALTER TABLE "+hiveTable+

" ADD PARTITION ("+partition.getPath().getName()

m6d.com (Media6Degrees) | 277

+") LOCATION '"+path+"'";

destHive.client.execute("SET hive.support.concurrency=false");

destHive.client.execute("USE "+db);

destHive.client.execute(HQL);

String [] results=destHive.client.fetchAll();

srcFs.delete(partition.getPath(),true);

} catch (Exception ex){

ex.printStackTrace();

}

} // for each partition

} catch (Exception ex) {

//error(ex);

ex.printStackTrace();

}

} // for each table

closeHiveService();

Thread.sleep(60L*1000L);

} // end run loop

} // end run

Outbrain

by David Funk

Outbrain is the leading content-discovery platform.

In-Site Referrer Identification

Sometimes, when you’re trying to aggregate your traffic, it can be tricky to tell where

it’s actually coming from, especially for traffic coming from elsewhere in your site. If

you have a site with a lot of URLs with different structures, you can’t simply check that

the referrer URLs match the landing page.

Cleaning up the URLs

What we want is to correctly group each referrer as either In-site, Direct, or Other. If

it’s Other, we’ll just keep the actual URL. That way you can tell your internal traffic

apart from Google searches to your site, and so on and so forth. If the referrer is blank

or null, we’ll label it as Direct.

From here on out, we’ll assume that all our URLs are already parsed down to the host

or domain, whatever level of granularity you’re aiming for. Personally, I like using the

domain because it’s a little simpler. That said, Hive only has a host function, but not

domain.

If you just have the raw URLs, there are a couple of options. The host, as given below,

gives the full host, like news.google.com or www.google.com, whereas the domain would

truncate it down to the lowest logical level, like google.com or google.co.uk.

Host = PARSE_URL(my_url, ‘HOST’’)

278 | Chapter 23: Case Studies

Or you could just use a UDF for it. Whatever, I don’t care. The important thing is that

we’re going to be using these to look for matches, so just make your choice based on

your own criteria.

Determining referrer type

So, back to the example. We have, let’s say, three sites: mysite1.com, mysite2.com, and

mysite3.com. Now, we can convert each pageview’s URL to the appropriate class. Let’s

imagine a table called referrer_identification:

ri_page_url STRING

ri_referrer_url STRING

Now, we can easily add in the referrer type with a query:

SELECT ri_page_url, ri_referrer_url,

CASE

WHEN ri_referrer_url is NULL or ri_referrer_url = ‘’ THEN ‘DIRECT’

WHEN ri_referrer_url is in (‘mysite1.com’,’mysite2.com’,’mysite3.com’) THEN ‘INSITE’

ELSE ri_referrer_url

END as ri_referrer_url_classed

FROM

referrer_identification;

Multiple URLs

This is all pretty simple. But what if we’re an ad network? What if we have hundreds

of sites? What if each of the sites could have any number of URL structures?

If that’s the case, we probably also have a table that has each URL, as well as what site

it belongs to. Let’s call it site_url, with a schema like:

su_site_id INT

su_url STRING

Let’s also add one more field to our earlier table, referrer_identification:

ri_site_id INT

Now we’re in business. What we want to do is go through each referrer URL and see

if it matches with anything of the same site ID. If anything matches, it’s an In-site

referrer. Otherwise, it’s something else. So, let’s query for that:

SELECT

c.c_page_url as ri_page_url,

c.c_site_id as ri_site_id,

CASE

WHEN c.c_referrer_url is NULL or c.c_referrer_url = ‘’ THEN ‘DIRECT’

WHEN c.c_insite_referrer_flags > 0 THEN ‘INSITE’

ELSE c.c_referrer_url

END as ri_referrer_url_classed

FROM

(SELECT

a.a_page_url as c_page_url,

a.a_referrer_url as c_referrer_url,

Outbrain | 279

a.a_site_id as c_site_id,

SUM(IF(b.b_url <> ‘’, 1, 0)) as c_insite_referrer_flags

FROM

(SELECT

ri_page_url as a_page_url,

ri_referrer_url as a_referrer_url,

ri_site_id as a_site_id

FROM

referrer_identification

) a

LEFT OUTER JOIN

(SELECT

su_site_id as b_site_id,

su_url as b_url

FROM

site_url

) b

a.a_site_id = b.b_site_id and

a.a_referrer_url = b.b_url

) c

A few small notes about this. We use the outer join in this case, because we expect there

to be some external referrers that won’t match, and this will let them through. Then,

we just catch any cases that did match, and if there were any, we know they came from

somewhere in the site.

Counting Uniques

Let’s say you want to calculate the number of unique visitors you have to your site/

network/whatever. We’ll use a ridiculously simple schema for our hypothetical table,

daily_users:

du_user_id STRING

du_date STRING

However, if you have too many users and not enough machines in your cluster, it might

begin to have trouble counting users over a month:

SELECT

COUNT(DISTINCT du_user_id)

FROM

daily_users

WHERE

du_date >= ‘2012-03-01’ and

du_date <= ‘2012-03-31’

In all likelihood, your cluster is probably able to make it through the map phase without

too much problems, but starts having issues around the reduce phase. The problem is

that it’s able to access all the records but it can’t count them all at once. Of course, you

can’t count them day by day, either, because there might be some redundancies.

280 | Chapter 23: Case Studies

Why this is a problem

Counting uniques is O(n), where n is the number of records, but it has a high constant

factor. We could maybe come up with some clever way to cut that down a little bit,

but it’s much easier to cut down your n. While it’s never good to have a high O(n),

most of the real problems happen further along. If you have something that takes

n1.1 time to run, who cares if you only have n=2 versus n=1. It’s slower, sure, but

nowhere near the difference between n=1 and n=100.

So, if each day has m entries, and an average of x redundancies, our first query would

have n= 31*m. We can reduce this to n=31*(m–x) by building a temp table to save

deduped versions for each day.

Load a temp table

First, create the temp table:

CREATE TABLE daily_users_deduped (dud_user_id STRING)

PARTITIONED BY (dud_date STRING)

ROW FORMAT DELIMITED

FIELDS TERMINATED BY ‘\t’;

Then we write a template version of a query to run over each day, and update it to our

temp table. I like to refer to these as “metajobs,” so let’s call this mj_01.sql:

INSERT OVERWRITE TABLE daily_users_deduped

PARTITION (dud_date = ‘:date:’)

SELECT DISTINCT

du_user_id

FROM

daily_users

WHERE

du_date = ‘:date:’

Next, we write a script that marks this file up, runs it, and repeats it for every date in a

range. For this, we have three functions, modify_temp_file, which replaces a variable

name with fire_query, which basically runs hive –f on a file, and then a function to

delete the file:

start_date = ‘2012-03-01’

end_date = ‘2012-03-31’

for date in date_range(start_date, end_date):

femp_file = modify_temp_file(‘mj_01.sql’,{‘:date:’:my_date})

fire_query(temp_file)

delete(temp_file)

Querying the temp table

Run the script, and you’ve got a table with a n=31*(m-x). Now, you can query the

deduped table without as big a reduce step to get through.

Outbrain | 281

SELECT

COUNT(DISTINCT (dud_uuid)

FROM

daily_users_deduped

If that’s not enough, you can then dedupe sets of dates, maybe two at a time, whatever

the interval that works for you. If you still have trouble, you could hash your user IDs

into different classes, maybe based on the first character, to shrink n even further.

The basic idea remains, if you limit the size of your n, a high O(n) isn’t as big of a deal.

Sessionization

For analyzing web traffic, we often want to be able to measure engagement based on

various criteria. One way is to break up user behavior into sessions, chunks of activity

that represent a single “use.” A user might come to your site several times a day, a few

days a month, but each visit is certainly not the same.

So, what is a session? One definition is a string of activity, not separated by more than

30 minutes. That is, if you go to your first page, wait five minutes, go to the second

page, it’s the same session. Wait 30 minutes exactly until the third page, still the same

session. Wait 31 minutes until that fourth page, and the session will be broken; rather

than the fourth pageview, it would be the first page of the second session.

Once we’ve got these broken out, we can look at properties of the session to see what

happened. The ubiquitous case is to compare referrers to your page by session length.

So, we might want to find out if Google or Facebook give better engagement on your

site, which we might measure by session length.

At first glance, this seems perfect for an iterative process. For each pageview, keep

counting backwards until you find the page that was first. But Hive isn’t iterative.

You can, however, figure it out. I like to break this into four phases.

1. Identify which pageviews are the session starters, or “origin” pages.

2. For every pageview, bucket it in with the correct origin page.

3. Aggregate all the pageviews for each origin page.

4. Label each origin page, then calculate engagement for each session.

This leaves a table where each row represents a full session, which you can then query

for whatever you want to find out.

Setting it up

Let’s define our table, session_test:

st_user_id STRING

st_pageview_id STRING

st_page_url STRING

282 | Chapter 23: Case Studies

st_referrer_url STRING

st_timestamp DOUBLE

Most of this is pretty straightforward, though I will mention that st_pageview_id is

basically a unique ID to represent each transaction, in this case a pageview. Otherwise,

it could be confusing if you happened to have multiple views of the same page. For the

purposes of this example, the timestamp will be in terms of seconds.

Finding origin pageviews

All right, let’s start with step one (shocking!). How do we find which pageviews are the

session starters? Well, if we assume any break of more than 30 minutes implies a new

session, than any session starter can’t have any activity that precedes it by 30 minutes

or less. This is a great case for conditional sums. What we want to do is count up how

many times, for each pageview. Then, anything with a count of zero must be an origin

case.

In order to do this, we need to compare every pageview that could precede it. This is a

pretty expensive move, as it requires performing a cross-product. To prevent this from

blowing up to unmanageable size, we should group everything on criteria that limits it

as much as possible. In this case, it’s just the user ID, but if you have a large network

of independent sites, you might also want to group based on each source, as well:

CREATE TABLE sessionization_step_one_origins AS

SELECT

c.c_user_id as ssoo_user_id,

c.c_pageview_id as ssoo_pageview_id,

c.c_timestamp as ssoo_timestamp

FROM

(SELECT

a.a_user_id as c_user_id,

a.a_pageview_id as c_pageview_id,

a.a_timestamp as c.c_timestamp,

SUM(IF(a.a_timestamp + 1800 >= b.b_timestamp AND

a.a_timestamp < b.b_timestamp,1,0)) AS c_nonorigin_flags

FROM

(SELECT

st_user_id as a_user_id,

st_pageview_id as a_pageview_id,

st_timestamp as a_timestamp

FROM

session_test

) a

JOIN

(SELECT

st_user_id as b_user_id,

st_timestamp as b_timestamp

FROM

session_test

) b

Outbrain | 283

a.a_user_id = b.b_user_id

GROUP BY

a.a_user_id,

a.a_pageview_id,

a.a_timestamp

) c

WHERE

c.c_nonorigin_flags

That’s a bit much, isn’t it? The important part is to count the flags that are not of a

session origin, which is where we define c_nonorigin_flags. Basically, counting up how

many reasons why it isn’t the session starter. Aka, this line:

SUM(IF(a.a_timestamp + 1800 >= b.b_timestamp AND

a.a_timestamp < b.b_timestamp,1,0)) as c_nonorigin_flags

Let’s break this up, part by part. First, everything is in terms of subquery a. We only

use b to qualify those candidates. So, the first part, the a.a_timestamp + 1800 >=

b.b_timestamp, is just asking if the candidate timestamp is no more than 30 minutes

prior to the qualifying timestamp. The second part, a.a_timestamp < b.b_timestamp

adds a check to make sure that it is earlier, otherwise every timestamp that occurred

later than it’s qualifier would trigger a false positive. Plus, since this is a cross-product,

it prevents a false positive by using the candidate as its own qualifier.

Now, we’re left with a table, sessionization_step_one_origins, with a schema of:

ssoo_user_id STRING

ssoo_pageview_id STRING

ssoo_timestamp DOUBLE

Bucketing PVs to origins

Which is probably a good reason to start on step two, finding which pageview belongs

to which origin. It’s pretty simple to do this, every pageview’s origin must be the one

immediately prior to it. For this, we take another big join to check for the minimum

difference between a pageview’s timestamp and all the potential origin pageviews:

CREATE TABLE sessionization_step_two_origin_identification AS

SELECT

c.c_user_id as sstoi_user_id,

c.c_pageview_id as sstoi_pageview_id,

d.d_pageview_id as sstoi_origin_pageview_id

FROM

(SELECT

a.a_user_id as c_user_id,

a.a_pageview_id as c_pageview_id,

MAX(IF(a.a_timestamp >= b.b_timestamp, b.b_timestamp, NULL)) as c_origin_timestamp

FROM

(SELECT

st_user_id as a_user_id,

st_pageview_id as a_pageview_id,

st_timestamp as a_timestamp

FROM

284 | Chapter 23: Case Studies

session_test

) a

JOIN

(SELECT

ssoo_user_id as b_user_id,

ssoo_timestamp as b_timestamp

FROM

sessionization_step_one_origins

) b

a.a_user_id = b.b_user_id

GROUP BY

a.a_user_id,

a.a_pageview_id

) c

JOIN

(SELECT

ssoo_usr_id as d_user_id,

ssoo_pageview_id as d_pageview_id,

ssoo_timestamp as d_timestamp

FROM

sessionization_step_one_origins

) d

c.c_user_id = d.d_user_id and

c.c_origin_timestamp = d.d_timestamp

There’s a lot to mention here. First, let’s look at this line:

MAX(IF(a.a_timestamp >= b.b_timestamp, b.b_timestamp, NULL)) as c_origin_timestamp

Again, we use the idea of qualifiers and candidates, in this case b are the candidates for

every qualifier a. An origin candidate can’t come later than the pageview, so for every

case like that, we want to find the absolute latest origin that meets that criteria. The

null is irrelevant, because we are guaranteed to have a minimum, because there is always

at least one possible origin (even if it’s itself). This doesn’t give us the origin, but it gives

us the timestamp, which we can use as a fingerprint for what the origin should be.

