MongoDB: The Definitive Guide, Second Edition Mongodb Guide

mongodb_%20the%20definitive%20guide%20-%20kristina%20chodorow_1401

mongodb_%20the%20definitive%20guide%20-%20kristina%20chodorow_1401

mongodb_%20the_definitive%20guide%20-%20kristina%20chodorow_1401

mongodb_guide

MongoDB:%20The%20Definitive%20Guide

User Manual: Pdf

Open the PDF directly: View PDF PDF.
Page Count: 432 [warning: Documents this large are best viewed by clicking the View PDF Link!]

Kristina Chodorow
SECOND EDITION
MongoDB: The Definitive Guide
MongoDB: The Definitive Guide, Second Edition
by Kristina Chodorow
Copyright © 2013 Kristina Chodorow. All rights reserved.
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.
Editor: Ann Spencer
Production Editor: Kara Ebrahim
Proofreader: Amanda Kersey
Indexer: Stephen Ingle, WordCo Indexing
Cover Designer: Randy Comer
Interior Designer: David Futato
Illustrator: Rebecca Demarest
May 2013: Second Edition
Revision History for the Second Edition:
2013-05-08: First release
See http://oreilly.com/catalog/errata.csp?isbn=9781449344689 for release details.
Nutshell Handbook, the Nutshell Handbook logo, and the O’Reilly logo are registered trademarks of O’Reilly
Media, Inc. MongoDB: The Definitive Guide, Second Edition, the image of a mongoose lemur, and related
trade dress are trademarks 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 trade‐
mark 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 author assume no
responsibility for errors or omissions, or for damages resulting from the use of the information contained
herein.
ISBN: 978-1-449-34468-9
[LSI]
Table of Contents
Foreword. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xiii
Preface. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xv
Part I. Introduction to MongoDB
1. Introduction. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
Ease of Use 3
Easy Scaling 3
Tons of Features… 4
…Without Sacrificing Speed 5
Let’s Get Started 5
2. Getting Started. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
Documents 7
Collections 8
Dynamic Schemas 8
Naming 9
Databases 10
Getting and Starting MongoDB 11
Introduction to the MongoDB Shell 12
Running the Shell 13
A MongoDB Client 13
Basic Operations with the Shell 14
Data Types 16
Basic Data Types 16
Dates 18
Arrays 18
Embedded Documents 19
_id and ObjectIds 20
iii
Using the MongoDB Shell 21
Tips for Using the Shell 22
Running Scripts with the Shell 23
Creating a .mongorc.js 25
Customizing Your Prompt 26
Editing Complex Variables 27
Inconvenient Collection Names 27
3. Creating, Updating, and Deleting Documents. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
Inserting and Saving Documents 29
Batch Insert 29
Insert Validation 30
Removing Documents 31
Remove Speed 31
Updating Documents 32
Document Replacement 32
Using Modifiers 34
Upserts 45
Updating Multiple Documents 47
Returning Updated Documents 48
Setting a Write Concern 51
4. Querying. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
Introduction to find 53
Specifying Which Keys to Return 54
Limitations 55
Query Criteria 55
Query Conditionals 55
OR Queries 56
$not 57
Conditional Semantics 57
Type-Specific Queries 58
null 58
Regular Expressions 58
Querying Arrays 59
Querying on Embedded Documents 63
$where Queries 65
Server-Side Scripting 66
Cursors 67
Limits, Skips, and Sorts 68
Avoiding Large Skips 70
Advanced Query Options 71
iv | Table of Contents
Getting Consistent Results 72
Immortal Cursors 75
Database Commands 75
How Commands Work 76
Part II. Designing Your Application
5. Indexing. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
Introduction to Indexing 81
Introduction to Compound Indexes 84
Using Compound Indexes 89
How $-Operators Use Indexes 91
Indexing Objects and Arrays 95
Index Cardinality 98
Using explain() and hint() 98
The Query Optimizer 102
When Not to Index 102
Types of Indexes 104
Unique Indexes 104
Sparse Indexes 106
Index Administration 107
Identifying Indexes 108
Changing Indexes 108
6. Special Index and Collection Types. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
Capped Collections 109
Creating Capped Collections 111
Sorting Au Naturel 112
Tailable Cursors 113
No-_id Collections 114
Time-To-Live Indexes 114
Full-Text Indexes 115
Search Syntax 118
Full-Text Search Optimization 119
Searching in Other Languages 119
Geospatial Indexing 120
Types of Geospatial Queries 120
Compound Geospatial Indexes 121
2D Indexes 122
Storing Files with GridFS 123
Getting Started with GridFS: mongofiles 124
Table of Contents | v
Working with GridFS from the MongoDB Drivers 124
Under the Hood 125
7. Aggregation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
The Aggregation Framework 127
Pipeline Operations 129
$match 129
$project 130
$group 135
$unwind 137
$sort 139
$limit 139
$skip 139
Using Pipelines 140
MapReduce 140
Example 1: Finding All Keys in a Collection 140
Example 2: Categorizing Web Pages 143
MongoDB and MapReduce 143
Aggregation Commands 146
count 146
distinct 147
group 147
8. Application Design. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153
Normalization versus Denormalization 153
Examples of Data Representations 154
Cardinality 157
Friends, Followers, and Other Inconveniences 158
Optimizations for Data Manipulation 160
Optimizing for Document Growth 160
Removing Old Data 162
Planning Out Databases and Collections 162
Managing Consistency 163
Migrating Schemas 164
When Not to Use MongoDB 165
Part III. Replication
9. Setting Up a Replica Set. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 169
Introduction to Replication 169
A One-Minute Test Setup 170
vi | Table of Contents
Configuring a Replica Set 174
rs Helper Functions 175
Networking Considerations 176
Changing Your Replica Set Configuration 176
How to Design a Set 178
How Elections Work 180
Member Configuration Options 181
Creating Election Arbiters 182
Priority 183
Hidden 184
Slave Delay 185
Building Indexes 185
10. Components of a Replica Set. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 187
Syncing 187
Initial Sync 188
Handling Staleness 190
Heartbeats 191
Member States 191
Elections 192
Rollbacks 193
When Rollbacks Fail 197
11. Connecting to a Replica Set from Your Application. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 199
Client-to-Replica-Set Connection Behavior 199
Waiting for Replication on Writes 200
What Can Go Wrong? 201
Other Options for “w” 202
Custom Replication Guarantees 202
Guaranteeing One Server per Data Center 202
Guaranteeing a Majority of Nonhidden Members 204
Creating Other Guarantees 204
Sending Reads to Secondaries 205
Consistency Considerations 205
Load Considerations 205
Reasons to Read from Secondaries 206
12. Administration. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 209
Starting Members in Standalone Mode 209
Replica Set Configuration 210
Creating a Replica Set 210
Changing Set Members 211
Table of Contents | vii
Creating Larger Sets 211
Forcing Reconfiguration 212
Manipulating Member State 213
Turning Primaries into Secondaries 213
Preventing Elections 213
Using Maintenance Mode 213
Monitoring Replication 214
Getting the Status 214
Visualizing the Replication Graph 216
Replication Loops 218
Disabling Chaining 218
Calculating Lag 219
Resizing the Oplog 220
Restoring from a Delayed Secondary 221
Building Indexes 222
Replication on a Budget 223
How the Primary Tracks Lag 224
Master-Slave 225
Converting Master-Slave to a Replica Set 226
Mimicking Master-Slave Behavior with Replica Sets 226
Part IV. Sharding
13. Introduction to Sharding. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 231
Introduction to Sharding 231
Understanding the Components of a Cluster 232
A One-Minute Test Setup 232
14. Configuring Sharding. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 241
When to Shard 241
Starting the Servers 242
Config Servers 242
The mongos Processes 243
Adding a Shard from a Replica Set 244
Adding Capacity 245
Sharding Data 245
How MongoDB Tracks Cluster Data 246
Chunk Ranges 247
Splitting Chunks 249
viii | Table of Contents
The Balancer 253
15. Choosing a Shard Key. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 257
Taking Stock of Your Usage 257
Picturing Distributions 258
Ascending Shard Keys 258
Randomly Distributed Shard Keys 261
Location-Based Shard Keys 263
Shard Key Strategies 264
Hashed Shard Key 264
Hashed Shard Keys for GridFS 266
The Firehose Strategy 267
Multi-Hotspot 268
Shard Key Rules and Guidelines 271
Shard Key Limitations 271
Shard Key Cardinality 271
Controlling Data Distribution 271
Using a Cluster for Multiple Databases and Collections 272
Manual Sharding 273
16. Sharding Administration. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 275
Seeing the Current State 275
Getting a Summary with sh.status 275
Seeing Configuration Information 277
Tracking Network Connections 283
Getting Connection Statistics 283
Limiting the Number of Connections 284
Server Administration 285
Adding Servers 285
Changing Servers in a Shard 285
Removing a Shard 286
Changing Config Servers 288
Balancing Data 289
The Balancer 289
Changing Chunk Size 290
Moving Chunks 291
Jumbo Chunks 292
Refreshing Configurations 295
Part V. Application Administration
Table of Contents | ix
17. Seeing What Your Application Is Doing. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 299
Seeing the Current Operations 299
Finding Problematic Operations 301
Killing Operations 301
False Positives 302
Preventing Phantom Operations 302
Using the System Profiler 302
Calculating Sizes 305
Documents 305
Collections 305
Databases 306
Using mongotop and mongostat 307
18. Data Administration. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 311
Setting Up Authentication 311
Authentication Basics 312
Setting Up Authentication 314
How Authentication Works 314
Creating and Deleting Indexes 315
Creating an Index on a Standalone Server 315
Creating an Index on a Replica Set 315
Creating an Index on a Sharded Cluster 316
Removing Indexes 316
Beware of the OOM Killer 317
Preheating Data 317
Moving Databases into RAM 317
Moving Collections into RAM 318
Custom-Preheating 318
Compacting Data 320
Moving Collections 321
Preallocating Data Files 322
19. Durability. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 323
What Journaling Does 323
Planning Commit Batches 324
Setting Commit Intervals 325
Turning Off Journaling 325
Replacing Data Files 325
Repairing Data Files 325
The mongod.lock File 326
Sneaky Unclean Shutdowns 327
What MongoDB Does Not Guarantee 327
x | Table of Contents
Checking for Corruption 327
Durability with Replication 329
Part VI. Server Administration
20. Starting and Stopping MongoDB. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 333
Starting from the Command Line 333
File-Based Configuration 336
Stopping MongoDB 336
Security 337
Data Encryption 338
SSL Connections 338
Logging 338
21. Monitoring MongoDB. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 341
Monitoring Memory Usage 341
Introduction to Computer Memory 341
Tracking Memory Usage 342
Tracking Page Faults 343
Minimizing Btree Misses 345
IO Wait 346
Tracking Background Flush Averages 346
Calculating the Working Set 348
Some Working Set Examples 350
Tracking Performance 350
Tracking Free Space 352
Monitoring Replication 353
22. Making Backups. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 357
Backing Up a Server 357
Filesystem Snapshot 357
Copying Data Files 358
Using mongodump 359
Backing Up a Replica Set 361
Backing Up a Sharded Cluster 362
Backing Up and Restoring an Entire Cluster 362
Backing Up and Restoring a Single Shard 362
Creating Incremental Backups with mongooplog 363
23. Deploying MongoDB. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 365
Designing the System 365
Table of Contents | xi
Choosing a Storage Medium 365
Recommended RAID Configurations 369
CPU 370
Choosing an Operating System 370
Swap Space 371
Filesystem 371
Virtualization 372
Turn Off Memory Overcommitting 372
Mystery Memory 372
Handling Network Disk IO Issues 373
Using Non-Networked Disks 374
Configuring System Settings 374
Turning Off NUMA 374
Setting a Sane Readahead 377
Disabling Hugepages 378
Choosing a Disk Scheduling Algorithm 379
Don’t Track Access Time 380
Modifying Limits 380
Configuring Your Network 382
System Housekeeping 383
Synchronizing Clocks 383
The OOM Killer 383
Turn Off Periodic Tasks 384
A. Installing MongoDB. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 385
B. MongoDB Internals. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 389
Index. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 393
xii | Table of Contents
Foreword
In the last 10 years, the Internet has challenged relational databases in ways nobody
could have foreseen. Having used MySQL at large and growing Internet companies
during this time, I’ve seen this happen firsthand. First you have a single server with a
small data set. Then you find yourself setting up replication so you can scale out reads
and deal with potential failures. And, before too long, you’ve added a caching layer,
tuned all the queries, and thrown even more hardware at the problem.
Eventually you arrive at the point when you need to shard the data across multiple
clusters and rebuild a ton of application logic to deal with it. And soon after that you
realize that you’re locked into the schema you modeled so many months before.
Why? Because there’s so much data in your clusters now that altering the schema will
take a long time and involve a lot of precious DBA time. It’s easier just to work around
it in code. This can keep a small team of developers busy for many months. In the end,
you’ll always find yourself wondering if there’s a better way—or why more of these
features are not built into the core database server.
Keeping with tradition, the Open Source community has created a plethora of “better
ways” in response to the ballooning data needs of modern web applications. They span
the spectrum from simple in-memory key/value stores to complicated SQL-speaking
MySQL/InnoDB derivatives. But the sheer number of choices has made finding the right
solution more difficult. I’ve looked at many of them.
I was drawn to MongoDB by its pragmatic approach. MongoDB doesn’t try to be ev‐
erything to everyone. Instead it strikes the right balance between features and com‐
plexity, with a clear bias toward making previously difficult tasks far easier. In other
words, it has the features that really matter to the vast majority of today’s web applica‐
tions: indexes, replication, sharding, a rich query syntax, and a very flexible data model.
All of this comes without sacrificing speed.
Like MongoDB itself, this book is very straightforward and approachable. New
MongoDB users can start with Chapter 1 and be up and running in no time. Experienced
xiii
users will appreciate this book’s breadth and authority. It’s a solid reference for advanced
administrative topics such as replication, backups, and sharding, as well as popular client
APIs.
Having recently started to use MongoDB in my day job, I have no doubt that this book
will be at my side for the entire journey—from the first install to production deployment
of a sharded and replicated cluster. It’s an essential reference to anyone seriously looking
at using MongoDB.
—Jeremy Zawodny
Craigslist Software Engineer
August 2010
xiv | Foreword
Preface
How This Book Is Organized
This book is split up into six sections, covering development, administration, and de‐
ployment information.
Getting Started with MongoDB
In Chapter 1 we provide background about MongoDB: why it was created, the goals it
is trying to accomplish, and why you might choose to use it for a project. We go into
more detail in Chapter 2, which provides an introduction to the core concepts and
vocabulary of MongoDB. Chapter 2 also provides a first look at working with MongoDB,
getting you started with the database and the shell. The next two chapters cover the
basic material that developers need to know to work with MongoDB. In Chapter 3, we
describe how to perform those basic write operations, including how to do them with
different levels of safety and speed. Chapter 4 explains how to find documents and create
complex queries. This chapter also covers how to iterate through results and gives op‐
tions for limiting, skipping, and sorting results.
Developing with MongoDB
Chapter 5 covers what indexing is and how to index your MongoDB collections. Chap‐
ter 6 explains how to use several special types of indexes and collections. Chapter 7
covers a number of techniques for aggregating data with MongoDB, including counting,
finding distinct values, grouping documents, the aggregation framework, and using
MapReduce. Finally, this section finishes with a chapter on designing your application:
Chapter 8 goes over tips for writing an application that works well with MongoDB.
xv
Replication
The replication section starts with Chapter 9, which gives you a quick way to set up a
replica set locally and covers many of the available configuration options. Chapter 10
then covers the various concepts related to replication. Chapter 11 shows how replica‐
tion interacts with your application and Chapter 12 covers the administrative aspects
of running a replica set.
Sharding
The sharding section starts in Chapter 13 with a quick local setup. Chapter 14 then gives
an overview of the components of the cluster and how to set them up. Chapter 15 has
advice on choosing a shard key for a variety of application. Finally, Chapter 16 covers
administering a sharded cluster.
Application Administration
The next two chapters cover many aspects of MongoDB administration from the per‐
spective of your application. Chapter 17 discusses how to introspect what MongoDB is
doing. Chapter 18 covers administrative tasks such as building indexes, and moving
and compacting data. Chapter 19 explains how MongoDB stores data durably.
Server Administration
The final section is focused on server administration. Chapter 20 covers common op‐
tions when starting and stopping MongoDB. Chapter 21 discusses what to look for and
how to read stats when monitoring. Chapter 22 describes how to take and restore back‐
ups for each type of deployment. Finally, Chapter 23 discusses a number of system
settings to keep in mind when deploying MongoDB.
Appendixes
Appendix A explains MongoDB’s versioning scheme and how to install it on Windows,
OS X, and Linux. Appendix B details ow MongoDB works internally: its storage engine,
data format, and wire protocol.
Conventions Used in This Book
The following typographical conventions are used in this book:
Italic
Indicates new terms, URLs, email addresses, collection names, database names,
filenames, and file extensions.
xvi | Preface
Constant width
Used for program listings, as well as within paragraphs to refer to program elements
such as variable or function names, command-line utilities, environment variables,
statements, and keywords.
Constant width bold
Shows commands or other text that should be typed literally by the user.
Constant width italic
Shows text that should be replaced with user-supplied values or by values deter‐
mined by context.
This icon signifies a tip, suggestion, or general note.
This icon indicates a warning or caution.
Using Code Examples
This book can help you get your job done. In general, you may use the code in this book
in your programs and documentation. You do not need to contact us for permission
unless you’re reproducing a significant portion of the code. For example, writing a pro‐
gram that uses several chunks of code from this book does not require permission.
Selling or distributing a CD-ROM of examples from O’Reilly books does require per‐
mission. Answering a question by citing this book and quoting example code does not
require permission. Incorporating a significant amount of example code from this book
into your product’s documentation does require permission.
We appreciate, but do not require, attribution. An attribution usually includes the
title, author, publisher, and ISBN. For example: “MongoDB: The Definitive Guide, Sec‐
ond Edition by Kristina Chodorow (O’Reilly). Copyright 2013 Kristina Chodorow,
978-1-449-34468-9.”
If you feel your use of code examples falls outside fair use or the permission given here,
feel free to contact us at permissions@oreilly.com.
Preface | xvii
Safari® Books Online
Safari Books Online (www.safaribooksonline.com) is an on-demand
digital library that delivers expert content in both book and video
form from the world’s leading authors in technology and business.
Technology professionals, software developers, web designers, and business and crea‐
tive professionals use Safari Books Online as their primary resource for research, prob‐
lem solving, learning, and certification training.
Safari Books Online offers a range of product mixes and pricing programs for organi‐
zations, government agencies, and individuals. Subscribers have access to thousands of
books, training videos, and prepublication manuscripts in one fully searchable database
from publishers like O’Reilly Media, Prentice Hall Professional, Addison-Wesley Pro‐
fessional, Microsoft Press, Sams, Que, Peachpit Press, Focal Press, Cisco Press, John
Wiley & Sons, Syngress, Morgan Kaufmann, IBM Redbooks, Packt, Adobe Press, FT
Press, Apress, Manning, New Riders, McGraw-Hill, Jones & Bartlett, Course Technol‐
ogy, and dozens more. For more information about Safari Books Online, please visit us
online.
How to Contact Us
Please address comments and questions concerning this book to the publisher:
O’Reilly Media, Inc.
1005 Gravenstein Highway North
Sebastopol, CA 95472
800-998-9938 (in the United States or Canada)
707-829-0515 (international or local)
707 829-0104 (fax)
We have a web page for this book, where we list errata, examples, and any additional
information. You can access this page at:
http://oreil.ly/mongodb-2e
To comment or ask technical questions about this book, send email to:
bookquestions@oreilly.com
For more information about our books, conferences, Resource Centers, and the O’Reilly
Network, see our website at:
http://www.oreilly.com
xviii | Preface
Acknowledgments
I would like to thank my tech reviewers, Adam Comerford, Eric Milke, and Greg Studer.
You guys made this book immeasurably better (and more correct). Thank you, Ann
Spencer, for being such a terrific editor and for helping me every step of the way. Thanks
to all of my coworkers at 10gen for sharing your knowledge and advice on MongoDB
as well as Eliot Horowitz and Dwight Merriman, for starting the MongoDB project. And
thank you, Andrew, for all of your support and suggestions.
Preface | xix
PART I
Introduction to MongoDB
CHAPTER 1
Introduction
MongoDB is a powerful, flexible, and scalable general-purpose database. It combines
the ability to scale out with features such as secondary indexes, range queries, sorting,
aggregations, and geospatial indexes. This chapter covers the major design decisions
that made MongoDB what it is.
Ease of Use
MongoDB is a document-oriented database, not a relational one. The primary reason
for moving away from the relational model is to make scaling out easier, but there are
some other advantages as well.
A document-oriented database replaces the concept of a “row” with a more flexible
model, the “document.” By allowing embedded documents and arrays, the document-
oriented approach makes it possible to represent complex hierarchical relationships
with a single record. This fits naturally into the way developers in modern object-
oriented languages think about their data.
There are also no predefined schemas: a document’s keys and values are not of fixed
types or sizes. Without a fixed schema, adding or removing fields as needed becomes
easier. Generally, this makes development faster as developers can quickly iterate. It is
also easier to experiment. Developers can try dozens of models for the data and then
choose the best one to pursue.
Easy Scaling
Data set sizes for applications are growing at an incredible pace. Increases in available
bandwidth and cheap storage have created an environment where even small-scale ap‐
plications need to store more data than many databases were meant to handle. A terabyte
of data, once an unheard-of amount of information, is now commonplace.
3
As the amount of data that developers need to store grows, developers face a difficult
decision: how should they scale their databases? Scaling a database comes down to the
choice between scaling up (getting a bigger machine) or scaling out (partitioning data
across more machines). Scaling up is often the path of least resistance, but it has draw‐
backs: large machines are often very expensive, and eventually a physical limit is reached
where a more powerful machine cannot be purchased at any cost. The alternative is to
scale out: to add storage space or increase performance, buy another commodity server
and add it to your cluster. This is both cheaper and more scalable; however, it is more
difficult to administer a thousand machines than it is to care for one.
MongoDB was designed to scale out. Its document-oriented data model makes it easier
for it to split up data across multiple servers. MongoDB automatically takes care of
balancing data and load across a cluster, redistributing documents automatically and
routing user requests to the correct machines. This allows developers to focus on pro‐
gramming the application, not scaling it. When a cluster need more capacity, new ma‐
chines can be added and MongoDB will figure out how the existing data should be
spread to them.
Tons of Features…
MongoDB is intended to be a general-purpose database, so aside from creating, reading,
updating, and deleting data, it provides an ever-growing list of unique features:
Indexing
MongoDB supports generic secondary indexes, allowing a variety of fast queries,
and provides unique, compound, geospatial, and full-text indexing capabilities as
well.
Aggregation
MongoDB supports an “aggregation pipeline” that allows you to build complex
aggregations from simple pieces and allow the database to optimize it.
Special collection types
MongoDB supports time-to-live collections for data that should expire at a certain
time, such as sessions. It also supports fixed-size collections, which are useful for
holding recent data, such as logs.
File storage
MongoDB supports an easy-to-use protocol for storing large files and file metadata.
Some features common to relational databases are not present in MongoDB, notably
joins and complex multirow transactions. Omitting these was an architectural decision
to allow for greater scalability, as both of those features are difficult to provide efficiently
in a distributed system.
4 | Chapter 1: Introduction
…Without Sacrificing Speed
Incredible performance is a major goal for MongoDB and has shaped much of its design.
MongoDB adds dynamic padding to documents and preallocates data files to trade extra
space usage for consistent performance. It uses as much of RAM as it can as its cache
and attempts to automatically choose the correct indexes for queries. In short, almost
every aspect of MongoDB was designed to maintain high performance.
Although MongoDB is powerful and attempts to keep many features from relational
systems, it is not intended to do everything that a relational database does. Whenever
possible, the database server offloads processing and logic to the client side (handled
either by the drivers or by a user’s application code). Maintaining this streamlined design
is one of the reasons MongoDB can achieve such high performance.
Let’s Get Started
Throughout the course of the book, we will take the time to note the reasoning or
motivation behind particular decisions made in the development of MongoDB.
Through those notes we hope to share the philosophy behind MongoDB. The best way
to summarize the MongoDB project, however, is through its main focus—to create a
full-featured data store that is scalable, flexible, and fast.
…Without Sacrificing Speed | 5
CHAPTER 2
Getting Started
MongoDB is powerful but easy to get started with. In this chapter we’ll introduce some
of the basic concepts of MongoDB:
• A document is the basic unit of data for MongoDB and is roughly equivalent to a
row in a relational database management system (but much more expressive).
• Similarly, a collection can be thought of as a table with a dynamic schema.
•A single instance of MongoDB can host multiple independent databases, each of
which can have its own collections.
•Every document has a special key, "_id", that is unique within a collection.
•MongoDB comes with a simple but powerful JavaScript shell, which is useful for
the administration of MongoDB instances and data manipulation.
Documents
At the heart of MongoDB is the document: an ordered set of keys with associated values.
The representation of a document varies by programming language, but most languages
have a data structure that is a natural fit, such as a map, hash, or dictionary. In JavaScript,
for example, documents are represented as objects:
{"greeting" : "Hello, world!"}
This simple document contains a single key, "greeting", with a value of "Hello,
world!". Most documents will be more complex than this simple one and often will
contain multiple key/value pairs:
{"greeting" : "Hello, world!", "foo" : 3}
As you can see from the example above, values in documents are not just “blobs.” They
can be one of several different data types (or even an entire embedded document—see
7
“Embedded Documents” on page 19). In this example the value for "greeting" is a string,
whereas the value for "foo" is an integer.
The keys in a document are strings. Any UTF-8 character is allowed in a key, with a few
notable exceptions:
•Keys must not contain the character \0 (the null character). This character is used
to signify the end of a key.
•The . and $ characters have some special properties and should be used only in
certain circumstances, as described in later chapters. In general, they should be
considered reserved, and drivers will complain if they are used inappropriately.
MongoDB is type-sensitive and case-sensitive. For example, these documents are
distinct:
{"foo" : 3}
{"foo" : "3"}
as are as these:
{"foo" : 3}
{"Foo" : 3}
A final important thing to note is that documents in MongoDB cannot contain duplicate
keys. For example, the following is not a legal document:
{"greeting" : "Hello, world!", "greeting" : "Hello, MongoDB!"}
Key/value pairs in documents are ordered: {"x" : 1, "y" : 2} is not the same as
{"y" : 2, "x" : 1}. Field order does not usually matter and you should not design
your schema to depend on a certain ordering of fields (MongoDB may reorder them).
This text will note the special cases where field order is important.
In some programming languages the default representation of a document does not
even maintain ordering (e.g., dictionaries in Python and hashes in Perl or Ruby 1.8).
Drivers for those languages usually have some mechanism for specifying documents
with ordering, when necessary.
Collections
A collection is a group of documents. If a document is the MongoDB analog of a row in
a relational database, then a collection can be thought of as the analog to a table.
Dynamic Schemas
Collections have dynamic schemas. This means that the documents within a single col‐
lection can have any number of different “shapes.” For example, both of the following
documents could be stored in a single collection:
8 | Chapter 2: Getting Started
{"greeting" : "Hello, world!"}
{"foo" : 5}
Note that the previous documents not only have different types for their values (string
versus integer) but also have entirely different keys. Because any document can be put
into any collection, the question often arises: “Why do we need separate collections at
all?” It’s a good question—with no need for separate schemas for different kinds of
documents, why should we use more than one collection? There are several good
reasons:
•Keeping different kinds of documents in the same collection can be a nightmare
for developers and admins. Developers need to make sure that each query is only
returning documents of a certain type or that the application code performing a
query can handle documents of different shapes. If we’re querying for blog posts,
it’s a hassle to weed out documents containing author data.
•It is much faster to get a list of collections than to extract a list of the types in a
collection. For example, if we had a "type" field in each document that specified
whether the document was a “skim,” “whole,” or “chunky monkey,” it would be much
slower to find those three values in a single collection than to have three separate
collections and query the correct collection.
•Grouping documents of the same kind together in the same collection allows for
data locality. Getting several blog posts from a collection containing only posts will
likely require fewer disk seeks than getting the same posts from a collection con‐
taining posts and author data.
•We begin to impose some structure on our documents when we create indexes.
(This is especially true in the case of unique indexes.) These indexes are defined per
collection. By putting only documents of a single type into the same collection, we
can index our collections more efficiently.
As you can see, there are sound reasons for creating a schema and for grouping related
types of documents together, even though MongoDB does not enforce it.
Naming
A collection is identified by its name. Collection names can be any UTF-8 string, with
a few restrictions:
•The empty string ("") is not a valid collection name.
•Collection names may not contain the character \0 (the null character) because
this delineates the end of a collection name.
•You should not create any collections that start with system., a prefix reserved for
internal collections. For example, the system.users collection contains the database’s
Collections | 9
users, and the system.namespaces collection contains information about all of the
database’s collections.
•User-created collections should not contain the reserved character $ in the name.
The various drivers available for the database do support using $ in collection names
because some system-generated collections contain it. You should not use $ in a
name unless you are accessing one of these collections.
Subcollections
One convention for organizing collections is to use namespaced subcollections sepa‐
rated by the . character. For example, an application containing a blog might have a
collection named blog.posts and a separate collection named blog.authors. This is for
organizational purposes only—there is no relationship between the blog collection (it
doesn’t even have to exist) and its “children.”
Although subcollections do not have any special properties, they are useful and incor‐
porated into many MongoDB tools:
• GridFS, a protocol for storing large files, uses subcollections to store file metadata
separately from content chunks (see Chapter 6 for more information about GridFS).
• Most drivers provide some syntactic sugar for accessing a subcollection of a given
collection. For example, in the database shell, db.blog will give you the blog col‐
lection, and db.blog.posts will give you the blog.posts collection.
Subcollections are a great way to organize data in MongoDB, and their use is highly
recommended.
Databases
In addition to grouping documents by collection, MongoDB groups collections into
databases. A single instance of MongoDB can host several databases, each grouping
together zero or more collections. A database has its own permissions, and each database
is stored in separate files on disk. A good rule of thumb is to store all data for a single
application in the same database. Separate databases are useful when storing data for
several application or users on the same MongoDB server.
Like collections, databases are identified by name. Database names can be any UTF-8
string, with the following restrictions:
• The empty string ("") is not a valid database name.
•A database name cannot contain any of these characters: /, \, ., ", *, <, >, :, |, ?, $, (a
single space), or \0 (the null character). Basically, stick with alphanumeric ASCII.
10 | Chapter 2: Getting Started
•Database names are case-sensitive, even on non-case-sensitive filesystems. To keep
things simple, try to just use lowercase characters.
• Database names are limited to a maximum of 64 bytes.
One thing to remember about database names is that they will actually end up as files
on your filesystem. This explains why many of the previous restrictions exist in the first
place.
There are also several reserved database names, which you can access but which have
special semantics. These are as follows:
admin
This is the “root” database, in terms of authentication. If a user is added to the admin
database, the user automatically inherits permissions for all databases. There are
also certain server-wide commands that can be run only from the admin database,
such as listing all of the databases or shutting down the server.
localThis database will never be replicated and can be used to store any collections that
should be local to a single server (see Chapter 9 for more information about repli‐
cation and the local database).
config
When MongoDB is being used in a sharded setup (see Chapter 13), it uses the config
database to store information about the shards.
By concatenating a database name with a collection in that database you can get a fully
qualified collection name called a namespace. For instance, if you are using the blog.posts
collection in the cms database, the namespace of that collection would be
cms.blog.posts. Namespaces are limited to 121 bytes in length and, in practice, should
be fewer than 100 bytes long. For more on namespaces and the internal representation
of collections in MongoDB, see Appendix B.
Getting and Starting MongoDB
MongoDB is almost always run as a network server that clients can connect to and
perform operations on. Download MongoDB and decompress it. To start the server,
run the mongod executable:
$ mongod
mongod --help for help and startup options
Thu Oct 11 12:36:48 [initandlisten] MongoDB starting : pid=2425 port=27017
dbpath=/data/db/ 64-bit host=spock
Thu Oct 11 12:36:48 [initandlisten] db version v2.4.0, pdfile version 4.5
Thu Oct 11 12:36:48 [initandlisten] git version:
3aaea5262d761e0bb6bfef5351cfbfca7af06ec2
Thu Oct 11 12:36:48 [initandlisten] build info: Darwin spock 11.2.0 Darwin Kernel
Getting and Starting MongoDB | 11
Version 11.2.0: Tue Aug 9 20:54:00 PDT 2011;
root:xnu-1699.24.8~1/RELEASE_X86_64 x86_64 BOOST_LIB_VERSION=1_48
Thu Oct 11 12:36:48 [initandlisten] options: {}
Thu Oct 11 12:36:48 [initandlisten] journal dir=/data/db/journal
Thu Oct 11 12:36:48 [initandlisten] recover : no journal files present, no
recovery needed
Thu Oct 11 12:36:48 [websvr] admin web console waiting for connections on
port 28017
Thu Oct 11 12:36:48 [initandlisten] waiting for connections on port 27017
Or if you’re on Windows, run this:
$ mongod.exe
For detailed information on installing MongoDB on your system, see
Appendix A.
When run with no arguments, mongod will use the default data directory, /data/db/ (or
\data\db\ on the current volume on Windows). If the data directory does not already
exist or is not writable, the server will fail to start. It is important to create the data
directory (e.g., mkdir -p /data/db/) and to make sure your user has permission to write
to the directory before starting MongoDB.
On startup, the server will print some version and system information and then begin
waiting for connections. By default MongoDB listens for socket connections on port
27017. The server will fail to start if the port is not available—the most common cause
of this is another instance of MongoDB that is already running.
mongod also sets up a very basic HTTP server that listens on a port 1,000 higher than
the main port, in this case 28017. This means that you can get some administrative
information about your database by opening a web browser and going to http://local
host:28017.
You can safely stop mongod by typing Ctrl-C in the shell that is running the server.
For more information on starting or stopping MongoDB, see Chap‐
ter 20.
Introduction to the MongoDB Shell
MongoDB comes with a JavaScript shell that allows interaction with a MongoDB in‐
stance from the command line. The shell is useful for performing administrative
12 | Chapter 2: Getting Started
functions, inspecting a running instance, or just playing around. The mongo shell is a
crucial tool for using MongoDB and is used extensively throughout the rest of the text.
Running the Shell
To start the shell, run the mongo executable:
$ mongo
MongoDB shell version: 2.4.0
connecting to: test
>
The shell automatically attempts to connect to a MongoDB server on startup, so make
sure you start mongod before starting the shell.
The shell is a full-featured JavaScript interpreter, capable of running arbitrary JavaScript
programs. To illustrate this, let’s perform some basic math:
> x = 200
200
> x / 5;
40
We can also leverage all of the standard JavaScript libraries:
> Math.sin(Math.PI / 2);
1
> new Date("2010/1/1");
"Fri Jan 01 2010 00:00:00 GMT-0500 (EST)"
> "Hello, World!".replace("World", "MongoDB");
Hello, MongoDB!
We can even define and call JavaScript functions:
> function factorial (n) {
... if (n <= 1) return 1;
... return n * factorial(n - 1);
... }
> factorial(5);
120
Note that you can create multiline commands. The shell will detect whether the Java‐
Script statement is complete when you press Enter. If the statement is not complete, the
shell will allow you to continue writing it on the next line. Pressing Enter three times in
a row will cancel the half-formed command and get you back to the >-prompt.
A MongoDB Client
Although the ability to execute arbitrary JavaScript is cool, the real power of the shell
lies in the fact that it is also a standalone MongoDB client. On startup, the shell connects
to the test database on a MongoDB server and assigns this database connection to the
Introduction to the MongoDB Shell | 13
global variable db. This variable is the primary access point to your MongoDB server
through the shell.
To see the database db is currently assigned to, type in db and hit Enter:
> db
test
The shell contains some add-ons that are not valid JavaScript syntax but were imple‐
mented because of their familiarity to users of SQL shells. The add-ons do not provide
any extra functionality, but they are nice syntactic sugar. For instance, one of the most
important operations is selecting which database to use:
> use foobar
switched to db foobar
Now if you look at the db variable, you can see that it refers to the foobar database:
> db
foobar
Because this is a JavaScript shell, typing a variable will convert the variable to a string
(in this case, the database name) and print it.
Collections can be accessed from the db variable. For example, db.baz returns the baz
collection in the current database. Now that we can access a collection in the shell, we
can perform almost any database operation.
Basic Operations with the Shell
We can use the four basic operations, create, read, update, and delete (CRUD) to ma‐
nipulate and view data in the shell.
Create
The insert function adds a document to a collection. For example, suppose we want
to store a blog post. First, we’ll create a local variable called post that is a JavaScript
object representing our document. It will have the keys "title", "content", and "date"
(the date that it was published):
> post = {"title" : "My Blog Post",
... "content" : "Here's my blog post.",
... "date" : new Date()}
{
"title" : "My Blog Post",
"content" : "Here's my blog post.",
"date" : ISODate("2012-08-24T21:12:09.982Z")
}
This object is a valid MongoDB document, so we can save it to the blog collection using
the insert method:
14 | Chapter 2: Getting Started
> db.blog.insert(post)
The blog post has been saved to the database. We can see it by calling find on the
collection:
> db.blog.find()
{
"_id" : ObjectId("5037ee4a1084eb3ffeef7228"),
"title" : "My Blog Post",
"content" : "Here's my blog post.",
"date" : ISODate("2012-08-24T21:12:09.982Z")
}
You can see that an "_id" key was added and that the other key/value pairs were saved
as we entered them. The reason for the sudden appearance of the "_id" field is explained
at the end of this chapter.
Read
find and findOne can be used to query a collection. If we just want to see one document
from a collection, we can use findOne:
> db.blog.findOne()
{
"_id" : ObjectId("5037ee4a1084eb3ffeef7228"),
"title" : "My Blog Post",
"content" : "Here's my blog post.",
"date" : ISODate("2012-08-24T21:12:09.982Z")
}
find and findOne can also be passed criteria in the form of a query document. This will
restrict the documents matched by the query. The shell will automatically display up to
20 documents matching a find, but more can be fetched. See Chapter 4 for more in‐
formation on querying.
Update
If we would like to modify our post, we can use update. update takes (at least) two
parameters: the first is the criteria to find which document to update, and the second is
the new document. Suppose we decide to enable comments on the blog post we created
earlier. We’ll need to add an array of comments as the value for a new key in our
document.
The first step is to modify the variable post and add a "comments" key:
> post.comments = []
[ ]
Then we perform the update, replacing the post titled “My Blog Post” with our new
version of the document:
> db.blog.update({title : "My Blog Post"}, post)
Introduction to the MongoDB Shell | 15
Now the document has a "comments" key. If we call find again, we can see the new key:
> db.blog.find()
{
"_id" : ObjectId("5037ee4a1084eb3ffeef7228"),
"title" : "My Blog Post",
"content" : "Here's my blog post.",
"date" : ISODate("2012-08-24T21:12:09.982Z"),
"comments" : [ ]
}
Delete
remove permanently deletes documents from the database. Called with no parameters,
it removes all documents from a collection. It can also take a document specifying
criteria for removal. For example, this would remove the post we just created:
> db.blog.remove({title : "My Blog Post"})
Now the collection will be empty again.
Data Types
The beginning of this chapter covered the basics of what a document is. Now that you
are up and running with MongoDB and can try things on the shell, this section will dive
a little deeper. MongoDB supports a wide range of data types as values in documents.
In this section, we’ll outline all the supported types.
Basic Data Types
Documents in MongoDB can be thought of as “JSON-like” in that they are conceptually
similar to objects in JavaScript. JSON is a simple representation of data: the specification
can be described in about one paragraph (their website proves it) and lists only six data
types. This is a good thing in many ways: it’s easy to understand, parse, and remember.
On the other hand, JSON’s expressive capabilities are limited because the only types are
null, boolean, numeric, string, array, and object.
Although these types allow for an impressive amount of expressivity, there are a couple
of additional types that are crucial for most applications, especially when working with
a database. For example, JSON has no date type, which makes working with dates even
more annoying than it usually is. There is a number type, but only one—there is no way
to differentiate floats and integers, never mind any distinction between 32-bit and 64-
bit numbers. There is no way to represent other commonly used types, either, such as
regular expressions or functions.
MongoDB adds support for a number of additional data types while keeping JSON’s
essential key/value pair nature. Exactly how values of each type are represented varies
16 | Chapter 2: Getting Started
by language, but this is a list of the commonly supported types and how they are rep‐
resented as part of a document in the shell. The most common types are:
nullNull can be used to represent both a null value and a nonexistent field:
{"x" : null}
boolean
There is a boolean type, which can be used for the values true and false:
{"x" : true}
number
The shell defaults to using 64-bit floating point numbers. Thus, these numbers look
“normal” in the shell:
{"x" : 3.14}
or:
{"x" : 3}
For integers, use the NumberInt or NumberLong classes, which represent 4-byte or
8-byte signed integers, respectively.
{"x" : NumberInt("3")}
{"x" : NumberLong("3")}
string
Any string of UTF-8 characters can be represented using the string type:
{"x" : "foobar"}
dateDates are stored as milliseconds since the epoch. The time zone is not stored:
{"x" : new Date()}
regular expression
Queries can use regular expressions using JavaScript’s regular expression syntax:
{"x" : /foobar/i}
array
Sets or lists of values can be represented as arrays:
{"x" : ["a", "b", "c"]}
embedded document
Documents can contain entire documents embedded as values in a parent
document:
{"x" : {"foo" : "bar"}}
Data Types | 17
object id
An object id is a 12-byte ID for documents. See the section “_id and ObjectIds” on
page 20 for details:
{"x" : ObjectId()}
There are also a few less common types that you may need, including:
binary data
Binary data is a string of arbitrary bytes. It cannot be manipulated from the shell.
Binary data is the only way to save non-UTF-8 strings to the database.
codeQueries and documents can also contain arbitrary JavaScript code:
{"x" : function() { /* ... */ }}
There are a few types that are mostly used internally (or superseded by other types).
These will be described in the text as needed.
For more information on MongoDB’s data format, see Appendix B.
Dates
In JavaScript, the Date class is used for MongoDB’s date type. When creating a new Date
object, always call new Date(...), not just Date(...). Calling the constructor as a
function (that is, not including new) returns a string representation of the date, not an
actual Date object. This is not MongoDB’s choice; it is how JavaScript works. If you are
not careful to always use the Date constructor, you can end up with a mishmash of
strings and dates. Strings do not match dates and vice versa, so this can cause problems
with removing, updating, querying…pretty much everything.
For a full explanation of JavaScript’s Date class and acceptable formats for the con‐
structor, see ECMAScript specification section 15.9.
Dates in the shell are displayed using local time zone settings. However, dates in the
database are just stored as milliseconds since the epoch, so they have no time zone
information associated with them. (Time zone information could, of course, be stored
as the value for another key.)
Arrays
Arrays are values that can be interchangeably used for both ordered operations (as
though they were lists, stacks, or queues) and unordered operations (as though they
were sets).
In the following document, the key "things" has an array value:
{"things" : ["pie", 3.14]}
18 | Chapter 2: Getting Started
As we can see from the example, arrays can contain different data types as values (in
this case, a string and a floating-point number). In fact, array values can be any of the
supported values for normal key/value pairs, even nested arrays.
One of the great things about arrays in documents is that MongoDB “understands” their
structure and knows how to reach inside of arrays to perform operations on their con‐
tents. This allows us to query on arrays and build indexes using their contents. For
instance, in the previous example, MongoDB can query for all documents where 3.14
is an element of the "things" array. If this is a common query, you can even create an
index on the "things" key to improve the query’s speed.
MongoDB also allows atomic updates that modify the contents of arrays, such as reach‐
ing into the array and changing the value pie to pi. We’ll see more examples of these
types of operations throughout the text.
Embedded Documents
Documents can be used as the value for a key. This is called an embedded document.
Embedded documents can be used to organize data in a more natural way than just a
flat structure of key/value pairs.
For example, if we have a document representing a person and want to store his address,
we can nest this information in an embedded "address" document:
{
"name" : "John Doe",
"address" : {
"street" : "123 Park Street",
"city" : "Anytown",
"state" : "NY"
}
}
The value for the "address" key in the previous example is an embedded document
with its own key/value pairs for "street", "city", and "state".
As with arrays, MongoDB “understands” the structure of embedded documents and is
able to reach inside them to build indexes, perform queries, or make updates.
We’ll discuss schema design in depth later, but even from this basic example we can
begin to see how embedded documents can change the way we work with data. In a
relational database, the previous document would probably be modeled as two separate
rows in two different tables (one for “people” and one for “addresses”). With MongoDB
we can embed the address document directly within the person document. When used
properly, embedded documents can provide a more natural representation of
information.
Data Types | 19
The flip side of this is that there can be more data repetition with MongoDB. Suppose
“addresses” were a separate table in a relational database and we needed to fix a typo in
an address. When we did a join with “people” and “addresses,” we’d get the updated
address for everyone who shares it. With MongoDB, we’d need to fix the typo in each
person’s document.
_id and ObjectIds
Every document stored in MongoDB must have an "_id" key. The "_id" key’s value
can be any type, but it defaults to an ObjectId. In a single collection, every document
must have a unique value for "_id", which ensures that every document in a collection
can be uniquely identified. That is, if you had two collections, each one could have a
document where the value for "_id" was 123. However, neither collection could contain
more than one document with an "_id" of 123.
ObjectIds
ObjectId is the default type for "_id". The ObjectId class is designed to be lightweight,
while still being easy to generate in a globally unique way across different machines.
MongoDB’s distributed nature is the main reason why it uses ObjectIds as opposed to
something more traditional, like an autoincrementing primary key: it is difficult and
time-consuming to synchronize autoincrementing primary keys across multiple
servers. Because MongoDB was designed to be a distributed database, it was important
to be able to generate unique identifiers in a sharded environment.
ObjectIds use 12 bytes of storage, which gives them a string representation that is 24
hexadecimal digits: 2 digits for each byte. This causes them to appear larger than they
are, which makes some people nervous. It’s important to note that even though an
ObjectId is often represented as a giant hexadecimal string, the string is actually twice
as long as the data being stored.