From here, it’s just a matter of matching up this timestamp with all the other potential

origins, and we know which origin each pageview belongs to. We’re left with the table

sessionization_step_two_origin_identification, with the following schema:

sstoi_user_id STRING

sstoi_pageview_id STRING

sstoi_origin_pageview_id STRING

It’s worth mentioning that this isn’t the only way to identify the origin pageviews. You

could do it based on the referrer, labeling any external referrer, homepage URL, or

blank referrer (indicating direct traffic) as a session origin. You could base it on an

action, only measuring activity after a click. There are plenty of options, but the im-

portant thing is simply to identify what the session origins are.

Outbrain | 285

Aggregating on origins

At this point, it’s all pretty easy. Step three, where we aggregate on origins, is really,

really simple. For each origin, count up how many pageviews match to it:

CREATE TABLE sessionization_step_three_origin_aggregation AS

SELECT

a.a_user_id as sstoa_user_id,

a.a_origin_pageview_id as sstoa_origin_pageview_id,

COUNT(1) as sstoa_pageview_count

FROM

(SELECT

ssoo_user_id as a_user_id

ssoo_pageview_id as a_origin_pageview_id

FROM

sessionization_step_one_origins

) a

JOIN

(SELECT

sstoi_user_id as b_user_id,

sstoi_origin_pageview_id as b_origin_pageview_id

FROM

sessionization_step_two_origin_identification

) b

a.a_user_id = b.b_user_id and

a.a_origin_pageview_id = b.b_origin_pageview_id

GROUP BY

a.a_user_id,

a.a_origin_pageview_id

Aggregating on origin type

Now, this last step we could have avoided by keeping all the qualitative info about a

pageview, particularly the origins, in one of the earlier steps. However, if you have a

lot of details you want to pay attention to, it can sometimes be easier to add it in at the

end. Which is step four:

CREATE TABLE sessionization_step_four_qualitative_labeling

SELECT

a.a_user_id as ssfql_user_id,

a.a_origin_pageview_id as ssfql_origin_pageview_id,

b.b_timestamp as ssfql_timestamp,

b.b_page_url as ssfql_page_url,

b.b_referrer_url as ssfql_referrer_url,

a.a_pageview_count as ssqfl_pageview_count

(SELECT

sstoa_user_id as a_user_id,

sstoa_origin_pageview_id as a_origin_pageview_id,

sstoa_pageview_count as a_pageview_count

FROM

sessionization_step_three_origin_aggregation

) a

286 | Chapter 23: Case Studies

JOIN

(SELECT

st_user_id as b_user_id,

st_pageview_id as b_pageview_id,

st_page_url as b_page_url,

st_referrer_url as b_referrer_url,

st_timestamp as b_timestamp

FROM

session_test

) b

a.a_user_id = b.b_user_id and

a.a_origin_pageview_id = b.b_pageview_id

Measure engagement

Now, with our final table, we can do whatever we want. Let’s say we want to check the

number of sessions, average pageviews per session, weighted average pageviews per

session, and the max or min. We could pick whatever criteria we want, or none at all,

but in this case, let’s do it by referrer URL so we can find out the answer to which traffic

source gives the best engagement. And, just for kicks, let’s also check who gives us the

most unique users:

SELECT

PARSE_URL(ssfql_referrer_url, ‘HOST’) as referrer_host,

COUNT(1) as session_count,

AVG(ssfql_pageview_count) as avg_pvs_per_session,

SUM(ssfq_pageview_count)/COUNT(1) as weighted_avg_pvs_per_session,

MAX(ssfql_pageview_count) as max_pvs_per_session,

MIN(ssfql_pageview_count) as min_pvs_per_session,

COUNT(DISTINCT ssfql_usr_id) as unique_users

FROM

sessionization_step_three_origin_aggregation

GROUP BY

PARSE_URL(ssfql_referrer_url, ‘HOST’) as referrer_host

And there we have it. We could check which page URL gives the best engagement,

figure out who the power users are, whatever. Once we’ve got it all in a temp table,

especially with a more complete set of qualitative attributes, we can answer all sorts of

questions about user engagement.

NASA’s Jet Propulsion Laboratory

The Regional Climate Model Evaluation System

by Chris A. Mattmann, Paul Zimdars, Cameron Goodale, Andrew F. Hart, Jinwon Kim,

Duane Waliser, Peter Lean

Since 2009, our team at NASA’s Jet Propulsion Laboratory (JPL) has actively led the

development of a Regional Climate Model Evaluation System (RCMES). The system,

NASA’s Jet Propulsion Laboratory | 287

originally funded under the American Recovery and Reinvestment Act (ARRA) has the

following goals:

• Facilitate the evaluation and analysis of regional climate model simulation outputs

via the availability of the reference data sets of quality-controlled observations and

assimilations especially from spaceborne sensors, an efficient database structure,

a collection of computational tools for calculating the metrics for model evaluation

metrics and diagnostics, and relocatable and friendly user interfaces.

• Easily bring together a number of complex, and heterogeneous software tools and

capability for data access, representation, regridding, reformatting, and visualiza-

tion so that the end product such as a bias plot can be easily delivered to the end

user.

• Support regional assessments of climate variability, and impacts, needed to inform

decision makers (e.g., local governments, agriculture, state government, hydrolo-

gists) so that they can make critical decisions with large financial and societal

impact.

• Overcome data format and metadata heterogeneity (e.g., NetCDF3/4, CF meta-

data conventions, HDF4/5, HDF-EOS metadata conventions).

• Deal with spatial and temporal differences, (e.g., line up the data alongside a

180/80 lat-lon grid—such as converting from, for example, a 360/360 lat-lon grid

—and making sure data, that may be originally daily, is properly comparable with

monthly data.

• Elastically scaling up, performing a regional study that requires specific remote

sensing data, and climate model output data, performing a series of analyses, and

then destroying that particular instance of the system. In other words, supporting

transient analyses, and rapid construction/deconstruction of RCMES instances.

Figure 23-2 shows the architecture and data flow of the Regional Climate Model Eval-

uation System

In support of these goals, we have constructed a multifaceted system shown in Fig-

ure 23-2. Reading the diagram from left to right, available reference data sets from

observations and assimilations, especially from satellite-based remote sensing, enters

the system according to the desired climate parameters useful for climate model eval-

uation. Those parameters are stored in various mission data sets, and those data sets

are housed in several external repositories, eventually fed into the database component

(RCMED: Regional Climate Model Evaluation Database) of RCMES.

As an example, AIRS is NASA’s Atmospheric Infrared Sounder and provides parame-

ters including surface air temperature, temperature, and geopotential; MODIS is

NASA’s Moderate Imaging Spectroradiometer and provides parameters including

cloud fraction; and TRMM is NASA’s Tropical Rainfall Measurement Mission and

provides parameters including monthly precipitation. This information is summarized

288 | Chapter 23: Case Studies

Downloa d f r o m W o w ! e B o o k < w w w.woweb o o k . c o m >

in our RCMES system website parameter table, http://rcmes.jpl.nasa.gov/rcmed/param

eters/, and shown in Figure 23-3.

Data sets are loaded into the RCMED using the Apache OODT extractor framework

and the desired parameters, their values, spatial and temporal constraints (and

optionally height) are loaded and potentially transformed (e.g., normalized, put on

same coordinate system, converted from unit values) into a MySQL database. The data

loaded into that MySQL database, RCMED, is exposed to external clients via a Space/

Time query and subsetting web service; the description of which is a topic of a separate

Figure 23-2. JPL Architecture Diagram.

NASA’s Jet Propulsion Laboratory | 289

study. For all intents and purposes, it provides the same capabilities that the OPeNDAP

technology does.

The right side of the diagram shows the Regional Climate Model Evaluation Toolkit

(RCMET). It provides users with the ability to take in the reference data from RCMED

and climate model output data produced elsewhere and to re-grid these datasets in

order to match them spatially and temporally in preparation for the comparison of the

reference and model data for the evaluation of model output against the user-selected

reference data. At that point, the system allows for seasonal cycle compositing (e.g., all

Januaries, or all summer months for N years), and for preparing the data for eventual

metrics calculation, that is, comparison of the values of the model output against the

remote sensing data observation parameters and their values. The system supports

several metrics, such as bias computation, Root Mean Squared Error (RMSE), and the

generation of relevant visualizations, including traditional plots and Taylor diagrams

for science use/decision making.

Our Experience: Why Hive?

So, where does Hive come in to play? After loading 6 billion rows of (latitude, longitude,

time, data point value, height) tuples into MySQL, the system fell down and

Figure 23-3. JPL Physical Architecture Diagram

290 | Chapter 23: Case Studies

experienced data loss. This is probably due in part to our naïve strategy of storing all

of the data points in a single table. Over time, we evolved this strategy to break tables

down by dataset and by parameter, which helped but added needless overhead that we

didn’t want to spend cycles engineering around.

Instead, we decided to experiment with the Apache Hive technology. We installed Hive

0.5+20 using CDHv3 and Apache Hadoop (0.20.2+320). CDHv3 came with a number

of other relevant tools including Sqoop, and Hue, which we leveraged in our architec-

ture, shown in the bottom portion of Figure 23-3.

We used Apache Sqoop to dump out the data into Hive, and then wrote an Apache

OODT wrapper that queried Hive for the data by Space/Time and provided it back to

the RCMET and other users (shown in the middle portion of Figure 23-2). The full

architecture for the RCMES cluster is shown in Figure 23-3. We had five machines,

including a master/slave configuration as shown in the diagram, connected by a private

network running GigE.

Some Challenges and How We Overcame Them

During the migration of data from MySQL to Hive, we experienced slow response times

while doing simple tasks such as a count DB query (e.g., hive> select count(data

point_id) from dataPoint;). We initially loaded up around 2.5 billion data points in

a single table and noticed that on our machine configuration, Hive took approximately

5–6 minutes to do a count of these 2.5 billion records (15–17 minutes for the full 6.8

billion records). The reduce portion was fast (we were experiencing a single reduce

phase since we were using a count * benchmark) but the map stage took the remainder

of the time (~95%). Our system at the time consisted of six (4 x quad-core) systems

with approximately 24 GB of RAM each (all of the machines shown in Figure 23-3,

plus another “borrowed machine” of similar class from another cluster).

We attempted to add more nodes, increase map tasktrackers (many different #s),

change DFS block size (32 M, 64 M, 128 MB, 256 M), leverage LZO compression, and

alter many other configuration variables (io.sort.factor, io.sort.mb) without much

success in lowering the time to complete the count. We did notice a high I/O wait on

the nodes no matter how many task trackers we ran. The size of the database was

approximately ~200GB and with MySQL it took a few seconds to do both the 2.5

billion and 6.7 billion count.

Members of the Hive community jumped in and provided us with insight, ranging from

mentioning that HDFS read speed is about 60 MB/sec comparing to about 1 GB/sec

on local disk, depending of course on network speed, and namenode workload. The

numbers suggested by the community member suggested that we needed roughly 16

mappers in the Hadoop job to match with the I/O performance of a local non-Hadoop

task. In addition, Hive community members suggested that we increase the number of

mappers (increase parallelism) by reducing the split size (input size) for each mapper,

noting we should examine the following parameters: mapred.min.split.size,

NASA’s Jet Propulsion Laboratory | 291

mapred.max.split.size, mapred.min.split.size.per.rack, and mapred.min.split.size

.per.node, and suggesting that the parameters should be set to a value of 64 MB. Finally,

the community suggested that we look at a benchmark that only counts rows by using

count(1) instead of count (datapoint_id), as the latter is faster since no column refer-

ence means no decompression and deserialization, e.g., if you store your table in RCFile

format.

Based on the above feedback, we were able to tune our Hive cluster for RCMES to

respond to a count query benchmark, and to a space/time query from RCMET for

billions of rows in under 15 seconds, using the above-mentioned resources, making

Hive a viable and great choice for our system architecture.

Conclusion

We have described our use of Apache Hive in the JPL Regional Climate Model Evalu-

ation System. We leveraged Hive during a case study wherein we wanted to explore

cloud-based technology alternatives to MySQL, and configuration requirements

needed to make it scale to the level of tens of billions of rows, and to elastically destroy

and re-create the data stored within.

Hive did a great job of meeting our system needs and we are actively looking for more

ways to closely integrate it into the RCMES system.

Photobucket

Photobucket is the largest dedicated photo-hosting service on the Internet. Started in

2003 by Alex Welch and Darren Crystal, Photobucket quickly became one of the most

popular sites on the Internet and attracted over one hundred million users and billions

of stored and shared media. User and system data is spread across hundreds of MySQL

instances, thousands of web servers, and petabytes of filesystem.

Big Data at Photobucket

Prior to 2008, Photobucket didn’t have a dedicated analytics system in-house. Ques-

tions from the business users were run across hundreds of MySQL instances and the

results aggregated manually in Excel.

In 2008, Photobucket embarked on implementing its first data warehouse dedicated

to answering the increasingly complex data questions being asked by a fast-growing

company.

The first iteration of the data warehouse was built using an open source system with a

Java SQL optimizer and a set of underlying PostGreSQL databases. The previous system

worked well into 2009, but the shortcomings in the architecture became quickly evi-

dent. Working data sets quickly became larger than the available memory; coupled

292 | Chapter 23: Case Studies

with the difficulty in repartitioning the data across the PostGreSQL nodes forced us to

scale up when we really wanted to scale out.

In 2009, we started to investigate systems that would allow us to scale out, as the

amount of data continued to grow and still meet our SLA with the business users.

Hadoop quickly became the favorite for consuming and analyzing the terabytes of data

generated daily by the system, but the difficulty of writing MapReduce programs for

simple ad hoc questions became a negative factor for full implementation. Thankfully,

Facebook open sourced Hive a few weeks later and the barriers to efficiently answering

ad hoc business questions were quickly smashed.

Hive demonstrates many advantages over the previous warehouse implementation.

Here are a few examples of why we chose Hadoop and Hive:

1. Ability to handle structured and unstructured data

2. Real-time streaming of data into HDFS from Flume, Scribe, or MountableHDFS

3. Extend functionality through UDFs

4. A well-documented, SQL-like interface specifically built for OLAP versus OLTP

What Hardware Do We Use for Hive?

Dell R410, 4 × 2 TB drives with 24 GB RAM for the data nodes, and Dell R610, 2 × 146

GB (RAID 10) drives with 24 GB RAM for the management hardware.

What’s in Hive?

The primary goal of Hive at Photobucket is to provide answers about business func-

tions, system performance, and user activity. To meet these needs, we store nightly

dumps of MySQL data sets from across hundreds of servers, terabytes of logfiles from

web servers and custom log formats ingested through Flume. This data helps support

many groups throughout the company, such as executive management, advertising,

customer support, product development, and operations just to name a few. For his-

torical data, we keep the partition of all data created on the first day of the month for

MySQL data and 30+ days of log files. Photobucket uses a custom ETL framework for

migrating MySQL data into Hive. Log file data is streamed into HDFS using Flume and

picked up by scheduled Hive processes.

Who Does It Support?

Executive management relies on Hadoop to provide reports surrounding the general

health of the business. Hive allows us to parse structured database data and unstruc-

tured click stream data and distill the data into a format requested by the business

stakeholder.

Photobucket | 293

Advertising operations uses Hive to sift through historical data for forecast and define

quotas for ad targeting.

Product development is far and away the group generating the largest number of ad

hoc queries. As with any user base, segments change and evolve over time. Hive is

important because it allows us to run A/B tests across current and historical data to

gauge relevancy of new products in a quickly changing user environment.

Providing our users with a best-in-class system is the most important goal at Photo-

bucket. From an operations perspective, Hive is used to generate rollup data partitioned

across multiple dimensions. Knowing the most popular media, users, and referring

domains is important for many levels across the company. Controlling expenses is

important to any organization. A single user can quickly consume large amounts of

system resources, significantly increasing monthly expenditures. Hive is used to iden-

tify and analyze rogue users; to determine which ones are within our Terms of Service

and which are not. Operations also uses Hive to run A/B tests defining new hardware

requirements and generating ROI calculations. Hive’s ability to abstract users from

underlying MapReduce code means questions can be answered in hours or days instead

of weeks.

SimpleReach

by Eric Lubow

At SimpleReach, we use Cassandra to store our raw data from all of our social network

polling. The format of the row key is an account ID (which is a MongoDB ObjectId)

and a content item ID (witha MD5 hash of the URL of the content item being tracked)

separated by an underscore which we split on to provide that data in the result set. The

columns in the row are composite columns that look like the ones below:

4e87f81ca782f3404200000a_8c825814de0ac34bb9103e2193a5b824

=> (column=meta:published-at, value=1330979750000, timestamp=1338919372934628)

=> (column=hour:1338876000000_digg-diggs, value=84, timestamp=1338879756209142)

=> (column=hour:1338865200000_googleplus-total, value=12, timestamp=1338869007737888)

In order for us to be able to query on composite columns, we need to know the hex

value of the column name. In our case, we want to know the hex value of the column

name (meta:'published-at').

The hex equivalent is below: 00046D65746100000C7075626C69736865642D617400 =

meta:published-at

Once the column name is converted to hexadecimal format, Hive queries are run

against it. The first part of the query is the LEFT SEMI JOIN, which is used to mimic a

SQL subselect. All the references to SUBSTR and INSTR are to handle the case of composite

columns. Since it is known in advance that characters 10–23 of the “hour:*” columns

(i.e., SUBSTR(r.column_name,10,13)) is a timestamp and therefore we can crop it out and

use it in the returned data or for matching. The INSTR is used to match column names

294 | Chapter 23: Case Studies

and ensure the result set always has the same columns in the same place in the output.

The SUBSTR is used for matching as part of the Ruby function. The SUBSTR returns a

timestamp (long) in milliseconds since epoch and the start_date and end_date are also

a timestamp in milliseconds since epoch. This means that the passed in values can be

matched to a part of the column name.

The goal of this query is to export our data from Cassandra into a CSV file to give

aggregated data dumps to our publishers. It is done via a Resque (offline) job that is

kicked off through our Rails stack. Having a full CSV file means that all columns in the

header must be accounted for in the Hive query (meaning that zeros need to be put to

fill places where there is no data). We do that by pivoting our wide rows into fixed

column tables using the CASE statement.

Here is the HiveQL for the CSV file:

SELECT CAST(SUBSTR(r.column_name, 10, 13) AS BIGINT) AS epoch,

SPLIT(r.row_key, '_')[0] AS account_id,

SPLIT(r.row_key, '_')[1] AS id,

SUM(CAST(CASE WHEN INSTR(r.column_name, 'pageviews-total') > 0

THEN r.value ELSE '0' END AS INT)) AS pageviews,

SUM(CAST(CASE WHEN INSTR(r.column_name, 'digg-digg') > 0

THEN r.value ELSE '0' END AS INT)) AS digg,

SUM(CAST(CASE WHEN INSTR(r.column_name, 'digg-referrer') > 0

THEN r.value ELSE '0' END AS INT)) AS digg_ref,

SUM(CAST(CASE WHEN INSTR(r.column_name, 'delicious-total') > 0

THEN r.value ELSE '0' END AS INT)) AS delicious,

SUM(CAST(CASE WHEN INSTR(r.column_name, 'delicious-referrer') > 0

THEN r.value ELSE '0' END AS INT)) AS delicious_ref,

SUM(CAST(CASE WHEN INSTR(r.column_name, 'googleplus-total') > 0

THEN r.value ELSE '0' END AS INT)) AS google_plus,

SUM(CAST(CASE WHEN INSTR(r.column_name, 'googleplus-referrer') > 0

THEN r.value ELSE '0' END AS INT)) AS google_plus_ref,

SUM(CAST(CASE WHEN INSTR(r.column_name, 'facebook-total') > 0

THEN r.value ELSE '0' END AS INT)) AS fb_total,

SUM(CAST(CASE WHEN INSTR(r.column_name, 'facebook-referrer') > 0

THEN r.value ELSE '0' END AS INT)) AS fb_ref,

SUM(CAST(CASE WHEN INSTR(r.column_name, 'twitter-tweet') > 0

THEN r.value ELSE '0' END AS INT)) AS tweets,

SUM(CAST(CASE WHEN INSTR(r.column_name, 'twitter-referrer') > 0

THEN r.value ELSE '0' END AS INT)) AS twitter_ref,

SUM(CAST(CASE WHEN INSTR(r.column_name, 'linkedin-share') > 0

THEN r.value ELSE '0' END AS INT)) AS linkedin,

SUM(CAST(CASE WHEN INSTR(r.column_name, 'linkedin-referrer') > 0

THEN r.value ELSE '0' END AS INT)) AS linkedin_ref,

SUM(CAST(CASE WHEN INSTR(r.column_name, 'stumbleupon-total') > 0

THEN r.value ELSE '0' END AS INT)) AS stumble_total,

SUM(CAST(CASE WHEN INSTR(r.column_name, 'stumbleupon-referrer') > 0

THEN r.value ELSE '0' END AS INT)) AS stumble_ref,

SUM(CAST(CASE WHEN INSTR(r.column_name, 'social-actions') > 0

THEN r.value ELSE '0' END AS INT)) AS social_actions,

SUM(CAST(CASE WHEN INSTR(r.column_name, 'referrer-social') > 0

THEN r.value ELSE '0' END AS INT)) AS social_ref,

MAX(CAST(CASE WHEN INSTR(r.column_name, 'score-realtime') > 0

SimpleReach | 295

THEN r.value ELSE '0.0' END AS DOUBLE)) AS score_rt

FROM content_social_delta r

LEFT SEMI JOIN (SELECT row_key

FROM content

WHERE HEX(column_name) = '00046D65746100000C7075626C69736865642D617400'

AND CAST(value AS BIGINT) >= #{start_date}

AND CAST(value AS BIGINT) <= #{end_date}

) c ON c.row_key = SPLIT(r.row_key, '_')[1]

WHERE INSTR(r.column_name, 'hour') > 0

AND CAST(SUBSTR(r.column_name, 10, 13) AS BIGINT) >= #{start_date}

AND CAST(SUBSTR(r.column_name, 10, 13) AS BIGINT) <= #{end_date}

GROUP BY CAST(SUBSTR(r.column_name, 10, 13) AS BIGINT),

SPLIT(r.row_key, '_')[0],

SPLIT(r.row_key, '_')[1]

The output of the query is a comma-separated value (CSV) file, an example of which

is below (wrapped for length with a blank line between each record for clarity):

epoch,account_id,id,pageviews,digg,digg_ref,delicious,delicious_ref,

google_plus,google_plus_ref,fb_total,fb_ref,tweets,twitter_ref,

linkedin,linkedin_ref,stumble_total,stumble_ref,social_actions,social_ref,score_rt

1337212800000,4eb331eea782f32acc000002,eaff81bd10a527f589f45c186662230e,

39,0,0,0,0,0,0,0,2,0,20,0,0,0,0,0,22,0

1337212800000,4f63ae61a782f327ce000007,940fd3e9d794b80012d3c7913b837dff,

101,0,0,0,0,0,0,44,63,11,16,0,0,0,0,55,79,69.64308064

1337212800000,4f6baedda782f325f4000010,e70f7d432ad252be439bc9cf1925ad7c,

260,0,0,0,0,0,0,8,25,15,34,0,0,0,0,23,59,57.23718477

1337216400000,4eb331eea782f32acc000002,eaff81bd10a527f589f45c186662230e,

280,0,0,0,0,0,0,37,162,23,15,0,0,0,2,56,179,72.45877173

1337216400000,4ebd76f7a782f30c9b000014,fb8935034e7d365e88dd5be1ed44b6dd,

11,0,0,0,0,0,0,0,1,1,4,0,0,0,0,0,5,29.74849901

Experiences and Needs from the Customer Trenches

A Karmasphere Perspective

By Nanda Vijaydev

Introduction

For over 18 months, Karmasphere has been engaged with a fast-growing number of

companies who adopted Hadoop and immediately gravitated towards Hive as the op-

timal way for teams of analysts and business users to use existing SQL skills with the

Hadoop environment. The first part of this chapter provides use case techniques that

we’ve seen used repeatedly in customer environments to advance Hive-based analytics.

296 | Chapter 23: Case Studies

The use case examples we cover are:

• Optimal data formatting for Hive

• Partitions and performance

• Text analytics with Hive functions including Regex, Explode and Ngram

As companies we’ve worked with plan for and move into production use of Hive, they

look for incremental capabilities that make Hive-based access to Hadoop even easier

to use, more productive, more powerful, and available to more people in their organi-

zation. When they wire Hadoop and Hive into their existing data architectures, they

also want to enable results from Hive queries to be systematized, shared and integrated

with other data stores, spreadsheets, BI tools, and reporting systems.

In particular, companies have asked for:

• Easier ways to ingest data, detect raw formats, and create metadata

• Work collaboratively in an integrated, multi-user environment

• Explore and analyze data iteratively

• Preserved and reusable paths to insight

• Finer-grain control over data, table, and column security, and compartmentalized

access to different lines of business

• Business user access to analytics without requiring SQL skills

• Scheduling of queries for automated result generation and export to non-Hadoop

data stores

• Integration with Microsoft Excel, Tableau, Spotfire, and other spreadsheet, re-

porting systems, dashboards, and BI tools

• Ability to manage Hive-based assets including queries, results, visualizations, and

standard Hive components such as UDFs and SerDes

Use Case Examples from the Customer Trenches

Customer trenches #1: Optimal data formatting for Hive

One recurring question from many Hive users revolves around the format of their data

and how to make that available in Hive.

Many data formats are supported out-of-the-box in Hive but some custom proprietary

formats are not. And some formats that are supported raise questions for Hive users

about how to extract individual components from within a row of data. Sometimes,

writing a standard Hive SerDe that supports a custom data format is the optimal ap-

proach. In other cases, using existing Hive delimiters and exploiting Hive UDFs is the

most convenient solution. One representative case we worked on was with a company

using Hadoop and Hive to provide personalization services from the analysis of mul-

tiple input data streams. They were receiving logfiles from one of their data providers

Experiences and Needs from the Customer Trenches | 297

in a format that could not easily be split into columns. They were trying to figure out

a way to parse the data and run queries without writing a custom SerDe.

The data had top header level information and multiple detailed information. The de-

tailed section was a JSON nested within the top level object, similar to the data set

below:

{ "top" : [

{"table":"user",

"data":{

"name":"John Doe","userid":"2036586","age":"74","code":"297994","status":1}},

{"table":"user",

"data":{

"name":"Mary Ann","userid":"14294734","age":"64","code":"142798","status":1}},

{"table":"user",

"data":{

"name":"Carl Smith","userid":"13998600","age":"36","code":"32866","status":1}},

{"table":"user",

"data":{

"name":"Anil Kumar":"2614012","age":"69","code":"208672","status":1}},

{"table":"user",

"data":{

"name":"Kim Lee","userid":"10471190","age":"53","code":"79365","status":1}}

]}

After talking with the customer, we realized they were interested in splitting individual

columns of the detailed information that was tagged with “data” identifier in the above

sample.