If you create multiple new ObjectIds in rapid succession, you can see that only the last
few digits change each time. In addition, a couple of digits in the middle of the Objec
tId will change (if you space the creations out by a couple of seconds). This is because
of the manner in which ObjectIds are created. The 12 bytes of an ObjectId are generated
as follows:
0 1 2 3 4 5 6 7 8 9 10 11
Timestamp Machine PID Increment
The first four bytes of an ObjectId are a timestamp in seconds since the epoch. This
provides a couple of useful properties:
•The timestamp, when combined with the next five bytes (which will be described
in a moment), provides uniqueness at the granularity of a second.
20 | Chapter 2: Getting Started
•Because the timestamp comes first, it means that ObjectIds will sort in roughly
insertion order. This is not a strong guarantee but does have some nice properties,
such as making ObjectIds efficient to index.
•In these four bytes exists an implicit timestamp of when each document was created.
Most drivers expose a method for extracting this information from an ObjectId.
Because the current time is used in ObjectIds, some users worry that their servers will
need to have synchronized clocks. Although synchronized clocks are a good idea for
other reasons (see “Synchronizing Clocks” on page 383), the actual timestamp doesn’t
matter to ObjectIds, only that it is often new (once per second) and increasing.
The next three bytes of an ObjectId are a unique identifier of the machine on which it
was generated. This is usually a hash of the machine’s hostname. By including these
bytes, we guarantee that different machines will not generate colliding ObjectIds.
To provide uniqueness among different processes generating ObjectIds concurrently
on a single machine, the next two bytes are taken from the process identifier (PID) of
the ObjectId-generating process.
These first nine bytes of an ObjectId guarantee its uniqueness across machines and
processes for a single second. The last three bytes are simply an incrementing counter
that is responsible for uniqueness within a second in a single process. This allows for
up to 2563 (16,777,216) unique ObjectIds to be generated per process in a single second.
Autogeneration of _id
As stated previously, if there is no "_id" key present when a document is inserted, one
will be automatically added to the inserted document. This can be handled by the Mon‐
goDB server but will generally be done by the driver on the client side. The decision to
generate them on the client side reflects an overall philosophy of MongoDB: work
should be pushed out of the server and to the drivers whenever possible. This philosophy
reflects the fact that, even with scalable databases like MongoDB, it is easier to scale out
at the application layer than at the database layer. Moving work to the client side reduces
the burden requiring the database to scale.
Using the MongoDB Shell
This section covers how to use the shell as part of your command line toolkit, customize
it, and use some of its more advanced functionality.
Although we connected to a local mongod instance above, you can connect your shell
to any MongoDB instance that your machine can reach. To connect to a mongod on a
different machine or port, specify the hostname, port, and database when starting the
shell:
Using the MongoDB Shell | 21
$ mongo some-host:30000/myDB
MongoDB shell version: 2.4.0
connecting to: some-host:30000/myDB
>
db will now refer to some-host:30000’s myDB database.
Sometimes it is handy to not connect to a mongod at all when starting the mongo shell.
If you start the shell with --nodb, it will start up without attempting to connect to
anything:
$ mongo --nodb
MongoDB shell version: 2.4.0
>
Once started, you can connect to a mongod at your leisure by running new Mon
go(hostname):
> conn = new Mongo("some-host:30000")
connection to some-host:30000
> db = conn.getDB("myDB")
myDB
After these two commands, you can use db normally. You can use these commands to
connect to a different database or server at any time.
Tips for Using the Shell
Because mongo is simply a JavaScript shell, you can get a great deal of help for it by simply
looking up JavaScript documentation online. For MongoDB-specific functionality, the
shell includes built-in help that can be accessed by typing help:
> help
db.help() help on db methods
db.mycoll.help() help on collection methods
sh.help() sharding helpers
...
show dbs show database names
show collections show collections in current database
show users show users in current database
...
Database-level help is provided by db.help() and collection-level help by
db.foo.help().
A good way of figuring out what a function is doing is to type it without the parentheses.
This will print the JavaScript source code for the function. For example, if we are curious
about how the update function works or cannot remember the order of parameters, we
can do the following:
22 | Chapter 2: Getting Started
> db.foo.update
function (query, obj, upsert, multi) {
assert(query, "need a query");
assert(obj, "need an object");
this._validateObject(obj);
this._mongo.update(this._fullName, query, obj,
upsert ? true : false, multi ? true : false);
}
Running Scripts with the Shell
Other chapters have used the shell interactively, but you can also pass the shell JavaScript
files to execute. Simply pass in your scripts at the command line:
$ mongo script1.js script2.js script3.js
MongoDB shell version: 2.4.0
connecting to: test
I am script1.js
I am script2.js
I am script3.js
$
The mongo shell will execute each script listed and exit.
If you want to run a script using a connection to a non-default host/port mongod, specify
the address first, then the script(s):
$ mongo --quiet server-1:30000/foo script1.js script2.js script3.js
This would execute the three scripts with db set to the foo database on server-1:30000.
As shown above, any command line options for running the shell go before the address.
You can print to stdout in scripts (as the scripts above did) using the print() function.
This allows you to use the shell as part of a pipeline of commands. If you’re planning to
pipe the output of a shell script to another command use the --quiet option to prevent
the “MongoDB shell version...” banner from printing.
You can also run scripts from within the interactive shell using the load() function:
> load("script1.js")
I am script1.js
>
Scripts have access to the db variable (as well as any other globals). However, shell helpers
such as "use db" or "show collections" do not work from files. There are valid Java‐
Script equivalents to each of these, as shown in Table 2-1.
Using the MongoDB Shell | 23
Table 2-1. JavaScript equivalents to shell helpers
Helper Equivalent
use foo db.getSisterDB("foo")
show dbs db.getMongo().getDBs()
show collections db.getCollectionNames()
You can also use scripts to inject variables into the shell. For example, we could have a
script that simply initializes helper functions that you commonly use. The script below,
for instance, may be helpful for the replication and sharding sections of the book. It
defines a function, connectTo(), that connects to the locally-running database on the
given port and sets db to that connection:
// defineConnectTo.js
/**
* Connect to a database and set db.
*/
var connectTo = function(port, dbname) {
if (!port) {
port = 27017;
}
if (!dbname) {
dbname = "test";
}
db = connect("localhost:"+port+"/"+dbname);
return db;
};
If we load this script in the shell, connectTo is now defined:
> typeof connectTo
undefined
> load('defineConnectTo.js')
> typeof connectTo
function
In addition to adding helper functions, you can use scripts to automate common tasks
and administrative activities.
By default, the shell will look in the directory that you started the shell in (use
run("pwd") to see what directory that is). If the script is not in your current directory,
you can give the shell a relative or absolute path to it. For example, if you wanted to put
your shell scripts in ~/my-scripts, you could load defineConnectTo.js with load("/home/
myUser/my-scripts/defineConnectTo.js"). Note that load cannot resolve ~.
You can use run() to run command-line programs from the shell. Pass arguments to
the function as parameters:
24 | Chapter 2: Getting Started
> run("ls", "-l", "/home/myUser/my-scripts/")
sh70352| -rw-r--r-- 1 myUser myUser 2012-12-13 13:15 defineConnectTo.js
sh70532| -rw-r--r-- 1 myUser myUser 2013-02-22 15:10 script1.js
sh70532| -rw-r--r-- 1 myUser myUser 2013-02-22 15:12 script2.js
sh70532| -rw-r--r-- 1 myUser myUser 2013-02-22 15:13 script3.js
This is of limited use, generally, as the output is formatted oddly and it doesn’t support
pipes.
Creating a .mongorc.js
If you have frequently-loaded scripts you might want to put them in your mongorc.js
file. This file is run whenever you start up the shell.
For example, suppose we would like the shell to greet us when we log in. Create a file
called .mongorc.js in your home directory, and then add the following lines to it:
// mongorc.js
var compliment = ["attractive", "intelligent", "like Batman"];
var index = Math.floor(Math.random()*3);
print("Hello, you're looking particularly "+compliment[index]+" today!");
Then, when you start the shell, you’ll see something like:
$ mongo
MongoDB shell version: 2.4.0-pre-
connecting to: test
Hello, you're looking particularly like Batman today!
>
More practically, you can use this script to set up any global variables you’d like to use,
alias long names to shorter ones, and override built-in functions. One of the most com‐
mon uses for .mongorc.js is remove some of the more “dangerous” shell helpers. You
can override functions like dropDatabase or deleteIndexes with no-ops or undefine
them altogether:
var no = function() {
print("Not on my watch.");
};
// Prevent dropping databases
db.dropDatabase = DB.prototype.dropDatabase = no;
// Prevent dropping collections
DBCollection.prototype.drop = no;
// Prevent dropping indexes
DBCollection.prototype.dropIndex = no;
Using the MongoDB Shell | 25
Make sure that, if you change any database functions, you do so on both the db variable
and the DB prototype (as shown in the example above). If you change only one, either
the db variable won’t see the change or all new databases you use (when you run use
anotherDB) won’t see your change.
Now if you try to call any of these functions, it will simply print an error message. Note
that this technique does not protect you against malicious users; it can only help with
fat-fingering.
You can disable loading your .mongorc.js by using the --norc option when starting the
shell.
Customizing Your Prompt
The default shell prompt can be overridden by setting the prompt variable to either a
string or a function. For example, if you are running a query that takes minutes to
complete, you may want to have a prompt that prints the current time when it is drawn
so you can see when the last operation finished:
prompt = function() {
return (new Date())+"> ";
};
Another handy prompt might show the current database you’re using:
prompt = function() {
if (typeof db == 'undefined') {
return '(nodb)> ';
}
// Check the last db operation
try {
db.runCommand({getLastError:1});
}
catch (e) {
print(e);
}
return db+"> ";
};
Note that prompt functions should return strings and be very cautious about catching
exceptions: it can be very confusing if your prompt turns into an exception!
In general, your prompt function should include a call to getLastError. This catches
errors on writes and reconnects you “automatically” if the shell gets disconnected (e.g.,
if you restart mongod).
The .mongorc.js file is a good place to set your prompt if you want to always use a custom
one (or set up a couple of custom prompts that you can switch between in the shell).
26 | Chapter 2: Getting Started
Editing Complex Variables
The multiline support in the shell is somewhat limited: you cannot edit previous lines,
which can be annoying when you realize that the first line has a typo and you’re currently
working on line 15. Thus, for larger blocks of code or objects, you may want to edit them
in an editor. To do so, set the EDITOR variable in the shell (or in your environment, but
since you’re already in the shell):
> EDITOR="/usr/bin/emacs"
Now, if you want to edit a variable, you can say "edit varname", for example:
> var wap = db.books.findOne({title: "War and Peace"})
> edit wap
When you’re done making changes, save and exit the editor. The variable will be parsed
and loaded back into the shell.
Add EDITOR="/path/to/editor"; to your .mongorc.js file and you won’t have to worry
about setting it again.
Inconvenient Collection Names
Fetching a collection with the db.collectionName syntax almost always works, unless
the collection name a reserved word or is an invalid JavaScript property name.
For example, suppose we are trying to access the version collection. We cannot say
db.version because db.version is a method on db (it returns the version of the running
MongoDB server):
> db.version
function () {
return this.serverBuildInfo().version;
}
To actually access the version collection, you must use the getCollection function:
> db.getCollection("version");
test.version
This can also be used for collection names with characters that aren’t valid in JavaScript
property names, such as foo-bar-baz and 123abc (JavaScript property names can only
contain letters, numbers, "$" and "_" and cannot start with a number).
Another way of getting around invalid properties is to use array-access syntax: in Java‐
Script, x.y is identical to x['y']. This means that subcollections can be accessed using
variables, not just literal names. Thus, if you needed to perform some operation on every
blog subcollection, you could iterate through them with something like this:
var collections = ["posts", "comments", "authors"];
Using the MongoDB Shell | 27
for (var i in collections) {
print(db.blog[collections[i]]);
}
instead of this:
print(db.blog.posts);
print(db.blog.comments);
print(db.blog.authors);
Note that you cannot do db.blog.i, which would be interpreted as test.blog.i, not
test.blog.posts. You must use the db.blog[i] syntax for i to be interpreted as a variable.
You can use this technique to access awkwardly-named collections:
> var name = "@#&!"
> db[name].find()
Attempting to query db.@#&! would be illegal, but db[name] would work.
28 | Chapter 2: Getting Started
CHAPTER 3
Creating, Updating, and
Deleting Documents
This chapter covers the basics of moving data in and out of the database, including the
following:
• Adding new documents to a collection
•Removing documents from a collection
•Updating existing documents
•Choosing the correct level of safety versus speed for all of these operations
Inserting and Saving Documents
Inserts are the basic method for adding data to MongoDB. To insert a document into a
collection, use the collection’s insert method:
> db.foo.insert({"bar" : "baz"})
This will add an "_id" key to the document (if one does not already exist) and store it
in MongoDB.
Batch Insert
If you have a situation where you are inserting multiple documents into a collection,
you can make the insert faster by using batch inserts. Batch inserts allow you to pass an
array of documents to the database.
In the shell, you can try this out using the batchInsert function, which is similar to
insert except that it takes an array of documents to insert:
29
> db.foo.batchInsert([{"_id" : 0}, {"_id" : 1}, {"_id" : 2}])
> db.foo.find()
{ "_id" : 0 }
{ "_id" : 1 }
{ "_id" : 2 }
Sending dozens, hundreds, or even thousands of documents at a time can make inserts
significantly faster.
Batch inserts are only useful if you are inserting multiple documents into a single col‐
lection: you cannot use batch inserts to insert into multiple collections with a single
request. If you are just importing raw data (for example, from a data feed or MySQL),
there are command-line tools like mongoimport that can be used instead of batch insert.
On the other hand, it is often handy to munge data before saving it to MongoDB (con‐
verting dates to the date type or adding a custom "_id") so batch inserts can be used
for importing data, as well.
Current versions of MongoDB do not accept messages longer than 48 MB, so there is
a limit to how much can be inserted in a single batch insert. If you attempt to insert
more than 48 MB, many drivers will split up the batch insert into multiple 48 MB batch
inserts. Check your driver documentation for details.
If you are importing a batch and a document halfway through the batch fails to be
inserted, the documents up to that document will be inserted and everything after that
document will not:
> db.foo.batchInsert([{"_id" : 0}, {"_id" : 1}, {"_id" : 1}, {"_id" : 2}])
Only the first two documents will be inserted, as the third will produce an error: you
cannot insert two documents with the same "_id".
If you want to ignore errors and make batchInsert attempt to insert the rest of the
batch, you can use the continueOnError option to continue after an insert failure. This
would insert the first, second, and fourth documents above. The shell does not support
this option, but all the drivers do.
Insert Validation
MongoDB does minimal checks on data being inserted: it check’s the document’s basic
structure and adds an "_id" field if one does not exist. One of the basic structure checks
is size: all documents must be smaller than 16 MB. This is a somewhat arbitrary limit
(and may be raised in the future); it is mostly to prevent bad schema design and ensure
consistent performance. To see the BSON size (in bytes) of the document doc, run
Object.bsonsize(doc) from the shell.
To give you an idea of how much data 16 MB is, the entire text of War and Peace is just
3.14 MB.
30 | Chapter 3: Creating, Updating, and Deleting Documents
These minimal checks also mean that it is fairly easy to insert invalid data (if you are
trying to). Thus, you should only allow trusted sources, such as your application servers,
to connect to the database. All of the drivers for major languages (and most of the minor
ones, too) do check for a variety of invalid data (documents that are too large, contain
non-UTF-8 strings, or use unrecognized types) before sending anything to the database.
Removing Documents
Now that there’s data in our database, let’s delete it:
> db.foo.remove()
This will remove all of the documents in the foo collection. This doesn’t actually remove
the collection, and any meta information about it will still exist.
The remove function optionally takes a query document as a parameter. When it’s given,
only documents that match the criteria will be removed. Suppose, for instance, that we
want to remove everyone from the mailing.list collection where the value for "opt-
out" is true:
> db.mailing.list.remove({"opt-out" : true})
Once data has been removed, it is gone forever. There is no way to undo the remove or
recover deleted documents.
Remove Speed
Removing documents is usually a fairly quick operation, but if you want to clear an
entire collection, it is faster to drop it (and then recreate any indexes on the empty
collection).
For example, suppose we insert a million dummy elements with the following:
> for (var i = 0; i < 1000000; i++) {
... db.tester.insert({"foo": "bar", "baz": i, "z": 10 - i})
... }
Now we’ll try to remove all of the documents we just inserted, measuring the time it
takes. First, here’s a simple remove:
> var timeRemoves = function() {
... var start = (new Date()).getTime();
...
... db.tester.remove();
... db.findOne(); // makes sure the remove finishes before continuing
...
... var timeDiff = (new Date()).getTime() - start;
... print("Remove took: "+timeDiff+"ms");
... }
> timeRemoves()
Removing Documents | 31
On a MacBook Air, this script prints “Remove took: 9676ms”.
If the remove and findOne are replaced by db.tester.drop(), the time drops to one
millisecond! This is obviously a vast improvement, but it comes at the expense of gran‐
ularity: we cannot specify any criteria. The whole collection is dropped, and all of its
metadata is deleted.
Updating Documents
Once a document is stored in the database, it can be changed using the update method.
update takes two parameters: a query document, which locates documents to update,
and a modifier document, which describes the changes to make to the documents found.
Updating a document is atomic: if two updates happen at the same time, whichever one
reaches the server first will be applied, and then the next one will be applied. Thus,
conflicting updates can safely be sent in rapid-fire succession without any documents
being corrupted: the last update will “win.”
Document Replacement
The simplest type of update fully replaces a matching document with a new one. This
can be useful to do a dramatic schema migration. For example, suppose we are making
major changes to a user document, which looks like the following:
{
"_id" : ObjectId("4b2b9f67a1f631733d917a7a"),
"name" : "joe",
"friends" : 32,
"enemies" : 2
}
We want to move the "friends" and "enemies" fields to a "relationships" subdo‐
cument. We can change the structure of the document in the shell and then replace the
database’s version with an update:
> var joe = db.users.findOne({"name" : "joe"});
> joe.relationships = {"friends" : joe.friends, "enemies" : joe.enemies};
{
"friends" : 32,
"enemies" : 2
}> joe.username = joe.name;
"joe"
> delete joe.friends;
true
> delete joe.enemies;
true
> delete joe.name;
true
> db.users.update({"name" : "joe"}, joe);
32 | Chapter 3: Creating, Updating, and Deleting Documents
Now, doing a findOne shows that the structure of the document has been updated:
{
"_id" : ObjectId("4b2b9f67a1f631733d917a7a"),
"username" : "joe",
"relationships" : {
"friends" : 32,
"enemies" : 2
}
}
A common mistake is matching more than one document with the criteria and then
creating a duplicate "_id" value with the second parameter. The database will throw an
error for this, and no documents will be updated.
For example, suppose we create several documents with the same value for "name", but
we don’t realize it:
> db.people.find()
{"_id" : ObjectId("4b2b9f67a1f631733d917a7b"), "name" : "joe", "age" : 65},
{"_id" : ObjectId("4b2b9f67a1f631733d917a7c"), "name" : "joe", "age" : 20},
{"_id" : ObjectId("4b2b9f67a1f631733d917a7d"), "name" : "joe", "age" : 49},
Now, if it’s Joe #2’s birthday, we want to increment the value of his "age" key, so we
might say this:
> joe = db.people.findOne({"name" : "joe", "age" : 20});
{
"_id" : ObjectId("4b2b9f67a1f631733d917a7c"),
"name" : "joe",
"age" : 20
}
> joe.age++;
> db.people.update({"name" : "joe"}, joe);
E11001 duplicate key on update
What happened? When you call update, the database will look for a document matching
{"name" : "joe"}. The first one it finds will be the 65-year-old Joe. It will attempt to
replace that document with the one in the joe variable, but there’s already a document
in this collection with the same "_id". Thus, the update will fail, because "_id" values
must be unique. The best way to avoid this situation is to make sure that your update
always specifies a unique document, perhaps by matching on a key like "_id". For the
example above, this would be the correct update to use:
> db.people.update({"_id" : ObjectId("4b2b9f67a1f631733d917a7c")}, joe)
Using "_id" for the criteria will also be faster than querying on random fields, as "_id"
is indexed. We’ll cover how indexing effects updates and other operations more in
Chapter 5.
Updating Documents | 33
Using Modifiers
Usually only certain portions of a document need to be updated. You can update specific
fields in a document using atomic update modifiers. Update modifiers are special keys
that can be used to specify complex update operations, such as altering, adding, or
removing keys, and even manipulating arrays and embedded documents.
Suppose we were keeping website analytics in a collection and wanted to increment a
counter each time someone visited a page. We can use update modifiers to do this
increment atomically. Each URL and its number of page views is stored in a document
that looks like this:
{
"_id" : ObjectId("4b253b067525f35f94b60a31"),
"url" : "www.example.com",
"pageviews" : 52
}
Every time someone visits a page, we can find the page by its URL and use the "$inc"
modifier to increment the value of the "pageviews" key:
> db.analytics.update({"url" : "www.example.com"},
... {"$inc" : {"pageviews" : 1}})
Now, if we do a find, we see that "pageviews" has increased by one:
> db.analytics.find()
{
"_id" : ObjectId("4b253b067525f35f94b60a31"),
"url" : "www.example.com",
"pageviews" : 53
}
When using modifiers, the value of "_id" cannot be changed. (Note that "_id" can be
changed by using whole-document replacement.) Values for any other key, including
other uniquely indexed keys, can be modified.
Getting started with the “$set” modifier
"$set" sets the value of a field. If the field does not yet exist, it will be created. This can
be handy for updating schema or adding user-defined keys. For example, suppose you
have a simple user profile stored as a document that looks something like the following:
> db.users.findOne()
{
"_id" : ObjectId("4b253b067525f35f94b60a31"),
"name" : "joe",
"age" : 30,
"sex" : "male",
"location" : "Wisconsin"
}
34 | Chapter 3: Creating, Updating, and Deleting Documents
This is a pretty bare-bones user profile. If the user wanted to store his favorite book in
his profile, he could add it using "$set":
> db.users.update({"_id" : ObjectId("4b253b067525f35f94b60a31")},
... {"$set" : {"favorite book" : "War and Peace"}})
Now the document will have a “favorite book” key:
> db.users.findOne()
{
"_id" : ObjectId("4b253b067525f35f94b60a31"),
"name" : "joe",
"age" : 30,
"sex" : "male",
"location" : "Wisconsin",
"favorite book" : "War and Peace"
}
If the user decides that he actually enjoys a different book, "$set" can be used again to
change the value:
> db.users.update({"name" : "joe"},
... {"$set" : {"favorite book" : "Green Eggs and Ham"}})
"$set" can even change the type of the key it modifies. For instance, if our fickle user
decides that he actually likes quite a few books, he can change the value of the “favorite
book” key into an array:
> db.users.update({"name" : "joe"},
... {"$set" : {"favorite book" :
... ["Cat's Cradle", "Foundation Trilogy", "Ender's Game"]}})
If the user realizes that he actually doesn’t like reading, he can remove the key altogether
with "$unset":
> db.users.update({"name" : "joe"},
... {"$unset" : {"favorite book" : 1}})
Now the document will be the same as it was at the beginning of this example.
You can also use "$set" to reach in and change embedded documents:
> db.blog.posts.findOne()
{
"_id" : ObjectId("4b253b067525f35f94b60a31"),
"title" : "A Blog Post",
"content" : "...",
"author" : {
"name" : "joe",
"email" : "joe@example.com"
}
}
> db.blog.posts.update({"author.name" : "joe"},
... {"$set" : {"author.name" : "joe schmoe"}})
Updating Documents | 35
> db.blog.posts.findOne()
{
"_id" : ObjectId("4b253b067525f35f94b60a31"),
"title" : "A Blog Post",
"content" : "...",
"author" : {
"name" : "joe schmoe",
"email" : "joe@example.com"
}
}
You must always use a $-modifier for adding, changing, or removing keys. A common
error people make when starting out is to try to set the value of "foo" to "bar" by doing
an update that looks like this:
> db.coll.update(criteria, {"foo" : "bar"})
This will not function as intended. It actually does a full-document replacement, re‐
placing the matched document with {"foo" : "bar"}. Always use $ operators for
modifying individual key/value pairs.
Incrementing and decrementing
The "$inc" modifier can be used to change the value for an existing key or to create a
new key if it does not already exist. It is very useful for updating analytics, karma, votes,
or anything else that has a changeable, numeric value.
Suppose we are creating a game collection where we want to save games and update
scores as they change. When a user starts playing, say, a game of pinball, we can insert
a document that identifies the game by name and user playing it:
> db.games.insert({"game" : "pinball", "user" : "joe"})
When the ball hits a bumper, the game should increment the player’s score. As points
in pinball are given out pretty freely, let’s say that the base unit of points a player can
earn is 50. We can use the "$inc" modifier to add 50 to the player’s score:
> db.games.update({"game" : "pinball", "user" : "joe"},
... {"$inc" : {"score" : 50}})
If we look at the document after this update, we’ll see the following:
> db.games.findOne()
{
"_id" : ObjectId("4b2d75476cc613d5ee930164"),
"game" : "pinball",
"user" : "joe",
"score" : 50
}
36 | Chapter 3: Creating, Updating, and Deleting Documents
The score key did not already exist, so it was created by "$inc" and set to the increment
amount: 50.
If the ball lands in a “bonus” slot, we want to add 10,000 to the score. This can be
accomplished by passing a different value to "$inc":
> db.games.update({"game" : "pinball", "user" : "joe"},
... {"$inc" : {"score" : 10000}})
Now if we look at the game, we’ll see the following:
> db.games.find()
{
"_id" : ObjectId("4b2d75476cc613d5ee930164"),
"game" : "pinball",
"user" : "joe",
"score" : 10050
}
The "score" key existed and had a numeric value, so the server added 10,000 to it.
"$inc" is similar to "$set", but it is designed for incrementing (and decrementing)
numbers. "$inc" can be used only on values of type integer, long, or double. If it is used
on any other type of value, it will fail. This includes types that many languages will
automatically cast into numbers, like nulls, booleans, or strings of numeric characters:
> db.foo.insert({"count" : "1"})
> db.foo.update({}, {"$inc" : {"count" : 1}})
Cannot apply $inc modifier to non-number
Also, the value of the "$inc" key must be a number. You cannot increment by a string,
array, or other non-numeric value. Doing so will give a “Modifier "$inc" allowed for
numbers only” error message. To modify other types, use "$set" or one of the following
array operations.
Array modifiers
An extensive class of modifiers exists for manipulating arrays. Arrays are common and
powerful data structures: not only are they lists that can be referenced by index, but they
can also double as sets.
Adding elements
"$push" adds elements to the end of an array if the array exists and creates a new array
if it does not. For example, suppose that we are storing blog posts and want to add a
"comments" key containing an array. We can push a comment onto the nonexistent
"comments" array, which will create the array and add the comment:
> db.blog.posts.findOne()
{
"_id" : ObjectId("4b2d75476cc613d5ee930164"),
Updating Documents | 37
"title" : "A blog post",
"content" : "..."
}
> db.blog.posts.update({"title" : "A blog post"},
... {"$push" : {"comments" :
... {"name" : "joe", "email" : "joe@example.com",
... "content" : "nice post."}}})
> db.blog.posts.findOne()
{
"_id" : ObjectId("4b2d75476cc613d5ee930164"),
"title" : "A blog post",
"content" : "...",
"comments" : [
{
"name" : "joe",
"email" : "joe@example.com",
"content" : "nice post."
}
]
}
Now, if we want to add another comment, we can simply use "$push" again:
> db.blog.posts.update({"title" : "A blog post"},
... {"$push" : {"comments" :
... {"name" : "bob", "email" : "bob@example.com",
... "content" : "good post."}}})
> db.blog.posts.findOne()
{
"_id" : ObjectId("4b2d75476cc613d5ee930164"),
"title" : "A blog post",
"content" : "...",
"comments" : [
{
"name" : "joe",
"email" : "joe@example.com",
"content" : "nice post."
},
{
"name" : "bob",
"email" : "bob@example.com",
"content" : "good post."
}
]
}
This is the “simple” form of push, but you can use it for more complex array operations
as well. You can push multiple values in one operation using the "$each" suboperator:
> db.stock.ticker.update({"_id" : "GOOG"},
... {"$push" : {"hourly" : {"$each" : [562.776, 562.790, 559.123]}}})
38 | Chapter 3: Creating, Updating, and Deleting Documents
This would push three new elements onto the array. Specify a single-element array to
get equivalent behavior to the non-$each form of "$push".
If you only want the array to grow to a certain length, you can also use the "$slice"
operator in conjunction with "$push" to prevent an array from growing beyond a certain
size, effectively making a “top N” list of items:
> db.movies.find({"genre" : "horror"},
... {"$push" : {"top10" : {
... "$each" : ["Nightmare on Elm Street", "Saw"],
... "$slice" : -10}}})
This example would limit the array to the last 10 elements pushed. Slices must always
be negative numbers.
If the array was smaller than 10 elements (after the push), all elements would be kept.
If the array was larger than 10 elements, only the last 10 elements would be kept. Thus,
"$slice" can be used to create a queue in a document.
Finally, you can "$sort" before trimming, so long as you are pushing subobjects onto
the array:
> db.movies.find({"genre" : "horror"},
... {"$push" : {"top10" : {
... "$each" : [{"name" : "Nightmare on Elm Street", "rating" : 6.6},
... {"name" : "Saw", "rating" : 4.3}],
... "$slice" : -10,
... "$sort" : {"rating" : -1}}}})
This will sort all of the objects in the array by their "rating" field and then keep the
first 10. Note that you must include "$each"; you cannot just "$slice" or "$sort" an
array with "$push".
Using arrays as sets
You might want to treat an array as a set, only adding values if they are not present. This
can be done using a "$ne" in the query document. For example, to push an author onto
a list of citations, but only if he isn’t already there, use the following:
> db.papers.update({"authors cited" : {"$ne" : "Richie"}},
... {$push : {"authors cited" : "Richie"}})
This can also be done with "$addToSet", which is useful for cases where "$ne" won’t
work or where "$addToSet" describes what is happening better.
For instance, suppose you have a document that represents a user. You might have a set
of email addresses that they have added:
> db.users.findOne({"_id" : ObjectId("4b2d75476cc613d5ee930164")})
{
"_id" : ObjectId("4b2d75476cc613d5ee930164"),
"username" : "joe",
Updating Documents | 39
"emails" : [
"joe@example.com",
"joe@gmail.com",
"joe@yahoo.com"
]
}
When adding another address, you can use "$addToSet" to prevent duplicates:
> db.users.update({"_id" : ObjectId("4b2d75476cc613d5ee930164")},
... {"$addToSet" : {"emails" : "joe@gmail.com"}})
> db.users.findOne({"_id" : ObjectId("4b2d75476cc613d5ee930164")})
{
"_id" : ObjectId("4b2d75476cc613d5ee930164"),
"username" : "joe",
"emails" : [
"joe@example.com",
"joe@gmail.com",
"joe@yahoo.com",
]
}
> db.users.update({"_id" : ObjectId("4b2d75476cc613d5ee930164")},
... {"$addToSet" : {"emails" : "joe@hotmail.com"}})
> db.users.findOne({"_id" : ObjectId("4b2d75476cc613d5ee930164")})
{
"_id" : ObjectId("4b2d75476cc613d5ee930164"),
"username" : "joe",
"emails" : [
"joe@example.com",
"joe@gmail.com",
"joe@yahoo.com",
"joe@hotmail.com"
]
}
You can also use "$addToSet" in conjunction with "$each" to add multiple unique
values, which cannot be done with the "$ne"/"$push" combination. For instance, we
could use these modifiers if the user wanted to add more than one email address:
> db.users.update({"_id" : ObjectId("4b2d75476cc613d5ee930164")}, {"$addToSet" :
... {"emails" : {"$each" :
... ["joe@php.net", "joe@example.com", "joe@python.org"]}}})
> db.users.findOne({"_id" : ObjectId("4b2d75476cc613d5ee930164")})
{
"_id" : ObjectId("4b2d75476cc613d5ee930164"),
"username" : "joe",
"emails" : [
"joe@example.com",
"joe@gmail.com",
"joe@yahoo.com",
"joe@hotmail.com"
"joe@php.net"
"joe@python.org"
40 | Chapter 3: Creating, Updating, and Deleting Documents
]
}
Removing elements
There are a few ways to remove elements from an array. If you want to treat the array
like a queue or a stack, you can use "$pop", which can remove elements from either
end. {"$pop" : {"key" : 1}} removes an element from the end of the array. {"$pop" :
{"key" : -1}} removes it from the beginning.
Sometimes an element should be removed based on specific criteria, rather than its
position in the array. "$pull" is used to remove elements of an array that match the
given criteria. For example, suppose we have a list of things that need to be done but
not in any specific order:
> db.lists.insert({"todo" : ["dishes", "laundry", "dry cleaning"]})
If we do the laundry first, we can remove it from the list with the following:
> db.lists.update({}, {"$pull" : {"todo" : "laundry"}})
Now if we do a find, we’ll see that there are only two elements remaining in the array:
> db.lists.find()
{
"_id" : ObjectId("4b2d75476cc613d5ee930164"),
"todo" : [
"dishes",
"dry cleaning"
]
}
Pulling removes all matching documents, not just a single match. If you have an array
that looks like [1, 1, 2, 1] and pull 1, you’ll end up with a single-element array, [2].
Array operators can be used only on keys with array values. For example, you cannot
push on to an integer or pop off of a string, for example. Use "$set" or "$inc" to modify
scalar values.
Positional array modifications
Array manipulation becomes a little trickier when we have multiple values in an array
and want to modify some of them. There are two ways to manipulate values in arrays:
by position or by using the position operator (the "$" character).
Arrays use 0-based indexing, and elements can be selected as though their index were
a document key. For example, suppose we have a document containing an array with a
few embedded documents, such as a blog post with comments:
> db.blog.posts.findOne()
{
"_id" : ObjectId("4b329a216cc613d5ee930192"),
Updating Documents | 41
"content" : "...",
"comments" : [
{
"comment" : "good post",
"author" : "John",
"votes" : 0
},
{
"comment" : "i thought it was too short",
"author" : "Claire",
"votes" : 3
},
{
"comment" : "free watches",
"author" : "Alice",
"votes" : -1
}
]
}
If we want to increment the number of votes for the first comment, we can say the
following:
> db.blog.update({"post" : post_id},
... {"$inc" : {"comments.0.votes" : 1}})
In many cases, though, we don’t know what index of the array to modify without query‐
ing for the document first and examining it. To get around this, MongoDB has a posi‐
tional operator, "$", that figures out which element of the array the query document
matched and updates that element. For example, if we have a user named John who
updates his name to Jim, we can replace it in the comments by using the positional
operator:
db.blog.update({"comments.author" : "John"},
... {"$set" : {"comments.$.author" : "Jim"}})
The positional operator updates only the first match. Thus, if John had left more than
one comment, his name would be changed only for the first comment he left.
Modifier speed
Some modifiers are faster than others. $inc modifies a document in place: it does not
have to change the size of a document, only a couple of bytes, so it is very efficient. On
the other hand, array modifiers might change the size of a document and can be slow.
("$set" can modify documents in place if the size isn’t changing but otherwise is subject
to the same performance limitations as array operators.)
When you start inserting documents into MongoDB, it puts each document right next
to the previous one on disk. Thus, if a document gets bigger, it will no longer fit in the
space it was originally written to and will be moved to another part of the collection.
42 | Chapter 3: Creating, Updating, and Deleting Documents
You can see this in action by creating a new collection with just a few documents and
then making a document that is sandwiched between two other documents larger. It
will be bumped to the end of the collection:
> db.coll.insert({"x" :"a"})
> db.coll.insert({"x" :"b"})
> db.coll.insert({"x" :"c"})
> db.coll.find()
{ "_id" : ObjectId("507c3581d87d6a342e1c81d3"), "x" : "a" }
{ "_id" : ObjectId("507c3583d87d6a342e1c81d4"), "x" : "b" }
{ "_id" : ObjectId("507c3585d87d6a342e1c81d5"), "x" : "c" }
> db.coll.update({"x" : "b"}, {$set: {"x" : "bbb"}})
> db.coll.find()
{ "_id" : ObjectId("507c3581d87d6a342e1c81d3"), "x" : "a" }
{ "_id" : ObjectId("507c3585d87d6a342e1c81d5"), "x" : "c" }
{ "_id" : ObjectId("507c3583d87d6a342e1c81d4"), "x" : "bbb" }
When MongoDB has to move a document, it bumps the collection’s padding factor,
which is the amount of extra space MongoDB leaves around new documents to give
them room to grow. You can see the padding factor by running db.coll.stats(). Be‐
fore doing the update above, the "paddingFactor" field will be 1: allocate exactly the
size of the document for each new document, as shown in Figure 3-1. If you run it again
after making one of the documents larger (as shown in Figure 3-2), you’ll see that it has
grown to around 1.5: each new document will be given half of its size in free space to
grow. If subsequent updates cause more moves, the padding factor will continue to grow
(although not as dramatically as it did on the first move). If there aren’t more moves,
the padding factor will slowly go down, as shown in Figure 3-3.
Figure 3-1. Initially, documents are inserted with no space between them
Figure 3-2. If a document grows and must be moved, free space is left behind and the
padding size in increased
Updating Documents | 43
Figure 3-3. Subsequent documents are inserted with the padding factor in free space be‐
tween them. If moves do not occur on subsequent inserts, this padding factor will
decrease.
Moving documents is slow. MongoDB has to free the space the document was in and
write the document somewhere else. Thus, you should try to keep the padding factor
as close to 1 as possible. You cannot manually set the padding factor (unless you’re
compacting the collection: see “Compacting Data” on page 320), but you can design a
schema that does not depend on documents growing arbitrarily large. See Chapter 8 for
more advice on schema design.
A simple program demonstrates the speed difference between in-place updates and
moves. The program below inserts a single key and increments its value 100,000 times:
> db.tester.insert({"x" : 1})
> var timeInc = function() {
... var start = (new Date()).getTime();
...
... for (var i=0; i<100000; i++) {
... db.tester.update({}, {"$inc" : {"x" : 1}});
... db.getLastError();
... }
...
... var timeDiff = (new Date()).getTime() - start;
... print("Updates took: "+timeDiff+"ms");
... }
> timeInc()
On a MacBook Air this took 7.33 seconds. That’s more than 13,000 updates per second.
Now, let’s try it with a document with a single array key, pushing new values onto that
array 100,000 times. Change the update call to this:
... db.tester.update({}, {"$push" : {"x" : 1}})
This program took 67.58 seconds to run, which is less than 1,500 updates per second.
Using "$push" and other array modifiers is encouraged and often necessary, but it is
good to keep in mind the trade-offs of such updates. If "$push" becomes a bottleneck,
it may be worth pulling an embedded array out into a separate collection, manually
padding, or using one of the other techniques discussed in Chapter 8.
As of this writing, MongoDB is not great at reusing empty space, so moving documents
around a lot can result in large swaths of empty data file. If you have a lot of empty space,
you’ll start seeing messages that look like this in the logs:
Thu Apr 5 01:12:28 [conn124727] info DFM::findAll(): extent a:7f18dc00 was
empty, skipping ahead
44 | Chapter 3: Creating, Updating, and Deleting Documents
That means that, while querying, MongoDB looked through an entire extent (see Ap‐
pendix B for a definition, but it’s basically a subset of your collection) without finding
any documents: it was just empty space. The message itself is harmless, but it indicates
that you have fragmentation and may wish to perform a compact.
If your schema requires lots of moves or lots of churn through inserts and deletes, you
can improve disk reuse by using the usePowerOf2Sizes option. You can set this with
the collMod command:
> db.runCommand({"collMod" : collectionName, "usePowerOf2Sizes" : true})
All subsequent allocations made by the collection will be in power-of-two-sized blocks.
Only use this option on high-churn collections, though, as this makes initial space al‐
location less efficient. Setting this on an insert- or in-place-update-only collection will
make writes slower.
Running this command with "usePowerOf2Sizes" : false turns off the special allo‐
cation. The option only affects newly allocated records, so there is no harm in running
it on an existing collection or toggling the value.
Upserts
An upsert is a special type of update. If no document is found that matches the update
criteria, a new document will be created by combining the criteria and updated docu‐
ments. If a matching document is found, it will be updated normally. Upserts can be
handy because they can eliminate the need to “seed” your collection: you can often have
the same code create and update documents.
Let’s go back to our example that records the number of views for each page of a website.
Without an upsert, we might try to find the URL and increment the number of views
or create a new document if the URL doesn’t exist. If we were to write this out as a
JavaScript program it might look something like the following:
// check if we have an entry for this page
blog = db.analytics.findOne({url : "/blog"})
// if we do, add one to the number of views and save
if (blog) {
blog.pageviews++;
db.analytics.save(blog);
}
// otherwise, create a new document for this page
else {
db.analytics.save({url : "/blog", pageviews : 1})
}
This means we are making a round trip to the database, plus sending an update or insert,
every time someone visits a page. If we are running this code in multiple processes, we
Updating Documents | 45
are also subject to a race condition where more than one document can be inserted for
a given URL.
We can eliminate the race condition and cut down on the amount of code by just sending
an upsert (the third parameter to update specifies that this should be an upsert):
db.analytics.update({"url" : "/blog"}, {"$inc" : {"pageviews" : 1}}, true)
This line does exactly what the previous code block does, except it’s faster and atomic!
The new document is created using the criteria document as a base and applying any
modifier documents to it.
For example, if you do an upsert that matches a key and has an increment to the value
of that key, the increment will be applied to the match:
> db.users.update({"rep" : 25}, {"$inc" : {"rep" : 3}}, true)
> db.users.findOne()
{
"_id" : ObjectId("4b3295f26cc613d5ee93018f"),
"rep" : 28
}
The upsert creates a new document with a "rep" of 25 and then increments that by 3,
giving us a document where "rep" is 28. If the upsert option were not specified, {"rep" :
25} would not match any documents, so nothing would happen.
If we run the upsert again (with the criteria {"rep" : 25}), it will create another new
document. This is because the criteria does not match the only document in the col‐
lection. (Its "rep" is 28.)
Sometimes a field needs to be seeded when a document is created, but not changed on
subsequent updates. This is what "$setOnInsert" is for. "$setOnInsert" is a modifier
that only sets the value of a field when the document is being inserted. Thus, we could
do something like this:
> db.users.update({}, {"$setOnInsert" : {"createdAt" : new Date()}}, true)
> db.users.findOne()
{
"_id" : ObjectId("512b8aefae74c67969e404ca"),
"createdAt" : ISODate("2013-02-25T16:01:50.742Z")
}
If we run this update again, it will match the existing document, nothing will be inserted,
and so the "createdAt" field will not be changed:
> db.users.update({}, {"$setOnInsert" : {"createdAt" : new Date()}}, true)
> db.users.findOne()
{
"_id" : ObjectId("512b8aefae74c67969e404ca"),
"createdAt" : ISODate("2013-02-25T16:01:50.742Z")
}
46 | Chapter 3: Creating, Updating, and Deleting Documents
Note that you generally do not need to keep a "createdAt" field, as ObjectIds contain
a timestamp of when the document was created. However, "$setOnInsert" can be
useful for creating padding, initializing counters, and for collections that do not use
ObjectIds.
The save shell helper
save is a shell function that lets you insert a document if it doesn’t exist and update it if
it does. It takes one argument: a document. If the document contains an "_id" key, save
will do an upsert. Otherwise, it will do an insert. save is really just a convenience function
so that programmers can quickly modify documents in the shell:
> var x = db.foo.findOne()
> x.num = 42
42
> db.foo.save(x)
Without save, the last line would have been a more cumbersome db.foo.up
date({"_id" : x._id}, x).
Updating Multiple Documents
Updates, by default, update only the first document found that matches the criteria. If
there are more matching documents, they will remain unchanged. To modify all of the
documents matching the criteria, you can pass true as the fourth parameter to update.
update’s behavior may be changed in the future (the server may update
all matching documents by default and update one only if false is
passed as the fourth parameter), so it is recommended that you always
specify whether you want a multiple update.
Not only is it more obvious what the update should be doing, but your
program also won’t break if the default is ever changed.
Multiupdates are a great way of performing schema migrations or rolling out new fea‐
tures to certain users. Suppose, for example, we want to give a gift to every user who
has a birthday on a certain day. We can use multiupdate to add a "gift" to their account:
> db.users.update({"birthday" : "10/13/1978"},
... {"$set" : {"gift" : "Happy Birthday!"}}, false, true)
This would add the "gift" field to all user documents with birthdays on October 13,
1978.
To see the number of documents updated by a multiupdate, you can run the
getLastError database command (which you can think of as “return information about
Updating Documents | 47
the last operation”). The "n" key will contain the number of documents affected by the
update:
> db.count.update({x : 1}, {$inc : {x : 1}}, false, true)
> db.runCommand({getLastError : 1})
{
"err" : null,
"updatedExisting" : true,
"n" : 5,
"ok" : true
}
"n" is 5, meaning that five documents were affected by the update. "updatedExist
ing" is true, meaning that the update modified existing documents.
Returning Updated Documents
You can get some limited information about what was updated by calling
getLastError, but it does not actually return the updated document. For that, you’ll
need the findAndModify command. It is handy for manipulating queues and perform‐
ing other operations that need get-and-set style atomicity.
Suppose we have a collection of processes run in a certain order. Each is represented
with a document that has the following form:
{
"_id" : ObjectId(),
"status" : state,
"priority" : N
}
"status" is a string that can be "READY", "RUNNING", or "DONE". We need to find the job
with the highest priority in the "READY" state, run the process function, and then update
the status to "DONE". We might try querying for the ready processes, sorting by priority,
and updating the status of the highest-priority process to mark it is "RUNNING". Once
we have processed it, we update the status to "DONE". This looks something like the
following:
var cursor = db.processes.find({"status" : "READY"});
ps = cursor.sort({"priority" : -1}).limit(1).next();
db.processes.update({"_id" : ps._id}, {"$set" : {"status" : "RUNNING"}});
do_something(ps);
db.processes.update({"_id" : ps._id}, {"$set" : {"status" : "DONE"}});
This algorithm isn’t great because it is subject to a race condition. Suppose we have two
threads running. If one thread (call it A) retrieved the document and another thread
(call it B) retrieved the same document before A had updated its status to "RUNNING",
then both threads would be running the same process. We can avoid this by checking
the status as part of the update query, but this becomes complex:
48 | Chapter 3: Creating, Updating, and Deleting Documents
var cursor = db.processes.find({"status" : "READY"});
cursor.sort({"priority" : -1}).limit(1);
while ((ps = cursor.next()) != null) {
ps.update({"_id" : ps._id, "status" : "READY"},
{"$set" : {"status" : "RUNNING"}});
var lastOp = db.runCommand({getlasterror : 1});
if (lastOp.n == 1) {
do_something(ps);
db.processes.update({"_id" : ps._id}, {"$set" : {"status" : "DONE"}})
break;
}
cursor = db.processes.find({"status" : "READY"});
cursor.sort({"priority" : -1}).limit(1);
}
Also, depending on timing, one thread may end up doing all the work while another
thread uselessly trails it. Thread A could always grab the process, and then B would try
to get the same process, fail, and leave A to do all the work.