To help them proceed, we used existing Hive function get_json_object as shown

below:

First step is to create a table using the sample data:

CREATE TABLE user (line string)

ROW FORMAT DELIMITED FIELDS TERMINATED BY '\n'

STORED AS TEXTFILE

LOCATION ‘hdfs://hostname/user/uname/tablefolder/’

Then using Hive functions such as get_json_object, we could get to the nested JSON

element and parse it using UDFs:

SELECT get_json_object(col0, '$.name') as name, get_json_object(col0, '$.userid') as uid,

get_json_object(col0, '$.age') as age, get_json_object(col0, '$.code') as code,

get_json_object(col0, '$.status') as status

FROM

(SELECT get_json_object(user.line, '$.data') as col0

FROM user

WHERE get_json_object(user.line, '$.data') is not null) temp;

Query details include:

• Extract the nested JSON object identified by data in the inner query as col0.

• Then the JSON object is split into appropriate columns using their names in the

name value pair.

298 | Chapter 23: Case Studies

The results of the query are given below, with header information saved, as a CSV file:

"name","uid","age","code","status"

"John Doe","2036586","74","297994","1"

"Mary Ann","14294734","64","142798","1"

"Carl Smith","13998600","36","32866","1"

"Kim Lee","10471190","53","79365","1"

Customer trenches #2: Partitions and performance

Using partitions with data being streamed or regularly added to Hadoop is a use case

we see repeatedly, and a powerful and valuable way of harnessing Hadoop and Hive

to analyze various kinds of rapidly additive data sets. Web, application, product, and

sensor logs are just some of the types of data that Hive users often want to perform ad

hoc, repeated, and scheduled queries on.

Hive partitions, when set up correctly, allow users to query data only in specific parti-

tions and hence improves performance significantly. To set up partitions for a table,

files should be located in directories as given in this example:

hdfs://user/uname/folder/"yr"=2012/"mon"=01/"day"=01/file1, file2, file3

/"yr"=2012/"mon"=01/"day"=02/file4, file5

…......

/"yr"=2012/"mon"=05/"day"=30/file100, file101

With the above structure, tables can be set up with partition by year, month, and day.

Queries can use yr, mon, and day as columns and restrict the data accessed to specific

values during query time. If you notice the folder names, partitioned folders have iden-

tifiers such as yr= , mon=, and day=.

Working with one high tech company, we discovered that their folders did not have

this explicit partition naming and they couldn’t change their existing directory struc-

ture. But they still wanted to benefit from having partitions. Their sample directory

structure is given below:

hdfs://user/uname/folder/2012/01/01/file1, file2, file3

/2012/01/02/file4, file5

…….

/2012/05/30/file100, file101

In this case, we can still add partitions by explicitly adding the location of the absolute

path to the table using ALTER TABLE statements. A simple external script can read the

directory and add the literal yr=, mon=, day= to an ALTER TABLE statement and provide

the value of the folder (yr=2012, mon=01,...) to ALTER TABLE statements. The output

of the script is a set of Hive SQL statements generated using the existing directory

structure and captured into a simple text file.

ALTER TABLE tablename

ADD PARTITION (yr=2012, mon=01, day=01) location '/user/uname/folder/2012/01/01/';

ALTER TABLE tablename

ADD PARTITION (yr=2012, mon=01, day=02) location '/user/uname/folder/2012/01/02/';

Experiences and Needs from the Customer Trenches | 299

...

ALTER TABLE tablename

ADD PARTITION (yr=2012, mon=05, day=30) location '/user/uname/folder/2012/05/30/';

When these statements are executed in Hive, the data in the specified directories au-

tomatically become available under defined logical partitions created using ALTER

TABLE statements.

You should make sure that your table is created with PARTITIONED BY

columns for year, month, and day.

Customer trenches #3: Text analytics with Regex, Lateral View Explode, Ngram, and other UDFs

Many companies we work with have text analytics use cases which vary from simple

to complex. Understanding and using Hive regex functions, n-gram functions and other

string functions can address a number of those use cases.

One large manufacturing customer we worked with had lot of machine-generated

compressed text data being ingested into Hadoop. The format of this data was:

1. Multiple rows of data in each file and a number of such files in time-partitioned

buckets.

2. Within each row there were a number of segments separated by /r/n (carriage

return and line feed).

3. Each segment was in the form of a “name: value” pair.

The use case requirement was to:

1. Read each row and separate individual segments as name-value pairs.

2. Zero in on specific segments and look for word counts and word patterns for an-

alyzing keywords and specific messages.

The sample below illustrates this customer’s data (text elided for space):

name:Mercury\r\ndescription:Mercury is the god of commerce, ...\r\ntype:Rocky planet

name:Venus\r\ndescription:Venus is the goddess of love...\r\ntype:Rocky planet

name:Earch\r\ndescription:Earth is the only planet ...\r\ntype:Rocky planet

name:Mars\r\ndescription: Mars is the god of War...\r\ntype:Rocky planet

name:Jupiter\r\ndescription:Jupiter is the King of the Gods...\r\ntype:Gas planet

name:Saturn\r\ndescription:Saturn is the god of agriculture...\r\ntype:Gas planet

name:Uranus\r\ndescription:Uranus is the God of the Heavens...\r\ntype:Gas planet

name:Neptune\r\ndescription:Neptune was the god of the Sea...\r\ntype:Gas planet

The data contains:

1. Planet names and their description with type.

2. Each row of data is separated by a delimiter.

300 | Chapter 23: Case Studies

3. Within each row there are three subsections, including “name,” “description,” and

“type” separated by /r/n.

4. Description is a large text.

First step is to create the initial table with this sample data:

CREATE TABLE planets (col0 string)

ROW FORMAT DELIMITED FIELDS TERMINATED BY '\n'

STORED AS TEXTFILE

LOCATION 'hdfs://hostname/user/uname/planets/'

In the following, we run a series of queries, starting with a simple query and adding

functions as we iterate. Note that the requirement can be met with queries written in

several different ways. The purpose of the queries shown below is to demonstrate some

of the key capabilities in Hive around text parsing.

First, we use a split function to separate each section of data into an array of individual

elements:

SELECT split(col0, '(\\\\r\\\\n)') AS splits FROM planets;

Next, we explode the splits (array) into individual lines using the LATERAL VIEW

EXPLODE function. Results of this query will have name-value pairs separated into indi-

vidual rows. We select only those rows that start with description. The function

LTRIM is also used to remove left spaces.

SELECT ltrim(splits) AS pairs FROM planets

LATERAL VIEW EXPLODE(split(col0, '(\\\\r\\\\n)')) col0 AS splits

WHERE ltrim(splits) LIKE 'desc%'

Now we separate the description line into name-value pair and select only the value

data. This can be done in different ways. We use split by : and choose the value pa-

rameter:

SELECT (split(pairs, ':'))[1] AS txtval FROM (

SELECT ltrim(splits) AS pairs FROM planets

LATERAL VIEW EXPLODE(split(col0, '(\\\\r\\\\n)')) col0 AS splits

WHERE ltrim(splits) LIKE 'desc%')tmp1;

Notice the use of temporary identifiers tmp1 for the inner query. This is required when

you use the output of a subquery as the input to outer query. At the end of step three,

we have the value of the description segment within each row.

In the next step, we use ngrams to show the top 10 bigrams (2-gram) words from the

description of planets. You could also use functions such as context_ngram,

find_in_set, regex_replace, and others to perform various text-based analyses:

SELECT ngrams(sentences(lower(txtval)), 2, 10) AS bigrams FROM (

SELECT (split(pairs, ':'))[1] AS txtval FROM (

SELECT ltrim(splits) AS pairs FROM planets

LATERAL VIEW EXPLODE(split(col0, '(\\\\r\\\\n)')) col0 AS splits

WHERE ltrim(splits) LIKE 'desc%') tmp1) tmp2;

Experiences and Needs from the Customer Trenches | 301

Downloa d f r o m W o w ! e B o o k < w w w.woweb o o k . c o m >

Notice that we have used functions such as lower to convert to lowercase and sentences

to tokenize each word in the text.

For additional information about the text analytics capabilities of Hive, see the func-

tions listed in Chapter 3.

Hive adoption continues to grow,

as outlined by the use cases defined above. Companies across different industry seg-

ments and various sizes have benefited immensely by leveraging Hive in their Hadoop

environments. A strong and active community of contributors and significant invest-

ments in Hive R&D efforts by leading Hadoop vendors ensures that Hive, already the

SQL-based standard for Hadoop, will become the SQL-based standard within organ-

izations that are leveraging Hadoop for Big Data analysis.

As companies invest significant resources and time in understanding and building Hive

resources, in many cases we find they look for additional capabilities that enable them

to build on their initial use of Hive and extend its reach faster and more broadly within

their organizations. From working with these customers looking to take Hive to the

next level, a common set of requirements have emerged.

These requirements include:

Collaborative multiuser environments

Hadoop enables new classes of analysis that were prohibitive computationally and

economically with traditional RDBMS technologies. Hadoop empowers organizations

to break down the data and people silos, performing analysis on every byte of data they

can get their hands on, doing this all in a way that enables them to share their queries,

results, and insights with other individuals, teams, and systems in the organization.

This model implies that users with deep understanding of these different data sets need

to collaborate in discovery, in the sharing of insights, and the availability of all Hive-

based analytic assets across the organization.

Productivity enhancements

The current implementation of Hive offers a serial batch environment on Hadoop to

run queries. This implies that once a user submits a query for job execution to the

Hadoop cluster, they have to wait for the query to complete execution before they can

execute another query against the cluster. This can limit user productivity.

One major reason for companies adopting Hive is that it enables their SQL-skilled data

professionals to move faster and more easily to working with Hadoop. These users are

usually familiar with graphical SQL editors in tools and BI products. They are looking

for similar productivity enhancements like syntax highlighting and code completion.

Managing Hive assets

A recent McKinsey report predicted significant shortage of skilled workers to enable

organizations to profit from their data. Technologies like Hive promise to help bridge

Apache Hive in production: Incremental needs and capabilities.

302 | Chapter 23: Case Studies

that skills shortage by allowing people with an SQL skillset to perform analysis on

Hadoop. However, organizations are realizing that just having Hive available to their

users is not enough. They need to be able to manage Hive assets like queries (history

and versions), UDFs, SerDes for later share and reuse. Organizations would like to

build this living knowledge repository of Hive assets that is easily searchable by users.

Extending Hive for advanced analytics

Many companies are looking to re-create analysis they perform in the traditional

RDBMS world in Hadoop. While not all capabilities in the SQL environment easily

translate into Hive functions, due to inherent limitations of how data is stored, there

are some advanced analytics functions like RANKING, etc., that are Hadoop-able. In ad-

dition, organizations have spent tremendous resources and time in building analytical

models using traditional tools like SAS and SPSS and would like the ability to score

these models on Hadoop via Hive queries.

Extending Hive beyond the SQL skill set

As Hadoop is gaining momentum in organizations and becoming a key fabric of data

processing and analytics within IT infrastructure, it is gaining popularity amongst users

with different skill sets and capabilities. While Hive is easily adopted by users with SQL

skill sets, other less SQL savvy users are also looking for drag-and-drop capabilities like

those available in traditional BI tools to perform analysis on Hadoop using Hive. The

ability to support interactive forms on top of Hive, where a user is prompted to provide

column values via simple web-based forms is an often-asked for capability.

Data exploration capabilities

Traditional database technologies provide data exploration capabilities; for example,

a user can view min, max values for an integer column. In addition, users can also view

visualizations of these columns to understand the data distribution before they perform

analysis on the data. As Hadoop stores hundreds of terabytes of data, and often peta-

bytes, similar capabilities are being requested by customers for specific use cases.

Schedule and operationalize Hive queries

As companies find insights using Hive on Hadoop, they are also looking to operation-

alize these insights and schedule them to run on a regular interval. While open source

alternatives are currently available, these sometimes fall short when companies also

want to manage the output of Hive queries; for example, moving result sets into a

traditional RDBMS system or BI stack. To manage certain use cases, companies often

have to manually string together various different open source tools or rely on poor

performing JDBC connectors.

Karmasphere is a software company, based in Silicon Valley Califor-

nia, focused exclusively on bringing native Hadoop Big Data Analytics capabilities to

teams of analysts and business users. Their flagship product, Karmasphere 2.0, is based

on Apache Hive, extending it in a multi-user graphical workspace to enable:

About Karmasphere.

Experiences and Needs from the Customer Trenches | 303

• Reuse of standard Hive-based tables, SerDes and UDFs

• Social, project-based big data analytics for teams of analysts and business users

• Easy data ingestion to the cluster

• Heuristic-based recognition and table creation of many popular data formats

• Visual and iterative data exploration and analysis

• Graphical exploration of all Hive-based analytic assets

• Sharing and scheduling of queries, results and visualizations

• Easy integration with traditional spreadsheets, reporting, dashboard, and BI tools

Figure 23-4 shows a screenshot of Karmasphere 2.0’s Hive-based Big Data Analytics

Environment.

Figure 23-4. Screenshot of Karmasphere 2.0

Hive features survey

We’d like to get feedback on the importance of these needs and share them back with

the growing Hive community. If you are interested in seeing what others think and

would like to participate, please visit:

http://karmasphere.com/hive-features-survey.html

304 | Chapter 23: Case Studies

Glossary

Amazon Elastic MapReduce

Amazon’s EMR is a hosted Hadoop service

on top of Amazon EC2 (Elastic Compute

Cloud).

Avro

Avro is a new serialization format developed

to address some of the common problems

associated with evolving other serialization

formats. Some of the benefits are: rich data

structures, fast binary format, support for

remote procedure calls, and built-in schema

evolution.

Bash

The “Bourne-Again Shell” that is the default

interactive command shell for Linux and

Mac OS X systems.

S3 Bucket

The term for the top-level container you

own and manage when using S3. A user may

have many buckets, analogous to the root of

a physical hard drive.

Command-Line Interface

The command-line interface (CLI) can run

“scripts” of Hive statements or all the user

to enter statements interactively.

Data Warehouse

A repository of structured data suitable for

analysis for reports, trends, etc. Warehouses

are batch mode or offline, as opposed to

providing real-time responsiveness for on-

line activity, like ecommerce.

Derby

A lightweight SQL database that can be em-

bedded in Java applications. It runs in the

same process and saves its data to local files.

It is used as the default SQL data store for

Hive’s metastore. See http://db.apache.org/

derby/ for more information.

Dynamic Partitions

A HiveQL extension to SQL that allows you

to insert query results into table partitions

where you leave one or more partition col-

umn values unspecified and they are deter-

mined dynamically from the query results

themselves. This technique is convenient for

partitioning a query result into a potentially

large number of partitions in a new table,

without having to write a separate query for

each partition column value.

Ephemeral Storage

In the nodes for a virtual Amazon EC2 clus-

ter, the on-node disk storage is called

ephemeral because it will vanish when the

cluster is shut down, in contrast to a physi-

cal cluster that is shut down. Hence, when

using an EC2 cluster, such as an Amazon

Elastic MapReduce cluster, it is important to

back up important data to S3.

ExternalTable

A table using a storage location and contents

that are outside of Hive’s control. It is con-

venient for sharing data with other tools, but

it is up to other processes to manage the life

cycle of the data. That is, when an external

table is created, Hive does not create the

305

external directory (or directories for parti-

tioned tables), nor are the directory and data

files deleted when an external table is

dropped.

Hadoop Distributed File System

(HDFS) A distributed, resilient file system

for data storage that is optimized for scan-

ning large contiguous blocks of data on hard

disks. Distribution across a cluster provides

horizontal scaling of data storage. Blocks of

HDFS files are replicated across the cluster

(by default, three times) to prevent data loss

when hard drives or whole servers fail.

HBase

The NoSQL database that uses HDFS for

durable storage of table data. HBase is a col-

umn-oriented, key-value store designed to

provide traditional responsiveness for quer-

ies and row-level updates and insertions.

Column oriented means that the data storage

is organized on disk by groups of columns,

called column families, rather than by row.

This feature promotes fast queries for sub-

sets of columns. Key-value means that rows

are stored and fetched by a unique key and

the value is the entire row. HBase does not

provide an SQL dialect, but Hive can be

used to query HBase tables.

Hive

A data warehouse tool that provides table

abstractions on top of data resident in

HDFS, HBase tables, and other stores. The

Hive Query Language is a dialect of the

Structured Query Language.

Hive Query Language

Hive’s own dialect of the Structured Query

Language (SQL). Abbreviated HiveQL or

HQL.

Input Format

The input format determines how input

streams, usually from files, are split into re-

cords. A SerDe handles parsing the record

into columns. A custom input format can be

specified when creating a table using the

INPUTFORMAT clause. The input format for the

default STORED AS TEXTFILE specification is

implemented by the Java object named

org.apache.hadoop.mapreduce.lib.input.

TextInputFormat. See also Output Format.

JDBC

The Java Database Connection API pro-

vides access to SQL systems, including Hive,

from Java code.

Job

In the Hadoop context, a job is a self-con-

tained workflow submitted to MapReduce.

It encompasses all the work required to per-

form a complete calculation, from reading

input to generating output. The MapRe-

duce JobTracker will decompose the job into

one or more tasks for distribution and exe-

cution around the cluster.

JobTracker

The top-level controller of all jobs using

Hadoop’s MapReduce. The JobTracker ac-

cepts job submissions, determines what

tasks to run and where to run them, moni-

tors their execution, restarts failed tasks as

needed, and provides a web console for

monitoring job and task execution, viewing

logs, etc.

Job Flow

A term used in Amazon Elastic MapReduce

(EMR) for the sequence of jobs executed on

a temporary EMR cluster to accomplish a

particular goal.

JSON

JSON (JavaScript Object Notation) is a

lightweight data serialization format used

commonly in web-based applications.

Map

The mapping phase of a MapReduce process

where an input set of key-value pairs are

converted into a new set of key-value pairs.

For each input key-value pair, there can be

zero-to-many output key-value pairs. The

input and output keys and the input and

output values can be completely different.

MapR

A commercial distribution of Hadoop that

replaces HDFS with the MapR File System

Hadoop Distributed File System

306 | Glossary

(MapR-FS), a high-performance, dis-

tributed file system.

MapReduce

A computation paradigm invented at

Google and based loosely on the common

data operations of mapping a collection of

data elements from one form to another (the

map phase) and reducing a collection to a

single value or a smaller collection (the

reduce phase). MapReduce is designed to

scale computation horizontally by decom-

posing map and reduce steps into tasks and

distributing those tasks across a cluster. The

MapReduce runtime provided by Hadoop

handles decomposition of a job into tasks,

distribution around the cluster, movement

of a particular task to the machine that holds

the data for the task, movement of data to

tasks (as needed), and automated re-

execution of failed tasks and other error

recovery and logging services.

Metastore

The service that maintains “metadata” in-

formation, such as table schemas. Hive re-

quires this service to be running. By default,

it uses a built-in Derby SQL server, which

provides limited, single-process SQL sup-

port. Production systems must use a full-

service relational database, such as MySQL.

NoSQL

An umbrella term for data stores that don’t

support the relational model for data man-

agement, dialects of the structured query

language, and features like transactional up-

dates. These data stores trade off these fea-

tures in order to provide more cost-effective

scalability, higher availability, etc.

ODBC

The Open Database Connection API pro-

vides access to SQL systems, including Hive,

from other applications. Java applications

typically use the JDBC API, instead.

Output Format

The output format determines how records

are written to output streams, usually to

files. A SerDe handles serialization of each

record into an appropriate byte stream. A

custom output format can be specified when

creating a table using the OUTPUTFORMAT

clause. The output format for the default

STORED AS TEXTFILE specification is im-

plemented by the Java object named org

.apache.hadoop.hive.ql.io.HiveIgnoreKey

TextOutputFormat. See also Input Format.

Partition

A subset of a table’s data set where one col-

umn has the same value for all records in the

subset. In Hive, as in most databases that

support partitioning, each partition is

stored in a physically separate location—in

Hive’s case, in a subdirectory of the root di-

rectory for the table. Partitions have several

advantages. The column value correspond-

ing to a partition doesn’t have to be repeated

in every record in the partition, saving space,

and queries with WHERE clauses that restrict

the result set to specific values for the parti-

tion columns can perform more quickly, be-

cause they avoid scanning the directories of

nonmatching partition values. See also dy-

namic partitions.

Reduce

The reduction phase of a MapReduce pro-

cess where the key-value pairs from the

map phase are processed. A crucial feature

of MapReduce is that all the key-value pairs

from all the map tasks that have the same

key will be sent together to the same reduce

task, so that the collection of values can be

“reduced” as appropriate. For example, a

collection of integers might be added or

averaged together, a collection of strings

might have all duplicates removed, etc.

Relational Model

The most common model for database man-

agement systems, it is based on a logical

model of data organization and manipula-

tion. A declarative specification of the data

structure and how it should be manipulated

is supplied by the user, most typically using

the Structured Query Language. The imple-

mentation translates these declarations into

Relational Model

Glossary | 307

procedures for storing, retrieving, and ma-

nipulating the data.

The distributed file system for Amazon Web

Services. It can be used with or instead of

HDFS when running MapReduce jobs.

SerDe

The Serializer/Deserializer or SerDe for

short is used to parse the bytes of a record

into columns or fields, the deserialization

process. It is also used to create those record

bytes (i.e., serialization). In contrast, the In-

put Format is used to split an input stream

into records and the Output Format is used

to write records to an output stream. A

SerDe can be specified when a Hive table is

created. The default SerDe supports the

field and collection separators discussed in

“Text File Encoding of Data Val-

ues” on page 45, as well as various opti-

mizations such as a lazy parsing.

Structured Query Language

A language that implements the relational

model for querying and manipulating data.

Abbreviated SQL. While there is an ANSI

standard for SQL that has undergone peri-

odic revisions, all SQL dialects in wide-

spread use add their own custom extensions

and variations.

Task

In the MapReduce context, a task is the

smallest unit of work performed on a single

cluster node, as part of an overall job. By de-

fault each task involves a separate JVM pro-

cess. Each map and reduce invocation will

have its own task.

Thrift

An RPC system invented by Facebook and

integrated into Hive. Remote processes can

send Hive statements to Hive through

Thrift.

User-Defined Aggregate Functions

User-defined functions that take multiple

rows (or columns from multiple rows) and

return a single “aggregation” of the data,

such as a count of the rows, a sum or average

of number values, etc. The term is abbrevi-

ated UDAF. See also user-defined functions

and user-defined table generating functions.

User-Defined Functions

Functions implemented by users of Hive to

extend their behavior. Sometimes the term

is used generically to include built-in func-

tions and sometimes the term is used for the

specific case of functions that work on a sin-

gle row (or columns in a row) and return a

single value, (i.e., which don’t change the

number of records). Abbreviated UDF. See

also user-defined aggregate functions and

user-defined table generating functions.

User-Defined Table Generating Functions

User-defined functions that take a column

from a single record and expand it into mul-

tiple rows. Examples include the explode

function that converts an array into rows of

single fields and, for Hive v0.8.0 and later,

converts a map into rows of key and value

fields. Abbreviated UDTF. See also user-

defined functions and user-defined aggregate

functions.

308 | Glossary

APPENDIX

References

Amazon Web Services, http://aws.amazon.com/.

Amazon DynamoDB, http://aws.amazon.com/dynamodb/.

Amazon Elastic MapReduce (Amazon EMR), http://aws.amazon.com/elasticmapreduce/.

Amazon Simple Storage Service (S3), http://aws.amazon.com/s3.

Cassandra Database, http://cassandra.apache.org/.

Apache HBase, http://hbase.apache.org/.

Apache Hive, http://hive.apache.org/.

Apache Hive Wiki: https://cwiki.apache.org/Hive/.

Apache Oozie, http://incubator.apache.org/oozie/.

Apache Pig, http://pig.apache.org/.

Apache Zookeeper, http://zookeeper.apache.org/.

Cascading, http://cascading.org.

Data processing on Hadoop without the hassle, https://github.com/nathanmarz/cascalog.

Easy, efficient MapReduce pipelines in Java and Scala, https://github.com/cloudera/

crunch.

Datalog, http://en.wikipedia.org/wiki/Datalog.

C.J. Date, The Relational Database Dictionary, O’Reilly Media, 2006, ISBN

978-0-596-52798-3.

Jeffrey Dean and Sanjay Ghemawat, MapReduce: simplified data processing on large

clusters, Proceeding OSDI ’04 Proceedings of the 6th conference on Symposium on

Operating Systems Design and Implementation - Volume 6, 2004.

Apache Derby, http://db.apache.org/derby/.

309

Jeffrey E.F. Friedl, Mastering Regular Expressions, 3rd Edition, O’Reilly Media, 2006,

ISBN 978-0-596-52812-6.

Alan Gates, Programming Pig, O’Reilly Media, 2011, ISBN 978-1-449-30264-1.

Lars George, HBase: The Definitive Guide, O’Reilly Media, 2011, ISBN

978-1-449-39610-7.

Sanjay Ghemawat, Howard Gobioff, and Shun-Tak Leung, The Google file system,

SOSP ’03 Proceedings of the nineteenth ACM symposium on Operating systems prin-

ciples, 2003.

Jan Goyvaerts and Steven Levithan, Regular Expressions Cookbook, 2nd Edition,

O’Reilly Media, 2009, ISBN 978-1-449-31943-4.

Eben Hewitt, Cassandra: The Definitive Guide, O’Reilly Media, 2010, ISBN

978-1-449-39041-9.

Ashish Thusoo, et al, Hive - a petabyte scale data warehouse using Hadoop, 2010 IEEE

26th International Conference on Data Engineering (ICDE).

JDK 1.6 java.util.regex.Pattern Javadoc, http://docs.oracle.com/javase/6/docs/api/java/

util/regex/Pattern.html.

The Java Tutorials, Lesson: Regular Expressions, http://docs.oracle.com/javase/tutorial/

essential/regex/.

JSON, http://json.org/.

Apache Kafka: A high-throughput, distributed messaging system, http://incubator.apache

.org/kafka/index.html.

Kerberos: The Network Authentication Protocol, http://web.mit.edu/kerberos.

MapR, the Next Generation Distribution for Apache Hadoop, http://mapr.com.

MarkLogic, http://www.marklogic.com/.

Wolfram Mathematica, http://www.wolfram.com/mathematica/.

Matlab: The Language of Technical Computing, http://www.mathworks.com/products/

matlab/index.html.

GNU Octave, http://www.gnu.org/software/octave/.

Oracle XML DB, http://www.oracle.com/technetwork/database/features/xmldb/index

.html.

The R Project for Statistical Computing, http://r-project.org/.

A Scala API for Cascading, https://github.com/twitter/scalding.

SciPy: Scientific Tools for Python, http://scipy.org.

Shark (Hive on Spark), http://shark.cs.berkeley.edu/.

310 | Appendix: References

Spark: Lightning-Fast Cluster Computing, http://www.spark-project.org/.

Storm: Distributed and fault-tolerant realtime computation: stream processing, continu-

ous computation, distributed RPC, and more, https://github.com/nathanmarz/storm.

Tony Stubblebine, Regular Expression Pocket Reference, O’Reilly Media, 2003, ISBN

978-0-596-00415-6.

Dean Wampler, Functional Programming for Java Developers, O’Reilly Media, 2011,

ISBN 978-1-449-31103-2.

Dean Wampler and Alex Payne, Programming Scala, O’Reilly Media, 2009, ISBN

978-0-596-15595-7.

Tom White, Hadoop: The Definitive Guide, 3nd Edition, O’Reilly Media, 2012, ISBN

978-1-449-31152-0.