Situations like this are perfect for findAndModify. findAndModify can return the item
and update it in a single operation. In this case, it looks like the following:
> ps = db.runCommand({"findAndModify" : "processes",
... "query" : {"status" : "READY"},
... "sort" : {"priority" : -1},
... "update" : {"$set" : {"status" : "RUNNING"}})
{
"ok" : 1,
"value" : {
"_id" : ObjectId("4b3e7a18005cab32be6291f7"),
"priority" : 1,
"status" : "READY"
}
}
Notice that the status is still "READY" in the returned document as findAndModify
defaults to returning the document in its pre-modified state. If you do a find on the
collection, though, you can see that the document’s "status" has been updated to
"RUNNING":
> db.processes.findOne({"_id" : ps.value._id})
{
"_id" : ObjectId("4b3e7a18005cab32be6291f7"),
"priority" : 1,
"status" : "RUNNING"
}
Thus, the program becomes the following:
ps = db.runCommand({"findAndModify" : "processes",
"query" : {"status" : "READY"},
"sort" : {"priority" : -1},
Updating Documents | 49
"update" : {"$set" : {"status" : "RUNNING"}}}).value
do_something(ps)
db.process.update({"_id" : ps._id}, {"$set" : {"status" : "DONE"}})
findAndModify can have either an "update" key or a "remove" key. A "remove" key
indicates that the matching document should be removed from the collection. For in‐
stance, if we wanted to simply remove the job instead of updating its status, we could
run the following:
ps = db.runCommand({"findAndModify" : "processes",
"query" : {"status" : "READY"},
"sort" : {"priority" : -1},
"remove" : true}).value
do_something(ps)
The findAndModify command has the following fields:
findAndModify
A string, the collection name
query
A query document; the criteria with which to search for documents
sort
Criteria by which to sort results (optional)
update
A modifier document; the update to perform on the document found (either this
or "remove" must be specified)
remove
Boolean specifying whether the document should be removed (either this or "up
date" must be specified)
new
Boolean specifying whether the document returned should be the updated docu‐
ment or the pre-update document, to which it defaults
fields
The fields of the document to return (optional)
upsert
Boolean specifying whether or not this should be an upsert, and which defaults to
false
Either "update" or "remove" must be included, but not both. If no matching document
is found, the command will return an error.
50 | Chapter 3: Creating, Updating, and Deleting Documents
Setting a Write Concern
Write concern is a client setting used to describe how safely a write should be stored
before the application continues. By default, inserts, removes, and updates wait for a
database response—did the write succeed or not?—before continuing. Generally, clients
will throw an exception (or whatever the language’s version of an exception is) on failure.
There are a number of options available to tune exactly what you want the application
to wait for. The two basic write concerns are acknowledged or unacknowledged writes.
Acknowledged writes are the default: you get a response that tells you whether or not
the database successfully processed your write. Unacknowledged writes do not return
any response, so you do not know if the write succeeded or not.
In general, applications should stick with acknowledged writes. However, for low-value
data (e.g., logs or bulk data loading), you may not want to wait for a response you don’t
care about. In these situations, use unacknowledged writes.
Although unacknowledged writes will not return database errors, they do not eliminate
the need for error checking in your application. If the socket was closed or there was an
error writing to it, attempting a write will cause an exception.
One type of error that is easy to miss when using unacknowledged writes is inserting
invalid data. For example, if we attempt to insert two documents with the same "_id",
the shell will throw an exception:
> db.foo.insert({"_id" : 1})
> db.foo.insert({"_id" : 1})
E11000 duplicate key error index: test.foo.$_id_ dup key: { : 1.0 }
Were the second write sent with “unacknowledged” write concern, the second insert
would not throw an exception. Duplicate key exceptions are a common source of errors,
but there are many others, from invalid $-modifiers to running out of disk space.
The shell does not actually support write concerns in the same way that the client li‐
braries do: it does unacknowledged writes and then checks that the last operation was
successful before drawing the prompt. Thus, if you do a series of invalid operations on
a collection, finishing with a valid operation, the shell will not complain:
> db.foo.insert({"_id" : 1}); db.foo.insert({"_id" : 1}); db.foo.count()
1
You can manually force a check in the shell by calling getLastError, which checks for an
error on the last operation:
> db.foo.insert({"_id" : 1}); db.foo.insert({"_id" : 1}); print(
... db.getLastError()); db.foo.count()
E11000 duplicate key error index: test.foo.$_id_ dup key: { : 1.0 }
1
This can be helpful when scripting for the shell.
Setting a Write Concern | 51
There are actually several other write concern options that are covered in later chapters:
Chapter 11 covers write concern for multiple servers and Chapter 19 covers committing
to disk on a per-write basis.
The default write concern was changed in 2012, so legacy code may
behave differently. Prior to the change, writes were unacknowledged by
default.
Fortunately, there is an easy way to tell if you’re using code written
before or after the write concern switch: all of the drivers began using
a class called MongoClient when they began defaulting to safe writes. If
your program is using a connection object called Mongo or Connec
tion or something else, you are using the old, default-unsafe API. No
language used MongoClient as a class name prior to the switch, so if
your code is using that, your writes are safe.
If you are using non-MongoClient connections, you should changed
unacknowledged writes to acknowledged writes wherever possible in
old code.
52 | Chapter 3: Creating, Updating, and Deleting Documents
CHAPTER 4
Querying
This chapter looks at querying in detail. The main areas covered are as follows:
•You can perform ad hoc queries on the database using the find or findOne functions
and a query document.
•You can query for ranges, set inclusion, inequalities, and more by using
$-conditionals.
•Queries return a database cursor, which lazily returns batches of documents as you
need them.
•There are a lot of metaoperations you can perform on a cursor, including skipping
a certain number of results, limiting the number of results returned, and sorting
results.
Introduction to find
The find method is used to perform queries in MongoDB. Querying returns a subset
of documents in a collection, from no documents at all to the entire collection. Which
documents get returned is determined by the first argument to find, which is a docu‐
ment specifying the query criteria.
An empty query document (i.e., {}) matches everything in the collection. If find isn’t
given a query document, it defaults to {}. For example, the following:
> db.c.find()
matches every document in the collection c (and returns these documents in batches).
When we start adding key/value pairs to the query document, we begin restricting our
search. This works in a straightforward way for most types: numbers match numbers,
booleans match booleans, and strings match strings. Querying for a simple type is as
53
easy as specifying the value that you are looking for. For example, to find all documents
where the value for "age" is 27, we can add that key/value pair to the query document:
> db.users.find({"age" : 27})
If we have a string we want to match, such as a "username" key with the value "joe",
we use that key/value pair instead:
> db.users.find({"username" : "joe"})
Multiple conditions can be strung together by adding more key/value pairs to the query
document, which gets interpreted as “condition1 AND condition2 AND … AND
conditionN.” For instance, to get all users who are 27-year-olds with the username “joe,”
we can query for the following:
> db.users.find({"username" : "joe", "age" : 27})
Specifying Which Keys to Return
Sometimes you do not need all of the key/value pairs in a document returned. If this is
the case, you can pass a second argument to find (or findOne) specifying the keys you
want. This reduces both the amount of data sent over the wire and the time and memory
used to decode documents on the client side.
For example, if you have a user collection and you are interested only in the "user
name" and "email" keys, you could return just those keys with the following query:
> db.users.find({}, {"username" : 1, "email" : 1})
{
"_id" : ObjectId("4ba0f0dfd22aa494fd523620"),
"username" : "joe",
"email" : "joe@example.com"
}
As you can see from the previous output, the "_id" key is returned by default, even if
it isn’t specifically requested.
You can also use this second parameter to exclude specific key/value pairs from the
results of a query. For instance, you may have documents with a variety of keys, and the
only thing you know is that you never want to return the "fatal_weakness" key:
> db.users.find({}, {"fatal_weakness" : 0})
This can also prevent "_id" from being returned:
> db.users.find({}, {"username" : 1, "_id" : 0})
{
"username" : "joe",
}
54 | Chapter 4: Querying
Limitations
There are some restrictions on queries. The value of a query document must be a con‐
stant as far as the database is concerned. (It can be a normal variable in your own code.)
That is, it cannot refer to the value of another key in the document. For example, if we
were keeping inventory and we had both "in_stock" and "num_sold" keys, we couldn’t
compare their values by querying the following:
> db.stock.find({"in_stock" : "this.num_sold"}) // doesn't work
There are ways to do this (see “$where Queries” on page 65), but you will usually get
better performance by restructuring your document slightly, such that a “normal” query
will suffice. In this example, we could instead use the keys "initial_stock" and
"in_stock". Then, every time someone buys an item, we decrement the value of the
"in_stock" key by one. Finally, we can do a simple query to check which items are out
of stock:
> db.stock.find({"in_stock" : 0})
Query Criteria
Queries can go beyond the exact matching described in the previous section; they can
match more complex criteria, such as ranges, OR-clauses, and negation.
Query Conditionals
"$lt", "$lte", "$gt", and "$gte" are all comparison operators, corresponding to <,
<=, >, and >=, respectively. They can be combined to look for a range of values. For
example, to look for users who are between the ages of 18 and 30, we can do this:
> db.users.find({"age" : {"$gte" : 18, "$lte" : 30}})
This would find all documents where the "age" field was greater than or equal to 18
AND less than or equal to 30.
These types of range queries are often useful for dates. For example, to find people who
registered before January 1, 2007, we can do this:
> start = new Date("01/01/2007")
> db.users.find({"registered" : {"$lt" : start}})
An exact match on a date is less useful, since dates are only stored with millisecond
precision. Often you want a whole day, week, or month, making a range query
necessary.
To query for documents where a key’s value is not equal to a certain value, you must use
another conditional operator, "$ne", which stands for “not equal.” If you want to find
all users who do not have the username “joe,” you can query for them using this:
Query Criteria | 55
> db.users.find({"username" : {"$ne" : "joe"}})
"$ne" can be used with any type.
OR Queries
There are two ways to do an OR query in MongoDB. "$in" can be used to query for a
variety of values for a single key. "$or" is more general; it can be used to query for any
of the given values across multiple keys.
If you have more than one possible value to match for a single key, use an array of criteria
with "$in". For instance, suppose we were running a raffle and the winning ticket
numbers were 725, 542, and 390. To find all three of these documents, we can construct
the following query:
> db.raffle.find({"ticket_no" : {"$in" : [725, 542, 390]}})
"$in" is very flexible and allows you to specify criteria of different types as well as values.
For example, if we are gradually migrating our schema to use usernames instead of user
ID numbers, we can query for either by using this:
> db.users.find({"user_id" : {"$in" : [12345, "joe"]})
This matches documents with a "user_id" equal to 12345, and documents with a
"user_id" equal to "joe".
If "$in" is given an array with a single value, it behaves the same as directly matching
the value. For instance, {ticket_no : {$in : [725]}} matches the same documents
as {ticket_no : 725}.
The opposite of "$in" is "$nin", which returns documents that don’t match any of the
criteria in the array. If we want to return all of the people who didn’t win anything in
the raffle, we can query for them with this:
> db.raffle.find({"ticket_no" : {"$nin" : [725, 542, 390]}})
This query returns everyone who did not have tickets with those numbers.
"$in" gives you an OR query for a single key, but what if we need to find documents
where "ticket_no" is 725 or "winner" is true? For this type of query, we’ll need to use
the "$or" conditional. "$or" takes an array of possible criteria. In the raffle case, using
"$or" would look like this:
> db.raffle.find({"$or" : [{"ticket_no" : 725}, {"winner" : true}]})
"$or" can contain other conditionals. If, for example, we want to match any of the three
"ticket_no" values or the "winner" key, we can use this:
> db.raffle.find({"$or" : [{"ticket_no" : {"$in" : [725, 542, 390]}},
{"winner" : true}]})
56 | Chapter 4: Querying
With a normal AND-type query, you want to narrow down your results as far as possible
in as few arguments as possible. OR-type queries are the opposite: they are most efficient
if the first arguments match as many documents as possible.
While "$or" will always work, use "$in" whenever possible as the query optimizer
handles it more efficiently.
$not
"$not" is a metaconditional: it can be applied on top of any other criteria. As an example,
let’s consider the modulus operator, "$mod". "$mod" queries for keys whose values, when
divided by the first value given, have a remainder of the second value:
> db.users.find({"id_num" : {"$mod" : [5, 1]}})
The previous query returns users with "id_num"s of 1, 6, 11, 16, and so on. If we want,
instead, to return users with "id_num"s of 2, 3, 4, 5, 7, 8, 9, 10, 12, etc., we can use "$not":
> db.users.find({"id_num" : {"$not" : {"$mod" : [5, 1]}}})
"$not" can be particularly useful in conjunction with regular expressions to find all
documents that don’t match a given pattern (regular expression usage is described in
the section “Regular Expressions” on page 58).
Conditional Semantics
If you look at the update modifiers in the previous chapter and previous query docu‐
ments, you’ll notice that the $-prefixed keys are in different positions. In the query,
"$lt" is in the inner document; in the update, "$inc" is the key for the outer document.
This generally holds true: conditionals are an inner document key, and modifiers are
always a key in the outer document.
Multiple conditions can be put on a single key. For example, to find all users between
the ages of 20 and 30, we can query for both "$gt" and "$lt" on the "age" key:
> db.users.find({"age" : {"$lt" : 30, "$gt" : 20}})
Any number of conditionals can be used with a single key. Multiple update modifiers
cannot be used on a single key, however. For example, you cannot have a modifier
document such as {"$inc" : {"age" : 1}, "$set" : {age : 40}} because it
modifies "age" twice. With query conditionals, no such rule applies.
There are a few “meta-operators” that go in the outer document: "$and“, "$or“, and
"$nor“. They all have a similar form:
> db.users.find({"$and" : [{"x" : {"$lt" : 1}}, {"x" : 4}]})
This query would match documents with an "x" field both less than 1 and equal to 4.
Although these seem like contradictory conditions, it is possible to fulfill if the "x" field
Query Criteria | 57
is an array: {"x" : [0, 4]} would match. Note that the query optimizer does not
optimize "$and" as well as other operators. This query would be more efficient to struc‐
ture as:
> db.users.find({"x" : {"$lt" : 1, "$in" : [4]}})
Type-Specific Queries
As covered in Chapter 2, MongoDB has a wide variety of types that can be used in a
document. Some of these types have special behavior when querying.
null
null behaves a bit strangely. It does match itself, so if we have a collection with the
following documents:
> db.c.find()
{ "_id" : ObjectId("4ba0f0dfd22aa494fd523621"), "y" : null }
{ "_id" : ObjectId("4ba0f0dfd22aa494fd523622"), "y" : 1 }
{ "_id" : ObjectId("4ba0f148d22aa494fd523623"), "y" : 2 }
we can query for documents whose "y" key is null in the expected way:
> db.c.find({"y" : null})
{ "_id" : ObjectId("4ba0f0dfd22aa494fd523621"), "y" : null }
However, null not only matches itself but also matches “does not exist.” Thus, querying
for a key with the value null will return all documents lacking that key:
> db.c.find({"z" : null})
{ "_id" : ObjectId("4ba0f0dfd22aa494fd523621"), "y" : null }
{ "_id" : ObjectId("4ba0f0dfd22aa494fd523622"), "y" : 1 }
{ "_id" : ObjectId("4ba0f148d22aa494fd523623"), "y" : 2 }
If we only want to find keys whose value is null, we can check that the key is null and
exists using the "$exists" conditional:
> db.c.find({"z" : {"$in" : [null], "$exists" : true}})
Unfortunately, there is no "$eq" operator, which makes this a little awkward, but "$in"
with one element is equivalent.
Regular Expressions
Regular expressions are useful for flexible string matching. For example, if we want to
find all users with the name Joe or joe, we can use a regular expression to do case-
insensitive matching:
> db.users.find({"name" : /joe/i})
58 | Chapter 4: Querying
Regular expression flags (for example, i) are allowed but not required. If we want to
match not only various capitalizations of joe, but also joey, we can continue to improve
our regular expression:
> db.users.find({"name" : /joey?/i})
MongoDB uses the Perl Compatible Regular Expression (PCRE) library to match reg‐
ular expressions; any regular expression syntax allowed by PCRE is allowed in
MongoDB. It is a good idea to check your syntax with the JavaScript shell before using
it in a query to make sure it matches what you think it matches.
MongoDB can leverage an index for queries on prefix regular expres‐
sions (e.g., /^joey/). Indexes cannot be used for case-insensitive
searches (/^joey/i).
Regular expressions can also match themselves. Very few people insert regular expres‐
sions into the database, but if you insert one, you can match it with itself:
> db.foo.insert({"bar" : /baz/})
> db.foo.find({"bar" : /baz/})
{
"_id" : ObjectId("4b23c3ca7525f35f94b60a2d"),
"bar" : /baz/
}
Querying Arrays
Querying for elements of an array is designed to behave the way querying for scalars
does. For example, if the array is a list of fruits, like this:
> db.food.insert({"fruit" : ["apple", "banana", "peach"]})
the following query:
> db.food.find({"fruit" : "banana"})
will successfully match the document. We can query for it in much the same way as we
would if we had a document that looked like the (illegal) document: {"fruit" : "ap
ple", "fruit" : "banana", "fruit" : "peach"}.
$all
If you need to match arrays by more than one element, you can use "$all". This allows
you to match a list of elements. For example, suppose we created a collection with three
elements:
> db.food.insert({"_id" : 1, "fruit" : ["apple", "banana", "peach"]})
> db.food.insert({"_id" : 2, "fruit" : ["apple", "kumquat", "orange"]})
> db.food.insert({"_id" : 3, "fruit" : ["cherry", "banana", "apple"]})
Type-Specific Queries | 59
Then we can find all documents with both "apple" and "banana" elements by querying
with "$all":
> db.food.find({fruit : {$all : ["apple", "banana"]}})
{"_id" : 1, "fruit" : ["apple", "banana", "peach"]}
{"_id" : 3, "fruit" : ["cherry", "banana", "apple"]}
Order does not matter. Notice "banana" comes before "apple" in the second result.
Using a one-element array with "$all" is equivalent to not using "$all". For instance,
{fruit : {$all : ['apple']} will match the same documents as {fruit : 'apple'}.
You can also query by exact match using the entire array. However, exact match will not
match a document if any elements are missing or superfluous. For example, this will
match the first document above:
> db.food.find({"fruit" : ["apple", "banana", "peach"]})
But this will not:
> db.food.find({"fruit" : ["apple", "banana"]})
and neither will this:
> db.food.find({"fruit" : ["banana", "apple", "peach"]})
If you want to query for a specific element of an array, you can specify an index using
the syntax key.index:
> db.food.find({"fruit.2" : "peach"})
Arrays are always 0-indexed, so this would match the third array element against the
string "peach".
$size
A useful conditional for querying arrays is "$size", which allows you to query for arrays
of a given size. Here’s an example:
> db.food.find({"fruit" : {"$size" : 3}})
One common query is to get a range of sizes. "$size" cannot be combined with another
$ conditional (in this example, "$gt"), but this query can be accomplished by adding a
"size" key to the document. Then, every time you add an element to the array, incre‐
ment the value of "size". If the original update looked like this:
> db.food.update(criteria, {"$push" : {"fruit" : "strawberry"}})
it can simply be changed to this:
> db.food.update(criteria,
... {"$push" : {"fruit" : "strawberry"}, "$inc" : {"size" : 1}})
Incrementing is extremely fast, so any performance penalty is negligible. Storing docu‐
ments like this allows you to do queries such as this:
60 | Chapter 4: Querying
> db.food.find({"size" : {"$gt" : 3}})
Unfortunately, this technique doesn’t work as well with the "$addToSet" operator.
The $slice operator
As mentioned earlier in this chapter, the optional second argument to find specifies the
keys to be returned. The special "$slice" operator can be used to return a subset of
elements for an array key.
For example, suppose we had a blog post document and we wanted to return the first
10 comments:
> db.blog.posts.findOne(criteria, {"comments" : {"$slice" : 10}})
Alternatively, if we wanted the last 10 comments, we could use −10:
> db.blog.posts.findOne(criteria, {"comments" : {"$slice" : -10}})
"$slice" can also return pages in the middle of the results by taking an offset and the
number of elements to return:
> db.blog.posts.findOne(criteria, {"comments" : {"$slice" : [23, 10]}})
This would skip the first 23 elements and return the 24th through 33th. If there were
fewer than 33 elements in the array, it would return as many as possible.
Unless otherwise specified, all keys in a document are returned when "$slice" is used.
This is unlike the other key specifiers, which suppress unmentioned keys from being
returned. For instance, if we had a blog post document that looked like this:
{
"_id" : ObjectId("4b2d75476cc613d5ee930164"),
"title" : "A blog post",
"content" : "...",
"comments" : [
{
"name" : "joe",
"email" : "joe@example.com",
"content" : "nice post."
},
{
"name" : "bob",
"email" : "bob@example.com",
"content" : "good post."
}
]
}
and we did a "$slice" to get the last comment, we’d get this:
> db.blog.posts.findOne(criteria, {"comments" : {"$slice" : -1}})
{
"_id" : ObjectId("4b2d75476cc613d5ee930164"),
Type-Specific Queries | 61
"title" : "A blog post",
"content" : "...",
"comments" : [
{
"name" : "bob",
"email" : "bob@example.com",
"content" : "good post."
}
]
}
Both "title" and "content" are still returned, even though they weren’t explicitly
included in the key specifier.
Returning a matching array element
"$slice" is helpful when you know the index of the element, but sometimes you want
whichever array element matched your criteria. You can return the matching element
with the $-operator. Given the blog example above, you could get Bob’s comment back
with:
> db.blog.posts.find({"comments.name" : "bob"}, {"comments.$" : 1})
{
"_id" : ObjectId("4b2d75476cc613d5ee930164"),
"comments" : [
{
"name" : "bob",
"email" : "bob@example.com",
"content" : "good post."
}
]
}
Note that this only returns the first match for each document: if Bob had left multiple
comments on this post, only the first one in the "comments" array would be returned.
Array and range query interactions
Scalars (non-array elements) in documents must match each clause of a query’s criteria.
For example, if you queried for {"x" : {"$gt" : 10, "$lt" : 20}}, "x" would have
to be both greater than 10 and less than 20. However, if a document’s "x" field is an array,
the document matches if there is an element of "x" that matches each part of the criteria
but each query clause can match a different array element.
The best way to understand this behavior is to see an example. Suppose we have the
following documents:
{"x" : 5}
{"x" : 15}
{"x" : 25}
{"x" : [5, 25]}
62 | Chapter 4: Querying
If we wanted to find all documents where "x" is between 10 and 20, one might naively
structure a query as db.test.find({"x" : {"$gt" : 10, "$lt" : 20}}) and expect
to get back one document: {"x" : 15}. However, running this, we get two:
> db.test.find({"x" : {"$gt" : 10, "$lt" : 20}})
{"x" : 15}
{"x" : [5, 25]}
Neither 5 nor 25 is between 10 and 20, but the document is returned because 25 matches
the first clause (it is greater than 25) and 5 matches the second clause (it is less than 20).
This makes range queries against arrays essentially useless: a range will match any multi-
element array. There are a couple of ways to get the expected behavior.
First, you can use "$elemMatch" to force MongoDB to compare both clauses with a single
array element. However, the catch is that "$elemMatch" won’t match non-array elements:
> db.test.find({"x" : {"$elemMatch" : {"$gt" : 10, "$lt" : 20}})
> // no results
The document {"x" : 15} no longer matches the query, because the "x" field is not an
array.
If you have an index over the field that you’re querying on (see Chapter 5), you can use
min() and max() to limit the index range traversed by the query to your "$gt" and "$lt"
values:
> db.test.find({"x" : {"$gt" : 10, "$lt" : 20}).min({"x" : 10}).max({"x" : 20})
{"x" : 15}
Now this will only traverse the index from 10 to 20, missing the 5 and 25 entries. You
can only use min() and max() when you have an index on the field you are querying
for, though, and you must pass all fields of the index to min() and max().
Using min() and max() when querying for ranges over documents that may include
arrays is generally a good idea: if you look at the index bounds for a "$gt“/”$lt" query
over an array, you can see that it’s horribly inefficient. It basically accepts any value, so
it will search every index entry, not just those in the range.
Querying on Embedded Documents
There are two ways of querying for an embedded document: querying for the whole
document or querying for its individual key/value pairs.
Querying for an entire embedded document works identically to a normal query. For
example, if we have a document that looks like this:
{
"name" : {
"first" : "Joe",
"last" : "Schmoe"
Type-Specific Queries | 63
},
"age" : 45
}
we can query for someone named Joe Schmoe with the following:
> db.people.find({"name" : {"first" : "Joe", "last" : "Schmoe"}})
However, a query for a full subdocument must exactly match the subdocument. If Joe
decides to add a middle name field, suddenly this query won’t work anymore; it doesn’t
match the entire embedded document! This type of query is also order-sensitive:
{"last" : "Schmoe", "first" : "Joe"} would not be a match.
If possible, it’s usually a good idea to query for just a specific key or keys of an embedded
document. Then, if your schema changes, all of your queries won’t suddenly break
because they’re no longer exact matches. You can query for embedded keys using dot-
notation:
> db.people.find({"name.first" : "Joe", "name.last" : "Schmoe"})
Now, if Joe adds more keys, this query will still match his first and last names.
This dot notation is the main difference between query documents and other document
types. Query documents can contain dots, which mean “reach into an embedded docu‐
ment.” Dot notation is also the reason that documents to be inserted cannot contain
the . character. Oftentimes people run into this limitation when trying to save URLs as
keys. One way to get around it is to always perform a global replace before inserting or
after retrieving, substituting a character that isn’t legal in URLs for the dot character.
Embedded document matches can get a little tricky as the document structure gets more
complicated. For example, suppose we are storing blog posts and we want to find com‐
ments by Joe that were scored at least a 5. We could model the post as follows:
> db.blog.find()
{
"content" : "...",
"comments" : [
{
"author" : "joe",
"score" : 3,
"comment" : "nice post"
},
{
"author" : "mary",
"score" : 6,
"comment" : "terrible post"
}
]
}
64 | Chapter 4: Querying
Now, we can’t query using db.blog.find({"comments" : {"author" : "joe",
"score" : {"$gte" : 5}}}). Embedded document matches have to match the whole
document, and this doesn’t match the "comment" key. It also wouldn’t work to do
db.blog.find({"comments.author" : "joe", "comments.score" : {"$gte" :
5}}), because the author criteria could match a different comment than the score cri‐
teria. That is, it would return the document shown above: it would match "author" :
"joe" in the first comment and "score" : 6 in the second comment.
To correctly group criteria without needing to specify every key, use "$elemMatch". This
vaguely-named conditional allows you to partially specify criteria to match a single
embedded document in an array. The correct query looks like this:
> db.blog.find({"comments" : {"$elemMatch" : {"author" : "joe",
"score" : {"$gte" : 5}}}})
"$elemMatch" allows us to “group” our criteria. As such, it’s only needed when you have
more than one key you want to match on in an embedded document.
$where Queries
Key/value pairs are a fairly expressive way to query, but there are some queries that they
cannot represent. For queries that cannot be done any other way, there are "$where"
clauses, which allow you to execute arbitrary JavaScript as part of your query. This allows
you to do (almost) anything within a query. For security, use of "$where" clauses should
be highly restricted or eliminated. End users should never be allowed to execute arbi‐
trary "$where" clauses.
The most common case for using "$where" is to compare the values for two keys in a
document. For instance, suppose we have documents that look like this:
> db.foo.insert({"apple" : 1, "banana" : 6, "peach" : 3})
> db.foo.insert({"apple" : 8, "spinach" : 4, "watermelon" : 4})
We’d like to return documents where any two of the fields are equal. For example, in
the second document, "spinach" and "watermelon" have the same value, so we’d like
that document returned. It’s unlikely MongoDB will ever have a $-conditional for this,
so we can use a "$where" clause to do it with JavaScript:
> db.foo.find({"$where" : function () {
... for (var current in this) {
... for (var other in this) {
... if (current != other && this[current] == this[other]) {
... return true;
... }
... }
... }
... return false;
... }});
$where Queries | 65
If the function returns true, the document will be part of the result set; if it returns
false, it won’t be.
"$where" queries should not be used unless strictly necessary: they are much slower
than regular queries. Each document has to be converted from BSON to a JavaScript
object and then run through the "$where" expression. Indexes cannot be used to satisfy
a "$where", either. Hence, you should use "$where" only when there is no other way of
doing the query. You can cut down on the penalty by using other query filters in com‐
bination with "$where". If possible, an index will be used to filter based on the non-
$where clauses; the "$where" expression will be used only to fine-tune the results.
Another way of doing complex queries is to use one of the aggregation tools, which are
covered in Chapter 7.
Server-Side Scripting
You must be very careful with security when executing JavaScript on the server. If done
incorrectly, server-side JavaScript is susceptible to injection attacks similar to those that
occur in a relational database. However, by following certain rules around accepting
input, you can use JavaScript safely. Alternatively, you can turn off JavaScript execution
altogether by running mongod with the --noscripting option.
The security issues with JavaScript are all related to executing user-provided programs
on the server. You want to avoid doing that, so make sure you aren’t accepting user input
and passing it directly to mongod. For example, suppose you want to print “Hello,
name!”, where name is provided by the user. A naive approach might be to write a Java‐
Script function such as the following:
> func = "function() { print('Hello, "+name+"!'); }"
If name is a user-defined variable, it could be the string "'); db.dropDatabase();
print('", which would turn the code into this:
> func = "function() { print('Hello, '); db.dropDatabase(); print('!'); }"
Now, if you run this code, your entire database will be dropped!
To prevent this, you should use a scope to pass in the name. In Python, for example, this
looks like this:
func = pymongo.code.Code("function() { print('Hello, '+username+'!'); }",
{"username": name})
Now the database will harmlessly print this:
Hello, '); db.dropDatabase(); print('!
Most drivers have a special type for sending code to the database, since code can actually
be a composite of a string and a scope. A scope is a document that maps variable names
to values. This mapping becomes a local scope for the JavaScript function being
66 | Chapter 4: Querying
executed. Thus, in the example above, the function would have access to a variable called
username, whose value would be the string that the user gave.
The shell does not have a code type that includes scope; you can only
use strings or JavaScript functions with it.
Cursors
The database returns results from find using a cursor. The client-side implementations
of cursors generally allow you to control a great deal about the eventual output of a
query. You can limit the number of results, skip over some number of results, sort results
by any combination of keys in any direction, and perform a number of other powerful
operations.
To create a cursor with the shell, put some documents into a collection, do a query on
them, and assign the results to a local variable (variables defined with "var" are local).
Here, we create a very simple collection and query it, storing the results in the cursor
variable:
> for(i=0; i<100; i++) {
... db.collection.insert({x : i});
... }
> var cursor = db.collection.find();
The advantage of doing this is that you can look at one result at a time. If you store the
results in a global variable or no variable at all, the MongoDB shell will automatically
iterate through and display the first couple of documents. This is what we’ve been seeing
up until this point, and it is often the behavior you want for seeing what’s in a collection
but not for doing actual programming with the shell.
To iterate through the results, you can use the next method on the cursor. You can use
hasNext to check whether there is another result. A typical loop through results looks
like the following:
> while (cursor.hasNext()) {
... obj = cursor.next();
... // do stuff
... }
cursor.hasNext() checks that the next result exists, and cursor.next() fetches it.
The cursor class also implements JavaScript’s iterator interface, so you can use it in a
forEach loop:
> var cursor = db.people.find();
> cursor.forEach(function(x) {
... print(x.name);
Cursors | 67
... });
adam
matt
zak
When you call find, the shell does not query the database immediately. It waits until
you start requesting results to send the query, which allows you to chain additional
options onto a query before it is performed. Almost every method on a cursor object
returns the cursor itself so that you can chain options in any order. For instance, all of
the following are equivalent:
> var cursor = db.foo.find().sort({"x" : 1}).limit(1).skip(10);
> var cursor = db.foo.find().limit(1).sort({"x" : 1}).skip(10);
> var cursor = db.foo.find().skip(10).limit(1).sort({"x" : 1});
At this point, the query has not been executed yet. All of these functions merely build
the query. Now, suppose we call the following:
> cursor.hasNext()
At this point, the query will be sent to the server. The shell fetches the first 100 results
or first 4 MB of results (whichever is smaller) at once so that the next calls to next or
hasNext will not have to make trips to the server. After the client has run through the
first set of results, the shell will again contact the database and ask for more results with
a getMore request. getMore requests basically contain an identifier for the query and
ask the database if there are any more results, returning the next batch if there are. This
process continues until the cursor is exhausted and all results have been returned.
Limits, Skips, and Sorts
The most common query options are limiting the number of results returned, skipping
a number of results, and sorting. All these options must be added before a query is sent
to the database.
To set a limit, chain the limit function onto your call to find. For example, to only
return three results, use this:
> db.c.find().limit(3)
If there are fewer than three documents matching your query in the collection, only the
number of matching documents will be returned; limit sets an upper limit, not a lower
limit.
skip works similarly to limit:
> db.c.find().skip(3)
This will skip the first three matching documents and return the rest of the matches. If
there are fewer than three documents in your collection, it will not return any
documents.
68 | Chapter 4: Querying
sort takes an object: a set of key/value pairs where the keys are key names and the values
are the sort directions. Sort direction can be 1 (ascending) or −1 (descending). If multiple
keys are given, the results will be sorted in that order. For instance, to sort the results
by "username" ascending and "age" descending, we do the following:
> db.c.find().sort({username : 1, age : -1})
These three methods can be combined. This is often handy for pagination. For example,
suppose that you are running an online store and someone searches for mp3. If you
want 50 results per page sorted by price from high to low, you can do the following:
> db.stock.find({"desc" : "mp3"}).limit(50).sort({"price" : -1})
If that person clicks Next Page to see more results, you can simply add a skip to the
query, which will skip over the first 50 matches (which the user already saw on page 1):
> db.stock.find({"desc" : "mp3"}).limit(50).skip(50).sort({"price" : -1})
However, large skips are not very performant; there are suggestions for how to avoid
them in the next section.
Comparison order
MongoDB has a hierarchy as to how types compare. Sometimes you will have a single
key with multiple types: for instance, integers and booleans, or strings and nulls. If you
do a sort on a key with a mix of types, there is a predefined order that they will be sorted
in. From least to greatest value, this ordering is as follows:
1. Minimum value
2. null
3. Numbers (integers, longs, doubles)
4. Strings
5. Object/document
6. Array
7. Binary data
8. Object ID
9. Boolean
10. Date
11. Timestamp
12. Regular expression
13. Maximum value
Cursors | 69
Avoiding Large Skips
Using skip for a small number of documents is fine. For a large number of results, skip
can be slow, since it has to find and then discard all the skipped results. Most databases
keep more metadata in the index to help with skips, but MongoDB does not yet support
this, so large skips should be avoided. Often you can calculate the next query based on
the result from the previous one.
Paginating results without skip
The easiest way to do pagination is to return the first page of results using limit and
then return each subsequent page as an offset from the beginning:
> // do not use: slow for large skips
> var page1 = db.foo.find(criteria).limit(100)
> var page2 = db.foo.find(criteria).skip(100).limit(100)
> var page3 = db.foo.find(criteria).skip(200).limit(100)
...
However, depending on your query, you can usually find a way to paginate without
skips. For example, suppose we want to display documents in descending order based
on "date". We can get the first page of results with the following:
> var page1 = db.foo.find().sort({"date" : -1}).limit(100)
Then, we can use the "date" value of the last document as the criteria for fetching the
next page:
var latest = null;
// display first page
while (page1.hasNext()) {
latest = page1.next();
display(latest);
}
// get next page
var page2 = db.foo.find({"date" : {"$gt" : latest.date}});
page2.sort({"date" : -1}).limit(100);
Now the query does not need to include a skip.
Finding a random document
One fairly common problem is how to get a random document from a collection. The
naive (and slow) solution is to count the number of documents and then do a find,
skipping a random number of documents between 0 and the size of the collection:
> // do not use
> var total = db.foo.count()
> var random = Math.floor(Math.random()*total)
> db.foo.find().skip(random).limit(1)
70 | Chapter 4: Querying
It is actually highly inefficient to get a random element this way: you have to do a count
(which can be expensive if you are using criteria), and skipping large numbers of ele‐
ments can be time-consuming.
It takes a little forethought, but if you know you’ll be looking up a random element on
a collection, there’s a much more efficient way to do so. The trick is to add an extra
random key to each document when it is inserted. For instance, if we’re using the shell,
we could use the Math.random() function (which creates a random number between 0
and 1):
> db.people.insert({"name" : "joe", "random" : Math.random()})
> db.people.insert({"name" : "john", "random" : Math.random()})
> db.people.insert({"name" : "jim", "random" : Math.random()})
Now, when we want to find a random document from the collection, we can calculate
a random number and use that as query criteria, instead of doing a skip:
> var random = Math.random()
> result = db.foo.findOne({"random" : {"$gt" : random}})
There is a slight chance that random will be greater than any of the "random" values in
the collection, and no results will be returned. We can guard against this by simply
returning a document in the other direction:
> if (result == null) {
... result = db.foo.findOne({"random" : {"$lt" : random}})
... }
If there aren’t any documents in the collection, this technique will end up returning
null, which makes sense.
This technique can be used with arbitrarily complex queries; just make sure to have an
index that includes the random key. For example, if we want to find a random plumber
in California, we can create an index on "profession", "state", and "random":
> db.people.ensureIndex({"profession" : 1, "state" : 1, "random" : 1})
This allows us to quickly find a random result (see Chapter 5 for more information on
indexing).
Advanced Query Options
There are two types of queries: wrapped and plain. A plain query is something like this:
> var cursor = db.foo.find({"foo" : "bar"})
There are a couple options that “wrap” the query. For example, suppose we perform a
sort:
> var cursor = db.foo.find({"foo" : "bar"}).sort({"x" : 1})
Cursors | 71
Instead of sending {"foo" : "bar"} to the database as the query, the query gets wrapped
in a larger document. The shell converts the query from {"foo" : "bar"} to
{"$query" : {"foo" : "bar"}, "$orderby" : {"x" : 1}}.
Most drivers provide helpers for adding arbitrary options to queries. Other helpful
options include the following:
$maxscan : integer
Specify the maximum number of documents that should be scanned for the query.
> db.foo.find(criteria)._addSpecial("$maxscan", 20)
This can be useful if you want a query to not to take too long but are not sure how
much of a collection will need to be scanned. This will limit your results to whatever
was found in the part of the collection that was scanned (i.e., you may miss other
documents that match).
$min : document
Start criteria for querying. document must exactly match the keys of an index used
for the query. This forces the given index to be used for the query.
This is used internally and you should generally use "$gt" instead of "$min". You
can use "$min" to force the lower bound on an index scan, which may be helpful
for complex queries.
$max : document
End criteria for querying. document must exactly match the keys of an index used
for the query. This forces the given index to be used for the query.
If this is used internally, you should generally use "$lt" instead of "$max". You can
use "$max" to force bounds on an index scan, which may be helpful for complex
queries.
$showDiskLoc : true
Adds a "$diskLoc" field to the results that shows where on disk that particular result
lives. For example:
> db.foo.find()._addSpecial('$showDiskLoc',true)
{ "_id" : 0, "$diskLoc" : { "file" : 2, "offset" : 154812592 } }
{ "_id" : 1, "$diskLoc" : { "file" : 2, "offset" : 154812628 } }
The file number shows which file the document is in. In this case, if we’re using the
test database, the document is in test.2. The second field gives the byte offset of each
document within the file.
Getting Consistent Results
A fairly common way of processing data is to pull it out of MongoDB, change it in some
way, and then save it again:
72 | Chapter 4: Querying
cursor = db.foo.find();
while (cursor.hasNext()) {
var doc = cursor.next();
doc = process(doc);
db.foo.save(doc);
}
This is fine for a small number of results, but MongoDB can return the same result
multiple times for a large result set. To see why, imagine how the documents are being
stored. You can picture a collection as a list of documents that looks something like
Figure 4-1. Snowflakes represent documents, since every document is beautiful and
unique.
Figure 4-1. A collection being queried
Now, when we do a find, the cursor starts returning results from the beginning of the
collection and moves right. Your program grabs the first 100 documents and processes
them. When you save them back to the database, if a document does not have the
padding available to grow to its new size, like in Figure 4-2, it needs to be relocated.
Usually, a document will be relocated to the end of a collection (Figure 4-3).
Figure 4-2. An enlarged document may not fit where it did before
Cursors | 73
Figure 4-3. MongoDB relocates updated documents that don’t fit in their original
position
Now our program continues to fetch batches of documents. When it gets toward the
end, it will return the relocated documents again (Figure 4-4)!
Figure 4-4. A cursor may return these relocated documents again in a later batch
The solution to this problem is to snapshot your query. If you add the option, the query
will be run by traversing the "_id" index, which guarantees that you’ll only return each
document once. For example, instead of db.foo.find(), you’d run:
> db.foo.find().snapshot()
Snapshotting makes queries slower, so only use snapshotted queries when necessary.
For example, mongodump (a backup utility covered in Chapter 22) uses snapshotted
queries by default.
All queries that return a single batch of results are effectively snapshotted. Inconsis‐
tencies arise only when the collection changes under a cursor while it is waiting to get
another batch of results.
74 | Chapter 4: Querying
Immortal Cursors
There are two sides to a cursor: the client-facing cursor and the database cursor that the
client-side one represents. We have been talking about the client-side one up until now,
but we are going to take a brief look at what’s happening on the server.
On the server side, a cursor takes up memory and resources. Once a cursor runs out of
results or the client sends a message telling it to die, the database can free the resources
it was using. Freeing these resources lets the database use them for other things, which
is good, so we want to make sure that cursors can be freed quickly (within reason).
There are a couple of conditions that can cause the death (and subsequent cleanup) of
a cursor. First, when a cursor finishes iterating through the matching results, it will clean
itself up. Another way is that, when a cursor goes out of scope on the client side, the
drivers send the database a special message to let it know that it can kill that cursor.
Finally, even if the user hasn’t iterated through all the results and the cursor is still in
scope, after 10 minutes of inactivity, a database cursor will automatically “die.” This way,
if a client crashes or is buggy, MongoDB will not be left with thousands of open cursors.
This “death by timeout” is usually the desired behavior: very few applications expect
their users to sit around for minutes at a time waiting for results. However, sometimes
you might know that you need a cursor to last for a long time. In that case, many drivers
have implemented a function called immortal, or a similar mechanism, which tells the
database not to time out the cursor. If you turn off a cursor’s timeout, you must iterate
through all of its results or kill it to make sure it gets closed. Otherwise, it will sit around
in the database hogging resources until the server is restarted.
Database Commands
There is one very special type of query called a database command. We’ve covered
creating, updating, deleting, and finding documents. Database commands do “every‐
thing else,” from administrative tasks like shutting down the server and cloning data‐
bases to counting documents in a collection and performing aggregations.
Commands are mentioned throughout this text, as they are useful for data manipula‐
tion, administration, and monitoring. For example, dropping a collection is done via
the "drop" database command:
> db.runCommand({"drop" : "test"});
{
"nIndexesWas" : 1,
"msg" : "indexes dropped for collection",
"ns" : "test.test",
"ok" : true
}
Database Commands | 75
You might be more familiar with the shell helper, which wraps the command and pro‐
vides a simpler interface:
> db.test.drop()
Often you can just use the shell helpers, but knowing the underlying commands can be
helpful if you’re stuck on a box with an old version of the shell and connected to a new
version of the database: the shell might not have the wrappers for new database com‐
mands, but you can still run them with runCommand().
We’ve already seen a couple of commands in the previous chapters; for instance, we
used the getLastError command in Chapter 3 to check the number of documents
affected by an update:
> db.count.update({x : 1}, {$inc : {x : 1}}, false, true)
> db.runCommand({getLastError : 1})
{
"err" : null,
"updatedExisting" : true,
"n" : 5,
"ok" : true
}
In this section, we’ll take a closer look at commands to see exactly what they are and
how they’re implemented. We’ll also describe some of the most useful commands that
are supported by MongoDB. You can see all commands by running the db.listCom
mands() command.
How Commands Work
A database command always returns a document containing the key "ok". If "ok" is 1,
the command was successful; and if it is 0, the command failed for some reason.
If "ok" is 0 then an additional key will be present, "errmsg". The value of "errmsg" is
a string explaining why the command failed. As an example, let’s try running the drop
command again, on the collection that was dropped in the previous section:
> db.runCommand({"drop" : "test"});
{ "errmsg" : "ns not found", "ok" : false }
Commands in MongoDB are implemented as a special type of query that gets performed
on the $cmd collection. runCommand just takes a command document and performs the
equivalent query, so our drop call becomes the following:
db.$cmd.findOne({"drop" : "test"});
When the MongoDB server gets a query on the $cmd collection, it handles it using
special logic, rather than the normal code for handling queries. Almost all MongoDB
drivers provide a helper method like runCommand for running commands, but com‐
mands can always be run using a simple query.
76 | Chapter 4: Querying
Some commands require administrator access and must be run on the admin database.
If such a command is run on any other database, it will return an "access denied"
error. If you’re working on another database and you need to run an admin command,
you can use the adminCommand function, instead of runCommand:
> use temp
switched to db temp
> db.runCommand({shutdown:1})
{ "errmsg" : "access denied; use admin db", "ok" : 0 }
> db.adminCommand({"shutdown" : 1})
Commands are one of the few places that are field-order-sensitive: the command name
must always be the first field in the command. Thus, {"getLastError" : 1, "w" :
2} will work, but {"w" : 2, "getLastError" : 1} will not.