XPath Specification, http://www.w3.org/TR/xpath/.

References | 311

Index

Symbols

" (single quotes)

regular expressions using, 55

*.domain.pvt, 27

.hiverc script, deploying, 249

; (semicolon)

at end of lines in Hive queries, 36

separating multiple queries, 34

^A (“Control” A) delimiters, 45, 47

AbstractGenericUDAFResolver, methods in,

174

ADD PARTITION clauses, ALTER TABLE ...,

aggregate functions, 85–87, 164–176

aggregate functions, UDF, 172–177

aggregates, calculating with streaming, 191–

192

algorithms (case study), creating machine

learning, 265–270

ALTER DATABASE command, 52

ALTER INDEX clauses, 118–119

ALTER TABLE ... ADD PARTITION clause,

ALTER TABLE ... ARCHIVE PARTITION

clause, 70

ALTER TABLE ... TOUCH clause, in triggering

execution hooks, 69

ALTER TABLE statements

adding, modifying, and dropping partitions,

altering storage properties, 68

changing schemas with, 53, 66

columns

adding, 68

changing, 67

deleting or replacing, 68

renaming tables, 66

table properties, altering, 68

Amazon Web Services (AWS)

DynamoDB, 225

Elastic MapReduce (EMR), 251

about, 245–246, 305

clusters, 246–248

logs on S3, 252

persistence and metastore on, 250–251

security groups for, 253

Thrift Server on, 247

vs. EC2 and Hive, 254

S3 system for

about, 308

accessing, 248

deploying .hiverc script to, 249

EMR logs on, 252

moving data to, 62

MySQL dump on, 251

putting resources on, 252

role on EMR cluster, 251

support for, 62

spot instances, 252–253

annotations, for use with functions, 184

ANSI SQL standard, 2, 49

Apache

Cassandra, 2, 224–225

DynamoDB, 2

Hadoop (see Hadoop)

HBase, 2, 8–9, 222–224

Hive (see Hive)

We’d like to hear your suggestions for improving our indexes. Send email to index@oreilly.com.

313

Downloa d f r o m W o w ! e B o o k < w w w.woweb o o k . c o m >

Incubator, 256

Oozie, 239–244

Zookeeper, 235–237

architecture, HCatalog, 262–263

arithmetic operators, 82

ARRAY

data types, 43, 114

return types for functions, 90

return types on aggregate functions, 86

ARRAY[ ... ], 80–81

atomic rebuilds, 119

authorization, in Hive, 229–234

autocomplete, in Hive CLI, 37

--auxpath option, for Hive services, 30

Avro ContainerOutputFormat, 209

AVRO Hive SerDe, 209

Avro Hive SerDe, 305

Avro schema, 209

^B delimiters, 45, 47

backup, metastore information, 28

bash shell, 305

bash shell commands, running without Hive

CLI, 37–38

Big Data systems, 44

Big Table, Google’s, 8

BIGINT data types

about, 41

as return type for aggregate functions, 85

as return type for mathematical functions,

/bin/sed program, as stream editor, 189

BINARY (v0.8.0+) data types

about, 42

casting, 109

Binary Output, 211

bitmap indexes, 118

BLOCK compression, in sequence files, 149

block sampling, of data, 111

BOOLEAN data types

about, 41

as return type for functions, 88, 90

bucket, S3, 305

bucketing

table data storage, 125–127

tables, input pruning for, 111–112

building Hive from source, 156

built-in functions vs. UDFs, 163

(see also functions)

^C delimiters, 45, 47

Cartesian product JOINs, 105

Cascading, alternative higher-level library for,

Cascalog, alternative higher-level library for, 9

CASE ... WHEN ... THEN clauses, 91–92, 91–

Cassandra, 2, 224–225, 294–296

cast() function, 109

-cat command, 21

character arrays (strings), in Hive, 42

cleanup scratch directory (HiveServer), 218

CLI (Command-Line Interface), Hive

about, 22, 305

autocomplete, 37

comments, in Hive scripts, 38

executing Hive queries from files, 35–36

for Hive services, 30

hadoop dfs commands within, 38

hive command in, 29

Hive web interface (HWI), 7

.hiverc file, 36

options, 31

pasting comments into, 38

prompt, 23

running bash shell commands without, 37–

running queries and exiting Hive, 34–35

scrolling through command history, 37

variables and properties, 31–34

Cloudera, Inc., virtual machines for VMWare,

CLUSTER BY clause

about, 108–109

in altering storage properties, 69

in streaming processes, 192–194

clusters

EMR, 246–248, 251

managing Hive data across multiple

MapReduce (case study), 274–

278

codecs

choosing, 146–147

DefaultCodec, 147

determining installed

SUB2 TEXT, 145

314 | Index

enabling intermediate compression, 147–

148

GZip compression, 148

SnappyCodec, 147

cogroups, calculating, 196

column headers, in Hive queries, 38–39

columnar tables, 128

columns

adding comments to, 53

adding to table, 127

aliases in queries for, 91

bug in using count(DISTINCT col) in with

partitioned, 86

changing, 67

computing with values from, 81–82

dynamic, transposed mapping for, 224

partition keys and, 59

specifying in queries with regular

expressions, 81

static mapping in Cassandra, 224

tuning using virtual, 142–143

Comma-Separated Values (CSVs)

as delimiters, 45, 48

SerDes, 206

command history, scrolling through Hive CLI,

command line tool, HCatalog, 261

Command-Line Interface (CLI) (see CLI

(Command-Line Interface), Hive)

comments, adding to columns, 53

comments, in Hive scripts, 38

compression of data

about, 128–129, 145

codecs

choosing, 146–147

determining installed, 145

enabling intermediate, 147–148

final output, 148

in action, 149–152

sequence file format, 148–149

--config directory option, for Hive services, 30

Core Instance Group, EMR, 247

count(DISTINCT col), bug in using in with

partitioned, 86

counting uniques (case study), 280–282

CREATE INDEX command, 118

CREATE TABLE statements

about, 53–56

demystifying, 199–201

CREATE VIEW command, 113, 115

Crunch, alternative higher-level library for, 9

CSVs (Comma-Separated Values)

as delimiters, 45, 48

SerDes, 206

customer experiences and needs, from

Karmasphere perspective (case

study), 296–304

cut command, 188–189

Cutting, Doug, 3

data

bucketing data storage on tables, 125–127

compression

about, 128–129, 145

choosing codec, 146–147

determining installed codecs, 145

enabling intermediate, 147–148

final output, 148

in action, 149–152

sequence file format, 148–149

denormalizing, 123

exporting, 76–77

inserting data into tables, 73–74

loading, 71–73, 75

making multiple passes over, 124

moving to S3 system, 62

partitioning, 59

queries (see queries, Hive)

queries that sample, 110–112

ranking by grouping like elements (case

study), 270–274

serializing and deserializing, 187

unique keys and normalizing, 123–124

using streaming for distributing and sorting

of, 192–194

views restricting data based on conditions,

114

data flow language, 8

data processing, using Hive on Hadoop for,

255

Data Science at M6D using Hive and R (case

study), 265–270

data types

changing using TRANSFORM(), 188

collection, 43–44

Hive handling wrong, 48

primitive, 41–43

Index | 315

data warehouse, 305

data warehouse applications, Hive databases

in, 2

databases, in HiveQL

about, 49–52

altering, 52

altering storage properties, 68

columns

adding, 68, 127

adding comments to, 53

bug in using count(DISTINCT col) in

with partitioned, 86

changing, 68

computing with values from, 81–82

deleting or replacing, 68

partition keys and, 59

specifying in queries with regular

expressions, 81

compression of data

about, 128–129, 145

choosing codec, 146–147

determining installed codecs, 145

enabling intermediate, 147–148

final output, 148

in action, 149–152

sequence file format, 148–149

creating directories for, 50, 54

indexing, 117–120

schema design, 121–129

setting property to print current, 52

table properties, altering, 68

table storage formats, customizing, 63–65

tables

adding, modifying, and dropping

partitions, 66

bucketing data storage on, 125–127

changing schemas, 53, 66

columnar, 128

copying schema, 54

creating, 53–56

creating and loading data, 75

creating in HBASE, 222–224

indexing, 117–120

input pruning for bucket, 111–112

JOINs, 98–105, 134

listing, 54–55

location of stored, 50

managed (see managed tables)

normalization and unique keys, 123–

124

partition tuning, 140–141

partitioning, 58–63, 74–75, 122–125

renaming, 66

schema design, 121–129

UNION ALL, 112

views and map type for dynamic, 114–

115

default database, directory of, 50

default record delimiters, 45

DefaultCodec, 147

--define key=value option, --hivevar key=value

option and, 31

delimiters (separators)

default field, 42

default record, 45

using, 42

using default, 47

denormalizing data, 123

Derby database, 305

Derby SQL server, Hive using, 22

Derby store, single node in pseudodistributed

mode using, 28

DESCRIBE command, 209

DESCRIBE DATABASE command, 50

DESCRIBE DATABASE EXTENDED

command, 51

DESCRIBE EXTENDED

command, 55, 60

log_messages, 63

table command, 65

DESCRIBE FUNCTION command, 164

DESCRIBE TABLE EXTENDED clause, 200

deserializing and serializing data, 187

Detailed Table Information, 55

deterministic annotations, 184

developing in Hive, 155–161

dfs -ls command, 20

directories

authentication of, 228–229

creating for databases, 50, 54

DISTINCT expressions, in aggregate functions,

86–87

DistinctLike annotations, 185

DISTRIBUTE BY clauses

GROUP BY clause and, 108

in streaming processes, 192, 194

with SORT BY clauses, 107–108

316 | Index

distributed cache

accessing from UDF, 182–184

using, 189–190

distributed data processing tools, not using

MapReduce, 10

distributed filesystems, 4

distributed mode

for Hadoop, 19

Hadoop in, 26–27

DNS names, 27, 51

Domain Specific Languages (DSLs), 9

*.domain.pvt, 27

DOUBLE data types, 43

about, 42

as return type for aggregate functions, 85–

as return type for mathematical functions,

83–85

DROP INDEX command, 119

DROP TABLE command, 66

DROP VIEW command, 116

dropping tables, 66

DSLs (Domain Specific Languages), 9

DualInputFormat, as custom input format,

203–205

dynamic partitions

about, 305

properties of, 75

tuning, 140–141

dynamic tables, views and map type for, 114–

115

DynamoDB, 2, 225

EC2 (Elastic Compute Cloud), 2

Eclipse, open source IDE

setting up, 158

starting Hive Command CLI Driver from

within, 159

Elastic MapReduce (EMR), 254

about, 245–246

clusters, 246–248, 251

logs on S3, 252

persistence and metastore on, 250–251

security groups for, 253

Thrift Server on, 247

vs. EC2 and Hive, 254

elastic-mapreduce Ruby client, for spot

instances, 253

Emacs text editor, 45

EMR (Elastic MapReduce), Amazon

about, 305

EMR API, 246

EMR AWS Management Console (Web-based

front-end), 246

EMR Command-Line Interface, 246

EMR Master node, Jobtracker and NameNode

User Interfaces accessible on, 253

env namespace option, for Hive CLI, 32, 34

Ephemeral Storage, 305

ETL (Extract, Transform, and Load) processes

partitioning tables and, 124–125

Pig as part of, 8, 255

User-Defined Functions in, 163

event stream processing, 10

exclusive and explicit locks, 238

execution hooks, 69

EXPLAIN clauses, tuning using, 131–134

EXPLAIN EXTENDED clauses, tuning using,

134

exporting data, 76–77

external tables

about, 56–57

partitioned, 61–63

ExternalTable, 305

fetchOne() method, 215

field delimiters, 45

file formats

about, 201

custom input, 203–205

HAR (Hadoop ARchive), 152–154

RCFILE format, 202

sequence, 148–149

SEQUENCEFILE, 201–202

text file encoding of data values, 45–48

vs. record formats, 199

files, authentication of, 228–229

filesystem, metadata in NameNode for, 122

final output compression, 148

FLOAT data types, 43

about, 41

as return type for mathematical functions,

gotchas with floating point comparisons,

94–96

Index | 317

Friedl, Jeffrey E.F., Mastering Regular

Expressions, 3rd Edition, 97

FROM clause, 79

full-outer JOINs, 104

functions

aggregate, 85–87, 164–176, 171, 172–177

annotations, for use with, 184

casting, 109

deterministic, 184

mathematical, 83–85

other built-in, 88–90

stateful, 184

table generating, 87–88, 165–166

User-Defined Functions (UDFs), 163

(see also User-Defined Functions

(UDFs))

XPath (XML Path Language), 207–208

Gates, Alan, Programming Pig, 8

Generalized Additive Models (GAM), 267

GenericMR Tools, for streaming to Java, 194–

196

GenericUDAs, 172–177

GenericUDFs vs. UDFs, 169–171

George, Lars, HBase: The Definitive Guide, 9

getClusterStatus method, 215

getQueryPlan() method, 216

getSchema() method, 215

Google Big Table, 8

Google File System, 3

Google Summer of Code project, JSON SerDe

and, 208

Goyvaerts, Jan, Regular Expression Pocket

Reference, 97

graphical interfaces, for interacting with Hive,

Groovy, setting up to connect to HiveService,

214

GROUP BY clause

about, 97

DISTRIBUTE BY clauses and, 108

HAVING clause and, 97

groups, granting and revoking privileges for

individual, 230–233

GZip compression, 148

Hadoop

about, 1

alternative higher-level libraries for, 9–10

alternatives to Hive for, 8–9

compression of data

about, 145

DefaultCodec, 147

SnappyCodec, 147

configuring, 24–29

HAR file format, 152–154

Hive in, 6–8

InputFormat API, 145

installing, 18–19

installing Java for, 16–18

JVM reuse as running parameter, 139–140

launching components of MapReduce for,

20–21

operating systems for, 15

runtime modes for, 19

speculative execution in, 141–142

testing installation of, 20

using Hive for data processing, 255

hadoop dfs commands, defining alias for, 20

Hadoop Distributed File System (HDFS)

about, 1, 4, 306

HBase and, 9

master node of, 51

NameNode and, 7

role on EMR cluster, 251

Sort and Shuffle phase in, 5

Hadoop Java API, implementing algorithms

using, 6

Hadoop JobTracker

getting cluster status from, 215

Hadoop security, 227

Hadoop Streaming API, 187

Hadoop: The Definitive Guide (White), 12, 24

HADOOP_HOME, Hive using, 21

HAR (Hadoop ARchive), 152–154

HAVING clause, 97

HBase, 2, 8–9, 222–224, 306

HBase: The Definitive Guide (George), 9

hcat (command line tool), options supported

by, 261

HCatalog

about, 255–256

architecture, 262–263

command line tool, 261

318 | Index

reading data in MapReduce and, 256–258

writing data in MapReduce and, 258–260

HCatInputFormat

HCatLoader atop, 262–263

reading data in MapReduce and, 256–258

HCatOutputFormat

HCatStorer atop, 262–263

writing data in MapReduce and, 258–260

HDFS (Hadoop Distributed File System)

about, 1, 4, 306

HBase and, 9

master node of, 51

NameNode and, 7

role on EMR cluster, 251

Sort and Shuffle phase in, 5

“Hello World” program, Word Count

algorithm as, 4–5

Hive

about, 306

alternatives to, 1, 8–9

core of binary distribution, 22–23

in Hadoop, 6–8

installing, 21–22

JDK for, 16–17

keys in, 44, 117–120

keywords, 24

limitations of, 2

list of built-in properties, 34

list of US states and territories used to query

in, 59

machine learning algorithms (case study),

using R in creating, 265–270

metastore requirement and Hadoop, 28–29

modules, 7

queries (see queries, Hive)

security, 227–234

starting, 23–24

strict mode, 137

using HADOOP_HOME, 21

Hive and EC2 vs. EMR, 254

Hive CLI (Command-Line Interface)

about, 22, 305

autocomplete, 37

comments, in Hive scripts, 38

executing Hive queries from files, 35–36

for Hive services, 30

hadoop dfs commands within, 38

hive command in, 29

Hive web interface (HWI), 7

.hiverc file, 36

options, 31

pasting comments into, 38

prompt, 23

running bash shell commands without, 37–

running queries and exiting Hive, 34–35

scrolling through command history, 37

variables and properties, 31–34

hive command, 29

Hive Command CLI Driver, starting from

within Eclipse, 159

Hive Thrift Service Action, 240

Hive Web Interface (HWI)

about, 23

as CLI, 7

Hive Wiki link, 3

Hive, in data warehouse applications, 2

hive-site.xml, deploying, 248

hiveconf namespace option, for Hive CLI, 32,

33–35

HiveQL (Hive Query Language)

about, 1, 306

altering storage properties, 68

columns

adding, 68, 127

adding comments to, 53

bug in using count(DISTINCT col) in

with partitioned, 86

changing, 67

computing with values from, 81–82

deleting or replacing, 68

partition keys and, 59

specifying in queries with regular

expressions, 81

databases, 54

about, 49–52

altering, 52

creating directories for, 50

Java vs., 10

queries (see queries, Hive)

SQL and, 2

table properties, altering, 68

table storage formats, customizing, 63–65

tables

adding, modifying, and dropping

partitions, 66

bucketing data storage on, 125–127

changing schemas, 53, 66

Index | 319

columnar, 128

copying schema, 54

creating, 53–56

creating and loading data, 75

creating in HBASE, 222–224

indexing, 117–120

input pruning for bucket, 111–112

JOINs, 98–105, 134

listing, 54–55

location of stored, 50

managed (see managed tables)

normalization and unique keys, 123–

124

partition tuning, 140–141

partitioning, 58–63, 74–75, 122–125

renaming, 65

schema design, 121–129

UNION ALL, 112

views and map type for dynamic, 114–

115

views, 113–116

.hiverc file, 36

HiveServer (HiveThrift)

about, 213

administering, 217–218

connecting to, 214–215

connecting to metastore, 216–217

fetching results, 215

getting cluster status, 215

getting schema of results, 215

on EMR Hive, 247

retrieving query plan, 216

setting up Groovy to connect to HiveService,

214

setting up ThriftMetastore, 219

starting, 213

hiveserver option, for Hive services, 30

HiveService, productionizing, 217–218

HiveService, setting up Groovy to connect to,

214

HiveStorageHandler, 222

--hivevar key=value option, --define key=value

option and, 31

hivevar namespace option, for Hive CLI, 32

hive_test, testing with, 159–161

Hortonworks, Inc., virtual machines for

VMWare, 16

Hue, graphical interface for interacting with

Hive, 6

HWI (Hive Web Interface)

about, 23

as CLI, 7

hwi option, for Hive services, 30

IDE (Integrated Development Environment),

setting up Eclipse open source, 158

identity transformation, using streaming, 188

IF EXISTS clause

in ALTER TABLE statements, 67

in dropping tables, 66

IF NOT EXISTS clause

in ALTER TABLE statements, 66

to CREATE TABLE statement, 53

implementation infrastructure, 4

IN ... EXISTS clauses, 104

IN database_name clause, and regular

expression, 55

Incubator, Apache, 256

indexes, Hive HQL, 117–120

tuning, 140

inner JOINs, 98–100

input

formats, 64, 203–205, 306

structure in MapReduce, 4–5

InputFormat

objects,coding, vs., streaming, 187

InputFormat API, Hadoop, 145

INPUTFORMAT clauses, 200

InputFormat clauses, 256

InputFormats, reading key-value pairs, 201

INSERT statement, in loading data into tables

using, 73

instance groups, on EMR, 247–248

INT data types, 41

as return type for functions, 88, 89, 90

as return type for mathematical functions,

internal tables, 56

INTO ... BUCKETS clause, in altering storage

properties, 69

INTO keyword, with INSERT statement, 73

JAR option, for Hive services, 30

Java

data types implemented in, 42

320 | Index

dotted name syntax in, 64

GenericMR Tools for streaming to, 194–

196

installing for Hadoop, 16–18

Java debugger, connecting to Hive, 156

Java MapReduce API, 10

Java MapReduce programs, 7

Java Virtual Machine (JVM) libraries, 9–10

Java vs. Hive, 10–13

JDBC (Java Database Connection API)

about, 306

compliant databases for using metastore,

JDK (Java Development Kit), for Hive, 16–17

job, 306

Job Flow, 306

JobTracker

about, 306

Hive communicating with, 7

in distributed mode, 26

security groups for, 253

JOINs

optimizing, 134

types of, 98–105, 134

JRE (Java Runtime Environment), for Hadoop,

16–17

JSON (JavaScript Object Notation)

about, 306

maps and structs in, 46

output from SELECT clause, 80

SerDe, 208–209

JUnit tests, 156, 160

JVM (Java Virtual Machine) libraries, 9–10

Kafka system, 10

Karmasphere perspective (case study), on

customer experiences and needs,

296–304

Karmasphere, graphical interface for

interacting with Hive, 6

key-value pairs

as structure for input and output in

MapReduce, 4–5

InputFormats reading, 201

keys, in Hive, 44, 117–120

keystrokes

for Hive CLI navigation, 37

tab key, for autocomplete, 37

last_modified_by table property, 56

last_modified_time table property, 56

left semi-JOINs, 104–105

left-outer JOINs, 101

LIKE predicate operator, 94, 96–97

LIMIT clause, 91, 134

LINES TERMINATED BY clause, 47

Linux

installing Hadoop on, 18–19

installing Java on, 17

load balancing, TCP, 218

LOAD DATA LOCAL statement, 72

local mode

confguration of Hadoop, 24–26

for Hadoop, 19

tuning using, 135

LOCATION clause, 54

locking, 235–238

Log4J Properties, changing, 155

Mac OSX, installing Java on, 17–18

machine learning algorithms (case study),

creating, 265–270

macros, in UDFs, 185

MAIN HEADING, 210

managed tables

about, 56

dropping, 66

partitioned, 58–60

managed tables, loading data into, 71–73

Managing Hive data across multiple map

reduce clusters (case study), m6d,

274–278

MAP data types

as return type for functions, 90

in Hive, 43, 46, 114

MAP(), streaming using, 187

Map, in MapReduce process, 306

map-side JOINs, 105–106

Mapper process, Word Count algorithm as

“Hello World” program for, 4

mappers and reducers, tuning by reducing

mappers and reducers, 138–139

MapR, 306

MapR, Inc., virtual machines for VMWare, 16

MapReduce

Index | 321

Downloa d f r o m W o w ! e B o o k < w w w.woweb o o k . c o m >

about, 3–6, 307

clusters, managing Hive data across

multiple (case study), 274–278

distributed data processing tools not using,

Hadoop and, 1

“Hello World” program for, 4–5

in running queries, 92

jobs for pairs of JOINs, 100

launching components in Hadoop for, 20–

metadata and, 255

multigroup into single job, 142

reading data in, 256–258

structure for input and output in, 4–5

writing data in, 258–260

Master Instance Group, EMR, 247

master node, of HDFS, 51

master security group, EMR modifying, 253

Mastering Regular Expressions, 3rd Edition

(Friedl), 97

Mathematica system, 10

mathematical functions, 83–85

Matlab system, 10

Maven project, Hive in, 158–159, 160

metadata

changing in database, 52

NameNode keeping filesystem, 122

Pig and MapReduce using, 255

metastore

about, 22, 307

backup information from, 28

connecting to metastore via HiveThrift,

216–217

database configuration, 28–29

on EMR, 250–251

option for Hive services, 30

setting up ThriftMetastore, 219

using JDBC, 28–29

MetaStore API, 216–217

methods, in AbstractGenericUDAFResolver,

174

MySQL

dump on S3, 251

server with EMR, 250

MySQL dialect vs. HiveQL, 49

MySQL, configuration for Hadoop

environment, 28

\n delimiters, 45

N rows return type, for table generating

functions, 87–88

NameNode

as HDFS master, 26

HDFS and, 7

metadata for filesystem in, 122

NameNode User Interfaces, security groups for,

253

namespace options, for variables and properties

of Hive CLI, 32

NASA’s Jet Propulsion Laboratory (case study),

287–292

nested SELECT statements, 91, 101–103

NONE compression, in sequence files, 149

normalizing data, unique keys and, 123–124

NoSQL, 307

about, 2

connecting using HiveStorageHandler to,

222

ObjectInspector, 206

Octave system, 10

ODBC (Open Database Connection), 307

OLAP (Online Analytic Processing), Hive

databases and, 2

OLTP (Online Transaction Processing), Hive

databases and, 2

Oozie, Apache, 239–244

operating systems for Hadoop, 15

operators

arithmetic, 82

predicate, 93–94

ORDER BY clauses, 107

Outbrain (case studies), 278–287

outer JOINs, 101–103

full, 104

gotcha in, 101–103

right, 103

output

capturing with Oozie, 243

compression , final, 148

format, 307

formats, 64

from SELECT clause, 80

structure in MapReduce, 4–5

322 | Index

output and input structure, in MapReduce, 4–

OUTPUTFORMAT clauses, 65, 200

OVERWRITE keyword

in loading data into managed tables, 72

with INSERT statement, 73

parallel execution, for tuning queries, 135

PARTITION clause, 60

in ALTER TABLE ... TOUCH clause, 69

in altering storage properties, 69

in loading data into managed tables, 71

partition keys

columns and, 59

showing, 60

partition-level privileges, 233

partitioned

external tables, 61–63

managed tables, 58–60

partitions

about, 307

archiving, 152–154

bug in using count(DISTINCT col) in

columns with, 86

dynamic and static, 74–75

dynamic partition

tuning, 140–141

dynamic partition properties, 75

schema design and, 122–123

PDK (Plugin Developer Kit), 161

permanent functions, UDFs as, 171

persistence, on EMR, 250–251

Photobucket (case study), 292–294

Pig, 8, 255

Plugin Developer Kit (PDK), 161

pom.xml file, for hive_test, 159, 160

PostHooks utility, 158

predicate operators

about, 93–94

LIKE, 94, 96–97

RLIKE, 94, 96–97

PreHooks utility, 158

primary keys, normalization and, 123–124

primitive data types, 41–43, 109

privileges

granting and revoking, 231–234

list of, 231

Programming Pig (Gates), 8

project specific fields, extracting, 188–189

property and variables, Hive CLI, 31–34

property names, trick for finding, 35

pseudodistributed mode

Hadoop in, 27–28

metastore in single node in, 28

pseudodistributed mode for Hadoop, running

in, 19

Python scripts, using CLUSTER BY clause,

192–194

.q files, testing from, 156–158

Qubole, graphical interface for interacting with

Hive, 6

queries, Hive HQL

aggregate functions, 85–87

Cartesian product, 105

CASE ... WHEN ... THEN clauses, 91–92

casting, 109

CLUSTER BY clause, 108–109

column aliases, 91

column headers in, 38–39

DISTRIBUTE BY clauses, 107–108

executing from files, 35–36

GROUP BY clauses, 97

HAVING clause, 97

JOINs

types of, 98–105

joins

optimizing, 134

LIMIT clause, 91

making small data sets run faster, 19

MapReduce in running, 92

mathematical functions, 83–85

nested SELECT statements, 91, 101–103

ORDER BY clauses, 107

sampling data using, 110–112

SELECT ... FROM clauses, 73–74

about, 79–81

computing with values from columns,

81–82

specify columns in with, 81

semicolon (;) at end of lines in, 36

separating multiple queries, 36

SORT BY clauses, 107–108

SUB1 TEXT, 82

table generating functions, 87–88

testing from .q files, 156–158

Index | 323

tuning

by optimizing JOINs, 134

by reducing mappers and reducers, 138–

139

EXPLAIN clauses for, 131–134

EXPLAIN EXTENDED clauses for, 134

LIMIT clause in, 134

using parallel execution, 135

using strict mode, 137

UNION ALL, 112

using src (“source”) in, 35–36

views reducing complexity of, 113

WHERE clause, 92

about, 92–93

gotchas with floating point comparisons,

94–96

join statements and, 103

LIKE predicate operator, 94, 96–97

predicate operators, 93–94

RLIKE predicate operator, 94, 96–97

R language, 10

R language(case study), machine learning

algorithms, 265–270

ranking data by grouping like elements (case

study), 270–274

RCFILE format, 64, 202–203

rcfilecat option, for Hive services, 30

RECORD compression, in sequence files, 149

record formats, 205–210, 205

(see also SerDe (Serializer Deserializer))