Database Commands | 77
PART II
Designing Your Application
CHAPTER 5
Indexing
This chapter introduces MongoDB’s indexing, which allows you to optimize your quer‐
ies and is even required for certain types of queries:
• What indexing is and why you’d want to use it
• How to choose which fields to index
•How to enforce and evaluate index usage
•Administrative details on creating and removing indexes
Choosing the right indexes for your collections is critical to performance.
Introduction to Indexing
A database index is similar to a book’s index. Instead of looking through the whole book,
the database takes a shortcut and just looks at an ordered list that points to the content,
which allows it to query orders of magnitude faster.
A query that does not use an index is called a table scan (a term inherited from relational
databases), which means that the server has to “look through the whole book” to find
a query’s results. This process is basically what you’d do if you were looking for infor‐
mation in a book without an index: you start at page 1 and read through the whole thing.
In general, you want to avoid making the server do table scans because it is very slow
for large collections.
For example, let’s create a collection with 1 million documents in it (or 10 million or
100 million, if you have the patience):
> for (i=0; i<1000000; i++) {
... db.users.insert(
... {
... "i" : i,
81
... "username" : "user"+i,
... "age" : Math.floor(Math.random()*120),
... "created" : new Date()
... }
... );
... }
If we do a query on this collection, we can use the explain() function to see what
MongoDB is doing when it executes the query. Try querying on a random username to
see an example.
> db.users.find({username: "user101"}).explain()
{
"cursor" : "BasicCursor",
"nscanned" : 1000000,
"nscannedObjects" : 1000000,
"n" : 1,
"millis" : 721,
"nYields" : 0,
"nChunkSkips" : 0,
"isMultiKey" : false,
"indexOnly" : false,
"indexBounds" : {
}
}
“Using explain() and hint()” on page 98 will explain the output fields; for now you can
ignore almost all of them. "nscanned" is the number of documents MongoDB looked at
while trying to satisfy the query, which, as you can see, is every document in the col‐
lection. That is, MongoDB had to look through every field in every document. This took
nearly a second to accomplish: the "millis" field shows the number of milliseconds it
took to execute the query.
The "n" field shows the number of results returned: 1, which makes sense because there
is only one user with the username "user101". Note that MongoDB had to look through
every document in the collection for matches because it did not know that usernames
are unique. To optimize this query, we could limit it to one result so that MongoDB
would stop looking after it found user101:
> db.users.find({username: "user101"}).limit(1).explain()
{
"cursor" : "BasicCursor",
"nscanned" : 102,
"nscannedObjects" : 102,
"n" : 1,
"millis" : 2,
"nYields" : 0,
"nChunkSkips" : 0,
"isMultiKey" : false,
"indexOnly" : false,
82 | Chapter 5: Indexing
"indexBounds" : {
}
}
The number scanned has now been cut way down and the query is almost instantaneous.
However, this is an impractical solution in general: what if we were looking for
user999999? Then we’d still have to traverse the entire collection and our service would
just get slower and slower as we added users.
Indexes are a great way to fix queries like this because they organize data by a given field
to let MongoDB find it quickly. Try creating an index on the username field:
> db.users.ensureIndex({"username" : 1})
Depending on your machine and how large you made the collection, creating an index
may take a few minutes. If the ensureIndex call does not return after a few seconds, run
db.currentOp() (in a different shell) or check your mongod’s log to see the index build’s
progress.
Once the index build is complete, try repeating the original query:
> db.users.find({"username" : "user101"}).explain()
{
"cursor" : "BtreeCursor username_1",
"nscanned" : 1,
"nscannedObjects" : 1,
"n" : 1,
"millis" : 3,
"nYields" : 0,
"nChunkSkips" : 0,
"isMultiKey" : false,
"indexOnly" : false,
"indexBounds" : {
"username" : [
[
"user101",
"user101"
]
]
}
}
This explain() output is more complex, but continue to ignore all the fields other than
"n", "nscanned", and "millis" for now. As you can see, the query is now almost in‐
stantaneous and, even better, has a similar runtime when querying for any username:
> db.users.find({username: "user999999"}).explain().millis
1
As you can see, an index can make a dramatic difference in query times. However,
indexes have their price: every write (insert, update, or delete) will take longer for every
Introduction to Indexing | 83
index you add. This is because MongoDB has to update all your indexes whenever your
data changes, as well as the document itself. Thus, MongoDB limits you to 64 indexes
per collection. Generally you should not have more than a couple of indexes on any
given collection. The tricky part becomes figuring out which fields to index.
MongoDB’s indexes work almost identically to typical relational data‐
base indexes, so if you are familiar with those, you can skim this section
for syntax specifics. We’ll go over some indexing basics, but keep in
mind that it’s an extensive topic and most of the material out there on
indexing for MySQL/Oracle/SQLite will apply equally well to Mon‐
goDB [such as the “Use the Index, Luke!” tutorial].
To choose which fields to create indexes for, look through your common queries and
queries that need to be fast and try to find a common set of keys from those. For instance,
in the example above, we were querying on "username". If that was a particularly com‐
mon query or was becoming a bottleneck, indexing "username" would be a good choice.
However, if this was an unusual query or one that was only done by administrators who
didn’t care how long it took, it would not be a good choice of index.
Introduction to Compound Indexes
An index keeps all of its values in a sorted order so it makes sorting documents by the
indexed key much faster. However, an index can only help with sorting if it is a prefix
of the sort. For example, the index on "username" wouldn’t help much for this sort:
> db.users.find().sort({"age" : 1, "username" : 1})
This sorts by "age" and then "username", so a strict sorting by "username" isn’t terribly
helpful. To optimize this sort, you could make an index on "age" and "username":
> db.users.ensureIndex({"age" : 1, "username" : 1})
This is called a compound index and is useful if your query has multiple sort directions
or multiple keys in the criteria. A compound index is an index on more than one field.
Suppose we have a users collection that looks something like this, if we run a query with
no sorting (called natural order):
> db.users.find({}, {"_id" : 0, "i" : 0, "created" : 0})
{ "username" : "user0", "age" : 69 }
{ "username" : "user1", "age" : 50 }
{ "username" : "user2", "age" : 88 }
{ "username" : "user3", "age" : 52 }
{ "username" : "user4", "age" : 74 }
{ "username" : "user5", "age" : 104 }
{ "username" : "user6", "age" : 59 }
{ "username" : "user7", "age" : 102 }
84 | Chapter 5: Indexing
{ "username" : "user8", "age" : 94 }
{ "username" : "user9", "age" : 7 }
{ "username" : "user10", "age" : 80 }
...
If we index this collection by {"age" : 1, "username" : 1}, the index will look roughly
like this:
[0, "user100309"] -> 0x0c965148
[0, "user100334"] -> 0xf51f818e
[0, "user100479"] -> 0x00fd7934
...
[0, "user99985" ] -> 0xd246648f
[1, "user100156"] -> 0xf78d5bdd
[1, "user100187"] -> 0x68ab28bd
[1, "user100192"] -> 0x5c7fb621
...
[1, "user999920"] -> 0x67ded4b7
[2, "user100141"] -> 0x3996dd46
[2, "user100149"] -> 0xfce68412
[2, "user100223"] -> 0x91106e23
...
Each index entry contains an age and a username and points to the location of a docu‐
ment on disk (represented by the hexadecimal numbers, which can be ignored). Note
that "age" fields are ordered to be strictly ascending and, within each age, "user
name"s are also in ascending order. As each age has approximately 8,000 usernames
associated with it, only those necessary to convey the general idea have been included.
The way MongoDB uses this index depends on the type of query you’re doing. These
are the three most common ways:
db.users.find({"age" : 21}).sort({"username" : -1})
This is a point query, which searches for a single value (although there may be
multiple documents with that value). Due to the second field in the index, the results
are already in the correct order for the sort: MongoDB can start with the last match
for {"age" : 21} and traverse the index in order:
[21, "user999977"] -> 0x9b3160cf
[21, "user999954"] -> 0xfe039231
[21, "user999902"] -> 0x719996aa
...
This type of query is very efficient: MongoDB can jump directly to the correct age
and doesn’t need to sort the results as traversing the index returns the data in the
correct order.
Note that sort direction doesn’t matter: MongoDB is comfortable traversing the
index in either direction.
Introduction to Indexing | 85
db.users.find({"age" : {"$gte" : 21, "$lte" : 30}})
This is a multi-value query, which looks for documents matching multiple values
(in this case, all ages between 21 and 30). MongoDB will use the first key in the
index, "age", to return the matching documents, like so:
[21, "user100000"] -> 0x37555a81
[21, "user100069"] -> 0x6951d16f
[21, "user1001"] -> 0x9a1f5e0c
[21, "user100253"] -> 0xd54bd959
[21, "user100409"] -> 0x824fef6c
[21, "user100469"] -> 0x5fba778b
...
[30, "user999775"] -> 0x45182d8c
[30, "user999850"] -> 0x1df279e9
[30, "user999936"] -> 0x525caa57
In general, if MongoDB uses an index for a query it will return the resulting docu‐
ments in index order.
db.users.find({"age" : {"$gte" : 21, "$lte" : 30}}).sort({"username" :
1})
This is a multi-value query, like the previous one, but this time it has a sort. As
before, MongoDB will use the index to match the criteria:
[21, "user100000"] -> 0x37555a81
[21, "user100069"] -> 0x6951d16f
[21, "user1001"] -> 0x9a1f5e0c
[21, "user100253"] -> 0xd54bd959
...
[22, "user100004"] -> 0x81e862c5
[22, "user100328"] -> 0x83376384
[22, "user100335"] -> 0x55932943
[22, "user100405"] -> 0x20e7e664
...
However, the index doesn’t return the usernames in sorted order and the query
requested that the results be sorted by username, so MongoDB has to sort the results
in memory before returning them. Thus, this query is usually less efficient than the
queries above.
Of course, the speed depends on how many results match your criteria: if your result
set is only a couple of documents, MongoDB won’t have much work to do to sort
them. If there are more results, it will be slower or may not work at all: if you have
more than 32 MB of results MongoDB will just error out, refusing to sort that much
data:
Mon Oct 29 16:25:26 uncaught exception: error: {
"$err" : "too much data for sort() with no index. add an index or
specify a smaller limit",
"code" : 10128
}
86 | Chapter 5: Indexing
One other index you can use in the last example is the same keys in reverse order:
{"username" : 1, "age" : 1}. MongoDB will then traverse all the index entries, but
in the order you want them back in. It will pick out the matching documents using the
"age" part of the index:
["user0", 69]
["user1", 50]
["user10", 80]
["user100", 48]
["user1000", 111]
["user10000", 98]
["user100000", 21] -> 0x73f0b48d
["user100001", 60]
["user100002", 82]
["user100003", 27] -> 0x0078f55f
["user100004", 22] -> 0x5f0d3088
["user100005", 95]
...
This is good in that it does not require any giant in-memory sorts. However, it does
have to scan the entire index to find all matches. Thus, putting the sort key first is
generally a good strategy when you’re using a limit so MongoDB can stop scanning the
index after a couple of matches.
You can diagnose how MongoDB defaults to processing db.users.find({"age" :
{"$gte" : 21, "$lte" : 30}}).sort({"username" : 1}) by using explain():
> db.users.find({"age" : {"$gte" : 21, "$lte" : 30}}).
... sort({"username" : 1}).
... explain()
{
"cursor" : "BtreeCursor age_1_username_1",
"isMultiKey" : false,
"n" : 83484,
"nscannedObjects" : 83484,
"nscanned" : 83484,
"nscannedObjectsAllPlans" : 83484,
"nscannedAllPlans" : 83484,
"scanAndOrder" : true,
"indexOnly" : false,
"nYields" : 0,
"nChunkSkips" : 0,
"millis" : 2766,
"indexBounds" : {
"age" : [
[
21,
30
]
],
"username" : [
[
Introduction to Indexing | 87
{
"$minElement" : 1
},
{
"$maxElement" : 1
}
]
]
},
"server" : "spock:27017"
}
You can ignore most of these fields; they will be covered later in this chapter. For now,
note that the "cursor" field indicates that this query would use the {"age" : 1, "user
name" : 1} index and looks at less than a tenth of the documents ("nscanned" is only
83484), but takes nearly three seconds to execute (the "millis" field). The "scanAnd
Order" field is true: this indicates that MongoDB had to sort the data in memory, as
mentioned above.
We can use a hint to force MongoDB to use a certain index, so try the same query again
using the {"username" : 1, "age" : 1} index, instead. This query scans more docu‐
ments, but does not have to do an in-memory sort:
> db.users.find({"age" : {"$gte" : 21, "$lte" : 30}}).
... sort({"username" : 1}).
... hint({"username" : 1, "age" : 1}).
... explain()
{
"cursor" : "BtreeCursor username_1_age_1",
"isMultiKey" : false,
"n" : 83484,
"nscannedObjects" : 83484,
"nscanned" : 984434,
"nscannedObjectsAllPlans" : 83484,
"nscannedAllPlans" : 984434,
"scanAndOrder" : false,
"indexOnly" : false,
"nYields" : 0,
"nChunkSkips" : 0,
"millis" : 14820,
"indexBounds" : {
"username" : [
[
{
"$minElement" : 1
},
{
"$maxElement" : 1
}
]
],
88 | Chapter 5: Indexing
"age" : [
[
21,
30
]
]
},
"server" : "spock:27017"
}
Note that this took nearly 15 seconds to run, making the first index the clear winner.
However, if we limit the number of results for each query, a new winner emerges:
> db.users.find({"age" : {"$gte" : 21, "$lte" : 30}}).
... sort({"username" : 1}).
... limit(1000).
... hint({"age" : 1, "username" : 1}).
... explain()['millis']
2031
> db.users.find({"age" : {"$gte" : 21, "$lte" : 30}}).
... sort({"username" : 1}).
... limit(1000).
... hint({"username" : 1, "age" : 1}).
... explain()['millis']
181
The first query is still hovering between two and three seconds, but the second query
now takes less than a fifth of a second! Thus, you should always run explain()s on
exactly the queries that your application is running. Excluding any options could make
the explain() output misleading.
The index pattern of {"sortKey" : 1, "queryCriteria" : 1} often works well in
applications, as most application do not want all possible results for a query but only
the first few. It also scales well because of the way indexes are organized internally.
Indexes are basically trees, with the smallest value on the leftmost leaf and the greatest
on the rightmost. If you have a "sortKey" that is a date (or any value that increases over
time) then as you traverse the tree from left to right, you’re basically travelling forward
in time. Thus, for applications that tend to use recent data more than older data, Mon‐
goDB only has to keep the rightmost (most recent) branches of the tree in memory, not
the whole thing. An index like this is called right-balanced and, whenever possible, you
should make your indexes right-balanced. The "_id" index is an example of a right-
balanced index.
Using Compound Indexes
In the section above, we’ve been using compound indexes, which are indexes with more
than one key in them. Compound indexes are a little more complicated to think about
than single-key indexes, but they are very powerful. This section covers them in more
depth.
Introduction to Indexing | 89
Choosing key directions
So far, all of our index entries have been sorted in ascending, or least-to-greatest, order.
However, if you need to sort on two (or more) criteria, you may need to have index keys
go in different directions. For example, suppose we want to sort the collection above by
youngest to oldest and usernames from Z-A. Our previous indexes would not be very
efficient for this problem: within each age group users were sorted by "username"
ascending A-Z, not Z-A. The compound indexes above do not hold the values in any
useful order for getting "age" ascending and "username" descending.
To optimize compound sorts in different directions, use an index with matching direc‐
tions. In this example, we could use {"age" : 1, "username" : -1}, which would
organize the data as follows:
[21, "user999977"] -> 0xe57bf737
[21, "user999954"] -> 0x8bffa512
[21, "user999902"] -> 0x9e1447d1
[21, "user999900"] -> 0x3a6a8426
[21, "user999874"] -> 0xc353ee06
...
[30, "user999936"] -> 0x7f39a81a
[30, "user999850"] -> 0xa979e136
[30, "user999775"] -> 0x5de6b77a
...
[30, "user100324"] -> 0xe14f8e4d
[30, "user100140"] -> 0x0f34d446
[30, "user100050"] -> 0x223c35b1
The ages are arranged from youngest to oldest and, within each age, usernames are
sorted from Z to A (or, rather, “9” to “0”, given our usernames).
If our application also needed to optimize sorting by {"age" : 1, "username" : 1},
we would have to create a second index with those directions. To figure out which
directions to use for an index, simply match the directions your sort is using. Note that
inverse indexes (multiplying each direction by -1) are equivalent: {"age" : 1, "user
name" : -1} suits the same queries that {"age" : -1, "username" : 1} does.
Index direction only really matters when you’re sorting based on multiple criteria. If
you’re only sorting by a single key, MongoDB can just as easily read the index in the
opposite order. For example, if you had a sort on {"age" : -1} and an index on
{"age" : 1}, MongoDB could optimize it just as well as if you had an index on {"age" :
-1} (so don’t create both!). The direction only matters for multikey sorts.
Using covered indexes
In the examples above, the query was always used to find the correct document, and
then follow a pointer back to fetch the actual document. However, if your query is only
looking for the fields that are included in the index, it does not need to fetch the docu‐
ment. When an index contains all the values requested by the user, it is considered to
90 | Chapter 5: Indexing
be covering a query. Whenever practical, use covered indexes in preference to going
back to documents. You can make your working set much smaller that way, especially
if you can combine this with a right-balanced index.
To make sure a query can use the index only, you should use projections (see “Specifying
Which Keys to Return” on page 54) to not return the "_id" field (unless it is part of the
index). You may also have to index fields that you aren’t querying on, so you should
balance your need for faster queries with the overhead this will add on writes.
If you run an explain on a covered query, the "indexOnly" field will be true.
If you index a field containing arrays, that index can never cover a query (due to the
way arrays are stored in indexes, this is covered in more depth in “Indexing Objects and
Arrays” on page 95). Even if you exclude the array field from the fields returned, you
cannot cover a query using such an index.
Implicit indexes
Compound index can do “double duty” and act like different indexes for different quer‐
ies. If we have an index on {"age" : 1, "username" : 1}, the "age" field is sorted
identically to the way it would be if you had an index on just {"age" : 1}. Thus, the
compound index can be used the way an index on {"age" : 1} by itself would be.
This can be generalized to as many keys as necessary: if an index has N keys, you get a
“free” index on any prefix of those keys. For example, if we have an index that looks like
{"a": 1, "b": 1, "c": 1, ..., "z": 1}, we effectively have indexes on {"a": 1},
{"a": 1, "b" : 1}, {"a": 1, "b": 1, "c": 1}, and so on.
Note that this doesn’t hold for any subset of keys: queries that would use the index {"b":
1} or {"a": 1, "c": 1} (for example) will not be optimized: only queries that can use
a prefix of the index can take advantage of it.
How $-Operators Use Indexes
Some queries can use indexes more efficiently than others; some queries cannot using
indexes at all. This section covers how various query operators are handled by
MongoDB.
Inefficient operators
There are a few queries that cannot use an index at all, such as “$where" queries and
checking if a key exists ({"key" : {"$exists" : true}}). There are several other
operators that use indexes but not very efficiently.
If there is a “vanilla” index on "x", querying for documents where "x" does not exist
can use the index ({"x" : {"$exists" : false}}). However, as nonexistent fields are
stored the same way as null fields in the index, the query must visit each document to
Introduction to Indexing | 91
check if the value is actually null or nonexistent. If you use a sparse index, it cannot be
used for either {"$exists" : true} nor {"$exists" : false}.
In general, negation is inefficient. "$ne" queries can use an index, but not very well.
They must look at all the index entries other than the one specified by the "$ne", so it
basically has to scan the entire index. For example, here are the index ranges traversed
for such a query:
> db.example.find({"i" : {"$ne" : 3}}).explain()
{
"cursor" : "BtreeCursor i_1 multi",
...,
"indexBounds" : {
"i" : [
[
{
"$minElement" : 1
},
3
],
[
3,
{
"$maxElement" : 1
}
]
]
},
...
}
This query looks at all index entries less than 3 and all index entries greater than 3. This
can be efficient if a large swath of your collection is 3, but otherwise it must check almost
everything.
"$not" can sometimes use an index but often does not know how. It can reverse basic
ranges ({"key" : {"$lt" : 7}} becomes {"key" : {"$gte" : 7}}) and regular
expressions. However, most other queries with "$not" will fall back to doing a table
scan. "$nin" always uses a table scan.
If you need to perform one of these types of queries quickly, figure out if there’s another
clause that you could add to the query that could use an index to filter the result set
down to a small number of documents before MongoDB attempts to do non-indexed
matching.
For example, suppose we were finding all users without a "birthday" field. If we knew
that the application started adding a birthday field on March 20th, we could limit the
query to users created before then:
> db.users.find({"birthday" : {"$exists" : false}, "_id" : {"$lt" : march20Id}})
92 | Chapter 5: Indexing
The order of fields in a query is irrelevant: MongoDB will find which fields it can use
indexes on regardless of ordering.
Ranges
Compound indexes can help MongoDB execute more effectively queries with multiple
clauses. When designing an index with multiple fields, put fields that will be used in
exact matches first (e.g., "x" : "foo") and ranges last (e.g., "y": {"$gt" : 3, "$lt" :
5}). This allows the query to find an exact value for the first index key and then search
within that for a second index range. For example, suppose we were querying for a
specific age and a range of usernames using an {"age" : 1, "username" : 1} index.
We would get fairly exact index bounds:
> db.users.find({"age" : 47,
... "username" : {"$gt" : "user5", "$lt" : "user8"}}).explain()
{
"cursor" : "BtreeCursor age_1_username_1",
"n" : 2788,
"nscanned" : 2788,
...,
"indexBounds" : {
"age" : [
[
47,
47
]
],
"username" : [
[
"user5",
"user8"
]
]
},
...
}
The query goes directly to "age" : 47 and then searches within that for usernames
between "user5" and "user8".
Conversely, suppose we use an index on {"username" : 1, "age" : 1}. This changes
the query plan, as the query must look at all users between "user5" and "user8" and
pick out the ones with "age" : 47:
> db.users.find({"age" : 47,
... "username" : {"$gt" : "user5", "$lt" : "user8"}}).explain()
{
"cursor" : "BtreeCursor username_1_age_1",
"n" : 2788,
"nscanned" : 319499,
...,
Introduction to Indexing | 93
"indexBounds" : {
"username" : [
[
"user5",
"user8"
]
],
"age" : [
[
47,
47
]
]
},
"server" : "spock:27017"
}
This forces MongoDB to scan 10 times the number of index entries as using the previous
index would. Using two ranges in a query basically always forces this less-efficient query
plan.
OR queries
As of this writing, MongoDB can only use one index per query. That is, if you create
one index on {"x" : 1} and another index on {"y" : 1} and then do a query on
{"x" : 123, "y" : 456}, MongoDB will use one of the indexes you created, not use
both. This only exception to this rule is "$or". "$or" can use one index per $or clause,
as $or preforms two queries and then merges the results:
> db.foo.find({"$or" : [{"x" : 123}, {"y" : 456}]}).explain()
{
"clauses" : [
{
"cursor" : "BtreeCursor x_1",
"isMultiKey" : false,
"n" : 1,
"nscannedObjects" : 1,
"nscanned" : 1,
"nscannedObjectsAllPlans" : 1,
"nscannedAllPlans" : 1,
"scanAndOrder" : false,
"indexOnly" : false,
"nYields" : 0,
"nChunkSkips" : 0,
"millis" : 0,
"indexBounds" : {
"x" : [
[
123,
123
]
94 | Chapter 5: Indexing
]
}
},
{
"cursor" : "BtreeCursor y_1",
"isMultiKey" : false,
"n" : 1,
"nscannedObjects" : 1,
"nscanned" : 1,
"nscannedObjectsAllPlans" : 1,
"nscannedAllPlans" : 1,
"scanAndOrder" : false,
"indexOnly" : false,
"nYields" : 0,
"nChunkSkips" : 0,
"millis" : 0,
"indexBounds" : {
"y" : [
[
456,
456
]
]
}
}
],
"n" : 2,
"nscannedObjects" : 2,
"nscanned" : 2,
"nscannedObjectsAllPlans" : 2,
"nscannedAllPlans" : 2,
"millis" : 0,
"server" : "spock:27017"
}
As you can see, this explain is the conglomerate of two separate queries. In general,
doing two queries and merging the results is much less efficient than doing a single
query; thus, whenever possible, prefer "$in" to "$or".
If you must use an $or, keep in mind that MongoDB needs to look through the query
results of both queries and remove any duplicates (documents that matched more than
one $or clause).
When running "$in" queries there is no way, other than sorting, to control the order
of documents returned. For example, {"x" : [1, 2, 3]} will return documents in the
same order as {"x" : [3, 2, 1]}.
Indexing Objects and Arrays
MongoDB allows you to reach into your documents and create indexes on nested fields
and arrays. Embedded object and array fields can be combined with top-level fields in
Introduction to Indexing | 95
compound indexes and although they are special in some ways, they mostly behave the
way “normal” index fields behave.
Indexing embedded docs
Indexes can be created on keys in embedded documents in the same way that they are
created on normal keys. If we had a collection where each document represented a user,
we might have an embedded document that described each user’s location:
{
"username" : "sid",
"loc" : {
"ip" : "1.2.3.4",
"city" : "Springfield",
"state" : "NY"
}
}
We could put an index on one of the subfields of "loc", say "loc.city", to speed up
queries using that field:
> db.users.ensureIndex({"loc.city" : 1})
You can go as deep as you’d like with these: you could index "x.y.z.w.a.b.c" (and so
on) if you wanted.
Note that indexing the embedded document itself ("loc") has very different behavior
than indexing a field of that embedded document ("loc.city"). Indexing the entire
subdocument will only help queries that are querying for the entire subdocument. In
the example above, the query optimizer could only use an index on "loc" for queries
that described the whole subdocument with fields in the correct order (e.g.,
db.users.find({"loc" : {"ip" : "123.456.789.000", "city" : "Shelbyville",
"state" : "NY"}}})). It could not use the index for queries that looked like
db.users.find({"loc.city" : "Shelbyville"}).
Indexing arrays
You can also index arrays, which allows you to use the index to search for specific array
elements efficiently.
Suppose we have a collection of blog posts where each document was a post. Each post
has a "comments" field, which is an array of comment subdocuments. If we want to be
able to find the most-recently-commented-on blog posts, we could create an index on
the "date" key in the array of embedded "comments" documents of our blog post
collection:
> db.blog.ensureIndex({"comments.date" : 1})
Indexing an array creates an index entry for each element of the array, so if a post had
20 comments, it would have 20 index entries. This makes array indexes more expensive
96 | Chapter 5: Indexing
than single-value ones: for a single insert, update, or remove, every array entry might
have to be updated (potentially thousands of index entries).
Unlike the "loc" example in the previous section, you cannot index an entire array as
a single entity: indexing an array field indexes each element of the array, not the array
itself.
Indexes on array elements do not keep any notion of position: you cannot use an index
for a query that is looking for a specific array element, such as "comments.4".
You can, incidentally, index a specific array entry, for example:
> db.blog.ensureIndex({"comments.10.votes": 1})
However, this index would only be useful for queries for exactly the 11th array element
(arrays start at index 0).
Only one field in an index entry can be from an array. This is to avoid the explosive
number of index entries you’d get from multiple multikey indexes: every possible pair
of elements would have to be indexed, causing indexes to be n*m entries per document.
For example, suppose we had an index on {"x" : 1, "y" : 1}:
> // x is an array - legal
> db.multi.insert({"x" : [1, 2, 3], "y" : 1})
>
> // y is an array - still legal
> db.multi.insert({"x" : 1, "y" : [4, 5, 6]})
>
> // x and y are arrays - illegal!
> db.multi.insert({"x" : [1, 2, 3], "y" : [4, 5, 6]})
cannot index parallel arrays [y] [x]
Were MongoDB to index the final example, it would have to create index entries for
{"x" : 1, "y" : 4}, {"x" : 1, "y" : 5}, {"x" : 1, "y" : 6}, {"x" : 2, "y" :
4}, {"x" : 2, "y" : 5}, {"x" : 2, "y" : 6}, {"x" : 3, "y" : 4}, {"x" : 3, "y" :
5}, and {"x" : 3, "y" : 6} (and these arrays are only three elements long).
Multikey index implications
If any document has an array field for the indexed key, the index immediately is flagged
as a multikey index. You can see whether an index is multikey from explain()’s output:
if a multikey index was used, the "isMultikey" field will be true. Once an index has
been flagged as multikey, it can never be un-multikeyed, even if all of the documents
containing arrays in that field are removed. The only way to un-multikey it is to drop
and recreate it.
Multikey indexes may be a bit slower than non-multikey indexes. Many index entries
can point at a single document so MongoDB may need to do some de-duping before
returning results.
Introduction to Indexing | 97
Index Cardinality
Cardinality refers to how many distinct values there are for a field in a collection. Some
fields, such as "gender" or "newsletter opt-out", might only have two possible values,
which is considered a very low cardinality. Others, such as "username" or "email",
might have a unique value for every document in the collection, which is high cardin‐
ality. Still others fall somewhere in between, such as "age" or "zip code".
In general, the greater the cardinality of a field, the more helpful an index on that field
can be. This is because the index can quickly narrow the search space to a much smaller
result set. For a low cardinality field, an index generally cannot eliminate as many pos‐
sible matches.
For example, suppose we had an index on "gender" and were looking for women named
Susan. We could only narrow down the result space by approximately 50% before re‐
ferring to individual documents to look up "name". Conversely, if we indexed by
"name", we could immediately narrow down our result set to the tiny fraction of users
named Susan and then we could refer to those documents to check the gender.
As a rule of thumb, try to create indexes on high-cardinality keys or at least put high-
cardinality keys first in compound indexes (before low-cardinality keys).
Using explain() and hint()
As you have seen above, explain() gives you lots of information about your queries. It
is one of the most important diagnostic tools there is for slow queries. You can find out
which indexes are being used and how by looking at a query’s explain. For any query,
you can add a call to explain() at the end (the way you would add a sort() or lim
it(), but explain() must be the last call).
There are two types of explain() output that you’ll see most commonly: indexed and
non-indexed queries. Special index types may create slightly different query plans, but
most fields should be similar. Also, sharding returns a conglomerate of explain()s (as
covered in Chapter 13), as it runs the query on multiple servers.
The most basic type of explain() is on a query that doesn’t use an index. You can tell
that a query doesn’t use an index because it uses a "BasicCursor". Conversely, most
queries that use an index use a BtreeCursor (some special types of indexes, such as
geospatial indexes, use their own type of cursor).
The output to an explain() on a query that uses an index varies, but in the simplest
case, it looks something like this:
> db.users.find({"age" : 42}).explain()
{
"cursor" : "BtreeCursor age_1_username_1",
"isMultiKey" : false,
98 | Chapter 5: Indexing
"n" : 8332,
"nscannedObjects" : 8332,
"nscanned" : 8332,
"nscannedObjectsAllPlans" : 8332,
"nscannedAllPlans" : 8332,
"scanAndOrder" : false,
"indexOnly" : false,
"nYields" : 0,
"nChunkSkips" : 0,
"millis" : 91,
"indexBounds" : {
"age" : [
[
42,
42
]
],
"username" : [
[
{
"$minElement" : 1
},
{
"$maxElement" : 1
}
]
]
},
"server" : "ubuntu:27017"
}
This output first tells you what index was used: age_1_username_1. "millis" reports
how fast the query was executed, from the server receiving the request to when it sent
a response. However, it may not always be the number you are looking for. If MongoDB
tried multiple query plans, "millis" will reflect how long it took all of them to run, not
the one chosen as the best.
Next is how many documents were actually returned as a result: "n". This doesn’t reflect
how much work MongoDB did to answer the query: how many index entries and
documents did it have to search? Index entries are described by "nscanned". The num‐
ber of documents scanned is reflected in "nscannedObjects". Finally, if you were using
a sort and MongoDB could not use an index for it, "scanAndOrder" would be true. This
means that MongoDB had to sort the results in memory, which is generally quite slow
and limited to a small number of results.
Now that you know the basics, here is a breakdown of the all of the fields in more detail:
"cursor" : "BtreeCursor age_1_username_1"
BtreeCursor means that an index was used, specifically, the index on age and user‐
name: {"age" : 1, "username" : 1}. You may also see reverse (if the query is
Using explain() and hint() | 99
traversing the index in reverse direction—say for a sort) or multi, if it is using a
multikey index.
"isMultiKey" : false
If this query used a multikey index (see “Indexing Objects and Arrays” on page 95).
"n" : 8332
This is the number of documents returned by the query.
"nscannedObjects" : 8332
This is a count of the number of times MongoDB had to follow an index pointer to
the actual document on disk. If the query contains criteria that is not part of the
index or requests fields back that aren’t contained in the index, MongoDB must
look up the document each index entry points to.
"nscanned" : 8332
The number of index entries looked at if an index was used. If this was a table scan,
it is the number of documents examined.
"scanAndOrder" : false
If MongoDB had to sort results in memory.
"indexOnly" : false
If MongoDB was able to fulfill this query using only the index entries (as discussed
in “Using covered indexes” on page 90).
In this example, MongoDB found all matching documents using the index, which
we know because "nscanned" is the same as "n". However, the query was told to
return every field in the matching documents and the index only contained the
"age" and "username" fields. If we changed the query to have a second argument,
{"_id" : 0, "age" : 1, "username" : 1}, then it would be covered by the index
and "indexOnly" would be true.
"nYields" : 0
The number of times this query yielded (paused) to allow a write request to proceed.
If there are writes waiting to go, queries will periodically release their lock and allow
them to do so. However, on this system, there were no writes waiting so the query
never yielded.
"millis" : 91
The number of milliseconds it took the database to execute the query. The lower
this number is, the better.
"indexBounds" : {...}
This field describes how the index was used, giving ranges of the index traversed.
As the first clause in the query was an exact match, the index only needed to look
at that value: 42. The second index key was a free variable, as the query didn’t specify
100 | Chapter 5: Indexing
any restrictions to it. Thus, the database looked for values between negative infinity
("$minElement" : 1) and infinity("$maxElement" : 1) for usernames within
"age" : 42.
Let’s take a slightly more complicated example: suppose you have an index on {"user
name" : 1, "age" : 1} and an index on {"age" : 1, "username" : 1}. What
happens if you query for "username" and "age"? Well, it depends on the query:
> db.c.find({age : {$gt : 10}, username : "sally"}).explain()
{
"cursor" : "BtreeCursor username_1_age_1",
"indexBounds" : [
[
{
"username" : "sally",
"age" : 10
},
{
"username" : "sally",
"age" : 1.7976931348623157e+308
}
]
],
"nscanned" : 13,
"nscannedObjects" : 13,
"n" : 13,
"millis" : 5
}
We are querying for an exact match on "username" and a range of values for "age", so
the database chooses to use the {"username" : 1, "age" : 1} index, reversing the
terms of the query. If, on the other hand, we query for an exact age and a range of names,
MongoDB will use the other index:
> db.c.find({"age" : 14, "username" : /.*/}).explain()
{
"cursor" : "BtreeCursor age_1_username_1 multi",
"indexBounds" : [
[
{
"age" : 14,
"username" : ""
},
{
"age" : 14,
"username" : {
}
}
],
[
{
Using explain() and hint() | 101
"age" : 14,
"username" : /.*/
},
{
"age" : 14,
"username" : /.*/
}
]
],
"nscanned" : 2,
"nscannedObjects" : 2,
"n" : 2,
"millis" : 2
}
If you find that Mongo is using different indexes than you want it to for a query, you
can force it to use a certain index by using hint(). For instance, if you want to make
sure MongoDB uses the {"username" : 1, "age" : 1} index on the previous query,
you could say the following:
> db.c.find({"age" : 14, "username" : /.*/}).hint({"username" : 1, "age" : 1})
If a query is not using the index that you want it to and you use a hint to change it, run
an explain() on the hinted query before deploying. If you force MongoDB to use an
index on a query that it does not know how to use an index for, you could end up making
the query less efficient than it was without the index.
The Query Optimizer
MongoDB’s query optimizer works a bit differently than any other database’s. Basically,
if an index exactly matches a query (you are querying for "x" and have an index on
"x"), the query optimizer will use that index. Otherwise, there might be a few possible
indexes that could work well for your query. MongoDB will select a subset of likely
indexes and run the query once with each plan, in parallel. The first plan to return 100
results is the “winner” and the other plans’ executions are halted.
This plan is cached and used subsequently for that query until the collection has seen a
certain amount of churn. Once the collection has changed a certain amount since the
initial plan evaluation, the query optimizer will re-race the possible plans. Plans will
also be reevaluated after index creation or every 1,000 queries.
The "allPlans" field in explain()’s output shows each plan the query tried running.
When Not to Index
Indexes are most effective at retrieving small subsets of data and some types of queries
are faster without indexes. Indexes become less and less efficient as you need to get
larger percentages of a collection because using an index requires two lookups: one to
102 | Chapter 5: Indexing
look at the index entry and one following the index’s pointer to the document. A table
scan only requires one: looking at the document. In the worst case (returning all of the
documents in a collection) using an index would take twice as many lookups and would
generally be significantly slower than a table scan.
Unfortunately, there isn’t a hard-and-fast rule about when an index helps and when it
hinders as it really depends on the size of your data, size of your indexes, size of your
documents, and the average result set size (Table 5-1). As a rule of thumb: if a query is
returning 30% or more of the collection, start looking at whether indexes or table scans
are faster. However, this number can vary from 2% to 60%.
Table 5-1. Properties that affect the effectiveness of indexes
Indexes often work well for Table scans often work well for
Large collections Small collections
Large documents Small documents
Selective queries Non-selective queries
Let’s say we have an analytics system that collects statistics. Your application queries the
system for all documents for a given account to generate a nice graph of all data from
an hour ago to the beginning of time:
> db.entries.find({"created_at" : {"$lt" : hourAgo}})
We index "created_at" to speed up this query.
When we first launch, this is a tiny result set and returns instantly. But after a couple
weeks, it starts being a lot of data, and after a month this query is already taking too
long to run.
For most applications, this is probably the “wrong” query: do you really want a query
that’s returning most of your data set? Most applications, particularly those with large
data sets, do not. However, there are some legitimate cases where you may want most
or all of your data: you might be exporting this data to a reporting system or using it for
a batch job. In these cases, you would like to return this large proportion of the data set
as fast as possible.
You can force it to do a table scan by hinting {"$natural" : 1}. As described in “Capped
Collections” on page 109, $natural specifies on-disk order when used in a sort. In par‐
ticular, $natural forces MongoDB to do a table scan:
> db.entries.find({"created_at" : {"$lt" : hourAgo}}).hint({"$natural" : 1})
One side effect of sorting by "$natural" is that it gives you results in on-disk order.
This is generally meaningless for an active collection: as documents grow and shrink
they’ll be moved around on disk and new documents will be written in the “holes” they
left. However, for insert-only workloads, $natural can be useful for giving you the latest
(or earliest) documents.
When Not to Index | 103
Types of Indexes
There are a few index options you can specify when building an index that change the
way the index behaves. The most common variations are described in the following
sections, and more advanced or special-case options are described in the next chapter.
Unique Indexes
Unique indexes guarantee that each value will appear at most once in the index. For
example, if you want to make sure no two documents can have the same value in the
"username" key, you can create a unique index:
> db.users.ensureIndex({"username" : 1}, {"unique" : true})
For example, suppose that we try to insert the following documents on the collection
above:
> db.users.insert({username: "bob"})
> db.users.insert({username: "bob"})
E11000 duplicate key error index: test.users.$username_1 dup key: { : "bob" }
If you check the collection, you’ll see that only the first "bob" was stored. Throwing
duplicate key exceptions is not very efficient, so use the unique constraint for the oc‐
casional duplicate, not to filter out zillions of duplicates a second.
A unique index that you are probably already familiar with is the index on "_id", which
is automatically created whenever you create a collection. This is a normal unique index
(aside from the fact that it cannot be dropped as other unique indexes can be).
If a key does not exist, the index stores its value as null for that docu‐
ment. This means that if you create a unique index and try to insert
more than one document that is missing the indexed field, the inserts
will fail because you already have a document with a value of null. See
“Sparse Indexes” on page 106 for advice on handling this.
In some cases a value won’t be indexed. Index buckets are of limited size and if an index
entry exceeds it, it just won’t be included in the index. This can cause confusion as it
makes a document “invisible” to queries that use the index. All fields must be smaller
than 1024 bytes to be included in an index. MongoDB does not return any sort of error
or warning if a document’s fields cannot be indexed due to size. This means that keys
longer than 8 KB will not be subject to the unique index constraints: you can insert
identical 8 KB strings, for example.
104 | Chapter 5: Indexing
Compound unique indexes
You can also create a compound unique index. If you do this, individual keys can have
the same values, but the combination of values across all keys in an index entry can
appear in the index at most once.
For example, if we had a unique index on {"username" : 1, "age" : 1}, the following
inserts would be legal:
> db.users.insert({"username" : "bob"})
> db.users.insert({"username" : "bob", "age" : 23})
> db.users.insert({"username" : "fred", "age" : 23})
However, attempting to insert a second copy of any of these documents would cause a
duplicate key exception.
GridFS, the standard method for storing large files in MongoDB (see “Storing Files with
GridFS” on page 123), uses a compound unique index. The collection that holds the file
content has a unique index on {"files_id" : 1, "n" : 1}, which allows documents
that look like (in part) the following:
{"files_id" : ObjectId("4b23c3ca7525f35f94b60a2d"), "n" : 1}
{"files_id" : ObjectId("4b23c3ca7525f35f94b60a2d"), "n" : 2}
{"files_id" : ObjectId("4b23c3ca7525f35f94b60a2d"), "n" : 3}
{"files_id" : ObjectId("4b23c3ca7525f35f94b60a2d"), "n" : 4}
Note that all of the values for "files_id" are the same, but "n" is different.
Dropping duplicates
If you attempt to build a unique index on an existing collection, it will fail to build if
there are any duplicate values:
> db.users.ensureIndex({"age" : 1}, {"unique" : true})
E11000 duplicate key error index: test.users.$age_1 dup key: { : 12 }
Generally, you’ll need to process your data (the aggregation framework can help) and
figure out where the duplicates are and what to do with them.
In a few rare cases, you may just want to delete documents with duplicate values. The
"dropDups" option will save the first document found and remove any subsequent
documents with duplicate values:
> db.people.ensureIndex({"username" : 1}, {"unique" : true, "dropDups" : true})
"dropDups" forces the unique index build, but it’s a very drastic option; you have no
control over which documents are dropped and which are kept (and MongoDB gives
you no indication of which documents were dropped, if any). If your data is of any
importance, do not use "dropDups".
Types of Indexes | 105
Sparse Indexes
As mentioned in an earlier section, unique indexes count null as a value, so you cannot
have a unique index with more than one document missing the key. However, there are
lots of cases where you may want the unique index to be enforced only if the key exists.
If you have a field that may or may not exist but must be unique when it does, you can
combine the unique option with the sparse option.
If you are familiar with sparse indexes on relational databases, Mon‐
goDB’s sparse indexes are a completely different concept. MongoDB
sparse indexes are basically indexes that need not include every docu‐
ment as an entry.
To create a sparse index, include the sparse option. For example, if providing an email
address was optional but, if provided, should be unique, we could do:
> db.ensureIndex({"email" : 1}, {"unique" : true, "sparse" : true})
Sparse indexes do not necessarily have to be unique. To make a non-unique sparse index,
simply do not include the unique option.
One thing to be aware of is that the same query can return different results depending
on whether or not it uses the sparse index. For example, suppose we had a collection
where most of the documents had "x" fields, but one does not:
> db.foo.find()
{ "_id" : 0 }
{ "_id" : 1, "x" : 1 }
{ "_id" : 2, "x" : 2 }
{ "_id" : 3, "x" : 3 }
When we do a query on "x", it will return all matching documents:
> db.foo.find({"x" : {"$ne" : 2}})
{ "_id" : 0 }
{ "_id" : 1, "x" : 1 }
{ "_id" : 3, "x" : 3 }
If we create a sparse index on "x", the "_id" : 0 document won’t be included in the
index. So now if we query on "x", MongoDB will use the index and not return the
{"_id" : 0} document:
> db.foo.find({"x" : {"$ne" : 2}})
{ "_id" : 1, "x" : 1 }
{ "_id" : 3, "x" : 3 }
You can use hint() to force it to do a table scan if you need documents with missing
fields.
106 | Chapter 5: Indexing
Index Administration
As shown in the previous section, you can create new indexes using the ensureIndex
function. An index only needs to be created once per collection. If you try to create the
same index again, nothing will happen.
All of the information about a database’s indexes is stored in the system.indexes collec‐
tion. This is a reserved collection, so you cannot modify its documents or remove
documents from it. You can manipulate it only through ensureIndex and the dropIn
dexes database command.
When you create an index, you can see its meta information in system.indexes. You can
also run db.collectionName.getIndexes() to see all index information about a given
collection:
> db.foo.getIndexes()
[
{
"v" : 1,
"key" : {
"_id" : 1
},
"ns" : "test.foo",
"name" : "_id_"
},
{
"v" : 1,
"key" : {
"y" : 1
},
"ns" : "test.foo",
"name" : "y_1"
},
{
"v" : 1,
"key" : {
"x" : 1,
"y" : 1
},
"ns" : "test.foo",
"name" : "x_1_y_1"
}
]
The important fields are the "key" and "name". The key can be used for hinting, max,
min, and other places where an index must be specified. This is a place where field order
matters: an index on {"x" : 1, "y" : 1} is not the same as an index on {"y" : 1,
"x" : 1}. The index name is used as identifier for a lot of administrative index opera‐
tions, such as dropIndex. Whether or not the index is multikey is not specified in its
spec.
Index Administration | 107
The "v" field is used internally for index versioning. If you have any indexes that do not
have a "v" : 1 field, they are being stored in an older, less efficient format. You can
upgrade them by ensuring that you’re running at least MongoDB version 2.0 and drop‐
ping and rebuilding the index.
Identifying Indexes
Each index in a collection has a name that uniquely identifies the index and is used by
the server to delete or manipulate it. Index names are, by default, key
name1_dir1_keyname2_dir2_..._keynameN_dirN, where keynameX is the index’s key
and dirX is the index’s direction (1 or -1). This can get unwieldy if indexes contain more
than a couple keys, so you can specify your own name as one of the options to ensur
eIndex:
> db.foo.ensureIndex({"a" : 1, "b" : 1, "c" : 1, ..., "z" : 1},
... {"name" : "alphabet"})
There is a limit to the number of characters in an index name, so complex indexes may
need custom names to be created. A call to getLastError will show if the index creation
succeeded or why it didn’t.
Changing Indexes
As your application grows and changes, you may find that your data or queries have
changed and that indexes that used to work well no longer do. You can remove unneeded
indexes using the dropIndex command:
> db.people.dropIndex("x_1_y_1")
{ "nIndexesWas" : 3, "ok" : 1 }
Use the "name" field from the index description to specify which index to drop.
Building new indexes is time-consuming and resource-intensive. By default, MongoDB
will build an index as fast as possible, blocking all reads and writes on a database until
the index build has finished. If you would like your database to remain somewhat re‐
sponsive to reads and writes, use the background option when building an index. This
forces the index build to occasionally yield to other operations, but may still have a
severe impact on your application (see “Building Indexes” on page 222 for more infor‐
mation). Background indexing is also much slower than foreground indexing.
If you have the choice, creating indexes on existing documents is slightly faster than
creating the index first and then inserting all documents.
There is more on the operational aspects of building indexes in Chapter 18.
108 | Chapter 5: Indexing
CHAPTER 6
Special Index and Collection Types
This chapter covers the special collections and index types MongoDB has available,
including:
• Capped collections for queue-like data
• TTL indexes for caches
• Full-text indexes for simple string searching
• Geospatial indexes for 2D and spherical geometries
• GridFS for storing large files
Capped Collections
“Normal” collections in MongoDB are created dynamically and automatically grow in
size to fit additional data. MongoDB also supports a different type of collection, called
a capped collection, which is created in advance and is fixed in size (see Figure 6-1).
Having fixed-size collections brings up an interesting question: what happens when we
try to insert into a capped collection that is already full? The answer is that capped
collections behave like circular queues: if we’re out of space, the oldest document will
be deleted, and the new one will take its place (see Figure 6-2). This means that capped
collections automatically age-out the oldest documents as new documents are inserted.
Certain operations are not allowed on capped collections. Documents cannot be re‐
moved or deleted (aside from the automatic age-out described earlier), and updates that
would cause documents to grow in size are disallowed. By preventing these two oper‐
ations, we guarantee that documents in a capped collection are stored in insertion order
and that there is no need to maintain a free list for space from removed documents.
109
Figure 6-1. New documents are inserted at the end of the queue
Figure 6-2. When the queue is full, the oldest element will be replaced by the newest
110 | Chapter 6: Special Index and Collection Types
Capped collections have a different access pattern than most MongoDB collections: data
is written sequentially over a fixed section of disk. This makes them tend to perform
writes quickly on spinning disk, especially if they can be given their own disk (so as not
to be “interrupted” by other collections’ random writes).
Capped collections cannot be sharded.
Capped collections tend to be useful for logging, although they lack flexibility: you
cannot control when data ages out, other than setting a size when you create the
collection.
Creating Capped Collections
Unlike normal collections, capped collections must be explicitly created before they are
used. To create a capped collection, use the create command. From the shell, this can
be done using createCollection:
> db.createCollection("my_collection", {"capped" : true, "size" : 100000});
{ "ok" : true }
The previous command creates a capped collection, my_collection, that is a fixed size
of 100,000 bytes.
createCollection can also specify a limit on the number of documents in a capped
collection in addition to the limit size:
> db.createCollection("my_collection2",
... {"capped" : true, "size" : 100000, "max" : 100});
{ "ok" : true }
You could use this to keep, say, the latest 10 news articles or limit a user to 1,000
documents.
Once a capped collection has been created, it cannot be changed (it must be dropped
and recreated if you wish to change its properties). Thus, you should think carefully
about the size of a large collection before creating it.
When limiting the number of documents in a capped collection, you
must specify a size limit as well. Age-out will be based on whichever
limit is reached first: it cannot hold more than "max" documents nor
take up more than "size" space.
Capped Collections | 111
Another option for creating a capped collection is to convert an existing, regular col‐
lection into a capped collection. This can be done using the convertToCapped
command—in the following example, we convert the test collection to a capped collec‐
tion of 10,000 bytes:
> db.runCommand({"convertToCapped" : "test", "size" : 10000});
{ "ok" : true }
There is no way to “uncap” a capped collection (other than dropping it).
Sorting Au Naturel
There is a special type of sort that you can do with capped collections, called a natural
sort. A natural sort returns the documents in the order that they appear on disk (see
Figure 6-3).
For most collections, this isn’t a very useful sort because documents move around.
However, documents in a capped collection are always kept in insertion order so that
natural order is the same as insertion order. Thus, a natural sort gives you documents
from oldest to newest. You can also sort from newest to oldest (see Figure 6-4):
> db.my_collection.find().sort({"$natural" : -1})
Figure 6-3. Sort by {"$natural” : 1}
112 | Chapter 6: Special Index and Collection Types
Figure 6-4. Sort by {"$natural” : -1}
Tailable Cursors
Tailable cursors are a special type of cursor that are not closed when their results are
exhausted. They were inspired by the tail -f command and, similar to the command,
will continue fetching output for as long as possible. Because the cursors do not die
when they run out of results, they can continue to fetch new results as documents are
added to the collection. Tailable cursors can be used only on capped collections, since
insert order is not tracked for normal collections.
Tailable cursors are often used for processing documents as they are inserted onto a
“work queue” (the capped collection). Because tailable cursors will time out after 10
minutes of no results, it is important to include logic to re-query the collection if they
die. The mongo shell does not allow you to use tailable cursors, but using one in PHP
looks something like the following:
$cursor = $collection->find()->tailable();
while (true) {
if (!$cursor->hasNext()) {
if ($cursor->dead()) {
break;
}
sleep(1);
}
else {
while ($cursor->hasNext()) {
Capped Collections | 113
do_stuff($cursor->getNext());
}
}
}
The cursor will process results or wait for more results to arrive until the cursor dies (it
will time out if there are no inserts for 10 minutes or someone kills the query operation).
No-_id Collections
By default, every collection has an "_id" index. However, you can create collections
with "_id" indexes by setting the autoIndexId option to false when calling createCol
lection. This is not recommended but can give you a slight speed boost on an insert-
only collection.
If you create a collection without an "_id" index, you will never be able
replicate the mongod it lives on. Replication requires the "_id" index
on every collection (it is important that replication can uniquely iden‐
tify each document in a collection).
Capped collections prior to version 2.2 did not have an "_id" index unless autoIndex
Id was explicitly set to true. If you are working with an “old” capped collection, ensure
that your application is populating the "_id" field (most drivers will do this automati‐
cally) and then create the "_id" index using ensureIndex.
Remember to make the "_id" index unique. Do a practice run before creating the index
in production, as unlike other indexes, the "_id" index cannot be dropped once created.
Thus, you must get it right the first time! If you do not, you cannot change it without
dropping the collection and recreating it.
Time-To-Live Indexes
As mentioned in the previous section, capped collections give you limited control over
when their contents are overwritten. If you need a more flexible age-out system, time-
to-live (TTL) indexes allow you to set a timeout for each document. When a document
reaches a preconfigured age, it will be deleted. This type of index is useful for caching
problems like session storage.
You can create a TTL index by specifying the expireAfterSecs option in the second
argument to ensureIndex:
> // 24-hour timeout
> db.foo.ensureIndex({"lastUpdated" : 1}, {"expireAfterSecs" : 60*60*24})
114 | Chapter 6: Special Index and Collection Types
This creates a TTL index on the "lastUpdated" field. If a document’s "lastUpdated"
field exists and is a date, the document will be removed once the server time is expir
eAfterSecs seconds ahead of the document’s time.
To prevent an active session from being removed, you can update the "lastUpdated"
field to the current time whenever there is activity. Once "lastUpdated" is 24 hours
old, the document will be removed.
MongoDB sweeps the TTL index once per minute, so you should not depend on to-
the-second granularity. You can change the expireAfterSecs using the collMod
command:
> db.runCommand({"collMod" : "someapp.cache", "expireAfterSecs" : 3600})
You can have multiple TTL indexes on a given collection. They cannot be compound
indexes but can be used like “normal” indexes for the purposes of sorting and query
optimization.
Full-Text Indexes
MongoDB has a special type of index for searching for text within documents. In pre‐
vious chapters, we’ve queried for strings using exact matches and regular expressions,
but these techniques have some limitations. Searching a large block of text for a regular
expression is slow and it’s tough to take linguistic issues into account (e.g., that “entry”
should match “entries”). Full-text indexes give you the ability to search text quickly, as
well as provide built-in support for multi-language stemming and stop words.
While all indexes are expensive to create, full-text indexes are particularly heavyweight.
Creating a full-text index on a busy collection can overload MongoDB, so adding this
type of index should always be done offline or at a time when performance does not
matter. You should be wary of creating full-text indexes that will not fit in RAM (unless
you have SSDs). See Chapter 18 for more information on creating indexes with minimal
impact on your application.
Full-text search will also incur more severe performance penalties on writes than “nor‐
mal” indexes, since all strings must be split, stemmed, and stored in a few places. Thus,
you will tend to see poorer write performance on full-text-indexed collections than on
others. It will also slow down data movement if you are sharding: all text must be rein‐
dexed when it is migrated to a new shard.
As of this writing, full text indexes are an “experimental” feature, so you must enable
them specifically. You can either start MongoDB with the --setParameter textSearch
Enabled=true option or set it at runtime by running the setParameter command:
> db.adminCommand({"setParameter" : 1, "textSearchEnabled" : true})
Full-Text Indexes | 115
Suppose we use the unofficial Hacker News JSON API to load some recent stories into
MongoDB.
To run a search over the text, we first need to create a "text" index:
> db.hn.ensureIndex({"title" : "text"})
Now, to use the index, we must use the text command (as of this writing, full text indexes
cannot be used with “normal” queries):
test> db.runCommand({"text" : "hn", "search" : "ask hn"})
{
"queryDebugString" : "ask|hn||||||",
"language" : "english",
"results" : [
{
"score" : 2.25,
"obj" : {
"_id" : ObjectId("50dcab296803fa7e4f000011"),
"title" : "Ask HN: Most valuable skills you have?",
"url" : "/comments/4974230",
"id" : 4974230,
"commentCount" : 37,
"points" : 31,
"postedAgo" : "2 hours ago",
"postedBy" : "bavidar"
}
},
{
"score" : 0.5625,
"obj" : {
"_id" : ObjectId("50dcab296803fa7e4f000001"),
"title" : "Show HN: How I turned an old book...",
"url" : "http://www.howacarworks.com/about",
"id" : 4974055,
"commentCount" : 44,
"points" : 95,
"postedAgo" : "2 hours ago",
"postedBy" : "AlexMuir"
}
},
{
"score" : 0.5555555555555556,
"obj" : {
"_id" : ObjectId("50dcab296803fa7e4f000010"),
"title" : "Show HN: ShotBlocker - iOS Screenshot detector...",
"url" : "https://github.com/clayallsopp/ShotBlocker",
"id" : 4973909,
"commentCount" : 10,
"points" : 17,
"postedAgo" : "3 hours ago",
"postedBy" : "10char"
}
116 | Chapter 6: Special Index and Collection Types
}
],
"stats" : {
"nscanned" : 4,
"nscannedObjects" : 0,
"n" : 3,
"timeMicros" : 89
},
"ok" : 1
}
The matching documents are returned in order of decreasing relevance: “Ask HN” is
first, then two “Show HN” partial matches. The "score" field before each object de‐
scribes how closely the result matched the query.
As you can see from the results, the search is case insensitive, at least for characters in
[a-zA-Z]. Full-text indexes use toLower to lowercase words, which is locale-dependant,
so users of other languages may find MongoDB unpredictably case sensitive, depending
on how toLower behaves on their character set. Better collation support is in the works.
Full text indexes only index string data: other data types are ignored and not included
in the index. Only one full-text index is allowed per collection, but it may contain mul‐
tiple fields:
> db.blobs.ensureIndex({"title" : "text", "desc" : "text", "author" : "text"})
This is not like “nomal” multikey indexes where there is an ordering on the keys: each
field is given equal consideration. You can control the relative importance MongoDB
attaches to each field by specifying a weight:
> db.hn.ensureIndex({"title" : "text", "desc" : "text", "author" : "text"},
... {"weights" : {"title" : 3, "author" : 2}})
The default weight is 1, and you may use weights from 1 to 1 billion. The weights above
would weight "title" fields the most, followed by "author" and then "desc" (not
specified in the weight list, so given a default weight of 1).
You cannot change field weights after index creation (without dropping the index and
recreating it), so you may want to play with weights on a sample data set before creating
the index on your production data.
For some collections, you may not know which fields a document will contain. You can
create a full-text index on all string fields in a document by creating an index on
"$**": this not only indexes all top-level string fields, but also searches embedded
documents and arrays for string fields:
> db.blobs.ensureIndex({"$**" : "text"})
You can also give "$**" a weight:
> db.hn.ensureIndex({"whatever" : "text"},
... {"weights" : {"title" : 3, "author" : 1, "$**" : 2}})
Full-Text Indexes | 117
"whatever" can be anything since it is not used. As the weights specify that you’re
indexing all fields, MongoDB does not require you to give a field list.
Search Syntax
By default, MongoDB queries for an OR of all the words: “ask OR hn”. This is the most
efficient way to perform a full text query, but you can also do exact phrase searches and
NOT. To search for the exact phrase “ask hn”, you can query for that by including the
query in quotes:
> db.runCommand({text: "hn", search: "\"ask hn\""})
{
"queryDebugString" : "ask|hn||||ask hn||",
"language" : "english",
"results" : [
{
"score" : 2.25,
"obj" : {
"_id" : ObjectId("50dcab296803fa7e4f000011"),
"title" : "Ask HN: Most valuable skills you have?",
"url" : "/comments/4974230",
"id" : 4974230,
"commentCount" : 37,
"points" : 31,
"postedAgo" : "2 hours ago",
"postedBy" : "bavidar"
}
}
],
"stats" : {
"nscanned" : 4,
"nscannedObjects" : 0,
"n" : 1,
"nfound" : 1,
"timeMicros" : 20392
},
"ok" : 1
}
This is slower than the OR-type match, since MongoDB first performs an OR match
and then post-processes the documents to ensure that they are AND matches, as well.
You can also make part of a query literal and part not:
> db.runCommand({text: "hn", search: "\"ask hn\" ipod"})
This will search for exactly "ask hn" and, optionally, "ipod".
You can also search for not including a certain string by using "-":
> db.runCommand({text: "hn", search: "-startup vc"})
118 | Chapter 6: Special Index and Collection Types
This will return results that match “vc” and don’t include the word “startup”.
Full-Text Search Optimization
There are a couple ways to optimize full text searches. If you can first narrow your search
results by other criteria, you can create a compound index with a prefix of the other
criteria and then the full-text fields:
> db.blog.ensureIndex({"date" : 1, "post" : "text"})
This is referred to as partitioning the full-text index, as it breaks it into several smaller
trees based on "date" (in the example above). This makes full-text searches for a certain
date much faster.
You can also use a postfix of other criteria to cover queries with the index. For example,
if we were only returning the "author" and "post" fields, we could create a compound
index on both:
> db.blog.ensureIndex({"post" : "text", "author" : 1})
These prefix and postfix forms can be combined:
> db.blog.ensureIndex({"date" : 1, "post" : "text", "author" : 1})
You cannot use a multikey field for any of the prefix or postfix index fields.
Creating a full-text index automatically enables the usePowerOf2Sizes option on the
collection, which controls how space is allocated. Do not disable this option, since it
should improve writes speed.
Searching in Other Languages
When a document is inserted (or the index is first created), MongoDB looks at the
indexes fields and stems each word, reducing it to an essential unit. However, different
languages stem words in different ways, so you must specify what language the index
or document is. Thus, text-type indexes allow a "default_language" option to be
specified, which defaults to "english" but can be set to a number of other languages
(see the online documentation for an up-to-date list).
For example, to create a French-language index, we could say:
> db.users.ensureIndex({"profil" : "text", "intérêts" : "text"},
... {"default_language" : "french"})
Then French would be used for stemming, unless otherwise specified. You can, on a
per-document basis, specify another stemming language by having a "language" field
that describes the document’s language:
> db.users.insert({"username" : "swedishChef",
... "profile" : "Bork de bork", language : "swedish"})
Full-Text Indexes | 119
Geospatial Indexing
MongoDB has a few types of geospatial indexes. The most commonly used ones are
2dsphere, for surface-of-the-earth-type maps, and 2d, for flat maps (and time series
data).
2dsphere allows you to specify points, lines, and polygons in GeoJSON format. A point
is given by a two-element array, representing [longitude, latitude]:
{
"name" : "New York City",
"loc" : {
"type" : "Point",
"coordinates" : [50, 2]
}
}
A line is given by an array of points:
{
"name" : "Hudson River",
"loc" : {
"type" : "Line",
"coordinates" : [[0,1], [0,2], [1,2]]
}
}
A polygon is specified the same way a line is (an array of points), but with a different
"type":
{
"name" : "New England",
"loc" : {
"type" : "Polygon",
"coordinates" : [[0,1], [0,2], [1,2]]
}
}
The "loc" field can be called anything, but the field names within its subobject are
specified by GeoJSON and cannot be changed.
You can create a geospatial index using the "2dsphere" type with ensureIndex:
> db.world.ensureIndex({"loc" : "2dsphere"})
Types of Geospatial Queries
There are several types of geospatial query that you can perform: intersection, within,
and nearness. To query, specify what you’re looking for as a GeoJSON object that looks
like {"$geometry" : geoJsonDesc}.
120 | Chapter 6: Special Index and Collection Types
For example, you can find documents that intersect the query’s location using the "$ge
oIntersects" operator:
> var eastVillage = {
... "type" : "Polygon",
... "coordinates" : [
... [-73.9917900, 40.7264100],
... [-73.9917900, 40.7321400],
... [-73.9829300, 40.7321400],
... [-73.9829300, 40.7264100]
... ]}
> db.open.street.map.find(
... {"loc" : {"$geoIntersects" : {"$geometry" : eastVillage}}})
This would find all point-, line-, and polygon-containing documents that had a point
in the East Village.
You can use "$within" to query for things that are completely contained in an area, for
instance: “What restaurants are in the East Village?”
> db.open.street.map.find({"loc" : {"$within" : {"$geometry" : eastVillage}}})
Unlike our first query, this would not return things that merely pass through the East
Village (such as streets) or partially overlap it (such as a polygon describing Manhattan).
Finally, you can query for nearby locations with "$near":
> db.open.street.map.find({"loc" : {"$near" : {"$geometry" : eastVillage}}})
Note that $near is the only geospatial operator that implies a sort: results from "$near"
are always returned in distance from closest to farthest.
One interesting thing about geospatial queries is that you do not need a geospatial index
to use "$geoIntersects" or "$within" ("$near" requires an index). However, having
an index on your geo field will speed up queries significantly, so it’s usually
recommended.
Compound Geospatial Indexes
As with other types of indexes, you can combine geospatial indexes with other fields to
optimize more complex queries. A possible query mentioned above was: “What res‐
taurants are in the East Village?” Using only a geospatial index, we could narrow the
field to everything in the East Village, but narrowing it down to only “restaurants” or
“pizza” would require another field in the index:
> db.open.street.map.ensureIndex({"tags" : 1, "location" : "2dsphere"})
Then we can quickly find a pizza place in the East Village:
> db.open.street.map.find({"loc" : {"$within" : {"$geometry" : eastVillage}},
... "tags" : "pizza"})
Geospatial Indexing | 121
We can have the “vanilla” index field either before or after the "2dsphere" field, de‐
pending on whether we’d like to filter by the vanilla field or the location first. Choose
whichever will filter out more results as the first index term.
2D Indexes
For non-spherical maps (video game maps, time series data, etc.) you can use a "2d"
index, instead of "2dsphere":
> db.hyrule.ensureIndex({"tile" : "2d"})
"2d" indexes assume a perfectly flat surface, instead of a sphere. Thus, "2d" indexes
should not be used with spheres unless you don’t mind massive distortion around the
poles.
Documents should use a two-element array for their 2d indexed field (which is not a
GeoJSON document, as of this writing). A sample document might look like this:
{
"name" : "Water Temple",
"tile" : [ 32, 22 ]
}
"2d" indexes can only index points. You can store an array of points, but it will be stored
as exactly that: an array of points, not a line. This is an important distinction for "$with
in" queries, in particular. If you store a street as an array of points, the document will
match $within if one of those points is within the given shape. However, the line created
by those points might not be wholly contained in the shape.
By default, geospatial indexing assumes that your values are going to range from -180
to 180. If you are expecting larger or smaller bounds, you can specify what the minimum
and maximum values will be as options to ensureIndex:
> db.star.trek.ensureIndex({"light-years" : "2d"}, {"min" : -1000, "max" : 1000})
This will create a spatial index calibrated for a 2,000 × 2,000 square.
"2d" predates "2dsphere", so querying is a bit simpler. You can only use "$near" or
"$within", and neither have a "$geometry" subobject. You just specify the coordinates:
> db.hyrule.find({"tile" : {"$near" : [20, 21]}})
This finds all of the documents in the hyrule collection, in order by distance from the
point (20, 21). A default limit of 100 documents is applied if no limit is specified. If you
don’t need that many results, you should set a limit to conserve server resources. For
example, the following code returns the 10 documents nearest to (20, 21):
> db.hyrule.find({"tile" : {"$near" : [20, 21]}}).limit(10)
"$within" can query for all points within a rectangle, circle, or polygon. To use a rec‐
tangle, use the "$box" option:
122 | Chapter 6: Special Index and Collection Types
> db.hyrule.find({"tile" : {"$within" : {"$box" : [[10, 20], [15, 30]]}}})
"$box" takes a two-element array: the first element specifies the coordinates of the
lower-left corner; the second element the upper right.
Similarly, you can find all points within a circle with "$center", which takes an array
with the center point and then a radius:
> db.hyrule.find({"tile" : {"$within" : {"$center" : [[12, 25], 5]}}})
Finally, you can specify a polygon as an array of points.
> db.hyrule.find(
... {"tile" : {"$within" : {"$polygon" : [[0, 20], [10, 0], [-10, 0]]}}})
This example would locate all documents containing points within the given triangle.
The final point in the list will be “connected to” the first point to form the polygon.
Storing Files with GridFS
GridFS is a mechanism for storing large binary files in MongoDB. There are several
reasons why you might consider using GridFS for file storage:
•Using GridFS can simplify your stack. If you’re already using MongoDB, you might
be able to use GridFS instead of a separate tool for file storage.
•GridFS will leverage any existing replication or autosharding that you’ve set up for
MongoDB, so getting failover and scale-out for file storage is easier.
•GridFS can alleviate some of the issues that certain filesystems can exhibit when
being used to store user uploads. For example, GridFS does not have issues with
storing large numbers of files in the same directory.
•You can get great disk locality with GridFS, because MongoDB allocates data files
in 2 GB chunks.
There are some downsides, too:
•Slower performance: accessing files from MongoDB will not be as fast as going
directly through the filesystem.
•You can only modify documents by deleting them and resaving the whole thing.
MongoDB stores files as multiple documents so it cannot lock all of the chunks in
a file at the same time.
GridFS is generally best when you have large files you’ll be accessing in a sequential
fashion that won’t be changing much.
Storing Files with GridFS | 123
Getting Started with GridFS: mongofiles
The easiest way to try out GridFS is by using the mongofiles utility. mongofiles is
included with all MongoDB distributions and can be used to upload, download, list,
search for, or delete files in GridFS.
As with any of the other command-line tools, run mongofiles --help to see the options
available for mongofiles.
The following session shows how to use mongofiles to upload a file from the filesystem
to GridFS, list all of the files in GridFS, and download a file that we’ve previously
uploaded:
$ echo "Hello, world" > foo.txt
$ ./mongofiles put foo.txt
connected to: 127.0.0.1
added file: { _id: ObjectId('4c0d2a6c3052c25545139b88'),
filename: "foo.txt", length: 13, chunkSize: 262144,
uploadDate: new Date(1275931244818),
md5: "a7966bf58e23583c9a5a4059383ff850" }
done!
$ ./mongofiles list
connected to: 127.0.0.1
foo.txt 13
$ rm foo.txt
$ ./mongofiles get foo.txt
connected to: 127.0.0.1
done write to: foo.txt
$ cat foo.txt
Hello, world
In the previous example, we perform three basic operations using mongofiles: put,
list, and get. The put operation takes a file in the filesystem and adds it to GridFS;
list will list any files that have been added to GridFS; and get does the inverse of put:
it takes a file from GridFS and writes it to the filesystem. mongofiles also supports two
other operations: search for finding files in GridFS by filename and delete for remov‐
ing a file from GridFS.
Working with GridFS from the MongoDB Drivers
All the client libraries have GridFS APIs. For example, with PyMongo (the Python driver
for MongoDB) you can perform the same series of operations as we did with mongo
files:
>>> from pymongo import Connection
>>> import gridfs
>>> db = Connection().test
>>> fs = gridfs.GridFS(db)
>>> file_id = fs.put("Hello, world", filename="foo.txt")
>>> fs.list()
124 | Chapter 6: Special Index and Collection Types
[u'foo.txt']
>>> fs.get(file_id).read()
'Hello, world'
The API for working with GridFS from PyMongo is very similar to that of mongo
files: we can easily perform the basic put, get, and list operations. Almost all the
MongoDB drivers follow this basic pattern for working with GridFS, while often ex‐
posing more advanced functionality as well. For driver-specific information on GridFS,
please check out the documentation for the specific driver you’re using.
Under the Hood
GridFS is a lightweight specification for storing files that is built on top of normal Mon‐
goDB documents. The MongoDB server actually does almost nothing to “special-case”
the handling of GridFS requests; all the work is handled by the client-side drivers and
tools.
The basic idea behind GridFS is that we can store large files by splitting them up into
chunks and storing each chunk as a separate document. Because MongoDB supports
storing binary data in documents, we can keep storage overhead for chunks to a mini‐
mum. In addition to storing each chunk of a file, we store a single document that groups
the chunks together and contains metadata about the file.
The chunks for GridFS are stored in their own collection. By default chunks will use the
collection fs.chunks, but this can be overridden. Within the chunks collection the struc‐
ture of the individual documents is pretty simple:
{
"_id" : ObjectId("..."),
"n" : 0,
"data" : BinData("..."),
"files_id" : ObjectId("...")
}
Like any other MongoDB document, the chunk has its own unique "_id". In addition,
it has a couple of other keys:
"files_id"
The "_id" of the file document that contains the metadata for this chunk.
"n"
The chunk’s position in the file, relative to the other chunks.
"data"
The bytes in this chunk of the file.
The metadata for each file is stored in a separate collection, which defaults to fs.files.
Each document in the files collection represents a single file in GridFS and can contain
Storing Files with GridFS | 125
any custom metadata that should be associated with that file. In addition to any user-
defined keys, there are a couple of keys that are mandated by the GridFS specification:
"_id"
A unique id for the file—this is what will be stored in each chunk as the value for
the "files_id" key.
"length"
The total number of bytes making up the content of the file.
"chunkSize"
The size of each chunk comprising the file, in bytes. The default is 256K, but this
can be adjusted if needed.
"uploadDate"
A timestamp representing when this file was stored in GridFS.
"md5"
An md5 checksum of this file’s contents, generated on the server side.
Of all of the required keys, perhaps the most interesting (or least self-explanatory) is
"md5". The value for "md5" is generated by the MongoDB server using the filemd5
command, which computes the md5 checksum of the uploaded chunks. This means
that users can check the value of the "md5" key to ensure that a file was uploaded
correctly.
As mentioned above, you are not limited to the required fields in fs.files: feel free to keep
any other file metadata in this collection as well. You might want to keep information
such as download count, mimetype, or user rating with a file’s metadata.
Once you understand the underlying GridFS specification, it becomes trivial to imple‐
ment features that the driver you’re using might not provide helpers for. For example,
you can use the distinct command to get a list of unique filenames stored in GridFS:
> db.fs.files.distinct("filename")
[ "foo.txt" , "bar.txt" , "baz.txt" ]
This allows your application a great deal of flexibility in loading and collecting infor‐
mation about files.
126 | Chapter 6: Special Index and Collection Types
CHAPTER 7
Aggregation
Once you have data stored in MongoDB, you may want to do more than just retrieve
it; you may want to analyze and crunch it in interesting ways. This chapter introduces
the aggregation tools MongoDB provides:
• The aggregation framework
• MapReduce support
•Several simple aggregation commands: count, distinct, and group
The Aggregation Framework
The aggregation framework lets you transform and combine documents in a collection.
Basically, you build a pipeline that processes a stream of documents through several
building blocks: filtering, projecting, grouping, sorting, limiting, and skipping.
For example, if you had a collection of magazine articles, you might want find out who
your most prolific authors were. Assuming that each article is stored as a document in
MongoDB, you could create a pipeline with several steps:
1. Project the authors out of each article document.
2. Group the authors by name, counting the number of occurrences.
3. Sort the authors by the occurrence count, descending.
4. Limit results to the first five.
Each of these steps maps to a aggregation framework operator:
1. {"$project" : {"author" : 1}}
This projects the author field in each document.
127
The syntax is similar to the field selector used in querying: you can select fields to
project by specifying "fieldname" : 1 or exclude fields with "fieldname" : 0.
After this operation, each document in the results looks like: {"_id" : id, "au
thor" : "authorName"}. These resulting documents only exists in memory and
are not written to disk anywhere.
2. {"$group" : {"_id" : "$author", "count" : {"$sum" : 1}}}
This groups the authors by name and increments "count" for each document an
author appears in.
First, we specify the field we want to group by, which is "author". This is indicated
by the "_id" : "$author" field. You can picture this as: after the group there will
be one result document per author, so "author" becomes the unique identifier
("_id").
The second field means to add 1 to a "count" field for each document in the group.
Note that the incoming documents do not have a "count" field; this is a new field
created by the "$group".
At the end of this step, each document in the results looks like: {"_id" : "author
Name", "count" : articleCount}.
3. {"$sort" : {"count" : -1}}
This reorders the result documents by the "count" field from greatest to least.
4. {"$limit" : 5}
This limits the result set to the first five result documents.
To actually run this in MongoDB, pass each operation to the aggregate() function:
> db.articles.aggregate({"$project" : {"author" : 1}},
... {"$group" : {"_id" : "$author", "count" : {"$sum" : 1}}},
... {"$sort" : {"count" : -1}},
... {"$limit" : 5})
{
"result" : [
{
"_id" : "R. L. Stine",
"count" : 430
},
{
"_id" : "Edgar Wallace",
"count" : 175
},
{
"_id" : "Nora Roberts",
"count" : 145
},
{
"_id" : "Erle Stanley Gardner",
128 | Chapter 7: Aggregation
"count" : 140
},
{
"_id" : "Agatha Christie",
"count" : 85
}
],
"ok" : 1
}
aggregate() returns an array of result documents, showing the five most prolific
authors.
To debug a pipeline that’s giving unexpected results, run the aggregation
with just the first pipeline operator. If that gives the expected result, add
the next. In the previous example, you’d first try aggregating with just
the "$project"; if that works, add the "$group"; if that works, add the
"$sort"; and finally add the "$limit". This can help you narrow down
which operator is causing issues.
As of this writing, the aggregation framework cannot write to collections, so all results
must be returned to the client. Thus, aggregation results are limited to 16 MB of data
(the maximum response size).
Pipeline Operations
Each operator receives a stream of documents, does some type of transformation on
these documents, and then passes on the results of the transformation. If it is the last
pipeline operator, these results are returned to the client. Otherwise, the results are
streamed to the next operator as input.
Operators can be combined in any order and repeated as many times as necessary. For
example, you could "$match", "$group", and then "$match" again with different
criteria.
$match
$match filters documents so that you can run an aggregation on a subset of documents.
For example, if you only want to find out stats about users in Oregon, you might add a
"$match" expression such as {$match : {"state" : "OR"}}. "$match" can use all of
the usual query operators ("$gt", "$lt", "$in", etc.). One notable exception is that you
cannot use geospatial operators in a "$match".
Generally, good practice is to put "$match" expressions as early as possible in the pipe‐
line. This has two benefits: it allows you to filter out unneeded documents quickly
Pipeline Operations | 129
(lightening the work the pipeline has to perform) and the query can use indexes if it is
run before any projections or groupings.
$project
Projection is much more powerful in the pipeline than it is in the “normal” query lan‐
guage. "$project" allows you to extract fields from subdocuments, rename fields, and
perform interesting operations on them.
The simplest operation "$project" can perform is simply selecting fields from your
incoming documents. To include or exclude a field, use the same syntax you would in
the second argument of a query. The following would return a result document con‐
taining one field, "author", for each document in the original collection:
> db.articles.aggregate({"$project" : {"author" : 1, "_id" : 0}})
By default, "_id" is always returned if it exists in the incoming document (some pipeline
operators remove the "_id" or it can be removed in a former projection). You can
exclude it as above. Inclusion and exclusion rules in general work the same way that
they do for “normal” queries.
You can also rename the projected field. For example, if you wanted to return the "_id"
of each user as "userId", you could do:
> db.users.aggregate({"$project" : {"userId" : "$_id", "_id" : 0}})
{
"result" : [
{
"userId" : ObjectId("50e4b32427b160e099ddbee7")
},
{
"userId" : ObjectId("50e4b32527b160e099ddbee8")
}
...
],
"ok" : 1
}
The "$fieldname" syntax is used to refer to fieldname’s value in the aggregation
framework. For example, "$age" would be replaced with the contents of the age field
(and probably be a number, not a string) and "$tags.3" would be replaced with the
fourth element of the tags array. Thus, "$_id" is replaced by the "_id" field of each
document coming through the pipeline.
Note that you must specifically exclude "_id" to prevent it from returning the field
twice, once labeled "userId" and once labelled "_id". You can use this technique to
make multiple copies of a field for later use in a "$group", say.
130 | Chapter 7: Aggregation
MongoDB does not track field name history when fields are renamed. Thus, if you had
an index on "originalFieldname", aggregation would be unable to use the index for
the sort below, even though to a human eye it is obvious that "newFieldname" is the
same field as "originalFieldname":
> db.articles.aggregate({"$project" : {"newFieldname" : "$originalFieldname"}},
... {"$sort" : {"newFieldname" : 1}})
Thus, try to utilize indexes before changing the names of fields.
Pipeline expressions
The simplest "$project" expressions are inclusion, exclusion, and field names
("$fieldname"). However, there are several other, more powerful options. You can also
use expressions, which allow you to combine multiple literals and variables into a single
value.
There are several expressions available with aggregation which you can combine and
nest to any depth to create more complex expressions.
Mathematical expressions. Arithmetic expressions let you manipulate numeric values.
You generally use these expressions by specifying an array of numbers to operate on.
For example, the following expression would sum the "salary" and "bonus" fields:
> db.employees.aggregate(
... {
... "$project" : {
... "totalPay" : {
... "$add" : ["$salary", "$bonus"]
... }
... }
... })
If we have a more complex expression, it can be nested. Suppose that we want to subtract
401k contributions from this total. We can add a "$subtract" expression:
> db.employees.aggregate(
... {
... "$project" : {
... "totalPay" : {
... "$subtract" : [{"$add" : ["$salary", "$bonus"]}, "$401k"]
... }
... }
... })
Expressions may be nested arbitrarily deep.
Here’s the syntax for each operator:
"$add" : [expr1[, expr2, ..., exprN]]
Takes one or more expressions and adds them together.
Pipeline Operations | 131
"$subtract" : [expr1, expr2]
Takes two expressions and subtracts the second from the first.
"$multiply" : [expr1[, expr2, ..., exprN]]
Takes one or more expressions and multiplies them together.
"$divide" : [expr1, expr2]
Takes two expressions and divides the first by the second.
"$mod" : [expr1, expr2]
Takes two expressions and returns the remainder of dividing the first by the second.
Date expressions. Many aggregations are time-based: What was happening last week?
Last month? Over the last year? Thus, aggregation has a set of expressions that can be
used to extract date information in more useful ways: "$year", "$month", "$week",
"$dayOfMonth", "$dayOfWeek", "$dayOfYear", "$hour", "$minute", and "$second".
You can only use date operations on fields stored with the date type, not numeric types.
Each of these date types is basically the same: it takes a date expression and returns a
number. This would return the month that each employee was hired in:
> db.employees.aggregate(
... {
... "$project" : {
... "hiredIn" : {"$month" : "$hireDate"}
... }
... })
You can also use literal dates. This would calculate many years each employee had
worked at the company:
> db.employees.aggregate(
... {
... "$project" : {
... "tenure" : {
... "$subtract" : [{"$year" : new Date()}, {"$year" : "$hireDate"}]
... }
... }
... })
String expressions. There are a few basic string operations you can do as well. Their
signatures are:
"$substr" : [expr, startOffset, numToReturn]
This returns a substring of the first argument, starting at the startOffset-th byte
and including the next numToReturn bytes (note that this is measured in bytes, not
characters, so multibytes encodings will have to be careful of this). expr must eval‐
uate to a string.
132 | Chapter 7: Aggregation
"$concat" : [expr1[, expr2, ..., exprN]]
Concatenates each string expression (or string) given.
"$toLower" : expr
Returns the string in lower case. expr must evaluate to a string.
"$toUpper" : expr
Returns the string in upper case. expr must evaluate to a string.
Case-affecting operations are only guaranteed to work on characters from the Roman
alphabet.
Here is an example that generates email addresses of the format j.doe@example.com. It
extracts the first initial and concatenates it with several constant strings and the "last
Name" field:
> db.employees.aggregate(
... {
... "$project" : {
... "email" : {
... "$concat" : [
... {"$substr" : ["$firstName", 0, 1]},
... ".",
... "$lastName",
... "@example.com"
... ]
... }
... }
... })
Logical expressions. There are several operators that you can use for control statements.
There are several comparison expressions:
"$cmp" : [expr1, expr2]
Compare expr1 and expr2. Return 0 if the two expressions are equal, a negative
number if expr1 is less than expr2, and a positive number if expr2 is less than expr1.
"$strcasecmp" : [string1, string2]
Case insensitive comparison between string1 and string2. Only works for Roman
characters.
"$eq"/"$ne"/"$gt"/"$gte"/"$lt"/"$lte" : [expr1, expr2]
Perform the comparison on expr1 and expr2, returning whether it evaluates to true
or false.
There are a few boolean expressions:
"$and" : [expr1[, expr2, ..., exprN]]
Returns true if all expressions are true.
Pipeline Operations | 133
"$or" : [expr1[, expr2, ..., exprN]]
Returns true if at least one expression is true.
"$not" : expr
Returns the boolean opposite of expr.
Finally, there are two control statements:
"$cond" : [booleanExpr, trueExpr, falseExpr]
If booleanExpr evaluates to true, trueExpr is returned; otherwise falseExpr is
returned.
"$ifNull" : [expr, replacementExpr]
If expr is null this returns replacementExpr; otherwise it returns expr.
These operators let you include more complex logic in your aggregations by following
different “code paths” depending on the shape of your data.
Pipelines are particular about getting properly formed input, so these operators can be
invaluable in filling in default values. The arithmetic operators will complain about non-
numeric values, date operators will complain about non-dates, string operators will
complain about non-strings, and everything will complain about missing fields. If your
data set is inconsistent, you can use this conditionals to detect missing values and pop‐
ulate them.
A projection example
Suppose a professor wanted to generate grades using a somewhat complex calculation:
the students are graded 10% on attendance, 30% on quizzes, and 60% on tests (unless
the student is a teacher’s pet, in which case the grade is set to 100). We could express
these rules as follows:
> db.students.aggregate(
... {
... "$project" : {
... "grade" : {
... "$cond" : [
... "$teachersPet",
... 100, // if
... { // else
... "$add" : [
... {"$multiply" : [.1, "$attendanceAvg"]},
... {"$multiply" : [.3, "$quizzAvg"]},
... {"$multiply" : [.6, "$testAvg"]}
... ]
... }
... ]
... }
... }
... })
134 | Chapter 7: Aggregation
$group
Grouping allows you to group documents based on certain fields and combine their
values. Some examples of groupings:
•If we had per-minute measurements and we wanted to find the average humidity
per day, we would group by the "day" field.
•If we had a collection of students and we wanted to organize student into groups
based on grade, we could group by their "grade" field.
•If we had a collection of users and we wanted to see how many users we had by city,
we could group by both the "state" and "city" fields, creating one group per city/
state pair. We wouldn’t want to just group by city, as there are many cities with
identical names in different states.
When you choose a field or fields to group by, you pass it to the "$group" function as
the group’s "_id" field. Thus, for the examples above, you’d have:
•{"$group" : {"_id" : "$day"}}
•{"$group" : {"_id" : "$grade"}}
•{"$group" : {"_id" : {"state" : "$state", "city" : "$city"}}}
As these stand now, the result would be a document for each group with a single field:
the grouping key. For example, the grading result might look something like this: {"re
sult" : [{"_id" : "A+"}, {"_id" : "A"}, {"_id" : "A-"}, ..., {"_id" :
"F"}], "ok" : 1}. This does a nice job of getting all of the distinct values for a field,
but all of the examples involve using these groups to calculate things. Thus, you can add
fields that use grouping operators to make computations based on the documents in
each group.
Grouping operators
These grouping operators allow you to compute results for each group. In the example
in “The Aggregation Framework” on page 127, we saw "$sum" used as a grouping op‐
erator: it added 1 to a total for each document in the group.
Arithmetic operators. There are two operators that can be used to compute numeric val‐
ues from fields: "$sum" and "$average".
"$sum" : value
This adds value for each document and returns the result. Note that, although the
example above used a literal (1), this can also take more complex values. For ex‐
ample, if we had a collection of sales made in a variety of countries, this would find
the total revenue by country:
Pipeline Operations | 135
> db.sales.aggregate(
... {
... "$group" : {
... "_id" : "$country",
... "totalRevenue" : {"$sum" : "$revenue"}
... }
... })
"$avg" : value
Returns an average of all input values seen during the group.
For example, this would return the average revenue per country, plus the number
of sales made:
> db.sales.aggregate(
... {
... "$group" : {
... "_id" : "$country",
... "totalRevenue" : {"$average" : "$revenue"},
... "numSales" : {"$sum" : 1}
... }
... })
Extreme operators. There are four operators to get the “edges” of a data set:
"$max" : expr
Returns the greatest value of any of the inputs.
"$min" : expr
Returns the smallest value of any of the inputs.
"$first" : expr
This returns the first value seen by group, ignoring subsequent values. This is only
sensible to use when you know the order that the data is being processed in: that is,
after a sort.
"$last" : expr
This is the opposite of the previous; it returns the last value seen by the group.
"$max" and "$min" look through each document and find the extreme values. Thus,
these operators work well when you do not have sorted data and are a bit wasteful when
data is sorted. For example, suppose that we had a set of student scores on a test. We
could find the outliers in each grade as follows:
> db.scores.aggregate(
... {
... "$group" : {
... "_id" : "$grade",
... "lowestScore" : {"$min" : "$score"},
... "highestScore" : {"$max" : "$score"}
136 | Chapter 7: Aggregation
... }
... })
Alternatively, "$first" and "$last" return useful results when your data is sorted by
the fields you are looking for. For example, to get the same result as before we could do:
> db.scores.aggregate(
... {
... "$sort" : {"score" : 1}
... },
... {
... "$group" : {
... "_id" : "$grade",
... "lowestScore" : {"$first" : "$score"},
... "highestScore" : {"$last" : "$score"}
... }
... })
If your data is already sorted, $first and $last will be more efficient than $min and
$max. If your data is not going to be sorted, $min and $max are more efficient than sorting
your data and then running $first and $last.
Array operators. There are two operators available for array manipulation:
"$addToSet" : expr
Keeps an array of values seen so far and, if expr is not present in the array, adds it.
Each value appears at most once in the resulting array and ordering is not
guaranteed.
"$push" : expr
Indiscriminately adds each value seen to the array. Returns an array of all values.
Grouping behavior
"$group" is one of the two “roadblock” operators that cannot be handled in the stream‐
ing fashion described earlier. While most operators can continuously process docu‐
ments as they arrive, "$group" must collect all documents, split them into groups, then
send them to the next operator in the pipeline. This means that, with sharding, "$group"
will first be run on each shard and then the individual shards’ groups will be sent to the
mongos to do the final grouping and the remainder of the pipeline will be run on the
mongos (not the shards).
$unwind
Unwinding turns each field of an array into a separate document. For example, if we
had a blog with comments, we could use unwind to turn each comment into its own
“document”:
Pipeline Operations | 137
> db.blog.findOne()
{
"_id" : ObjectId("50eeffc4c82a5271290530be"),
"author" : "k",
"post" : "Hello, world!",
"comments" : [
{
"author" : "mark",
"date" : ISODate("2013-01-10T17:52:04.148Z"),
"text" : "Nice post"
},
{
"author" : "bill",
"date" : ISODate("2013-01-10T17:52:04.148Z"),
"text" : "I agree"
}
]
}
> db.blog.aggregate({"$unwind" : "$comments"})
{
"results" :
{
"_id" : ObjectId("50eeffc4c82a5271290530be"),
"author" : "k",
"post" : "Hello, world!",
"comments" : {
"author" : "mark",
"date" : ISODate("2013-01-10T17:52:04.148Z"),
"text" : "Nice post"
}
},
{
"_id" : ObjectId("50eeffc4c82a5271290530be"),
"author" : "k",
"post" : "Hello, world!",
"comments" : {
"author" : "bill",
"date" : ISODate("2013-01-10T17:52:04.148Z"),
"text" : "I agree"
}
}
],
"ok" : 1
}
This is particularly useful if you want to return certain subdocuments from a query:
"$unwind" the subdocuments and then "$match" the ones you want. For example, it is
impossible in the normal query language to return all comments by a certain user and
only those comments, not the posts they commented on. However, by projecting, un‐
winding, and matching, it becomes trivial:
138 | Chapter 7: Aggregation
> db.blog.aggregate({"$project" : {"comments" : "$comments"}},
... {"$unwind" : "$comments"},
... {"$match" : {"comments.author" : "Mark"}})
You might want to do a final projection to format the output more nicely, as all of the
comments will still be in a "comments" subdocument.
$sort
You can sort by any field or fields, using the same syntax you would with the “normal”
query language. If you are sorting a non-trivial number of documents, it is highly rec‐
ommended that you do the sort at the beginning of the pipeline and have an index it
can use. Otherwise, the sort may be slow and take a lot of memory.