vs. file formats, 199

record parsing, 64

record-level updates, 2

Reduce Operator Tree, 133, 134

REDUCE(), streaming using, 187

Reduce, in MapReduce process, 307

reducers

and mappers, tuning by reducing, 138–139

in MapReduce process, 4

Regional Climate Model Evaluation System

(RCMES) (case study), 287–292

regular expression features, support for, 55

Regular Expression Pocket Reference

(Goyvaerts and Stubblebine), 97

regular expressions

specifying columns in queries with, 81

using single quotes ("), 55

relational databases, collection data types and,

relational model, 307

renaming tables, 66

replicating, metastore information, 28

right-outer JOINs, 103

RLIKE predicate operator, 94, 96–97

roles, granting and revoking privileges

for individual, 230–233

ROW FORMAT DELIMITED clauses, 47

ROW FORMAT SERDE clauses, 65

runtime modes, for Hadoop, 19

S3 bucket, 305

S3 system for AWS

about, 308

accessing, 248

deploying .hiverc script to, 249

EMR logs on, 252

moving data to, 62

MySQL dump on, 251

putting resources on, 252

role on EMR cluster, 251

support for, 62

s3n “scheme,” for accessing S3, 248

Safari Books Online, xiv–xv

sampling data, using queries, 110–112

schema

Avro, 209

changing with ALTER TABLE statements,

53, 66

copying, 54

defining from URI, 210

design, 121–129

using getSchema() method, 215

schema on write vs. schema on read, 48

SciPy, NumPy system, 10

security

groups, EMR, 253

Hive, 227–234

model for HCatalog, 261–262

SELECT ... FROM clauses

about, 79–81

computing with values from columns, 81–

specify columns in with, 81

SELECT ... WHERE clauses, 73–74

SELECT clause, 79

324 | Index

SELECT statements, nested, 91, 101–103

semicolon (;)

at end of lines in Hive queries, 36

separating queries, 34

sequence file format, 148–149

SEQUENCEFILE format, 64, 149, 201–202

SerDe (Serializer Deserializer)

about, 205, 308

AVRO Hive, 209

Avro Hive, 305

Cassandra SerDe Storage Handler

properties, 224

columnar, 128

CSVs (Comma-Separated Values), 206

extraction of data from input with, 127

JSON, 208–209

record parsing by, 64–65, 127

TSV (Tab-Separated Values), 206

using SERDEPROPERTIES, 69

serializing and deserializing data, 187

Service List, for hive command, 30

sessionization (case study), 282–287

SET command

Hive CLI variables using, 32–33

to disable authentication, 230

shell execution, running bash shell commands

without Hive CLI, 37–38

shell “pipes,” bash shell commands, 37–38

SHOW FORMATTED INDEX command,

119

SHOW FUNCTIONS command, 163

SHOW PARTITIONS command, 60, 62–63

SHOW TABLES command, 52, 54

SHOW TBLPROPERTIES table_name

command, 54

Shuffle and Sort phase, in HDFS, 5

SimpleReach (case study), using Cassandra to

store social network polling at, 294–

296

single MapReduce job, multigroup by, 142

Single Point of Failure (SPOF), metastore

information as, 28

SMALLINT data types, 41

SnappyCodec, 147

social network polling (case study), using

Cassandra to store, 294–296

Sort and Shuffle phase, in HDFS, 5

SORT BY clauses

DISTRIBUTE BY clauses with, 107–108

in streaming processes, 192, 194

ORDER BY clauses and, 107

SORTED BY clauses, in altering storage

properties, 69

Spark system, 10

splittable files, 148–149

SPOF (Single Point of Failure), metastore

information as, 28

spot instances, 252–253

SQL (Structured Query Language)

about, 308

HiveQL and, 1

src (“source”), using in queries, 35–36

STAGE PLANS command, Hive job consisting

of, 132–133

standard functions, 164

stateful annotations, 184

static column mapping, in Cassandra, 224

Storage Handlers, 221–225

storage properties, altering, 68

STORED AS INPUTFORMAT ...

OUTPUTFORMAT clause, 65

STORED AS SEQUENCEFILE clause, 149,

200

STORED AS TEXTFILE clause, 47, 63–64

Storm system, 10

streaming

about, 187–188

calculating

aggregates with streaming, 191–192

cogroups, 196

changing data types, 188

distributing and sorting of data, 192–194

editor for manipulating transformations,

189

extracting project specific fields, 188–189

identity transformation using, 188

producing multiple rows from single row,

189–190

to Java using GenericMR Tools, 194–196

using distributed cache, 189–190

strict mode, 137

STRING data types

about, 42

as return type for functions, 88–90

as return type for mathematical functions,

STRUCT data types

in Hive, 43, 46, 114

Index | 325

in JSON, 46

Structured Query Language (SQL), 308

HiveQL and, 1

Stubblebine, Tony, Regular Expression Pocket

Reference, 97

sudo (“super user do something”), running in

Linux, 17

system namespace option, for Hive CLI, 32

tab key, for autocomplete, 37

Tab-Separated Values (TSV) SerDes, 206

Tab-separated Values (TSVs), as delimiters, 45,

Table Generating Functions (UDTFs), User-

Defined, 87–88, 165–166, 177–182,

308

table storage formats

customizing, 63–65

STORED AS TEXTFILE clause, 47, 64

tables

altering

changing schemas, 53, 66

properties of, 68

renaming, 66

bucketing data storage on, 125–127

columnar, 128

columns in

adding, 68, 127

adding comments to, 53

bug in using count(DISTINCT col) in

with partitioned, 86

changing, 67

computing with values from, 81–82

deleting or replacing, 68

partition keys and, 59

specifying in queries with regular

expressions, 81

copying schema, 54

creating, 53–56

creating and loading data, 75

creating in HBASE, 222–224

dropping, 66

external

about, 56–57

partitioned, 61–63

indexing, 117–120

input pruning for bucket, 111–112

JOINs

optimizing, 134

types of, 98–105

listing, 54–55

location of stored, 50, 54

managed

about, 56

loading data into, 71–73

partitioned, 58–60

normalization and unique keys, 123–124

partitioned, 58–63, 74–75, 122–125

partitions in

adding, modifying, and dropping, 66

tuning dynamic, 140–141

schema design, 121–129

Table-by-day pattern, 121–122

tables

views and map type for dynamic, 114–

115

UNION ALL, 112

views and map type for dynamic, 114–115

Task Instance Group, EMR, 248

Task, in MapReduce context, 308

TaskTrackers, 141, 240, 253

TBLPROPERTIES, adding additional

documentation in key-value format,

TCP load balancing, 218

test cases, running Hive, 156–158

testing, with hive_test, 159–161

text editor, Emacs, 45

text files, encoding of data values, 45–48

TEXTFILE clause

implication of, 64

STORED AS, 47

Think Big Analytics

JSON SerDe and, 208

ObjectInspector, 206

Think Big Analytics, Inc., virtual machines for

VMWare, 16

Thrift Server (HiveServer)

about, 213

administering, 217–218

connecting to, 214–215

connecting to metastore, 216–217

fetching results, 215

getting cluster status, 215

getting schema of results, 215

on EMR Hive, 247

retrieving query plan, 216

326 | Index

setting up Groovy to connect to HiveService,

214

setting up ThriftMetastore, 219

starting, 213

Thrift Service Action, Hive, 240

Thrift service component, 22, 308

ThriftMetastore, Hive, setting up, 219

time-range partitioning, 122

TIMESTAMP (v0.8.0+) data types, 42

TINYINT data types, 41

TRANSFORM()

changing data types, 188

doing multiple transformations, 192

in producing multiple rows from single row,

191

streaming using, 187

using with CLUSTER BY clause, 192–194

TSV (Tab-Separated Values) SerDes, 206

tuning

dynamic partition, 140–141

indexes, 140

multigroup MapReduce into single job,

142

optimizing JOINs, 134

queries

by reducing mappers and reducers, 138–

139

EXPLAIN clauses for, 131–134

EXPLAIN EXTENDED clauses for, 134

LIMIT clause in, 134

using parallel execution, 135

using strict mode, 137

using JVM reuse, 139–140

using local mode, 135

using speculative execution, 141–142

using virtual columns, 142–143

tuple return type, for table generating

functions, 87

UDAFs (User-Defined Aggregate Functions),

172–177, 185, 308

UDF Pseudo Rank (case study), M6d, 270–

274

UDFs (User-Defined Functions) (see User-

Defined Functions (UDFs))

UDTFs (User-Defined Table Generating

Functions), 87–88, 165–166, 177–

182, 308

UNION ALL, 112

unique keys, normalization and, 123–124

URI, defining schema from, 210

URLs (case study, identifying, 278–280

USE command, 52

User-Defined Aggregate Functions (UDAFs),

164–176, 171, 172–177, 172–177,

185, 308

User-Defined Functions (UDFs)

about, 163, 308

accessing distributed cache from, 182–184

aggregate functions, user-defined, 164–176,

171, 172–177

annotations for use with, 184

as standard functions, 164

calling, 164

describing, 164

discovering, 163

table generating functions, 87–88, 165–166,

177–182

vs. built-in functions, 163

writing example of, 166–169

XML UDF, 207

User-Defined Functions (UDFs), vs.

GenericUDFs, 169–171

User-Defined Table Generating Functions

(UDTFs), 87–88, 165–166, 177–182,

308

users, granting and revoking privileges for

individual, 230–233

/usr/bin/sed program, as stream editor, 189

value of type, as return type for functions, 89

variables

and properties, Hive CLI, 31–34

capturing with Oozie output to, 243–244

in Oozie workflows, 242–243

views, 113–116

virtual columns, tuning using, 142–143

Virtual Machine (VM), installing

preconfigured, 15–16

VirtualBox, 15

VMWare, 15

web console, Oozie, 242

web traffic (case study), analyzing, 282–287

Index | 327

Downloa d f r o m W o w ! e B o o k < w w w.woweb o o k . c o m >

WHERE clause

about, 92–93

gotchas with floating point comparisons,

94–96

join statements and, 103

LIKE predicate operator, 94, 96–97

predicate operators, 93–94

RLIKE predicate operator, 94, 96–97

White, Tom, Hadoop: The Definitive Guide,

White, Tom, “Hadoop: The Definitive Guide”,

Windows

installing Java on, 16–17

running Hadoop within, 15

WITH DEFERRED REBUILD clauses, 118

WITH SERDEPROPERTIES, 208

Word Count algorithm

as “Hello World” program for MapReduce,

4–5

using Hadoop command to launch, 20–21

using HiveQL, 10

using Java MapReduce API, 10–12

workflow, creating two-query, 242

XML UDF, 207

XPath (XML Path Language)

about, 207

functions, 207–208

Zookeeper, 235–237

328 | Index

About the Authors

Edward Capriolo is currently System Administrator at Media6degrees, where he helps

design and maintain distributed data storage systems for the Internet advertising

industry.

Edward is a member of the Apache Software Foundation and a committer for the

Hadoop-Hive project. He has experience as a developer, as well as a Linux and network

administrator, and enjoys the rich world of open source software.

Dean Wampler is a Principal Consultant at Think Big Analytics, where he specializes

in “Big Data” problems and tools like Hadoop and Machine Learning. Besides Big Data,

he specializes in Scala, the JVM ecosystem, JavaScript, Ruby, functional and object-

oriented programming, and Agile methods. Dean is a frequent speaker at industry and

academic conferences on these topics. He has a Ph.D. in Physics from the University

of Washington.

Jason Rutherglen is a software architect at Think Big Analytics and specializes in Big

Data, Hadoop, search, and security.

Colophon

The animal on the cover of Programming Hive is a European hornet (Vespa cabro) and

its hive. The European hornet is the only hornet in North America, introduced to the

continent when European settlers migrated to the Americas. This hornet can be found

throughout Europe and much of Asia, adapting its hive-building techniques to different

climates when necessary.

The hornet is a social insect, related to bees and ants. The hornet’s hive consists of one

queen, a few male hornets (drones), and a large quantity of sterile female workers. The

chief purpose of drones is to reproduce with the hornet queen, and they die soon after.

It is the female workers who are responsible for building the hive, carrying food, and

tending to the hornet queen’s eggs.

The hornet’s nest itself is the consistency of paper, since it is constructed out of wood

pulp in several layers of hexagonal cells. The end result is a pear-shaped nest attached

to its shelter by a short stem. In colder areas, hornets will abandon the nest in the winter

and take refuge in hollow logs or trees, or even human houses, where the queen and

her eggs will stay until the warmer weather returns. The eggs form the start of a new

colony, and the hive can be constructed once again.

The cover image is from Johnson’s Natural History. The cover font is Adobe ITC Ga-

ramond. The text font is Linotype Birka; the heading font is Adobe Myriad Condensed;

and the code font is LucasFont’s TheSansMonoCondensed.

Programming Hive Guide

Navigation menu

Versions of this User Manual:

Views

Navigation