You can use both existing fields and projected fields in a sort:
> db.employees.aggregate(
... {
... "$project" : {
... "compensation" : {
... "$add" : ["$salary", "$bonus"]
... },
... "name" : 1
... }
... },
... {
... "$sort" : {"compensation" : -1, "name" : 1}
... })
This example would sort employees by compensation, from highest to lowest, and then
name from A-Z.
Possible sorts are 1 (for ascending) and −1 (for descending).
"$sort" is another roadblock operation, like "$group" earlier. "$sort" must collect all
documents to properly sort them and, with sharding, sends the individual shards’ sorted
results to the mongos for further processing.
$limit
$limit takes a number, n, and returns the first n resulting documents.
$skip
$skip takes a number, n, and discards the first n documents from the result set. As with
“normal” querying, it isn’t efficient for large skips, as it must find all of the matches that
must be skipped and then discard them.
Pipeline Operations | 139
Using Pipelines
Attempt to filter out as many documents (and as many fields from the documents) as
possible at the beginning of your pipeline before hitting any "$project", "$group", or
"$unwind" operations. Once the pipeline isn’t using the data directly from the collection,
indexes can no longer be used to help filter and sort. The aggregation pipeline will
attempt to reorder operations for you, if possible, to be able to use indexes.
MongoDB won’t allow a single aggregation to use more than a fraction of the system’s
memory: if it calculates that an aggregation has used more than 20% of the memory,
the aggregation will simply error out. Allowing output to be piped to a collection (which
would minimize the amount of memory required) is planned for the future.
If you can quickly whittle down the result set size with a selective "$match", you can
use the pipeline for real-time aggregations. As pipelines need to include more docu‐
ments and become more complex, it is less likely that you’ll be able to get realtime results
from them.
MapReduce
MapReduce is a powerful and flexible tool for aggregating data. It can solve some prob‐
lems that are too complex to express using the aggregation framework’s query language.
MapReduce uses JavaScript as its “query language” so it can express arbitrarily complex
logic. However, this power comes at a price: MapReduce tends to be fairly slow and
should not be used for real-time data analysis.
MapReduce can be easily parallelized across multiple servers. It splits up a problem,
sends chunks of it to different machines, and lets each machine solve its part of the
problem. When all the machines are finished, they merge all the pieces of the solution
back into a full solution.
MapReduce has a couple of steps. It starts with the map step, which maps an operation
onto every document in a collection. That operation could be either “do nothing” or
“emit these keys with X values.” There is then an intermediary stage called the shuffle
step: keys are grouped and lists of emitted values are created for each key. The reduce
takes this list of values and reduces it to a single element. This element is returned to the
shuffle step until each key has a list containing a single value: the result.
We’ll go through a couple examples because MapReduce is an incredibly useful and
powerful, but also somewhat complex, tool.
Example 1: Finding All Keys in a Collection
Using MapReduce for this problem might be overkill, but it is a good way to get familiar
with how MapReduce works. If you already understand MapReduce, feel free to skip
140 | Chapter 7: Aggregation
ahead to the last part of this section, where we cover MongoDB-specific MapReduce
considerations.
MongoDB assumes that your schema is dynamic, so it does not keep track of the keys
in each document. The best way, in general, to find all the keys across all the documents
in a collection is to use MapReduce. In this example, we’ll also get a count of how many
times each key appears in the collection. This example doesn’t include keys for embed‐
ded documents, but it would be a simple addition to the map function to do so.
For the mapping step, we want to get every key of every document in the collection. The
map function uses a special function to “return” values that we want to process later:
emit. emit gives MapReduce a key (like the one used by group earlier) and a value. In
this case, we emit a count of how many times a given key appeared in the document
(once: {count : 1}). We want a separate count for each key, so we’ll call emit for every
key in the document. this is a reference to the current document we are mapping:
> map = function() {
... for (var key in this) {
... emit(key, {count : 1});
... }};
Now we have a ton of little {count : 1} documents floating around, each associated
with a key from the collection. An array of one or more of these {count : 1} documents
will be passed to the reduce function. The reduce function is passed two arguments:
key, which is the first argument from emit, and an array of one or more {count : 1}
documents that were emitted for that key:
> reduce = function(key, emits) {
... total = 0;
... for (var i in emits) {
... total += emits[i].count;
... }
... return {"count" : total};
... }
reduce must be able to be called repeatedly on results from either the map phase or
previous reduce phases. Therefore, reduce must return a document that can be re-sent
to reduce as an element of its second argument. For example, say we have the key x
mapped to three documents: {count : 1, id : 1}, {count : 1, id : 2}, and
{count : 1, id : 3}. (The ID keys are just for identification purposes.) MongoDB
might call reduce in the following pattern:
> r1 = reduce("x", [{count : 1, id : 1}, {count : 1, id : 2}])
{count : 2}
> r2 = reduce("x", [{count : 1, id : 3}])
{count : 1}
> reduce("x", [r1, r2])
{count : 3}
MapReduce | 141
You cannot depend on the second argument always holding one of the initial documents
({count : 1} in this case) or being a certain length. reduce should be able to be run
on any combination of emit documents and reduce return values.
Altogether, this MapReduce function would look like this:
> mr = db.runCommand({"mapreduce" : "foo", "map" : map, "reduce" : reduce})
{
"result" : "tmp.mr.mapreduce_1266787811_1",
"timeMillis" : 12,
"counts" : {
"input" : 6
"emit" : 14
"output" : 5
},
"ok" : true
}
The document MapReduce returns gives you a bunch of metainformation about the
operation:
"result" : "tmp.mr.mapreduce_1266787811_1"
This is the name of the collection the MapReduce results were stored in. This is a
temporary collection that will be deleted when the connection that did the Map‐
Reduce is closed. We will go over how to specify a nicer name and make the col‐
lection permanent in a later part of this chapter.
"timeMillis" : 12
How long the operation took, in milliseconds.
"counts" : { ... }
This embedded document is mostly used for debugging and contains three keys:
"input" : 6
The number of documents sent to the map function.
"emit" : 14
The number of times emit was called in the map function.
"output" : 5
The number of documents created in the result collection.
If we do a find on the resulting collection, we can see all the keys and their counts from
our original collection:
> db[mr.result].find()
{ "_id" : "_id", "value" : { "count" : 6 } }
{ "_id" : "a", "value" : { "count" : 4 } }
{ "_id" : "b", "value" : { "count" : 2 } }
{ "_id" : "x", "value" : { "count" : 1 } }
{ "_id" : "y", "value" : { "count" : 1 } }
142 | Chapter 7: Aggregation
Each of the key values becomes an "_id", and the final result of the reduce step(s)
becomes the "value".
Example 2: Categorizing Web Pages
Suppose we have a site where people can submit links to other pages, such as reddit.
Submitters can tag a link as related to certain popular topics, e.g., “politics,” “geek,” or
“icanhascheezburger.” We can use MapReduce to figure out which topics are the most
popular, as a combination of recent and most-voted-for.
First, we need a map function that emits tags with a value based on the popularity and
recency of a document:
map = function() {
for (var i in this.tags) {
var recency = 1/(new Date() - this.date);
var score = recency * this.score;
emit(this.tags[i], {"urls" : [this.url], "score" : score});
}
};
Now we need to reduce all of the emitted values for a tag into a single score for that tag:
reduce = function(key, emits) {
var total = {urls : [], score : 0}
for (var i in emits) {
emits[i].urls.forEach(function(url) {
total.urls.push(url);
}
total.score += emits[i].score;
}
return total;
};
The final collection will end up with a full list of URLs for each tag and a score showing
how popular that particular tag is.
MongoDB and MapReduce
Both of the previous examples used only the mapreduce, map, and reduce keys. These
three keys are required, but there are many optional keys that can be passed to the
MapReduce command:
"finalize" : function
A final step to send reduce’s output to.
"keeptemp" : boolean
If the temporary result collection should be saved when the connection is closed.
MapReduce | 143
"out" : string
Name for the output collection. Setting this option implies keeptemp : true.
"query" : document
Query to filter documents by before sending to the map function.
"sort" : document
Sort to use on documents before sending to the map (useful in conjunction with
the limit option).
"limit" : integer
Maximum number of documents to send to the map function.
"scope" : document
Variables that can be used in any of the JavaScript code.
"verbose" : boolean
Whether or not to use more verbose output in the server logs.
The finalize function
As with the previous group command, MapReduce can be passed a finalize function
that will be run on the last reduce’s output before it is saved to a temporary collection.
Returning large result sets is less critical with MapReduce than group because the whole
result doesn’t have to fit in 4 MB. However, the information will be passed over the wire
eventually, so finalize is a good chance to take averages, chomp arrays, and remove
extra information in general.
Keeping output collections
By default, Mongo creates a temporary collection while it is processing the MapReduce
with a name that you are unlikely to choose for a collection: a dot-separated string
containing mr, the name of the collection you’re MapReducing, a timestamp, and the
job’s ID with the database. It ends up looking something like mr.stuff.18234210220.2.
MongoDB will automatically destroy this collection when the connection that did the
MapReduce is closed. (You can also drop it manually when you’re done with it.) If you
want to persist this collection even after disconnecting, you can specify keeptemp :
true as an option.
If you’ll be using the temporary collection regularly, you may want to give it a better
name. You can specify a more human-readable name with the out option, which takes
a string. If you specify out, you need not specify keeptemp : true, since it is implied.
Even if you specify a “pretty” name for the collection, MongoDB will use the autogen‐
erated collection name for intermediate steps of the MapReduce. When it has finished,
it will automatically and atomically rename the collection from the autogenerated name
144 | Chapter 7: Aggregation
to your chosen name. This means that if you run MapReduce multiple times with the
same target collection, you will never be using an incomplete collection for operations.
The output collection created by MapReduce is a normal collection, which means that
there is no problem with doing a MapReduce on it or a MapReduce on the results from
that MapReduce, ad infinitum!
MapReduce on a subset of documents
Sometimes you need to run MapReduce on only part of a collection. You can add a
query to filter the documents before they are passed to the map function.
Every document passed to the map function needs to be deserialized from BSON into a
JavaScript object, which is a fairly expensive operation. If you know that you will need
to run MapReduce only on a subset of the documents in the collection, adding a filter
can greatly speed up the command. The filter is specified by the "query", "limit", and
"sort" keys.
The "query" key takes a query document as a value. Any documents that would ordi‐
narily be returned by that query will be passed to the map function. For example, if we
have an application tracking analytics and want a summary for the last week, we can
use MapReduce on only the most recent week’s documents with the following
command:
> db.runCommand({"mapreduce" : "analytics", "map" : map, "reduce" : reduce,
"query" : {"date" : {"$gt" : week_ago}}})
The sort option is mostly useful in conjunction with limit. limit can be used on its
own, as well, to simply provide a cutoff on the number of documents sent to the map
function.
If, in the previous example, we wanted an analysis of the last 10,000 page views (instead
of the last week), we could use limit and sort:
> db.runCommand({"mapreduce" : "analytics", "map" : map, "reduce" : reduce,
"limit" : 10000, "sort" : {"date" : -1}})
query, limit, and sort can be used in any combination, but sort isn’t useful if limit
isn’t present.
Using a scope
MapReduce can take a code type for the map, reduce, and finalize functions, and, in
most languages, you can specify a scope to be passed with code. However, MapReduce
ignores this scope. It has its own scope key, "scope", and you must use that if there are
client-side values you want to use in your MapReduce. You can set them using a plain
document of the form variable_name : value, and they will be available in your map,
reduce, and finalize functions. The scope is immutable from within these functions.
MapReduce | 145
For instance, in the example in the previous section, we calculated the recency of a page
using 1/(new Date() - this.date). We could, instead, pass in the current date as part
of the scope with the following code:
> db.runCommand({"mapreduce" : "webpages", "map" : map, "reduce" : reduce,
"scope" : {now : new Date()}})
Then, in the map function, we could say 1/(now - this.date).
Getting more output
There is also a verbose option for debugging. If you would like to see the progress of
your MapReduce as it runs, you can specify "verbose" : true.
You can also use print to see what’s happening in the map, reduce, and finalize
functions. print will print to the server log.
Aggregation Commands
There are several commands that MongoDB provides for basic aggregation tasks over
a collection. These commands were added before the aggregation framework and have
been superceded by it, for the most part. However, complex groups may still require
JavaScript and counts and distincts can be simpler to run as non-framework commands.
count
The simplest aggregation tool is count, which returns the number of documents in the
collection:
> db.foo.count()
0
> db.foo.insert({"x" : 1})
> db.foo.count()
1
Counting the total number of documents in a collection is fast regardless of collection
size.
You can also pass in a query and Mongo will count the number of results for that query:
> db.foo.insert({"x" : 2})
> db.foo.count()
2
> db.foo.count({"x" : 1})
1
This can be useful for getting a total for pagination: “displaying results 0–10 of 439.”
Adding criteria does make the count slower. Counts can use indexes, but indexes do not
contain enough metadata to make counting any more efficient than actually doing a
query for the criteria.
146 | Chapter 7: Aggregation
distinct
The distinct command finds all of the distinct values for a given key. You must specify
a collection and key:
> db.runCommand({"distinct" : "people", "key" : "age"})
For example, suppose we had the following documents in our collection:
{"name" : "Ada", "age" : 20}
{"name" : "Fred", "age" : 35}
{"name" : "Susan", "age" : 60}
{"name" : "Andy", "age" : 35}
If you call distinct on the "age" key, you will get back all of the distinct ages:
> db.runCommand({"distinct" : "people", "key" : "age"})
{"values" : [20, 35, 60], "ok" : 1}
A common question at this point is if there’s a way to get all of the distinct keys in a
collection. There is no built-in way of doing this, although you can write something to
do it yourself using MapReduce (described in “MapReduce” on page 140).
group
group allows you to perform more complex aggregation. You choose a key to group by,
and MongoDB divides the collection into separate groups for each value of the chosen
key. For each group, you can create a result document by aggregating the documents
that are members of that group.
If you are familiar with SQL, group is similar to SQL’s GROUP BY.
Suppose we have a site that keeps track of stock prices. Every few minutes from 10 a.m.
to 4 p.m., it gets the latest price for a stock, which it stores in MongoDB. Now, as part
of a reporting application, we want to find the closing price for the past 30 days. This
can be easily accomplished using group.
The collection of stock prices contains thousands of documents with the following form:
{"day" : "2010/10/03", "time" : "10/3/2010 03:57:01 GMT-400", "price" : 4.23}
{"day" : "2010/10/04", "time" : "10/4/2010 11:28:39 GMT-400", "price" : 4.27}
{"day" : "2010/10/03", "time" : "10/3/2010 05:00:23 GMT-400", "price" : 4.10}
{"day" : "2010/10/06", "time" : "10/6/2010 05:27:58 GMT-400", "price" : 4.30}
{"day" : "2010/10/04", "time" : "10/4/2010 08:34:50 GMT-400", "price" : 4.01}
Aggregation Commands | 147
You should never store money amounts as floating-point numbers be‐
cause of inexactness concerns, but for simplicity we’ll do it in this ex‐
ample.
We want our results to be a list of the latest time and price for each day, something
like this:
[
{"time" : "10/3/2010 05:00:23 GMT-400", "price" : 4.10},
{"time" : "10/4/2010 11:28:39 GMT-400", "price" : 4.27},
{"time" : "10/6/2010 05:27:58 GMT-400", "price" : 4.30}
]
We can accomplish this by splitting the collection into sets of documents grouped by
day then finding the document with the latest timestamp for each day and adding it to
the result set. The whole function might look something like this:
> db.runCommand({"group" : {
... "ns" : "stocks",
... "key" : "day",
... "initial" : {"time" : 0},
... "$reduce" : function(doc, prev) {
... if (doc.time > prev.time) {
... prev.price = doc.price;
... prev.time = doc.time;
... }
... }}})
Let’s break this command down into its component keys:
"ns" : "stocks"
This determines which collection we’ll be running the group on.
"key" : "day"
This specifies the key on which to group the documents in the collection. In this
case, that would be the "day" key. All the documents with a "day" key of a given
value will be grouped together.
"initial" : {"time" : 0}
The first time the reduce function is called for a given group, it will be passed the
initialization document. This same accumulator will be used for each member of a
given group, so any changes made to it can be persisted.
"$reduce" : function(doc, prev) { ... }
This will be called once for each document in the collection. It is passed the current
document and an accumulator document: the result so far for that group. In this
example, we want the reduce function to compare the current document’s time with
the accumulator’s time. If the current document has a later time, we’ll set the ac‐
cumulator’s day and price to be the current document’s values. Remember that there
148 | Chapter 7: Aggregation
is a separate accumulator for each group, so there is no need to worry about different
days using the same accumulator.
In the initial statement of the problem, we said that we wanted only the last 30 days
worth of prices. Our current solution is iterating over the entire collection, however.
This is why you can include a "condition" that documents must satisfy in order to be
processed by the group command at all:
> db.runCommand({"group" : {
... "ns" : "stocks",
... "key" : "day",
... "initial" : {"time" : 0},
... "$reduce" : function(doc, prev) {
... if (doc.time > prev.time) {
... prev.price = doc.price;
... prev.time = doc.time;
... }},
... "condition" : {"day" : {"$gt" : "2010/09/30"}}
... }})
Some documentation refers to a "cond" or "q" key, both of which are
identical to the "condition" key (just less descriptive).
Now the command will return an array of 30 documents, each of which is a group. Each
group has the key on which the group was based (in this case, "day" : string) and the
final value of prev for that group. If some of the documents do not contain the key, these
will be grouped into a single group with a day : null element. You can eliminate this
group by adding "day" : {"$exists" : true} to the "condition". The group com‐
mand also returns the total number of documents used and the number of distinct values
for "key":
> db.runCommand({"group" : {...}})
{
"retval" :
[
{
"day" : "2010/10/04",
"time" : "Mon Oct 04 2010 11:28:39 GMT-0400 (EST)"
"price" : 4.27
},
...
],
"count" : 734,
"keys" : 30,
"ok" : 1
}
Aggregation Commands | 149
We explicitly set the "price" for each group, and the "time" was set by the initializer
and then updated. The "day" is included because the key being grouped by is included
by default in each "retval" embedded document. If you don’t want to return this key,
you can use a finalizer to change the final accumulator document into anything, even a
nondocument (e.g., a number or string).
Using a finalizer
Finalizers can be used to minimize the amount of data that needs to be transferred from
the database to the user, which is important because the group command’s output needs
to fit in a single database response. To demonstrate this, we’ll take the example of a blog
where each post has tags. We want to find the most popular tag for each day. We can
group by day (again) and keep a count for each tag. This might look something like this:
> db.posts.group({
... "key" : {"day" : true},
... "initial" : {"tags" : {}},
... "$reduce" : function(doc, prev) {
... for (i in doc.tags) {
... if (doc.tags[i] in prev.tags) {
... prev.tags[doc.tags[i]]++;
... } else {
... prev.tags[doc.tags[i]] = 1;
... }
... }
... }})
This will return something like this:
[
{"day" : "2010/01/12", "tags" : {"nosql" : 4, "winter" : 10, "sledding" : 2}},
{"day" : "2010/01/13", "tags" : {"soda" : 5, "php" : 2}},
{"day" : "2010/01/14", "tags" : {"python" : 6, "winter" : 4, "nosql": 15}}
]
Then we could find the largest value in the "tags" document on the client side. However,
sending the entire tags document for every day is a lot of extra overhead to send to the
client: an entire set of key/value pairs for each day, when all we want is a single string.
This is why group takes an optional "finalize" key. "finalize" can contain a function
that is run on each group once, right before the result is sent back to the client. We can
use a "finalize" function to trim out all of the cruft from our results:
> db.runCommand({"group" : {
... "ns" : "posts",
... "key" : {"day" : true},
... "initial" : {"tags" : {}},
... "$reduce" : function(doc, prev) {
... for (i in doc.tags) {
... if (doc.tags[i] in prev.tags) {
... prev.tags[doc.tags[i]]++;
150 | Chapter 7: Aggregation
... } else {
... prev.tags[doc.tags[i]] = 1;
... }
... },
... "finalize" : function(prev) {
... var mostPopular = 0;
... for (i in prev.tags) {
... if (prev.tags[i] > mostPopular) {
... prev.tag = i;
... mostPopular = prev.tags[i];
... }
... }
... delete prev.tags
... }}})
Now, we’re only getting the information we want; the server will send back something
like this:
[
{"day" : "2010/01/12", "tag" : "winter"},
{"day" : "2010/01/13", "tag" : "soda"},
{"day" : "2010/01/14", "tag" : "nosql"}
]
finalize can either modify the argument passed in or return a new value.
Using a function as a key
Sometimes you may have more complicated criteria that you want to group by, not just
a single key. Suppose you are using group to count how many blog posts are in each
category. (Each blog post is in a single category.) Post authors were inconsistent, though,
and categorized posts with haphazard capitalization. So, if you group by category name,
you’ll end up with separate groups for “MongoDB” and “mongodb.” To make sure any
variation of capitalization is treated as the same key, you can define a function to de‐
termine documents’ grouping key.
To define a grouping function, you must use a $keyf key (instead of "key"). Using
"$keyf" makes the group command look something like this:
> db.posts.group({"ns" : "posts",
... "$keyf" : function(x) { return x.category.toLowerCase(); },
... "initializer" : ... })
"$keyf" allows you can group by arbitrarily complex criteria.
Aggregation Commands | 151
CHAPTER 8
Application Design
This chapter covers designing applications to work effectively with MongoDB. It
discusses:
• Trade-offs when deciding whether to embed data or to reference it
• Tips for optimizations
•Consistency considerations
•How to migrate schemas
•When MongoDB isn’t a good choice of data store
Normalization versus Denormalization
There are many ways of representing data and one of the most important issues is how
much you should normalize your data. Normalization is dividing up data into multiple
collections with references between collections. Each piece of data lives in one collection
although multiple documents may reference it. Thus, to change the data, only one
document must be updated. However, MongoDB has no joining facilities, so gathering
documents from multiple collections will require multiple queries.
Denormalization is the opposite of normalization: embedding all of the data in a single
document. Instead of documents containing references to one definitive copy of the
data, many documents may have copies of the data. This means that multiple documents
need to be updated if the information changes but that all related data can be fetched
with a single query.
Deciding when to normalize and when to denormalize can be difficult: typically, nor‐
malizing makes writes faster and denormalizing makes reads faster. Thus, you need to
find what trade-offs make sense for your application.
153
Examples of Data Representations
Suppose we are storing information about students and the classes that they are taking.
One way to represent this would be to have a students collection (each student is one
document) and a classes collection (each class is one document). Then we could have a
third collection (studentClasses) that contains references to the student and classes he
is taking:
> db.studentClasses.findOne({"studentId" : id})
{
"_id" : ObjectId("512512c1d86041c7dca81915"),
"studentId" : ObjectId("512512a5d86041c7dca81914"),
"classes" : [
ObjectId("512512ced86041c7dca81916"),
ObjectId("512512dcd86041c7dca81917"),
ObjectId("512512e6d86041c7dca81918"),
ObjectId("512512f0d86041c7dca81919")
]
}
If you are familiar with relational databases, you may have seen this type of join table
before, although typically you’d have one student and one class per document (instead
of a list of class "_id"s). It’s a bit more MongoDB-ish to put the classes in an array, but
you usually wouldn’t want to store the data this way because it requires a lot of querying
to get to the actual information.
Suppose we wanted to find the classes a student was taking. We’d query for the student
in the students collection, query studentClasses for the course "_id"s, and then query
the classes collection for the class information. Thus, finding this information would
take three trips to the server. This is generally not the way you want to structure data in
MongoDB, unless the classes and students are changing constantly and reading the data
does not need to be done quickly.
We can remove one of the dereferencing queries by embedding class references in the
student’s document:
{
"_id" : ObjectId("512512a5d86041c7dca81914"),
"name" : "John Doe",
"classes" : [
ObjectId("512512ced86041c7dca81916"),
ObjectId("512512dcd86041c7dca81917"),
ObjectId("512512e6d86041c7dca81918"),
ObjectId("512512f0d86041c7dca81919")
]
}
The "classes" field keeps an array of "_id"s of classes that John Doe is taking. When
we want to find out information about those classes, we can query the classes collection
154 | Chapter 8: Application Design
with those "_id"s. This only takes two queries. This is fairly popular way to structure
data that does not need to be instantly accessible and changes, but not constantly.
If we need to optimize reads further, we can get all of the information in a single query
by fully denormalizing the data and storing each class as an embedded document in the
"classes" field:
{
"_id" : ObjectId("512512a5d86041c7dca81914"),
"name" : "John Doe",
"classes" : [
{
"class" : "Trigonometry",
"credits" : 3,
"room" : "204"
},
{
"class" : "Physics",
"credits" : 3,
"room" : "159"
},
{
"class" : "Women in Literature",
"credits" : 3,
"room" : "14b"
},
{
"class" : "AP European History",
"credits" : 4,
"room" : "321"
}
]
}
The upside of this is that it only takes one query to get the information. The downsides
are that it takes up more space and is more difficult to keep in sync. For example, if it
turns out that physics was supposed to be four credits (not three) every student in the
physics class would need to have her document updated (instead of just updating a
central “Physics” document).
Finally, you can use a hybrid of embedding and referencing: create an array of subdo‐
cuments with the frequently used information, but with a reference to the actual docu‐
ment for more information:
{
"_id" : ObjectId("512512a5d86041c7dca81914"),
"name" : "John Doe",
"classes" : [
{
"_id" : ObjectId("512512ced86041c7dca81916"),
"class" : "Trigonometry"
Normalization versus Denormalization | 155
},
{
"_id" : ObjectId("512512dcd86041c7dca81917"),
"class" : "Physics"
},
{
"_id" : ObjectId("512512e6d86041c7dca81918"),
"class" : "Women in Literature"
},
{
"_id" : ObjectId("512512f0d86041c7dca81919"),
"class" : "AP European History"
}
]
}
This approach is also a nice option because the amount of information embedded can
change over time as your requirements changes: if you want to include more or less
information on a page, you could embed more or less of it in the document.
Another important consideration is how often this information will change versus how
often it’s read. If it will be updated regularly, then normalizing it is a good idea. However,
if it changes infrequently, then there is little benefit to optimize the update process at
the expense of every read your application performs.
For example, a textbook normalization use case is to store a user and his address in
separate collections. However, people almost never change their address, so you gen‐
erally shouldn’t penalize every read on the off chance that someone’s moved. Your ap‐
plication should embed the address in the user document.
If you decide to use embedded documents and you need to update them, you should
set up a cron job to ensure that any updates you do are successfully propagated to every
document. For example, you might attempt to do a multiupdate but the server crashes
before all of the documents have been updated. You need a way to detect this and retry
the update.
To some extent, the more information you are generating the less of it you should embed.
If the embedded fields or number of embedded fields is supposed to grow without bound
then they should generally be referenced, not embedded. Things like comment trees or
activity lists should be stored as their own documents, not embedded.
Finally, fields should be included that are integral to the data in the document. If a field
is almost always excluded from your results when you query for this document, it’s a
good sign that it may belong in another collection. These guidelines are summarized in
Table 8-1.
156 | Chapter 8: Application Design
Table 8-1. Comparison of embedding versus references
Embedding is better for... References are better for...
Small subdocuments Large subdocuments
Data that does not change regularly Volatile data
When eventual consistency is acceptable When immediate consistency is necessary
Documents that grow by a small amount Documents that grow a large amount
Data that you’ll often need to perform a second query to fetch Data that you’ll often exclude from the results
Fast reads Fast writes
Suppose we had a users collection. Here are some example fields we might have and
whether or not they should be embedded:
Account preferences
They are only relevant to this user document, and will probably be exposed with
other user information in this document. Account preferences should generally be
embedded.
Recent activity
This one depends on how much recent activity grows and changes. If it is a fixed-
size field (last 10 things), it might be useful to embed.
Friends
Generally this should not be embedded, or at least not fully. See the section below
on advice on social networking.
All of the content this user has produced
No.
Cardinality
Cardinality is how many references a collection has to another collection. Common
relationships are one-to-one, one-to-many, or many-to-many. For example, suppose we
had a blog application. Each post has a title, so that’s a one-to-one relationship. Each
author has many posts, so that’s a one-to-many relationship. And posts have many tags
and tags refer to many posts, so that’s a many-to-many relationship.
When using MongoDB, it can be conceptually useful to split “many” into subcategories:
“many” and “few.” For example, you might have a one-to-few cardinality between
authors and posts: each author only writes a few posts. You might have many-to-few
relation between blog posts and tags: your probably have many more blog posts than
you have tags. However, you’d have a one-to-many relationship between blog posts and
comments: each post has many comments.
Normalization versus Denormalization | 157
When you’ve determined few versus many relations, it can help you decide what to
embed versus what to reference. Generally, “few” relationships will work better with
embedding, and “many” relationships will work better as references.
Friends, Followers, and Other Inconveniences
Keep your friends close and your enemies embedded.
Many social applications need to link people, content, followers, friends, and so on.
Figuring out how to balance embedding and referencing this highly connected infor‐
mation can be tricky. This section covers considerations for social graph data. But gen‐
erally following, friending, or favoriting can be simplified to a publication-subscription
system: one user is subscribing to notifications from another. Thus, there are two basic
operations that need to be efficient: how to store subscribers and how to notify all
interested parties of an event.
There are three ways people typically implement subscribing. The first option is that
you can put the producer in the subscriber’s document, which looks something like this:
{
"_id" : ObjectId("51250a5cd86041c7dca8190f"),
"username" : "batman",
"email" : "batman@waynetech.com"
"following" : [
ObjectId("51250a72d86041c7dca81910"),
ObjectId("51250a7ed86041c7dca81936")
]
}
Now, given a user’s document, you can query for something like db.activi
ties.find({"user" : {"$in" : user["following"]}}) to find all of the activities
that have been published that she’d be interested in. However, if you need to find ev‐
eryone who is interested in a newly published activity, you’d have to query the "follow
ing" field across all users.
Alternatively, you could append the followers to the producer’s document, like so:
{
"_id" : ObjectId("51250a7ed86041c7dca81936"),
"username" : "joker",
"email" : "joker@mailinator.com"
"followers" : [
ObjectId("512510e8d86041c7dca81912"),
ObjectId("51250a5cd86041c7dca8190f"),
ObjectId("512510ffd86041c7dca81910")
]
}
158 | Chapter 8: Application Design
Whenever this user does something, all the users we need to notify are right there. The
downside is that now you need to query the whole users collection to find everyone a
user follows (the opposite limitation as above).
Either of these options comes with an additional downside: they make your user docu‐
ment larger and more volatile. The "following" (or "followers") field often won’t
even need to be returned: how often do you want to list every follower? If users are
frequently followed or unfollowed, this can result in a lot of fragmentation, as well. Thus,
the final option neutralizes these downsides by normalizing even further and storing
subscriptions in another collection. Normalizing this far is often overkill, but it can be
useful for an extremely volatile field that often isn’t returned with the rest of the docu‐
ment. "followers" may be a sensible field to normalize this way.
Keep a collection that matches publishers to subscribers, with documents that look
something like this:
{
"_id" : ObjectId("51250a7ed86041c7dca81936"), // followee's "_id"
"followers" : [
ObjectId("512510e8d86041c7dca81912"),
ObjectId("51250a5cd86041c7dca8190f"),
ObjectId("512510ffd86041c7dca81910")
]
}
This keeps your user documents svelte but takes an extra query to get the followers. As
"followers" arrays will generally change size a lot, this allows you to enable the usePo
werOf2Sizes on this collection while keeping the users collection as small as possible.
If you put this followers collection in another database, you can also compact it without
affecting the users collection too much.
Dealing with the Wil Wheaton effect
Regardless of which strategy you use, embedding only works with a limited number of
subdocuments or references. If you have celebrity users, they may overflow any docu‐
ment that you’re storing followers in. The typical way of compensating this is to have a
“continuation” document, if necessary. For example, you might have:
> db.users.find({"username" : "wil"})
{
"_id" : ObjectId("51252871d86041c7dca8191a"),
"username" : "wil",
"email" : "wil@example.com",
"tbc" : [
ObjectId("512528ced86041c7dca8191e"),
ObjectId("5126510dd86041c7dca81924")
]
"followers" : [
ObjectId("512528a0d86041c7dca8191b"),
Normalization versus Denormalization | 159
ObjectId("512528a2d86041c7dca8191c"),
ObjectId("512528a3d86041c7dca8191d"),
...
]
}
{
"_id" : ObjectId("512528ced86041c7dca8191e"),
"followers" : [
ObjectId("512528f1d86041c7dca8191f"),
ObjectId("512528f6d86041c7dca81920"),
ObjectId("512528f8d86041c7dca81921"),
...
]
}
{
"_id" : ObjectId("5126510dd86041c7dca81924"),
"followers" : [
ObjectId("512673e1d86041c7dca81925"),
ObjectId("512650efd86041c7dca81922"),
ObjectId("512650fdd86041c7dca81923"),
...
]
}
Then add application logic to support fetching the documents in the “to be continued”
("tbc") array.
Optimizations for Data Manipulation
To optimize your application, you must first know what its bottleneck is by evaluating
its read and write performance. Optimizing reads generally involves having the correct
indexes and returning as much of the information as possible in a single document.
Optimizing writes usually involves minimizing the number of indexes you have and
making updates as efficient as possible.
There is often a trade-off between schemas that are optimized for writing quickly and
those that are optimized for reading quickly, so you may have to decide which is a more
important for your application. Factor in not only the importance of reads versus writes,
but also their proportions: if writes are more important but you’re doing a thousand
reads to every write, you may still want to optimize reads first.
Optimizing for Document Growth
If you’re going to need to update data, determine whether or not your documents are
going to grow and by how much. If it is by a predictable amount, manually padding
your documents will prevent moves, making writes faster. Check your padding factor:
if it is about 1.2 or greater, consider using manual padding.
160 | Chapter 8: Application Design
When you manually pad a document, you create the document with a large field that
will later be removed. This preallocates the space that the document will eventually need.
For example, suppose you had a collection of restaurant reviews and your documents
looked like this:
{
"_id" : ObjectId(),
"restaurant" : "Le Cirque",
"review" : "Hamburgers were overpriced."
"userId" : ObjectId(),
"tags" : []
}
The "tags" field will grow as users add tags, so the application will often have to perform
an update like this:
> db.reviews.update({"_id" : id},
... {"$push" : {"tags" : {"$each" : ["French", "fine dining", "hamburgers"]}}}})
If "tags" generally doesn’t grow to more than 100 bytes, you could manually pad the
document to prevent any unwanted moves. If you leave the document without padding,
moves will definitely occur as "tags" grows. To pad, add a final field to the document
with whatever field name you’d like:
{
"_id" : ObjectId(),
"restaurant" : "Le Cirque",
"review" : "Hamburgers were overpriced."
"userId" : ObjectId(),
"tags" : [],
"garbage" : "........................................................"+
"................................................................"+
"................................................................"
}
You can either do this on insert or, if the document is created with an upsert, use
"$setOnInsert" to create the field when the document is first inserted.
When you update the document, always "$unset" the "garbage" field:
> db.reviews.update({"_id" : id},
... {"$push" : {"tags" : {"$each" : ["French", "fine dining", "hamburgers"]}}},
... "$unset" : {"garbage" : true}})
The "$unset" will remove the "garbage" field if it exists and be a no-op if it does not.
If your document has one field that grows, try to keep is as the last field in the document
(but before "garbage"). It is slightly more efficient for MongoDB not to have to rewrite
fields after "tags" if it grows.
Optimizations for Data Manipulation | 161
Removing Old Data
Some data is only important for a brief time: after a few weeks or months it is just wasting
storage space. There are three popular options for removing old data: capped collections,
TTL collections, and dropping collections per time period.
The easiest option is to use a capped collection: set it to a large size and let old data “fall
off” the end. However, capped collections pose certain limitations on the operations
you can do and are vulnerable to spikes in traffic, temporarily lowering the length of
time that they can hold. See “Capped Collections” on page 109 for more information.
The second option is TTL collections: this gives you a finer-grain control over when
documents are removed. However, it may not be fast enough for very high-write-volume
collections: it removes documents by traversing the TTL index the same way a user-
requested remove would. If TTL collections can keep up, though, they are probably the
easiest solution. See “Time-To-Live Indexes” on page 114 for more information about
TTL indexes.
The final option is to use multiple collections: for example, one collection per month.
Every time the month changes, your application starts using this month’s (empty) col‐
lection and searching for data in both the current and previous months’ collections.
Once a collection is older than, say, six months, you can drop it. This can keep up with
nearly any volume of traffic, but it is more complex to build an application around, since
it has to use dynamic collection (or database) names and possibly query multiple
databases.
Planning Out Databases and Collections
Once you have sketched out what your documents look like, you must decide what
collections or databases to put them in. This is often a fairly intuitive process, but there
are some guidelines to keep in mind.
In general, documents with a similar schema should be kept in the same collection.
MongoDB generally disallows combining data from multiple collections, so if there are
documents that need to be queried or aggregated together, those are good candidates
for putting in one big collection. For example, you might have documents that are fairly
different “shapes,” but if you’re going to be aggregating them, they all need to live in the
same collection.
For databases, the big issues to consider are locking (you get a read/write lock per da‐
tabase) and storage. Each database resides in its own files and often its own directory
on disk, which means that you could mount different databases to different volumes.
Thus, you may want all items within a database to be of similar “quality,” similar access
pattern, or similar traffic levels.
162 | Chapter 8: Application Design
For example, suppose we have an application with several components: a logging com‐
ponent that creates a huge amount of not-very-valuable data, a user collection, and a
couple of collections for user-generated data. The user collections are high-value: it is
important that user data is safe. There is also a high-traffic collection for social activities,
which is of lower importance but not quite as unimportant as the logs. This collection
is mainly used for user notifications, so it is almost an append-only collection.
Splitting these up by importance, we might end up with three databases: logs, activi
ties, and users. The nice thing about this strategy is that you may find that your highest-
value data is also your smallest (for instance, users probably don’t generate as much data
as your logging does). You might not be able to afford an SSD for your entire data set,
but you might be able to get one for your users. Or use RAID10 for users and RAID0
for logs and activities.
Be aware that there are some limitations when using multiple databases: MongoDB
generally does not allow you to move data directly from one database to another. For
example, you cannot store the results of a MapReduce in a different database than you
ran the MapReduce on and you cannot move a collection from one database to another
with the renameCollection command (e.g., you can rename foo.bar as foo.baz, but not
foo2.baz).
Managing Consistency
You must figure out how consistent your application’s reads need to be. MongoDB sup‐
ports a huge variety in consistency levels, from always reading your own writes to read‐
ing data of unknown oldness. If you’re reporting on the last year of activity, you might
only need data that’s correct to the last couple of days. Conversely, if you’re doing real-
time trading, you might need to immediately read the latest writes.
To understand how to achieve these varying levels of consistency, it is important to
understand what MongoDB is doing under the hood. The server keeps a queue of re‐
quests for each connection. When the client sends a request, it will be placed at the end
of its connection’s queue. Any subsequent requests on the connection will occur after
the enqueued operation is processed. Thus, a single connection has a consistent view
of the database and can always read its own writes.
Note that this is a per-connection queue: if we open two shells, we will have two con‐
nections to the database. If we perform an insert in one shell, a subsequent query in the
other shell might not return the inserted document. However, within a single shell, if
we query for the document after inserting, the document will be returned. This behavior
can be difficult to duplicate by hand, but on a busy server interleaved inserts and queries
are likely to occur. Often developers run into this when they insert data in one thread
and then check that it was successfully inserted in another. For a moment or two, it looks
like the data was not inserted, and then it suddenly appears.
Managing Consistency | 163
This behavior is especially worth keeping in mind when using the Ruby, Python, and
Java drivers, because all three use connection pooling. For efficiency, these drivers open
multiple connections (a pool) to the server and distribute requests across them. They
all, however, have mechanisms to guarantee that a series of requests is processed by a
single connection. There is detailed documentation on connection pooling in various
languages on the MongoDB wiki.
When you send reads to a replica set secondary (see Chapter 11), this becomes an even
larger issue. Secondaries may lag behind the primary, leading to reading data from
seconds, minutes, or even hours ago. There are several ways of dealing with this, the
easiest being to simply send all reads to the primary if you care about staleness. You
could also set up an automatic script to detect lag on a secondary and put it into main‐
tenance mode if it lags too far behind. If you have a small set, it might be worth using
"w" : setSize as a write concern and sending subsequent reads to the primary if
getLastError does not return successfully.
Migrating Schemas
As your application grows and your needs change, your schema may have to grow and
change as well. There are a couple of ways of accomplishing this, and regardless of the
method you chose, you should carefully document each schema that your application
has used.
The simplest method is to simply have your schema evolve as your application requires,
making sure that your application supports all old versions of the schema (e.g., accepting
the existence or non-existence of fields or dealing with multiple possible field types
gracefully). This technique can become messy, particularly if you have conflicting ver‐
sions. For instance, one version might require a "mobile" field and one version might
require not having a "mobile" field but does require another field, and yet another
version thinks that the "mobile" field is optional. Keeping track of these requirements
can gradually turn code into spaghetti.
To handle changing requirements in a slightly more structured way you can include a
"version" field (or just "v") in each document and use that to determine what your
application will accept for document structure. This enforces your schema more rig‐
orously: a document has to be valid for some version of the schema, if not the current
one. However, it still requires supporting old versions.
The final option is to migrate all of your data when the schema changes. Generally this
is not a good idea: MongoDB allows you to have a dynamic schema in order to avoid
migrates because they put a lot of pressure on your system. However, if you do decide
to change every document, you will need to ensure that all documents were successfully
updated. MongoDB does not support atomic multiupdates (either they all happen or
164 | Chapter 8: Application Design
they all don’t across multiple documents). If MongoDB crashes in the middle of a mi‐
grate, you could end up with some updated and some non-updated documents.
When Not to Use MongoDB
While MongoDB is a general-purpose database that works well for most applications,
it isn’t good at everything. Here are some tasks that MongoDB is not designed to do:
•MongoDB does not support transactions, so systems that require transactions
should use another data store. There are a couple of ways to hack in simple
transaction-like semantics, particularly on a single document, but there is no da‐
tabase enforcement. Thus, you can make all of your clients agree to obey whatever
semantics you come up with (e.g., “Check the "locks" field before doing any op‐
eration”) but there is nothing stopping an ignorant or malicious client from messing
things up.
•Joining many different types of data across many different dimensions is something
relational databases are fantastic at. MongoDB isn’t supposed to do this well and
most likely never will.
•Finally, one of the big (if hopefully temporary) reasons to use a relational database
over MongoDB is if you’re using tools that don’t support MongoDB. From SQLAl‐
chemy to Wordpress, there are thousands of tools that just weren’t built to support
MongoDB. The pool of tools that support MongoDB is growing but is hardly the
size of relational databases’ ecosystem, yet.
When Not to Use MongoDB | 165
PART III
Replication
CHAPTER 9
Setting Up a Replica Set
This chapter introduces MongoDB’s replication system: replica sets. It covers:
• What replica sets are
• How to set up a replica set
• What configuration options are available for replica set members
Introduction to Replication
Since the first chapter, we’ve been using a standalone server, a single mongod server. It’s
an easy way to get started but a dangerous way to run in production: what if your server
crashes or becomes unavailable? Your database will at least be unavailable for a little
while. If there are problems with the hardware, you may have to move your data to
another machine. In the worst case, disk or network issues could leave you with corrupt
or inaccessible data.
Replication is a way of keeping identical copies of your data on multiple servers and is
recommended for all production deployments. Replication keeps your application run‐
ning and your data safe, even if something happens to one or more of your servers.
With MongoDB, you set up replication by creating a replica set. A replica set is a group
of servers with one primary, the server taking client requests, and multiple seconda‐
ries, servers that keep copies of the primary’s data. If the primary crashes, the secondaries
can elect a new primary from amongst themselves.
If you are using replication and a server goes down, you can still access your data from
the other servers in the set. If the data on a server is damaged or inaccessible, you can
make a new copy of the data from one the other members of the set.
This chapter introduces replica sets and covers how to set up replication on your system.
169
A One-Minute Test Setup
This section will get you started quickly by setting up a three-member replica set on
your local machine. This setup is obviously not suitable for production, but it’s a nice
way to familiarize yourself with replication and play around with configuration.
This quick-start method stores data in /data/db, so make sure that di‐
rectory exists and is writable by your user before running this code.
Start up a mongo shell with the --nodb option, which allows you to start a shell that is
not connected to any mongod:
$ mongo --nodb
Create a replica set by running the following command:
> replicaSet = new ReplSetTest({"nodes" : 3})
This tells the shell to create a new replica set with three servers: one primary and two
secondaries. However, it doesn’t actually start the mongod servers until you run the
following two commands:
> // starts three mongod processes
> replicaSet.startSet()
>
> // configures replication
> replicaSet.initiate()
You should now have three mongod processes running locally on ports 31000, 31001,
and 31002. They will all be dumping their logs into the current shell, which is very noisy,
so put this shell aside and open up a new one.
In the second shell, connect to the mongod running on port 31000:
> conn1 = new Mongo("localhost:31000")
connection to localhost:31000
testReplSet:PRIMARY>
testReplSet:PRIMARY> primaryDB = conn1.getDB("test")
test
Notice that, when you connect to a replica set member, the prompt changes to
testReplSet:PRIMARY>. "PRIMARY" is the state of the member and "testReplSet" is
an identifier for this set. You’ll learn how to choose your own identifier later; testRepl
Set is the default name ReplSetTest uses.
Examples from now on will just use > for the prompt instead of testReplSet:PRI
MARY> to keep things more readable.
170 | Chapter 9: Setting Up a Replica Set
Use your connection to the primary to run the isMaster command. This will show you
the status of the set:
> primaryDB.isMaster()
{
"setName" : "testReplSet",
"ismaster" : true,
"secondary" : false,
"hosts" : [
"wooster:31000",
"wooster:31002",
"wooster:31001"
],
"primary" : "wooster:31000",
"me" : "wooster:31000",
"maxBsonObjectSize" : 16777216,
"localTime" : ISODate("2012-09-28T15:48:11.025Z"),
"ok" : 1
}
There are a bunch of fields in the output from isMaster, but the important ones indicate
that you can see that this node is primary (the "ismaster" : true field) and that there
is a list of hosts in the set.
If this server says "ismaster" : false, that’s fine. Look at the "pri
mary" field to see which node is primary and then repeat the connection
steps above for that host/port.
Now that you’re connected to the primary, let’s try doing some writes and see what
happens. First, insert 1,000 documents:
> for (i=0; i<1000; i++) { primaryDB.coll.insert({count: i}) }
>
> // make sure the docs are there
> primaryDB.coll.count()
1000
Now check one of the secondaries and verify that they have a copy of all of these docu‐
ments. Connect to either of the secondaries:
> conn2 = new Mongo("localhost:31001")
connection to localhost:31001
> secondaryDB = conn2.getDB("test")
test
Secondaries may fall behind the primary (or lag) and not have the most current writes,
so secondaries will refuse read requests by default to prevent applications from acci‐
dentally reading stale data. Thus, if you attempt to query a secondary, you’ll get an error
that it’s not primary:
A One-Minute Test Setup | 171
> secondaryDB.coll.find()
error: { "$err" : "not master and slaveok=false", "code" : 13435 }
This is to protect your application from accidentally connecting to a secondary and
reading stale data. To allow queries on the secondary, we set an “I’m okay with reading
from secondaries” flag, like so:
> conn2.setSlaveOk()
Note that slaveOk is set on the connection (conn2), not the database (secondaryDB).
Now you’re all set to read from this member. Query it normally:
> secondaryDB.coll.find()
{ "_id" : ObjectId("5037cac65f3257931833902b"), "count" : 0 }
{ "_id" : ObjectId("5037cac65f3257931833902c"), "count" : 1 }
{ "_id" : ObjectId("5037cac65f3257931833902d"), "count" : 2 }
...
{ "_id" : ObjectId("5037cac65f3257931833903c"), "count" : 17 }
{ "_id" : ObjectId("5037cac65f3257931833903d"), "count" : 18 }
{ "_id" : ObjectId("5037cac65f3257931833903e"), "count" : 19 }
Type "it" for more
>
> secondaryDB.coll.count()
1000
You can see that all of our documents are there.
Now, try to write to a secondary:
> secondaryDB.coll.insert({"count" : 1001})
> secondaryDB.runCommand({"getLastError" : 1})
{
"err" : "not master",
"code" : 10058,
"n" : 0,
"lastOp" : Timestamp(0, 0),
"connectionId" : 5,
"ok" : 1
}
You can see that the secondary does not accept the write. The secondary will only per‐
form writes that it gets through replication, not from clients.
There is one other interesting feature that you should try out: automatic failover. If the
primary goes down, one of the secondaries will automatically be elected primary. To try
this out, stop the primary:
> primaryDB.adminCommand({"shutdown" : 1})
Run isMaster on the secondary to see who has become the new primary:
> secondaryDB.isMaster()
It should look something like this:
172 | Chapter 9: Setting Up a Replica Set
{
"setName" : "testReplSet",
"ismaster" : true,
"secondary" : false,
"hosts" : [
"wooster:31001",
"wooster:31000",
"wooster:31002"
],
"primary" : "wooster:31001",
"me" : "wooster:31001",
"maxBsonObjectSize" : 16777216,
"localTime" : ISODate("2012-09-28T16:52:07.975Z"),
"ok" : 1
}
Your primary may be the other server; whichever secondary noticed that the primary
was down first will be elected. Now you can send writes to the new primary.
isMaster is a very old command, predating replica sets to when MongoDB only sup‐
ported master-slave replication. Thus, it does not use the replica set terminology con‐
sistently: it still calls the primary a “master.” You can generally think of “master” as
equivalent to “primary” and “slave” as equivalent to “secondary.”
When you’re done working with the set, shut down the servers from your first shell.
This shell will be full of log output from the members of the set, so hit Enter a few times
to get back to a prompt. To shutdown the set, run:
> replicaSet.stopSet()
Congratulations! You just set up, used, and tore down replication.
There are a few key concepts to remember:
•Clients can send a primary all the same operations they could send a standalone
server (reads, writes, commands, index builds, etc.).
•Clients cannot write to secondaries.
•Clients, by default, cannot read from secondaries. By explicitly setting an “I know
I’m reading from a secondary” setting, clients can read from secondaries.
Now that you understand the basics, the rest of this chapter focuses on configuring a
replica set under more realistic circumstances. Remember that you can always go back
to ReplSetTest if you want to quickly try out a configuration or option.
A One-Minute Test Setup | 173
Configuring a Replica Set
For actual deployments, you’ll need to set up replication across multiple machines. This
section takes you through setting up a real replica set that could be used by your
application.
Let’s say that you already have a standalone mongod on server-1:27017 with some data
on it. (If you do not have any pre-existing data, this will work the same way, just with
an empty data directory.) The first thing you need to do is choose a name for your set.
Any string whatsoever will do, so long as it’s UTF-8.
Once you have a name for your replica set, restart server-1 with the --replSet name
option. For example:
$ mongod --replSet spock -f mongod.conf --fork
Now start up two more mongod servers with the replSet option and the same identifier
(spock): these will be the other members of the set:
$ ssh server-2
server-2$ mongod --replSet spock -f mongod.conf --fork
server-2$ exit
$
$ ssh server-3
server-3$ mongod --replSet spock -f mongod.conf --fork
server-3$ exit
Each of the other members should have an empty data directory, even if the first member
had data. They will automatically clone the first member’s data to their machines once
they have been added to the set.
For each member, add the replSet option to its mongod.conf file so that it will be used
on startup from now on.
Once you’ve started the mongods, you should have three mongods running on three
separate servers. However, each mongod does not yet know that the others exist. To tell
them about one another, you have to create a configuration that lists each of the members
and send this configuration to server-1. It will take care of propagating it to the other
members.
First we’ll create the configuration. In the shell, create a document that looks like this:
> config = {
"_id" : "spock",
"members" : [
{"_id" : 0, "host" : "server-1:27017"},
{"_id" : 1, "host" : "server-2:27017"},
{"_id" : 2, "host" : "server-3:27017"}
]
}
174 | Chapter 9: Setting Up a Replica Set
There are several important parts of config. The config’s "_id" is the name of the set
that you passed in on the command line (in this example, "spock"). Make sure that this
name matches exactly.
The next part of the document is an array of members of the set. Each of these needs
two fields: a unique "_id" that is an integer and a hostname (replace the hostnames
with whatever your servers are called).
This config object is your replica set configuration, so now you have to send it to a
member of the set. To do so, connect to the server with data on it (server-1:27017) and
initiate the set with this configuration:
> // connect to server-1
> db = (new Mongo("server-1:27017")).getDB("test")
>
> // initiate replica set
> rs.initiate(config)
{
"info" : "Config now saved locally. Should come online in about a minute.",
"ok" : 1
}
server-1 will parse the configuration and send messages to the other members, alerting
them of the new configuration. Once they have all loaded the configuration, they will
elect a primary and start handling reads and writes.
Unfortunately, you cannot convert a standalone server to a replica set
without some downtime for restarting it and initializing the set. Thus,
even if you only have one server to start out with, you may want to
configure it as a one-member replica set. That way, if you want to add
more members later, you can do so without downtime.
If you are starting a brand-new set, you can send the configuration to any member in
the set. If you are starting with data on one of the members, you must send the config‐
uration to the member with data. You cannot initiate a set with data on more than one
member.
You must use the mongo shell to configure replica sets. There is no way
to do file-based replica set configuration.
rs Helper Functions
Note the rs in the rs.initiate() command above. rs is a global variable that contains
replication helper functions (run rs.help() to see the helpers it exposes). These
Configuring a Replica Set | 175
functions are almost always just wrappers around database commands. For example,
the following database command is equivalent to rs.initiate(config):
> db.adminCommand({"replSetInitiate" : config})
It is good to have a passing familiarity with both the helpers and the underlying com‐
mands, as it may sometimes be easier to use the command form instead of the helper.
Networking Considerations
Every member of a set must be able to make connections to every other member of the
set (including itself). If you get errors about members not being able to reach other
members that you know are running, you may have to change your network configu‐
ration to allow connections between them.
Also, replica sets configurations shouldn’t use localhost as a hostname. There isn’t much
point to running a replica set on one machine and localhost won’t resolve correctly from
a foreign machine. MongoDB allows all-localhost replica sets for testing locally but will
protest if you try to mix localhost and non-localhost servers in a config.
Changing Your Replica Set Configuration
Replica set configurations can be changed at any time: members can be added, removed,
or modified. There are shell helpers for some common operations; for example, to add
a new member to the set, you can use rs.add:
> rs.add("server-4:27017")
Similarly, you can remove members;
> rs.remove("server-1:27017")
Fri Sep 28 16:44:46 DBClientCursor::init call() failed
Fri Sep 28 16:44:46 query failed : admin.$cmd { replSetReconfig: {
_id: "testReplSet", version: 2, members: [ { _id: 0, host: "ubuntu:31000" },
{ _id: 2, host: "ubuntu:31002" } ] } } to: localhost:31000
Fri Sep 28 16:44:46 Error: error doing query:
failed src/mongo/shell/collection.js:155
Fri Sep 28 16:44:46 trying reconnect to localhost:31000
Fri Sep 28 16:44:46 reconnect localhost:31000 ok
Note that when you remove a member (or do almost any configuration change other
than adding a member), you will get a big, ugly error about not being able to connect
to the database in the shell. This is okay; it actually means the reconfiguration succeeded!
When you reconfigure a set, the primary closes all connections as the last step in the
reconfiguration process. Thus, the shell will briefly be disconnected but will automati‐
cally reconnect on your next operation.
176 | Chapter 9: Setting Up a Replica Set
The reason that the primary closes all connections is that it briefly steps down whenever
you reconfigure the set. It should step up again immediately, but be aware that your set
will not have a primary for a moment or two after reconfiguring.
You can check that a reconfiguration succeeded by run rs.config() in the shell. It will
print the current configuration:
> rs.config()
{
"_id" : "testReplSet",
"version" : 2,
"members" : [
{
"_id" : 1,
"host" : "server-2:27017"
},
{
"_id" : 2,
"host" : "server-3:27017"
},
{
"_id" : 3,
"host" : "server-4:27017"
}
]
}
Each time you change the configuration, the "version" field will increase. It starts at
version 1.
You can also modify existing members, not just add and remove them. To make mod‐
ifications, create the configuration document that you want in the shell and call rs.re
config. For example, suppose we have a configuration such as the one shown here:
> rs.config()
{
"_id" : "testReplSet",
"version" : 2,
"members" : [
{
"_id" : 0,
"host" : "server-1:27017"
},
{
"_id" : 1,
"host" : "10.1.1.123:27017"
},
{
"_id" : 2,
"host" : "server-3:27017"
}
Changing Your Replica Set Configuration | 177
]
}
Someone accidentally added member 1 by IP, instead of its hostname. To change that,
first we load the current configuration in the shell and then we change the relevant fields:
> var config = rs.config()
> config.members[1].host = "server-2:27017"
Now that the config document is correct, we need to send it to the database using the
rs.reconfig helper:
> rs.reconfig(config)
rs.reconfig is often more useful that rs.add and rs.remove for complex operations,
such as modifying members’ configuration or adding/removing multiple members at
once. You can use it to make any legal configuration change you need: simply create the
config document that represents your desired configuration and pass it to rs.reconfig.
How to Design a Set
To plan out your set, there are certain replica set concepts that you must be familiar
with. The next chapter goes into more detail about these, but the most important is that
replica sets are all about majorities: you need a majority of members to elect a primary,
a primary can only stay primary so long as it can reach a majority, and a write is safe
when it’s been replicated to a majority. This majority is defined to be “more than half of
all members in the set,” as shown in Table 9-1.
Table 9-1. What is a majority?
Number of members in the set Majority of the set
1 1
2 2
3 2
4 3
5 3
6 4
7 4
Note that it doesn’t matter how many members are down or unavailable, as majority is
based on the set’s configuration.
For example, suppose that we have a five-member set and three members go down, as
shown in Figure 9-1. There are still two members up. These two members cannot reach
a majority of the set (at least three members), so they cannot elect a primary. If one of
them were primary, it would step down as soon as it noticed that it could not reach a
178 | Chapter 9: Setting Up a Replica Set
majority. After a few seconds, your set would consist of two secondaries and three un‐
reachable members.
Figure 9-1. With a minority of the set available, all members will be secondaries
Many users find this frustrating: why can’t the two remaining members elect a primary?
The problem is that it’s possible that the other three members didn’t go down, and that
it was the network that went down, as shown in Figure 9-2. In this case, the three mem‐
bers on the left will elect a primary, since they can reach a majority of the set (three
members out of five).
In the case of a network partition, we do not want both sides of the partition to elect a
primary: otherwise the set would have two primaries. Then both primaries would be
writing to the data and the data sets would diverge. Requiring a majority to elect or stay
primary is a neat way of avoiding ending up with more than one primary.
Figure 9-2. For the members, a network partition looks identical to servers on the other
side of the partition going down
It is important to configure your set in such a way that you’ll usually be able to have one
primary. For example, in the five-member set described above, if members 1, 2, and 3
are in one data center and members 4 and 5 are in another, there should almost always
be a majority available in the first data center (it’s more likely to have a network break
between data centers than within them).
How to Design a Set | 179
One common setup that usually isn’t what you want is a two member set: one primary
and one secondary. Suppose one member becomes unavailable: the other member can‐
not see it, as shown in Figure 9-3. In this situation, neither side of the network partition
has a majority so you’ll end up with two secondaries. For this reason, this type of con‐
figuration is not generally recommended.
Figure 9-3. With an even number of members, neither side of a partition has a majority
There are a couple of configurations that are recommended:
•A majority of the set in one data center, as in Figure 9-2. This is a good design if
you have a primary data center where you always want your replica set’s primary
to be located. So long as your primary data center is healthy, you will have a primary.
However, if that data center becomes unavailable, your secondary data center will
not be able to elect a new primary.
•An equal number of servers in each data center, plus a tie-breaking server in a third
location. This is a good design if your data centers are “equal” in preference, since
generally servers from either data center will be able to see a majority of the set.
However, it involves having three separate locations for servers.
More complex requirements might require different configurations, but you should
keep in mind how your set will acquire a majority under adverse conditions.
All of these complexities would disappear if MongoDB supported having more than
one primary. However, multimaster would bring its own host of complexities. With two
primaries, you would have to handle conflicting writes (for example, someone updates
a document on one primary and someone deletes it on another primary). There are two
popular ways of handling conflicts in systems that support multiple writers: manual
reconciliation or having the system arbitrarily pick a “winner.” Neither of these options
is a very easy model for developers to code against, seeing that you can’t be sure that the
data you’ve written won’t change out from under you. Thus, MongoDB chose to only
support having a single primary. This makes development easier but can result in pe‐
riods when the replica set is read-only.
How Elections Work
When a secondary cannot reach a primary, it will contact all the other members and
request that it be elected primary. These other members do several sanity checks: Can
they reach a primary that the member seeking election cannot? Is the member seeking
180 | Chapter 9: Setting Up a Replica Set
election up to date with replication? Is there anyone with a higher priority available who
should be elected instead?
If a member seeking election receives “ayes” from a majority of the set, it becomes
primary. If even one server vetoes the election, the election is canceled. A member vetoes
an election when it knows any reason that the member seeking election shouldn’t be‐
come primary.
You may see a very large negative number in the logs, since a veto is registered as 10,000
votes. Often you’ll see messages about election results being 9,999 or similar if one
member voted for a member and another member vetoed the election:
Wed Jun 20 17:44:02 [rsMgr] replSet info electSelf 1
Wed Jun 20 17:44:02 [rsMgr] replSet couldn't elect self, only received -9999 votes
If two members vetoed and one voted for, the election results would be 19,999, and so
on. These messages are normal and nothing to worry about.
The member seeking election (the candidate) must be up to date with replication, as far
as the members it can reach know. All replicated operations are strictly ordered by
ascending timestamp, so the candidate must have operations later than or equal to any
member it can reach.
For example, suppose that the latest operation that the candidate has replicated is op
123. It contacts the other members of the set and one of them has replicated up to
operation 124. That member will veto the candidate’s election. The candidate will con‐
tinue syncing and once it has synced operation 124, it will call for an election again (if
no one else has become primary in that time). This time around, assuming nothing else
is wrong with candidate, the member that previously vetoed the election will vote for
the candidate.
Assuming that the candidate receives “ayes” from a majority of voters, it will transition
into primary state.
A common point of confusion is that members always seek election for themselves. For
simplicity’s sake, neighbors cannot “nominate” another server to be primary, they can
only vote for it if it is seeking election.
Member Configuration Options
The replica sets we have set up so far have been fairly uniform in that every member
has the same configuration as every other member. However, there are many situations
when you don’t want members to be identical: you might want one member to prefer‐
entially be primary or make a member invisible to clients so that no read requests can
be routed to it. These and many other configuration options can be specified in the
member subdocuments of the replica set configuration. This section outlines the mem‐
ber options that you can set.
Member Configuration Options | 181
Creating Election Arbiters
The example above shows the disadvantages two-member sets have for majority re‐
quirements. However, many people with small deployments do not want to keep three
copies of their data, feeling that two is enough and keeping a third copy is not worth
the administrative, operational, and financial costs.
For these deployments, MongoDB supports a special type of member called an arbiter,
whose only purpose is to participate in elections. Arbiters hold no data and aren’t used
by clients: they just provide a majority for two-member sets.
As arbiters don’t have any of the traditional responsibilities of a mongod server, you can
run an arbiter as a lightweight process on a wimpier server than you’d generally use for
MongoDB. It’s often a good idea, if possible, to run an arbiter in a separate failure domain
from the other members, so that it has an “outside perspective” on the set, as described
in the recommended deployments listed in “How to Design a Set” on page 178.
You start up an arbiter in the same way that you start a normal mongod, using the --
replSet name option and an empty data directory. You can add it to the set using the
rs.addArb() helper:
> rs.addArb("server-5:27017")
Equivalently, you can specify the arbiterOnly option in the member configuration:
> rs.add({"_id" : 4, "host" : "server-5:27017", "arbiterOnly" : true})
An arbiter, once added to the set, is an arbiter forever: you cannot reconfigure an arbiter
to become a nonarbiter, or vice versa.
One other thing that arbiters are good for is breaking ties in larger clusters. If you have
an even number of nodes, you may have half the nodes vote for one member and half
for another. Adding an arbiter can add a deciding vote.
Use at most one arbiter
Note that, in both of the use cases above, you need at most one arbiter. You do not need
an arbiter if you have an odd number of nodes. A common misconception seems to be
that you should add extra arbiters “just in case.” However, it doesn’t help elections go
any faster or provide any data safety to add extra arbiters.
Suppose you have a three members set. Two members are required to elect a primary.
If you add an arbiter, you’ll have a four member set, so three members will be required
to choose a primary. Thus, your set is potentially less stable: instead of requiring 67%
of your set to be up, you’re now requiring 75%.
Having extra members can also make elections take longer. If you have an even number
of nodes because you added an arbiter, your arbiters can cause ties, not prevent them.
182 | Chapter 9: Setting Up a Replica Set
The downside to using an arbiter
If you have a choice between a data node and an arbiter, choose a data node. Using an
arbiter instead of a data node in a small set can make some operational tasks more
difficult. For example, suppose you are running a replica set with two “normal” members
and one arbiter, and one of the data-holding members goes down. If that member is
well and truly dead (the data is unrecoverable), you will have to get a copy of the data
from the current primary to the new server you’ll be using as a secondary. Copying data
can put a lot of stress on a server and, thus, slow down your application. (Generally,
copying a few gigabytes to a new server is trivial but more than a hundred starts be‐
coming impractical.)
Conversely, if you have three data-holding members, there’s more “breathing room” if
a server completely dies. You can use the remaining secondary to bootstrap a new server
instead of depending on your primary.
In the two-member-plus-arbiter scenario, the primary is the last remaining good copy
of your data and the one trying to handle load from your application while you’re trying
to get another copy of your data online.
Thus, if possible, use an odd number of “normal” members instead of an arbiter.
Priority
Priority is how strongly this member “wants” to become primary. Priority can range
from 0 to 100 and the default is 1. Setting priority to 0 has a special meaning: members
with 0 priority can never become primary. These members are called passive members.
The highest-priority member will always be elected primary (so long as they can reach
a majority of the set and have the most up-to-date data). For example, suppose you add
a member with priority of 1.5 to the set, like so:
> rs.add({"_id" : 4, "host" : "server-4:27017", "priority" : 1.5})
Assuming the other members of the set have priority 1, once server-4 caught up with
the rest of the set, the current primary would automatically step down and server-4
would elect itself. If server-4 was, for some reason, unable to catch up, the current pri‐
mary would stay primary. Setting priorities will never cause your set to go primary-less.
It will also never cause a member who is behind to become primary (until it has caught
up).
One interesting wrinkle with priority is that reconfigurations must always be sent to a
member that could be primary with the new configuration. Therefore, you cannot set
the current primary’s priority to 0 with a single reconfig (and you cannot configure a
set where all of the priorities are 0).
Member Configuration Options | 183
The absolute value of a priority only matters in relation to whether it is greater or less
than the other priorities in the set: members with priorities of 500, 1, and 1 will behave
the same way as another set with priorities 2, 1, and 1.
Hidden
Clients do not route requests to hidden members and hidden members are not preferred
as replication sources (although they will be used if more desirable sources are not
available). Thus, many people will hide less powerful or backup servers.
For example, suppose you had a set that looked like this:
> rs.isMaster()
{
...
"hosts" : [
"server-1:27107",
"server-2:27017",
"server-3:27017"
],
...
}
To hide server-3, add the hidden: true field to its configuration. A member must have
a priority of 0 to be hidden (you can’t have a hidden primary):
> var config = rs.config()
> config.members[2].hidden = 0
0
> config.members[2].priority = 0
0
> rs.reconfig(config)
Now running isMaster() will show:
> rs.isMaster()
{
...
"hosts" : [
"server-1:27107",
"server-2:27017"
],
...
}
rs.status() and rs.config() will still show the member; it only disappears from
isMaster(). When clients connect to a replica set, they call isMaster() to determine
the members of the set. Thus, hidden members will never be used for read requests.
To unhide a member, change the hidden option to false or remove the option entirely.
184 | Chapter 9: Setting Up a Replica Set
Slave Delay
It’s always possible for your data to be nuked by human error: someone might acciden‐
tally drop your main database or a newly deployed version of your application might
have a bug that replaces all of your data with garbage. To defend against that type of
problem, you can set up a delayed secondary using the slaveDelay setting.
A delayed secondary purposely lags by the specified number of seconds. This way, if
someone fat-fingers away your main collection, you can restore it from an identical copy
of the data from earlier. This is covered in “Restoring from a Delayed Secondary” on
page 221.
slaveDelay requires the member’s priority to be 0. If your application is routing reads
to secondaries, you should make slave delayed members hidden so that reads are not
routed to them.
Building Indexes
Sometimes a secondary does not need to have the same (or any) indexes that exist on
the primary. If you are using a secondary only for backup data or offline batch jobs, you
might want to specify "buildIndexes" : false in the member’s configuration. This
option prevents the secondary from building any indexes.
This is a permanent setting: members that have "buildIndexes" : false specified can
never be reconfigured to be “normal” index-building members again. If you want to
change a non-index-building member to an index-building one, you must remove it
from the set, delete all of its data, re-add it to the set, and allow it to resync from scratch.
Again, this option requires the member’s priority to be 0.
Member Configuration Options | 185
CHAPTER 10
Components of a Replica Set
This chapter covers how the pieces of a replica set fit together, including:
• How replica set members replicate new data
• How bringing up new members works
• How elections work
•Possible server and network failure scenarios
Syncing
Replication is concerned with keeping an identical copy of data on multiple servers. The
way MongoDB accomplishes this is by keeping a log of operations, or oplog, containing
every write that a primary performs. This is a capped collection that lives in the local
database on the primary. The secondaries query this collection for operations to
replicate.
Each secondary maintains its own oplog, recording each operation it replicates from
the primary. This allows any member to be used as a sync source for any other member,
as shown in Figure 10-1. Secondaries fetch operations from the member they are syncing
from, apply the operations to their data set, and then write the operations to the oplog.
If applying an operation fails (which should only happen if the underlying data has been
corrupted or in some way differs from the primary), the secondary will exit.
187
Figure 10-1. Oplog keep an ordered list of write operations that have occurred. Each
member has its own copy of the oplog, which should be identical to the primary’s (mod‐
ulo some lag).
If a secondary goes down for any reason, when it restarts it will start syncing from the
last operation in its oplog. As operations are applied to data and then written to the
oplog, the secondary may replay operations that it has already applied to its data. Mon‐
goDB is designed for it to handle this correctly: replaying oplog ops multiple times yields
the same result as replaying them once.
As the oplog is a fixed size, it can only hold a certain number of operations. In general,
the oplog will use space at approximately the same rate as writes come into the system:
if you’re writing 1 KB/minute on the primary, your oplog is probably going to fill up at
about 1 KB/minute. However, there are a few exceptions: operations that effect multiple
documents, such as removes or a multi-updates, that will be exploded into many oplog
entries. The single operation on the primary will be split into one oplog op per document
affected. Thus, if you remove 1,000,000 documents from a collection with db.coll.re
move(), it will become 1,000,000 oplog entries removing one document at a time. If you
are doing lots of bulk operations, this can fill up your oplog more quickly than you
might expect.
Initial Sync
When a member of the set starts up, it will check if it is in a valid state to begin syncing
from someone. If not, it will attempt to make a full copy of data from another member
of the set. This is called initial syncing and there are several steps to the process, which
you can follow in the mongod’s log:
1. First, the member does some preliminary bookkeeping: it chooses a member to
sync from, creates an identifier for itself in local.me, and drops all existing databases
to start with a clean slate:
Mon Jan 30 11:09:18 [rsSync] replSet initial sync pending
Mon Jan 30 11:09:18 [rsSync] replSet syncing to: server-1:27017
188 | Chapter 10: Components of a Replica Set
Mon Jan 30 11:09:18 [rsSync] build index local.me { _id: 1 }
Mon Jan 30 11:09:18 [rsSync] build index done 0 records 0 secs
Mon Jan 30 11:09:18 [rsSync] replSet initial sync drop all databases
Mon Jan 30 11:09:18 [rsSync] dropAllDatabasesExceptLocal 1
Note that any existing data will be dropped at this point. Only do an initial sync if
you do not want the data in your data directory or have moved it elsewhere, as
mongod’s first action is to delete it all.
2. Cloning is the initial data copy of all records from the sync source. This is usually
the most time-consuming part of the process:
Mon Jan 30 11:09:18 [rsSync] replSet initial sync clone all databases
Mon Jan 30 11:09:18 [rsSync] replSet initial sync cloning db: db1
Mon Jan 30 11:09:18 [FileAllocator] allocating new datafile /data/db/db1.ns,
filling with zeroes...
3. Then the first oplog application occurs, which applies any operations that happened
during the clone. This may have to reclone certain documents that were moved
and, therefore, missed by the cloner:
Mon Jan 30 15:38:36 [rsSync] oplog sync 1 of 3
Mon Jan 30 15:38:36 [rsBackgroundSync] replSet syncing to: server-1:27017
Mon Jan 30 15:38:37 [rsSyncNotifier] replset setting oplog notifier to
server-1:27017
Mon Jan 30 15:38:37 [repl writer worker 2] replication update of non-mod
failed:
{ ts: Timestamp 1352215827000|17, h: -5618036261007523082, v: 2, op: "u",
ns: "db1.someColl", o2: { _id: ObjectId('50992a2a7852201e750012b7') },
o: { $set: { count.0: 2, count.1: 0 } } }
Mon Jan 30 15:38:37 [repl writer worker 2] replication info
adding missing object
Mon Jan 30 15:38:37 [repl writer worker 2] replication missing object
not found on source. presumably deleted later in oplog
This is roughly what the logs will look like if some documents had to be recloned.
Depending on the level of traffic and the types of operations that where happening
on the sync source, you may or may not have missing objects.
4. Then the second oplog application occurs, which applies operations that happened
during the first oplog application:
Mon Jan 30 15:39:41 [rsSync] oplog sync 2 of 3
This one generally passes without much fanfare. It is only distinct from the first
application in that there should no longer be anything to reclone.
5. At this point, the data should exactly match the data set as it existed at some point
on the primary so that the secondary can start building indexes. This can be quite
time-consuming if you have large collections or lots of indexes:
Mon Jan 30 15:39:43 [rsSync] replSet initial sync building indexes
Mon Jan 30 15:39:43 [rsSync] replSet initial sync cloning indexes for : db1
Mon Jan 30 15:39:43 [rsSync] build index db.allObjects { someColl: 1 }
Syncing | 189
Mon Jan 30 15:39:44 [rsSync] build index done. scanned 209844 total records.
1.96 secs
6. Then the final oplog application occurs; this final step is merely to prevent the
member from becoming a secondary while it is still far behind the sync source. It
applies all of the operations that happened while indexes were building:
Tue Nov 6 16:05:59 [rsSync] oplog sync 3 of 3
7. At this point, the member finishes the initial sync process and transitions to normal
syncing, which allows it to become a secondary:
Mon Jan 30 16:07:52 [rsSync] replSet initial sync done
Mon Jan 30 16:07:52 [rsSync] replSet syncing to: server-1:27017
Mon Jan 30 16:07:52 [rsSync] replSet SECONDARY
The best way to track an initial sync’s progress is by watching the server’s log.
Doing an initial sync is very easy from an operator perspective: start up a mongod with
a clean data directory. However, it is often preferable to restore from backup instead, as
covered in Chapter 22. Restoring from backup is often faster than copying all of your
data through mongod.
Also, cloning can ruin the sync source’s working set. Many deployments end up with a
subset of their data that’s frequently accessed and always in memory (because the OS is
accessing it frequently). Performing an initial sync forces the member to page all of its
data into memory, evicting the frequently-used data. This can slow down a member
dramatically as requests that were being handled by data in RAM are suddenly forced
to go to disk. However, for small data sets and servers with some breathing room, initial
syncing is a good, easy option.
One of the most common issues people run into with initial sync is when step 2 (cloning)
or step 5 (building indexes) takes too long. In these cases, the new member can “fall off”
the end of sync source’s oplog: the new member gets so far behind the sync source that
it can no longer catch up because the sync source’s oplog has overwritten the data the
member would need to use to continue replicating.
There is no way to fix this other than attempting the initial sync at a less-busy time or
restoring from a backup. The initial sync cannot proceed if the member has fallen off
of the sync source’s oplog. The next section covers this in more depth.
Handling Staleness
If a secondary falls too far behind the actual operations being performed on the sync
source, the secondary will go stale. A stale secondary is unable to continue catch up
because every operation in the sync source’s oplog is too far ahead: it would be skipping
operations if it continued to sync. This could happen if the slave has had downtime, has
more writes than it can handle, or is too busy handling reads.
190 | Chapter 10: Components of a Replica Set
When a secondary goes stale, it will attempt to replicate from each member of the set
in turn to see if there’s anyone with a longer oplog that it can bootstrap from. If there is
no one with a long-enough oplog, replication on that member will halt and it will need
to be fully resynced (or restored from a more recent backup).
To avoid out-of-sync secondaries, it’s important to have a large oplog so that the primary
can store a long history of operations. A larger oplog will obviously use more disk space.
But in general this is a good trade-off to make because the disk space tends to be cheap
and little of the oplog is usually in use, and therefore it doesn’t take up much RAM. For
more information on sizing the oplog, see “Resizing the Oplog” on page 220.
Heartbeats
Members need to know about the other members’ states: who’s primary, who they can
sync from, and who’s down. To keep an up-to-date view of the set a member sends out
a heartbeat request to every other member of the set every two seconds. A heartbeat
request is a short message that checks everyone’s state.
One of the most important functions of heartbeats is to let the primary know if it can
reach a majority of the set. If a primary can no longer reach a majority of the servers, it
will demote itself and become a secondary.
Member States
Members also communicate what state they are in via heartbeats. We’ve already dis‐
cussed two states: primary and secondary. There are several other normal states that
you’ll often see members be in:
STARTUP
This is the state MongoDB goes into when you first start a member. It’s the state
when MongoDB is attempting to load a member’s replica set configuration. Once
the configuration has been loaded, it transitions to STARTUP2.
STARTUP2
This state will last throughout the initial sync process but on a normal member, it
should only ever last a few seconds. It just forks off a couple of threads to handle
replication and elections and then transitions into the next state: RECOVERING.
RECOVERING
This state means that the member is operating correctly but is not available for
reads. This state is a bit overloaded: you may see it in a variety of situations.
On startup, a member has to make a couple checks to make sure it’s in a valid state
before accepting reads; therefore, all members will go through recovering state
briefly on startup before becoming secondaries. A member can also go into
RECOVERING state during long-running operations such as compact or in re‐
Heartbeats | 191
sponse to the replSetMaintenance command (see “Using Maintenance Mode” on
page 213).
A member will also go into RECOVERING state if it has fallen too far behind the
other members to catch up. This is, generally, a failure state that requires resyncing
the member. The member does not go into an error state at this point because it
lives in hope that someone will come online with a long-enough oplog that it can
bootstrap itself back to non-staleness.
ARBITER
Arbiters have a special state and should always be in state ARBITER during normal
operation.
There are also a few states that indicate a problem with the system. These include:
DOWN
If a member was up but then becomes unreachable. Note that a member reported
as “down” might, in fact, still be up, just unreachable due to network issues.
UNKNOWN
If a member has never been able to reach another member, it will not know what
state it’s in, so it will report it as unknown. This generally indicates that the unknown
member is down or that there are network problems between the two members.
REMOVED
This is the state of a member that has been removed from the set. If a removed
member is added back into the set, it will transition back into its “normal” state.
ROLLBACK
This state is used when a member is rolling back data, as described in “Rollbacks”
on page 193. At the end of the rollback process, a server will transition back into the
recovering state and then become a secondary.
FATAL
Something uncorrectable has gone wrong and this member has given up trying to
function normally. You should take a look at the log to figure out what has caused
it to go into this state (grep for "replSet FATAL" to find the point where it went into
the FATAL state). You generally will have to shut down the server and resync it or
restore from backup once it’s in this state.
Elections
A member will seek election if it cannot reach a primary (and is itself eligible to become
primary). A member seeking election will send out a notice to all of the members it can
reach. These members may know of reasons that this member is an unsuitable primary:
it may be behind in replication or there may already be a primary that the member
192 | Chapter 10: Components of a Replica Set
seeking election cannot reach. In these cases, the other members will not allow the
election to proceed.
Assuming that there is no reason to object, the other members will vote for the member
seeking election. If the member seeking election receives votes from a majority of the
set, the election was successful and will transition into primary state. If it did not receive
a majority if votes, it will remain a secondary and may try to become a primary again
later. A primary will remain primary until it cannot reach a majority of members, goes
down, is stepped down, or the set is reconfigured.
Assuming that the network is healthy and a majority of the servers are up, elections
should be fast. It will take a member up to two seconds to notice that a primary has gone
down (due to the heartbeats mentioned earlier) and it will immediately start an election,
which should only take a few milliseconds. However, the situation is often non-optimal:
an election may be triggered due to networking issues or overloaded servers responding
too slowly. In these cases, heartbeats will take up to 20 seconds to timeout. If, at that
point, the election results in a tie, everyone will have to wait 30 seconds to attempt
another election. Thus, if everything goes wrong, an election may take a few minutes.
Rollbacks
The election process described in the previous section means that if a primary does a
write and goes down before the secondaries have a chance to replicate it, the next pri‐
mary elected may not have the write. For example, suppose we have two data centers,
one with the primary and a secondary, and the other with three secondaries, as shown
in Figure 10-2.
Figure 10-2. A possible two-data-center configuration
Suppose that there is a network partition between the two data centers, as shown in
Figure 10-3. The servers in the first data center are up to operation 126, but that data
center hasn’t yet replicated to the servers in the other data center.
Rollbacks | 193
Figure 10-3. Replication across data centers can be slower than within a single data
center
The servers in the other data center can still reach a majority of the set (three out of five
servers). Thus, one of them may be elected primary. This new primary begins taking its
own writes, as shown in Figure 10-4.
Figure 10-4. Unreplicated writes won’t match writes on the other side of a network
partition
When the network is repaired, the servers in the first data center will look for operation
126 to start syncing from the other servers but will not be able to find it. When this
happens, A and B will begin a process called rollback. Rollback is used to undo ops that
were not replicated before failover. The servers with 126 in their oplogs will look back
through the oplogs of the servers in the other data center for a common point. They’ll
find that operation 125 is the latest operation that matches. Figure 10-5 shows what the
oplogs would look like.
194 | Chapter 10: Components of a Replica Set
Figure 10-5. Two members with conflicting oplogs: A apparently crashed before repli‐
cating ops 126−128, so these operations are not present on B, which has more recent
operations. A will have to rollback these three operations before resuming syncing.
At this point, the server will go through the ops it has and write its version of each
document affected by those ops to a .bson file in a rollback directory of your data di‐
rectory. Thus, if (for example) operation 126 was an update, it will write the document
updated by 126 to collectionName.bson. Then it will copy the version of that document
from the current primary.
The following is a paste of the log entries generated from a typical rollback:
Fri Oct 7 06:30:35 [rsSync] replSet syncing to: server-1
Fri Oct 7 06:30:35 [rsSync] replSet our last op time written: Oct 7
06:30:05:3
Fri Oct 7 06:30:35 [rsSync] replset source's GTE: Oct 7 06:30:31:1
Fri Oct 7 06:30:35 [rsSync] replSet rollback 0
Fri Oct 7 06:30:35 [rsSync] replSet ROLLBACK
Fri Oct 7 06:30:35 [rsSync] replSet rollback 1
Fri Oct 7 06:30:35 [rsSync] replSet rollback 2 FindCommonPoint
Fri Oct 7 06:30:35 [rsSync] replSet info rollback our last optime: Oct 7
06:30:05:3
Fri Oct 7 06:30:35 [rsSync] replSet info rollback their last optime: Oct 7
06:30:31:2
Fri Oct 7 06:30:35 [rsSync] replSet info rollback diff in end of log times:
-26 seconds
Fri Oct 7 06:30:35 [rsSync] replSet rollback found matching events at Oct 7
06:30:03:4118
Fri Oct 7 06:30:35 [rsSync] replSet rollback findcommonpoint scanned : 6
Fri Oct 7 06:30:35 [rsSync] replSet replSet rollback 3 fixup
Fri Oct 7 06:30:35 [rsSync] replSet rollback 3.5
Fri Oct 7 06:30:35 [rsSync] replSet rollback 4 n:3
Fri Oct 7 06:30:35 [rsSync] replSet minvalid=Oct 7 06:30:31 4e8ed4c7:2
Fri Oct 7 06:30:35 [rsSync] replSet rollback 4.6
Fri Oct 7 06:30:35 [rsSync] replSet rollback 4.7
Fri Oct 7 06:30:35 [rsSync] replSet rollback 5 d:6 u:0
Fri Oct 7 06:30:35 [rsSync] replSet rollback 6
Fri Oct 7 06:30:35 [rsSync] replSet rollback 7
Fri Oct 7 06:30:35 [rsSync] replSet rollback done
Fri Oct 7 06:30:35 [rsSync] replSet RECOVERING
Fri Oct 7 06:30:36 [rsSync] replSet syncing to: server-1
Fri Oct 7 06:30:36 [rsSync] replSet SECONDARY
Rollbacks | 195
The server begins syncing from another member (server-1, in this case) and realizes that
it cannot find its latest operation on the sync source. At that point, it starts the rollback
process by going into rollback state ("replSet ROLLBACK").
At step 2, it finds the common point between the two oplogs, which was 26 seconds ago.
It then begins undoing the operations from the last 26 seconds from its oplog. Once the
rollback is complete, it transitions into recovering state and begins syncing normally
again.
To apply operations that have been rolled back to the current primary, first use mongor
estore to load them into a temporary collection:
$ mongorestore --db stage --collection stuff \
> /data/db/rollback/important.stuff.2012-12-19T18-27-14.0.bson
Now you should examine the documents (using the shell) and compare them to the
current contents of the collection from whence they came. For example, if someone had
created a “normal” index on the rollback member and a unique index on current pri‐
mary, you’d want to make sure that there weren’t any duplicates in the rolled-back data
and resolve them if there were.
Once you have a version of the documents that you like in your staging collection, load
it into your main collection:
> staging.stuff.find().forEach(function(doc) {
... prod.stuff.insert(doc);
... })
If you have any insert-only collections, you can directly load the rollback documents
into the collection. However, if you are doing updates on the collection you will need
to be more careful about how you merge rollback data.
One often-misused member configuration option is the number of votes each member
has. Manipulating the number of votes is almost always not what you want and causes
a lot of rollbacks (which is why it was not included in the list of member properties in
the last chapter). Do not change the number of votes unless you are prepared to deal
with regular rollbacks.
For more information on preventing rollbacks, see Chapter 11.
196 | Chapter 10: Components of a Replica Set
When Rollbacks Fail
In some cases, MongoDB decides that the rollback is too large to undertake. Rollback
can fail if there are more than 300 MB of data or about 30 minutes of operations to roll
back. In these cases, you must resync the node that is stuck in rollback.
The most common cause of this is when secondaries are lagging and the primary goes
down. If one of the secondaries becomes primary, it will be missing a lot of operations
from the old primary. The best way to make sure you don’t get a member stuck in rollback
is to keep your secondaries as up to date as possible.
Rollbacks | 197
CHAPTER 11
Connecting to a Replica Set from
Your Application
This chapter covers how applications interact with replica sets, including:
• How connections and failovers work
• Waiting for replication on writes
•Routing reads to the correct member
Client-to-Replica-Set Connection Behavior
From an application’s point of view, a replica set behaves much like a standalone server.
By default, client libraries will connect to the primary and route all traffic to it. Your
application can perform reads and writes as though it were talking to a standalone server
while your replica set quietly keeps hot standbys ready in the background.
Connections to a replica set are similar to connections to a single server. Use the
MongoClient-equivalent in your driver and provide a list of seeds for the driver to con‐
nect to. Seeds are members of the replica set. You do not have to list all members (al‐
though you can): when the driver connects to the seeds, it will discover the other mem‐
bers from them. A connection string usually looks something like this:
"mongodb://server-1:27017,server-2:27017"
See your driver’s documentation for details.
When a primary goes down, the driver will automatically find the new primary (once
one is elected) and will route requests to it as soon as possible. However, while there is
no reachable primary your application will be unable to perform writes.
199
There may be no primary available for a brief time (during an election) or for an ex‐
tended period of time (if no reachable member can become primary). By default, the
driver will not service any requests—read or write—during this period. However, you
can optionally use secondaries for read requests.
A common desire is to have the driver hide the entire election process (the primary
going away and a new primary being elected) from the user. However, this is not possible
or desirable in many cases, so no driver handles failover this way. First, a driver can only
hide a lack of primary for so long: a set could exist forever with no primary. Second, a
driver often finds out that the primary went down because an operation failed, which
means that the driver doesn’t know whether or not the primary processed the operation
before going down. Thus, the driver leaves it to the user: Do you want to retry the
operation on the new primary, if one is elected quickly? Assume it got through on the
old primary? Check and see if the new primary has the operation? The strategy that
makes sense will depend on your application.
There isn’t a general way to discover whether an operation succeeded or not before a
server crashed, although depending on your application you might be able to craft a
custom solution. For example, if the driver just inserted the document {"_id" : 1}
and received an error that the primary crashed, when it reconnects to the newly elected
primary it could check if {"_id" : 1} exists or not.
Waiting for Replication on Writes
To ensure that writes will be persisted no matter what happens to the set, you must
ensure that the write propagates to a majority of the members of the set, as mentioned
in the previous chapter.
Earlier, we used the getLastError command to check that writes were successful. We can
use that same command to ensure that a write has been replicated to secondaries. The
“w” parameter forces getLastError to wait until the given number of members has the
last write. MongoDB has a special keyword that you can pass to "w" for this: "majori
ty". In the shell, this looks like:
> db.runCommand({"getLastError" : 1, "w" : "majority"})
{
"n" : 0,
"lastOp" : Timestamp(1346790783000, 1),
"connectionId" : 2,
"writtenTo" : [
{ "_id" : 0 , "host" : "server-0" },
{ "_id" : 1 , "host" : "server-1" },
{ "_id" : 3 , "host" : "server-3" }
],
"wtime" : 76,
"err" : null,
200 | Chapter 11: Connecting to a Replica Set from Your Application
"ok" : 1
}
Notice the new field in getLastError’s output, "writtenTo". It is only present if you use
the "w" option and it is a list of servers that the last operation was replicated to.
Suppose that we run the command above, but only the primary and an arbiter are up:
the primary cannot replicate the write to any other member of the set. getLastError
doesn’t know how long to wait for replication, so it will wait forever. Thus, you should
always set wtimeout to a reasonable value. "wtimeout" is another option to getLastEr‐
ror and specifies in milliseconds how long the command should wait before returning
failure: MongoDB was unable to replicate to "w" members in the time specified.
The code below will wait for one second before giving up:
> db.runCommand({"getLastError" : 1, "w" : "majority", "wtimeout" : 1000})
This may fail for a variety of reasons: the other members may be down or lagging or
unavailable due to network issues. If getLastError times out, your application has to
decide what to do next. Note that getLastError timing out doesn’t mean that the write
failed. It merely means that it failed to replicate far enough in the time specified. The
write is still present on any servers it made it to and will continue to propagate to the
other members of the set as quickly as possible.
A common use for "w" is to throttle writes. MongoDB lets you write “too fast.” It will
let you write to the primary so quickly that the secondaries cannot keep up. A common
way of preventing this is to periodically call getLastError with "w" set to something
higher than 1. This forces writes on that particular connection to wait for replication.
Note that it only blocks writes on that connection: writes can still occur on any other
connection.
If you wish to make your application behave sensibly and robustly, regularly call get‐
LastError with "majority" and a reasonable timeout. If this begins timing out, look
into what’s wrong with your set.
What Can Go Wrong?
Suppose your application sends a write to the primary. It calls getLastError (without the
"majority" option) and receives confirmation that the write was written, but the pri‐
mary crashes before any secondaries have had a chance to replicate that write.
Now your application thinks that it’ll be able to access that write (getLastError confirmed
that the write succeeded) and the current members of the replica set don’t have a copy
of it.
At some point, a secondary may be elected primary and start taking new writes. When
the former primary comes back up, it will discover that it has writes that the current
primary does not. To correct this, it will undo any writes that do not match the sequence
Waiting for Replication on Writes | 201
of operations on the current primary. These operations are not lost, but they are written
to special rollback files that have to be manually applied to the current primary. Mon‐
goDB cannot automatically apply these writes, since they may conflict with other writes
that have happened since the crash. Thus, the write essentially disappears until an admin
gets a chance to apply the rollback files to the current primary.
Writing to a majority prevents this situation: if the application initially used "w" :
"majority" and gets a confirmation that the write succeeded, then the new primary
would have to have a copy of the write to be elected (a member must be up to date to
be elected primary). If getLastError failed, then the application would know to try again,
given that the write had not been propagated to a majority of the set before the primary
crashed.
See Chapter 10 for more details on rollbacks.
Other Options for “w”
"majority" is not the only option that you can pass to getLastError, MongoDB also lets
you specify an arbitrary number of servers to replicate to by passing "w" a number, as
below:
> db.runCommand({"getLastError" : 1, "w" : 2, "wtimeout" : 500})
This would wait until two members (the primary and one secondary) had the write.
Note that the “w” value includes the primary. If you want the write propagated to n
secondaries, you should set "w" to n+1 (to include the primary). Setting "w" : 1 is the
same as not passing the "w" option at all because it just checks that the write was suc‐
cessful on the primary, which is what getLastError does anyway.
The downside to using a literal number is that you have to change your application if
your replica set configuration changes.
Custom Replication Guarantees
Writing to a majority of a set is considered “safe.” However, some sets may have more
complex requirements: you may want to make sure that a write makes it to at least one
server in each data center or a majority of the nonhidden nodes. Replica sets allows you
to create custom rules that you can pass to getLastError to guarantee replication to
whatever combination of servers you need.
Guaranteeing One Server per Data Center
Network issues between data centers are much more common than within data centers
and it is more likely for an entire data center to go dark than an equivalent smattering
of servers across multiple data centers. Thus, you might want some data-center-specific
202 | Chapter 11: Connecting to a Replica Set from Your Application
logic for writes. Guaranteeing a write to every data center before confirming success
means that, in the case of a write followed by the data center going offline, every other
data center will have at least one local copy.
To set this up, first classify the members by data center. You do this by adding a "tags"
field to their replica set configuration:
> var config = rs.config()
> config.members[0].tags = {"dc" : "us-east"}
> config.members[1].tags = {"dc" : "us-east"}
> config.members[2].tags = {"dc" : "us-east"}
> config.members[3].tags = {"dc" : "us-east"}
> config.members[4].tags = {"dc" : "us-west"}
> config.members[5].tags = {"dc" : "us-west"}
> config.members[6].tags = {"dc" : "us-west"}
The "tags" field is an object, as each member can have multiple tags. It might be a
"high quality" server in the "us-east" data center, for example, in which case we’d want
a tags field such as {"dc": "us-east", "quality" : "high"}.
The second step is to add a rule by creating a "getLastErrorMode" field in our replica
set config. Each rule is of the form "name" : {"key" : number}}. "name" is the name
for the rule, which should describe what the rule does in a way that clients can under‐
stand, as they’ll be using this name when they call getLastError. In this example, we
might call this rule "eachDC" or something more abstract such as "user-level safe".
The "key" field is the key field from the tags, so in this example it will be "dc". The
number is the number groups that are needed to fulfil this rule. In this case, number is 2
(because we want at least one server from "us-east" and one from "us-west"). num
ber always means “at least one server from each of number groups.”
We add "getLastErrorModes" to the replica set config and reconfigure to create the
rule:
> config.settings = {}
> config.settings.getLastErrorModes = [{"eachDC" : {"dc" : 2}}]
> rs.reconfig(config)
"getLastErrorModes" lives in the "settings" subobject of a replica set config, which
contains a few set-level optional settings.
Now we can use this rule for writes:
> db.foo.insert({"x" : 1})
> db.runCommand({"getLastError" : 1, "w" : "eachDC", "wtimeout" : 1000})
Note that rules are somewhat abstracted away from the application developer: they don’t
have to know which servers are in “eachDC” to use the rule, and the rule can change
without their application having to change. We could add a datacenter or change set
members and the application would not have to know.
Custom Replication Guarantees | 203
Guaranteeing a Majority of Nonhidden Members
Often, hidden members are somewhat second-class citizens: you’re never going to fail
over to them and they certainly aren’t taking any reads. Thus, you may only care that
nonhidden members received a write and let the hidden members sort it out for
themselves.
Suppose we have five members, host0 through host4, host4 being a hidden member. We
want to make sure that a majority of the nonhidden members have a write, that is, at
least three of host0, host1, host2, and host3. To create a rule for this, first we tag each of
the nonhidden members with its own tag:
> var config = rs.config()
> config.members[0].tags = [{"normal" : "A"}]
> config.members[1].tags = [{"normal" : "B"}]
> config.members[2].tags = [{"normal" : "C"}]
> config.members[3].tags = [{"normal" : "D"}]
The hidden member, host4, is not given a tag.
Now we add a rule for the majority of these servers:
> config.settings.getLastErrorModes = [{"visibleMajority" : {"normal" : 3}}]
> rs.reconfig(config)
Finally, you can use this rule in your application:
> db.foo.insert({"x" : 1})
> db.runCommand({"getLastError" : 1, "w" : "visibleMajority", "wtimeout": 1000})
This will wait until at least three of the nonhidden member have the write.
Creating Other Guarantees
The rules you can create are limitless. Remember that there are two steps to creating a
custom replication rule:
1. Tag members by assigning them key-value pairs. The keys describe classifications;
for example, you might have keys such as "data_center" or "region" or "server
Quality". Values determine which group a server belongs to within a classification.
For example, for the key "data_center", you might have some servers tagged "us-
east", some "us-west", and others "aust".
2. Create a rule based on the classifications you create. Rules are always of the form
{"name" : {"key" : number}}, where at least one server from number groups must
have a write before it has succeeded. For example, you could create a rule
{"twoDCs" : {"data_center" : 2}}, which would mean that at least one server
in two of the data centers tagged must confirm a write before it is successful.
Then you can use this rule in getLastError.
204 | Chapter 11: Connecting to a Replica Set from Your Application
Rules are immensely powerful ways to configure replication, although they are complex
to understand and set up. Unless you have fairly involved replication requirements, you
should be perfectly safe sticking with "w" : "majority".
Sending Reads to Secondaries
By default, drivers will route all requests to the primary. This is generally what you want,
but you can configure other options by setting read preferences in your driver. Read
preferences let you specify the types of servers queries should be sent to.
Sending read requests to secondaries is generally a bad idea. There are some specific
situations in which it makes sense, but you should generally send all traffic to the pri‐
mary. If you are considering sending reads to secondaries, make sure to weigh the pros
and cons very carefully before allowing it. This section covers why it’s a bad idea and
the specific conditions when it makes sense to do so.
Consistency Considerations
Applications that require strongly consistent reads should not read from secondaries.
Secondaries should usually be within a few milliseconds of the primary. However, there
is no guarantee of this. Sometimes secondaries can fall behind by minutes, hours, or
even days due to load, misconfiguration, network errors, or other issues. Client libraries
cannot tell how up to date a secondary is, so clients will cheerfully send queries to
secondaries that are far behind. Hiding a secondary from client reads can be done but
is a manual process. Thus, if your application needs data that is predictably up to date,
it should not read from secondaries.
If your application needs to read its own writes (e.g., insert a document and then query
for it and find it) you should not send the read to a secondary (unless the write waits
for replication to all secondaries using "w" as shown earlier). Otherwise, an application
may perform a successful write, attempt to read the value, and not be able to find it
(because it sent the read to a secondary, which hadn’t replicated yet). Clients can issue
requests faster than replication can copy operations.
To always send read requests to the primary, set your read preference to Primary (or
leave it alone, since Primary is the default). If there is no primary, queries will error out.
This means that your application cannot perform queries if the primary goes down.
However, it is certainly an acceptable option if your application can deal with downtime
during failovers or network partitions or if getting stale data is unacceptable.
Load Considerations
Many users send reads to secondaries to distribute load. For example, if your servers
can only handle 10,000 queries a second and you need to handle 30,000, you might set
Sending Reads to Secondaries | 205
up a couple of secondaries and have them take some of the load. However, this is a
dangerous way to scale because it’s easy to accidentally overload your system and dif‐
ficult to recover from once you do.
For example, suppose that you have the situation above: 30,000 reads per second. You
decide to create a replica set with four members to handle this: each secondary is well
below it’s maximum load and the system works perfectly.
Until one of the secondaries crashes.
Now each of the remaining members are handling 100% of their possible load. If you
need to rebuild the member that crashed, it may need to copy data from one of the other
servers, overwhelming the remaining servers. Overloading a server often makes it per‐
form slower, lowering the set’s capacity even further and forcing other members to take
more load, causing them to slow down in a death spiral.
Overloading can also cause replication to slow down, making the remaining secondaries
fall behind. Suddenly you have a member down, a member lagging, and everything is
too overloaded to have any wiggle room.
If you have a good idea of how much load a server can take, you might feel like you can
plan this out better: use five servers instead of four and the set won’t be overloaded if
one goes down. However, even if you plan it out perfectly (and only lose the number of
servers you expected), you still have to fix the situation with the other servers under
more stress than they would be otherwise.
A better choice is to use sharding to distribute load. We’ll cover how to set sharding up
in Chapter 13.
Reasons to Read from Secondaries
There are a few cases in which it’s reasonable to send application reads to secondaries.
For instance, you may want your application to still be able to perform reads if the
primary goes down (and you do not care if those reads are somewhat stale). The is the
most common case for distributing reads to secondaries: you’d like a temporary read-
only mode when your set loses a primary. This read preference is called Primary pre
ferred.
One common argument for reading from secondaries is to get low-latency reads. You
can specify Nearest as your read preference to route requests to the lowest-latency
member based on average ping time from the driver to the replica set member. If your
application needs to access the same document with low latency in multiple data centers,
this is the only way to do it. If, however, your documents are more location-based (ap‐
plication servers in this data center need low-latency access to some of your data, or
application servers in another data center need low-latency access to other data), this
should be done with sharding. Note that you must use sharding if your application
206 | Chapter 11: Connecting to a Replica Set from Your Application
requires low-latency reads and low-latency writes: replica sets only allow writes to one
location (wherever the primary is).
You must be willing to sacrifice consistency if you are reading from members that may
not have replicated all the writes yet. Alternatively, you could sacrifice write speed if you
wanted to wait until writes had been replicated to all members.
If your application can truly function acceptably with arbitrarily stale data, you can use
Secondary or Secondary preferred read preferences. Secondary will always send read
requests to a secondary. If there are no secondaries available, this will error out rather
than send reads to the primary. It can be used for applications that do not care about
stale data and want to use the primary for writes only. If you have any concerns about
staleness of data, this is not recommended.
Secondary preferred will send read requests to a secondary, if one is available. If no
secondaries are available, requests will be sent to the primary.
Sometimes, read load is drastically different than write load: you’re reading entirely
different data than you’re writing. You might want dozens of indexes for offline pro‐
cessing that you don’t want to have on the primary. In this case, you might want to set
up a secondary with different indexes than the primary. If you’d like to use a secondary
for this purpose, you’d probably create a connection directly to it from the driver, instead
of using a replica set connection.
Consider which of the options makes sense for your application. You can also combine
options: if some read requests must be from the primary, use Primary for those. If you
are OK with other reads not having the most up-to-date data, use Primary prefer
red for those. And if certain requests require low latency over consistency, use Near
est for those.
Sending Reads to Secondaries | 207
CHAPTER 12
Administration
This chapter covers replica set administration, including:
• Techniques for performing maintenance on individual members
• Configuring sets under a variety of circumstances
• Getting information about and resizing your oplog
•Doing some more exotic set configurations
•Converting from master-slave to a replica set
Starting Members in Standalone Mode
A lot of maintenance tasks cannot be performed on secondaries (because they involve
writes) and shouldn’t be performed on primaries. Thus, the following sections fre‐
quently mention starting up a server in standalone mode. This means restarting the
member so that it is a standalone server, not a member of a replica set (temporarily).
To start up a member in standalone mode, first look at the command line argument.
Suppose it looks something like this:
> db.serverCmdLineOpts()
{
"argv" : [ "mongod", "-f", "/var/lib/mongod.conf" ],
"parsed" : {
"replSet": "mySet",
"port": "27017",
"dbpath": "/var/lib/db"
},
"ok" : 1
}
209
To perform maintenance on this server we can restart it without the replSet option.
This will allow us to read and write to it as a normal standalone mongod. We don’t want
the other servers in the set to be able to contact it, so we’ll make it listen on a different
port (so that the other members won’t be able to find it). Finally, we want to keep the
dbpath the same, as we are presumably starting it up this way to manipulate the server’s
data somehow. Thus, we start up this server with the following arguments:
$ mongod --port 30000 --dbpath /var/lib/db
It will now be a running as a standalone server, listening on port 30000 for connections.
The other members of the set will attempt to connect to it on port 27017 and assume
that it is down.
When we have finished performing maintenance on the server, we can shut it down and
restart it with its original options. It will automatically sync up with the rest of the set,
replicating any operations that it missed while it was “away.”
Replica Set Configuration
Replica set configuration is always kept in a document in the local.system.replset col‐
lection. This document is the same on all members of the set. Never update this docu‐
ment using update. Always use an rs helper or the replSetReconfig command.
Creating a Replica Set
You create a replica set by starting up the mongods that you want to be members and
then passing one of them a configuration through rs.initiate:
> var config = {
... "_id" : setName,
... "members" : [
... {"_id" : 0, "host" : host1},
... {"_id" : 1, "host" : host2},
... {"_id" : 2, "host" : host3}
... ]}
> rs.initiate(config)
You should always pass a config object to rs.initiate. If you do not, MongoDB will
attempt to automatically generate a config for a one-member replica set. It may not use
the hostname that you want or correctly configure the set.
You only call rs.initiate on one member of the set. The member that receives the
initiate will pass the configuration on to the other members.
210 | Chapter 12: Administration
Changing Set Members
When you add a new set member, it should either have nothing in its data directory (in
which case it will initial sync) or have a copy of the data from another member. See
Chapter 22 for more information about backing up and restoring replica set members.
Connect to the primary and add the new member:
> rs.add("spock:27017")
Alternatively, you can specify a more complex member config as a document:
> rs.add({"_id" : 5, "host" : "spock:27017", "priority" : 0, "hidden" : true})
You can also remove members by their "host" field:
> rs.remove("spock:27017")
You can change a member’s settings by reconfiguring. There are a few restrictions in
changing a member’s settings:
• You cannot change a member’s "_id".
•You cannot make the member you’re sending the reconfig to (generally the primary)
priority 0.
•You cannot turn an arbiter into a nonarbiter and visa versa.
•You cannot change a member with "buildIndexes" : false to "buildIndexes" :
true.
Notably, you can change a member’s "host" field. Thus, if you incorrectly specify a host
(say, use a public IP instead of a private one) you can later go back and simply change
the config to use the correct IP.
To change a hostname, you could do something like this:
> var config = rs.config()
> config.members[0].host = "spock:27017"
spock:27017
> rs.reconfig(config)
This same strategy applies to change any other option: fetch the config with rs.con
fig(), modify any parts of it that you wish, and reconfigure the set by passing rs.re
config() the new configuration.
Creating Larger Sets
Replica sets are limited to 12 members and only 7 voting members. This is to reduce
the amount of network traffic required for everyone to heartbeat everyone else and to
limit the amount of time elections take. However, it can be too restrictive: see “Master-
Slave” on page 225 if you require more that 11 secondaries.
Replica Set Configuration | 211
If you are creating a replica set that has more than 7 members, every additional member
must be given 0 votes. You can do this by specifying it in the member’s config:
> rs.add({"_id" : 7, "host" : "server-7:27017", "votes" : 0})
This prevents these members from casting positive votes in elections, although they can
still veto.
Please do not alter votes if you can possibly avoid it. Votes have weird, non-intuitive
implications for elections and consistency guarantees. Only use votes if you are creating
a set with more than seven members or you want to prevent automatic failover, as
described in “Mimicking Master-Slave Behavior with Replica Sets” on page 226. Often,
developers mistakenly think that a member having more votes will make it more likely
to be primary (which it won’t). If you wish a member to be preferentially chosen as
primary, use priorities (see “Priority” on page 183).
Forcing Reconfiguration
When you permanently lose a majority of the set, you may want to reconfigure the set
while it doesn’t have a primary. This is a little tricky: usually you’d send the reconfig to
the primary. In this case, you can force reconfigure the set by sending a reconfig com‐
mand to a secondary. Connect to a secondary in the shell and pass it a reconfig with the
"force" option:
> rs.reconfig(config, {"force" : true})
Forced reconfigurations follow the same rules as a normal reconfiguration: you must
send a valid, well-formed configuration with the correct options. The "force" option
doesn’t allow invalid configs; it just allows a secondary to accept a reconfig.
Forced reconfigurations bump the replica set "version" number by a large amount.
You may see it jump by thousands. This is normal: it is to prevent version number
collisions (just in case there’s a reconfig on either side of a network partition).
When the secondary receives the reconfig, it will update its configuration and pass the
new config along to the other members. The other members of the set will only pick up
on a change of config if they recognize the sending server as a member of their current
config. Thus, if some of your members have changed hostnames, you should force
reconfig from a member that kept its old hostname. If everyone has a new hostname,
you should shut down each member of the set, start it up in standalone mode, change
its local.system.replset document manually, and then restart the member.
212 | Chapter 12: Administration
Manipulating Member State
There are several ways to manually change member state for maintenance or in response
to load. Note that there is no way to force a member to become primary other than
configuring the set appropriately.
Turning Primaries into Secondaries
You can demote a primary to a secondary using the stepDown function:
> rs.stepDown()
This makes the primary step down into secondary state for 60 seconds. If no other
primary has been elected in that time period, it will be able to attempt a reelection. If
you would like it to remain a secondary for a longer or shorter amount of time, you can
specify your own number of seconds for it to stay in SECONDARY state:
> rs.stepDown(600) // 10 minutes
Preventing Elections
If you need to do some maintenance on the primary but don’t want any of the other
eligible members to become primary in the interim, you can force them to stay secon‐
daries by running freeze on each of them:
> rs.freeze(10000)
Again, this takes a number of seconds to remain secondary.
When you have finished whatever maintenance you are doing and want to “unfreeze”
the other members you can run the command again, giving a timeout of 0 seconds:
> rs.freeze(0)
This will allow the member to hold an election, if it chooses.
You can also unfreeze primaries that have been stepped down by running rs.freeze(0).
Using Maintenance Mode
Maintenance mode occurs when you perform a long-running op on a replica set mem‐
ber: it forces the member into RECOVERING state. Sometimes, a member will go into
maintenance mode automatically, for example, if you run a compact on it. When the
compact begins, the member will go into RECOVERING state so that reads will no
longer go to that member. Clients will stop using it for reads (if they were) and it should
no longer be used as a replication source.
You can force a member to go into maintenance mode by running the replSetMainte‐
nanceMode command on it. You might want to do this if a member begins to fall behind
Manipulating Member State | 213
and you don’t want any read load on it. For example, you could have a script like this,
which detects if a member is behind and then puts it in maintenance mode:
function maybeMaintenanceMode() {
var local = db.getSisterDB("local");
// Return if this member isn't a secondary (it might be a primary
// or already in recovering)
if (!local.isMaster().secondary) {
return;
}
// Find the last optime written on this member
var last = local.oplog.rs.find().sort({"$natural" : -1}).next();
var lastTime = last['ts']['t'];
// If more than 30 seconds behind
if (lastTime < (new Date()).getTime()-30) {
db.adminCommand({"replSetMaintenanceMode" : true});
}
};
To get out of maintenance mode, pass the command false:
> db.adminCommand({"replSetMaintenanceMode" : false});
Monitoring Replication
It is important to be able to monitor the status of a set: not only that everyone is up, but
what states they are in and how up-to-date the replication is. There are several com‐
mands you can use to see replica set information. MMS (see Chapter 21) also keeps
some useful stats on replication.
Often issues with replication are transient: a server could not reach another server but
now it can. The easiest way to see issues like this is to look at the logs. Make sure you
know where the logs are being stored (and that they are being stored) and that you can
access them.
Getting the Status
One of the most useful commands you can run is replSetGetStatus, which gets the cur‐
rent information about every member of the set (from the of view of the member you’re
running it on). There is a helper for this command in the shell:
> rs.status()
{
"set" : "spock",
"date" : ISODate("2012-10-17T18:17:52Z"),
"myState" : 2,
"syncingTo" : "server-1:27017",
214 | Chapter 12: Administration
"members" : [
{
"_id" : 0,
"name" : "server-1:27017",
"health" : 1,
"state" : 1,
"stateStr" : "PRIMARY",
"uptime" : 74824,
"optime" : { "t" : 1350496621000, "i" : 1 },
"optimeDate" : ISODate("2012-10-17T17:57:01Z"),
"lastHeartbeat" : ISODate("2012-10-17T17:57:00Z"),
"pingMs" : 3,
},
{
"_id" : 1,
"name" : "server-2:27017",
"health" : 1,
"state" : 2,
"stateStr" : "SECONDARY",
"uptime" : 161989,
"optime" : { "t" : 1350377549000, "i" : 500 },
"optimeDate" : ISODate("2012-10-17T17:57:00Z"),
"self" : true
},
{
"_id" : 2,
"name" : "server-3:27017",
"health" : 1,
"state" : 3,
"stateStr" : "RECOVERING",
"uptime" : 24300,
"optime" : { "t" : 1350411407000, "i" : 739 },
"optimeDate" : ISODate("2012-10-16T18:16:47Z"),
"lastHeartbeat" : ISODate("2012-10-17T17:57:01Z"),
"pingMs" : 12,
"errmsg" : "still syncing, not yet to minValid optime 507e9a30:851"
}
],
"ok" : 1
}
These are some of the most useful fields:
self
This field is only present in the member rs.status() was run on, in this case,
server-2.
stateStr
A string describing the state of the server. See “Member States” on page 191 to see
descriptions of the various states.
Monitoring Replication | 215
uptime
The number of seconds a member has been reachable (or the time since this server
was started for the "self" member). Thus, server-2 has been up for 161989 seconds,
or about 45 hours. server-1 has been available for the last 21 hours and server-3 has
been available for the last 7 hours.
optimeDate
The last optime in each member’s oplog (where that member is synced to). Note
that this is the state of each member as reported by the heartbeat, so the optime
reported here may be off by a couple of seconds.
lastHeartbeat
The time this server last received a heartbeat from the member. If there have been
network issues or the server has been busy, this may be longer than two seconds
ago.
pingMs
The running average of how long heartbeats to this server have taken. This is used
in determining which member to sync from.
errmsg
Any status message that the member chose to return in the heartbeat request. These
are often merely informational, not error messages. For example, the "errmsg" field
in server-3 indicates that this server is in the process of initial syncing. The hexa‐
decimal number 507e9a30:851 is the timestamp of the operation this member needs
to get to to complete the initial sync.
There are a several fields that give overlapping information: "state" is the same as
"stateStr", it’s simply the internal id for the state. "health" merely reflects whether a
given server is reachable (1) or unreachable (0), which is also shown by "state" and
"stateStr" (they’ll be UNKNOWN or DOWN if the server is unreachable). Similarly,
"optime" and "optimeDate" are the same value represented in two ways: one represents
milliseconds since the epoch ("t" : 135...) and the other is a more human-readable
date.
Note that this is report is from the point of view of whichever member of the set you
run it on: it may be incorrect or out of date due to network issues.
Visualizing the Replication Graph
If you run rs.status() on a secondary, there will be a top-level field called "syncing
To". This gives the host that this member is replicating from. By running the replSet‐
GetStatus command on each member of the set, you can figure out the replication graph.
For example, assuming server1 was a connection to server1, server2 was a connection
to server2, and so on, you might have something like:
216 | Chapter 12: Administration
> server1.adminCommand({replSetGetStatus: 1})['syncingTo']
server0:27017
> server2.adminCommand({replSetGetStatus: 1})['syncingTo']
server1:27017
> server3.adminCommand({replSetGetStatus: 1})['syncingTo']
server1:27017
> server4.adminCommand({replSetGetStatus: 1})['syncingTo']
server2:27017
Thus, server0 is the source for server1, server1 is the replication source for server2 and
server3, and server2 is the replication source for server4.
MongoDB determines who to sync to based on ping time. When one member heartbeats
another, it times how long that request took. MongoDB keeps a running average of these
times. When it has to choose a member to sync from, it looks for the member that is
closest to it and ahead of it in replication (thus, you cannot end up with a replication
cycle: members will only replicate from the primary or secondaries that are strictly
further ahead).
Thus, if you bring up a new member in a secondary data center, it is more likely to sync
from the other members in that data center than a member in your primary data center
(thus minimizing WAN traffic), as show in Figure 12-1.
Figure 12-1. New secondaries will generally choose to sync from a member in the same
data center
However, there are some downsides to automatic replication chaining: more replication
hops mean that it takes a bit longer to replicate writes to all servers. For example, let’s
say that everything is in one data center but, due to the vagaries of network speeds when
you added members, MongoDB ends up replicating in a line, as shown in Figure 12-2.
Monitoring Replication | 217
Figure 12-2. As replication chains get longer, it takes longer for all members to get a
copy of the data
This is highly unlikely, but not impossible. It is, however, probably undesirable: each
secondary in the chain will have to be a bit further behind than the secondary “in front”
of it. You can fix this by modifying the replication source for a member using the
replSetSyncFrom command (or the rs.syncFrom() helper).
Connect to the secondary whose replication source you want to change and run this
command, passing it the server you’d prefer this member to sync from:
> secondary.adminCommand({"replSetSyncFrom" : "server0:27017"})
It may take a few seconds to switch sync sources, but if you run rs.status() on that
member again, you should see that the "syncingTo" field now says "server0:27017".
At this point, server4 will continue replicating from server0 until server0 becomes un‐
available or, if it happened to be a secondary, falls significantly behind the other
members.
Replication Loops
A replication loop is when members end up replicating from one another, for example,
A is syncing from B who is syncing from C who is syncing from A. As none of the
members in a replication loop can be a primary, the members will not receive any new
operations to replicate and fall behind. On the plus side, replication loops should be
impossible when members choose who to sync from automatically.
However, you can force replication loops using the replSetSyncFrom command. Inspect
the rs.status() output careful before manually changing sync targets and be careful
not to create loops. The replSetSyncFrom command will warn you if you do not choose
to sync from a member who is strictly ahead, but it will allow it.
Disabling Chaining
Chaining is when a secondary syncs from another secondary (instead of the primary).
As mentioned earlier, members may decide to sync from other members automatically.
218 | Chapter 12: Administration
You can disable chaining, forcing everyone to sync from the primary, by changing the
"allowChaining" setting to false (if not specified, it defaults to true):
> var config = rs.config()
> // create the settings subobject, if it does not already exist
> config.settings = config.settings || {}
> config.settings.allowChaining = false
> rs.reconfig(config)
With allowChaining set to false, all members will sync from the primary. If the primary
becomes unavailable, they will fall back to syncing from secondaries.
Calculating Lag
One of the most important metrics to track for replication is how well the secondaries
are keeping up with the primary. Lag is how far behind a secondary is, which means the
difference in timestamp between the last operation the primary has performed and the
timestamp of the last operation the secondary has applied.
You can use rs.status() to see a member’s replication state, but you can also get a
quick summary (along with oplog size) by running db.printReplicationInfo() (on
a primary) and db.printSlaveReplicationInfo() on a secondary. Note that these are
both functions of db, not rs.
db.printReplicationInfo gives a summary of the primary’s oplog:
> db.printReplicationInfo();
configured oplog size: 10.48576MB
log length start to end: 34secs (0.01hrs)
oplog first event time: Tue Mar 30 2010 16:42:57 GMT-0400 (EDT)
oplog last event time: Tue Mar 30 2010 16:43:31 GMT-0400 (EDT)
now: Tue Mar 30 2010 16:43:37 GMT-0400 (EDT)
This gives information about the size of the oplog and the date ranges of operations
contained in the oplog. In this example, the oplog is about 10 MB and is only able to fit
about 30 seconds of operations.
If this were a real deployment, the oplog should be much larger (see “Resizing the
Oplog” on page 220 for instructions on changing oplog size). We want the log length to
be at least as long as the time it takes to do a full resync. That way, we don’t run into a
case where a secondary falls off the end of the oplog before finishing its initial sync.
Monitoring Replication | 219
The log length is computed by taking the time difference between the
first and last operation in the oplog once the oplog has filled up. If the
server has just started with nothing in the oplog, then the earliest op‐
eration will be relatively recent. In that case, the log length will be small,
even though the oplog probably still has free space available. The length
is a more useful metric for servers that have been operating long enough
to write through their entire oplog at least once.
You can also use the db.printSlaveReplicationInfo() function on a secondary to get
information about who it is syncing from and how far behind it is:
> db.printSlaveReplicationInfo();
source: server-0:27017
syncedTo: Tue Mar 30 2012 16:44:01 GMT-0400 (EDT)
= 12secs ago (0hrs)
This will show who the slave is syncing from. In this case, the secondary is 12 seconds
behind the primary.
Remember that a replica set member’s lag is calculated relative to the primary, not
against “wall time.” This usually is irrelevant, but on very low-write systems, this can
cause phantom replication lag “spikes.” For example, suppose you do a write once an
hour. Right after that write, before it’s replicated, the secondary will look like it’s an hour
behind the primary. However, it’ll be able to catch up with that “hour” of operations in
a few milliseconds. This can sometimes cause confusion when monitoring a low-
throughput system.
Resizing the Oplog
Your primary’s oplog should be thought of as your maintenance window. If your primary
has an oplog that is an hour long, then you only have one hour to fix anything that goes
wrong before your secondaries fall too far behind and must be resynced from scratch.
Thus, you generally want to have an oplog that can hold a couple days to a week’s worth
of data, to give yourself some breathing room if something goes wrong.
Unfortunately, there’s no easy way to tell how long your oplog is going to be before it
fills up and there’s no way to resize it while your server is running. However, it is possible
to cycle through you servers, taking each one offline, making its oplog larger, and then
adding it back into the set. Remember that each server that could become a primary
should have a large enough oplog to give you a sane maintenance window.
To increase the size of your oplog, perform the following steps:
1. If this is currently the primary, step it down and wait for the other servers to catch
up.
220 | Chapter 12: Administration
2. Shut down the server.
3. Start it up as a standalone server.
4. Temporarily store the last insert in the oplog in another collection:
> use local
> // op: "i" finds the last insert
> var cursor = db.oplog.rs.find({"op" : "i"})
> var lastInsert = cursor.sort({"$natural" : -1}).limit(1).next()
> db.tempLastOp.save(lastInsert)
>
> // make sure it was saved! It's very important that you don't lose this op
> db.tempLastOp.findOne()
We could use the last update or delete, but $-operators cannot be inserted into a
collection.
5. Drop the current oplog:
> db.oplog.rs.drop()
6. Create a new oplog:
> db.createCollection("oplog.rs", {"capped" : true, "size" : 10000})
7. Put the last op back in the oplog:
> var temp = db.tempLastOp.findOne()
> db.oplog.rs.insert(temp)
>
> // make sure that this was actually inserted
> db.oplog.rs.findOne()
Make sure that the last op was inserted into the oplog. If it was not, the server will
drop all of its data and resync when you add it back into the set.
8. Finally, restart the server as a member of the replica set. Remember that it only has
one op in the oplog to start out with, so you won’t be able to see its true oplog length
(how long it is in time) for a while. Also, it won’t be a very good sync source if other
members are behind.
You generally should not decrease the size of your oplog: although it may be months
long, there is usually ample disk space for it and it does not use up any valuable resources
like RAM or CPU.
Restoring from a Delayed Secondary
Suppose someone accidentally drops a database but, luckily, you had a delayed secon‐
dary. Now you need to get rid of the data on the other members and use the delayed
slave as your definitive source of data. There are a couple of ways to do this.
This is the simplest way:
Monitoring Replication | 221
1. Shut down all the other members.
2. Delete all the data in their data directories. Make sure every member (other than
the delayed secondary) has an empty data directory.
3. Restart all the members. They will begin making a copy of the delayed secondary’s
data.
This is certainly easy, but your replica set will essentially be one rather overloaded sec‐
ondary for however long it takes the other members to initial sync.
The other option may or may not work better, depending on your amount of data:
1. Shut down all the members, including the delayed secondary.
2. Delete the data files from the non-delayed servers.
3. Copy the delayed secondary’s data files to the other servers.
4. Start up everyone.
Note that this will mean all the servers will have the same oplog size as the delayed
secondary, which may not be what you want.
Building Indexes
If you send an index build to the primary, the primary will build the index normally
and then the secondaries will build the index when they replicate the “build index”
operation. Although this is the easiest way to build an index, index builds are resource-
intensive operations that can make members unavailable. If all of your secondaries start
building an index at the same time, almost every member of your set will be offline until
the index build completes.
Therefore, you may want to build an index on one member at a time to minimize impact
on your application. To accomplish this, do the following:
1. Shut down a secondary.
2. Restart it as a standalone server.
3. Build the index on the standalone server.
4. When the index build is complete, restart the server as a member of the replica set.
5. Repeat steps 1 through 4 for each secondary in the replica set.
You should now have a set where every member other than the primary has the index
built. Now there are two options, and you should choose the one that will impact your
production system the least:
222 | Chapter 12: Administration
1. Build the index on the primary. If you have an “off” time when you have less traffic,
that would probably be a good time to build it. You also might want to modify read
preferences to temporarily shunt more load onto secondaries while the build is in
progress.
The primary will replicate the index build to the secondaries, but they will already
have the index so it will be a no-op for them.
2. Step down the primary, then follow steps 1 through 4 as outlined earlier. This re‐
quires a failover, but you will have a normally-functioning primary while the old
primary is building its index. After its index build is complete, you can reintroduce
it to the set.
Note that you could also use this technique to build different indexes on a secondary
than you have on the rest of the set. This could be useful for offline processing, but make
sure a member with different indexes can never become primary: its priority should
always be 0.
If you are building a unique index, make sure that the primary is not inserting duplicates
or that you build the index on the primary first. Otherwise, the primary could be in‐
serting duplicates that would then cause replication errors on secondaries. If this occurs,
the secondary will shut itself down. You will have to restart it as a stand alone, remove
the unique index, and restart it.
Replication on a Budget
If it is difficult get more than one high-quality server, consider getting a secondary server
that is strictly for disaster recovery, with less RAM, CPU, slower disk IO, etc. The good
server will always be your primary and the cheaper server will never handle any client
traffic (configure your clients to send all reads to the primary). Here are all the options
to set for the cheaper box:
"priority" : 0
You do not want this server to ever become primary.
"hidden" : true
You do not want clients ever sending reads to this secondary.
"buildIndexes" : false
This is optional, but it can decrease the load this server has to handle considerably.
If you ever need to restore from this server, you’ll need to rebuild indexes.
"votes" : 0
If you only have two machines, set the votes on this secondary to 0 so that the
primary can stay primary if this machine goes down. If you have a third server (even
just your application server), run an arbiter on that instead of setting votes to 0.
Monitoring Replication | 223
This will give you the safety and security of having a secondary without having to invest
in two high-performance servers.
How the Primary Tracks Lag
Each member that has ever been a sync source keeps a collection called local.slaves,
which holds information about which servers are syncing from it and how up to date
they are. When you run a query using w, MongoDB looks through this information to
decide if enough secondaries are up to date enough to proceed.
The local.slaves collection is actually an “echo” of an in-memory data structure, so it
may be a few seconds out of date:
> db.slaves.find()
{ "_id" : ObjectId("4c1287178e00e93d1858567c"), "host" : "10.4.1.100",
"ns" : "local.oplog.rs", "syncedTo" : { "t" : 1276282710000, "i" : 1 } }
{ "_id" : ObjectId("4c128730e6e5c3096f40e0de"), "host" : "10.4.1.101",
"ns" : "local.oplog.rs", "syncedTo" : { "t" : 1276282710000, "i" : 1 } }
The "_id" of each server is important: it is an identifier for the syncing member. You
can see what a member’s "_id" is by connecting to it and looking at the local.me col‐
lection:
> db.me.findOne()
{ "_id" : ObjectId("50e6edb517c789e46695212f"), "host" : "server-1" }
Occasionally, due to configuration issues, you may end up with multiple servers with
the same "_id". If this happens, only one will be able to report how far it has replicated
to the primary. This, in turn, can cause issues with your application (if you’re depending
on a certain number of servers reporting that they got a write) and sharding (migrates
cannot proceed until a majority of secondaries have replicated the migration). If mul‐
tiple machines have the same "_id", you can fix it by logging into each machine, drop‐
ping the local.me collection, and restarting the mongod. On startup, mongod will repo‐
pulate local.me with a new "_id".
If a server’s address changes, you may get errors in the logs about duplicate key excep‐
tions in the local database, given that the slave’s "_id" will be the same but its hostname
will have changed. If this occurs, you can drop the local.slaves collection and the errors
will stop (this is simpler than the previous case because you just need to clear the old
data, not resolve conflicting data).
The local.slaves collection is never cleaned up by mongod, so it may list servers that
haven’t used the member as a sync source in months or members that aren’t even part
of the set anymore. As the collection is merely a dumping ground for MongoDB to
report on replication status, there’s no harm in leaving old entries in there. However, if
you find the old entries confusing or cluttered, you can drop the whole collection. It
224 | Chapter 12: Administration
will be repopulated after a few seconds with updated entries from the servers that are
currently syncing from it.
If you have secondaries that are chained, you may notice that the primary has several
local.slaves documents for a certain server. This is because each secondary will “forward”
any replication requests it gets to its sync target so that the primary knows where each
secondary has synced to. These are called “ghost syncs” because the requests don’t ac‐
tually request any ops back; they just inform the primary of where the secondaries have
synced to.
The local database is used for replication information because it is not
replicated. Thus, if you have data that you want to be local to a certain
machine, you can load it into collections in the local database.
Master-Slave
MongoDB originally supported a more traditional master-slave setup (no automatic
failover and you declare who the master and slaves are). There are exactly two reasons
to consider using master-slave instead of replica sets: you need more than 11 secondaries
or you need to replicate individual databases. Unless your application unavoidably re‐
quires it, use replica sets. They are much better maintained and fully featured. Master-
slave will be deprecated at some point, probably once replica sets support unlimited
members.
However, there are still times when you may need more than 11 slaves or need to rep‐
licate a single database. In these cases, you still do need master-slave.
To set up a master, start your server with --master. To start a slave, use two options: --
slave and --source master. --source specifies the sync source: the host and port of
the master. Note that you do not use the --replSet option: you’re not setting up a replica
set.
For example, if you have two servers, server-0 and server-1, you can do:
$ # server-0
$ mongod --master
$
$ # server-1
$ mongod --slave --source server-0:27017
At that point, master-slave is set up, and there is no further configuration necessary. You
can begin writing to the master and the slave will replicate the changes.
Master-slave can also be used to replicate a single database. You can use --only option
to select a database to replicate:
Master-Slave | 225
$ mongod --slave --source server-1:27017 --only super-important-db
Drivers will not automatically distribute reads to slaves if you set a read preference. You
must make an explicit connection to a slave to read from it.
Converting Master-Slave to a Replica Set
Converting from a master-slave setup to replica sets requires some downtime. Here are
the steps:
1. Stop all writes on your system. This is very important, as your former slave will
briefly not have an oplog, so it won’t be able to catch up with any writes it misses
during the upgrade.
2. Shut down all mongod servers.
3. Restart the master with the --replSet option instead of --master.
4. Initialize the set with one member: the former master. It will become primary.
5. Start up the slaves with the --replSet option and the --fastsync option. Ordi‐
narily, if you added a member to the set without an oplog, it would go through the
full initial sync process. The fastsync option tells the member to not worry that it
doesn’t have an oplog, just start syncing from the latest time on the master.
6. Add the former slave to the set with rs.add().
7. Repeat steps 4 and 5 for each slave.
8. Once all slaves have been converted into secondaries, you can turn on writes to the
system.
9. Remove fastsync from your config file, command line alias, and long-term mem‐
ory. It is an extremely dangerous option to use because it makes the member skip
operations on restart. Its only use is to convert from master-slave to replica sets.
Now that you’ve converted, you no longer need it.
Your master-slave setup should now be a replica set.
Mimicking Master-Slave Behavior with Replica Sets
You usually want a primary to be available as often as possible, thus you should allow
automatic failover if the primary becomes unavailable. However, for some sets you may
want to require an operator to manually promote a new primary and never allow an
automatic failover. This makes replica sets behave the same way as master-slave does
(and is preferable to using master-slave to get this behavior).
To achieve this, reconfigure the set so that every member except the primary has a
priority of 0 and votes of 0. This way, no member will seek election if the primary
226 | Chapter 12: Administration
goes down. Also, the current primary can remain primary (as it has the only vote in the
system) even if all of the other members go down.
For example, this configuration would create a 5-member set where server-0 was always
primary and the other four members were always secondaries:
{
"_id" : "spock",
"members" : [
{"_id" : 0, "host" : "server-0:27017"},
{"_id" : 1, "host" : "server-1:27017", "priority" : 0, "votes" : 0},
{"_id" : 2, "host" : "server-2:27017", "priority" : 0, "votes" : 0},
{"_id" : 3, "host" : "server-3:27017", "priority" : 0, "votes" : 0},
{"_id" : 4, "host" : "server-4:27017", "priority" : 0, "votes" : 0}
]
}
If the primary goes down, an operator must manually intervene to select a new primary.
To manually promote a new primary, connect to the secondary that you would like to
become primary and run a forced reconfig, changing its priority and votes to 1 and
the former primary’s to 0.
For example, if server-0 went down, connect to the server that you would like to be the
new primary, say, server-1. Then change the config as follows:
> var config = rs.config()
> config.members[1].priority = 1
> config.members[1].votes = 1
> config.members[0].priority = 0
> config.members[0].votes = 0
> rs.reconfig(config, {"force" : true})
Now, if you run rs.config(), it should look like this:
> rs.config()
{
"_id" : "spock",
"version" : 3
"members" : [
{"_id" : 0, "host" : "server-0:27017", "priority" : 0, "votes" : 0},
{"_id" : 1, "host" : "server-1:27017"},
{"_id" : 2, "host" : "server-2:27017", "priority" : 0, "votes" : 0},
{"_id" : 3, "host" : "server-3:27017", "priority" : 0, "votes" : 0},
{"_id" : 4, "host" : "server-4:27017", "priority" : 0, "votes" : 0}
]
}
If the new primary fails, repeat the process.
Master-Slave | 227
PART IV
Sharding
CHAPTER 13
Introduction to Sharding
This chapter covers how to scale with MongoDB:
• What sharding is and the components of a cluster
• How to configure sharding
• The basics of how sharding interacts with your application
Introduction to Sharding
Sharding refers to the process of splitting data up across machines; the term partition‐
ing is also sometimes used to describe this concept. By putting a subset of data on each
machine, it becomes possible to store more data and handle more load without requiring
larger or more powerful machines, just a larger quantity of less-powerful machines.
Manual sharding can be done with almost any database software. Manual sharding is
when an application maintains connections to several different database servers, each
of which are completely independent. The application manages storing different data
on different servers and querying against the appropriate server to get data back. This
approach can work well but becomes difficult to maintain when adding or removing
nodes from the cluster or in the face of changing data distributions or load patterns.
MongoDB supports autosharding, which tries to both abstract the architecture away
from the application and simplify the administration of such a system. MongoDB allows
your application to ignore the fact that it isn’t talking to a standalone MongoDB server,
to some extent. On the operations side, MongoDB automates balancing data across
shards and makes it easier to add and remove capacity.
Sharding is the most difficult and complex way of configuring MongoDB, both from a
development and operational point of view. There are many components to configure
231
and monitor and data moves around the cluster automatically. You should be comfort‐
able with standalone servers and replica sets before attempting to deploy or use a shar‐
ded cluster.
Understanding the Components of a Cluster
MongoDB’s sharding allows you to create a cluster of many machines (shards) and break
up your collection across them, putting a subset of data on each shard. This allows your
application to grow beyond the resource limits of a standalone server or replica set.
Many people are confused about the difference between replication and
sharding. Remember that replication creates an exact copy of your data
on multiple servers, so every server is a mirror-image of every other
server. Conversely, every shard contains a different subset of data.
One of the goals of sharding is to make a cluster of 5, 10, or 1,000 machines look like a
single machine to your application. To hide these details from the application, we run
a routing process called mongos in front of the shards. This router keeps a “table of
contents” that tells it which shard contains which data. Applications can connect to this
router and issue requests normally, as shown in Figure 13-1. The router, knowing what
data is on which shard, is able to forward the requests to the appropriate shard(s). If
there are responses to the request, the router collects them, merges them, and sends
them back to the application. As far as the application knows, it’s connected to a stand-
alone mongod, as in Figure 13-2.
A One-Minute Test Setup
As in the replication section, we will start by setting up a quick cluster on a single
machine. First, start a mongo shell with the --nodb option:
$ mongo --nodb
To create a cluster, use the ShardingTest class:
> cluster = new ShardingTest({"shards" : 3, "chunksize" : 1})
The chunksize option is covered in Chapter 16. For now, simply set it to 1.
Running this command creates a new cluster with three shards (mongod processes)
running on ports 30000, 30001, and 30002. By default, ShardingTest starts a mongos
on port 30999. We will to connect to this mongos to play around with the cluster.
232 | Chapter 13: Introduction to Sharding
Figure 13-1. Sharded client connection
Figure 13-2. Nonsharded client connection
Your entire cluster will be dumping its logs to your current shell, so open up a second
shell and use that to connect to your cluster’s mongos:
> db = (new Mongo("localhost:30999")).getDB("test")
Now you are in the situation show in Figure 13-1: the shell is the client and is connected
to a mongos. You can start passing requests to the mongos and it’ll route them to the
A One-Minute Test Setup | 233
shards. You don’t really have to know anything about the shards, like how many their
are or what their addresses are. So long as there are some shards out there, you can pass
the requests to the mongos and allow it to forward them appropriately.
Start by inserting some data:
> for (var i=0; i<100000; i++) {
... db.users.insert({"username" : "user"+i, "created_at" : new Date()});
... }
> db.users.count()
100000
As you can see, interacting with mongos works the same way as interacting with a stand-
alone server does.
You can get an overall view of your cluster by running sh.status(). It will give you a
summary of your shards, databases, and collections:
> sh.status()
--- Sharding Status ---
sharding version: { "_id" : 1, "version" : 3 }
shards:
{ "_id" : "shard0000", "host" : "localhost:30000" }
{ "_id" : "shard0001", "host" : "localhost:30001" }
{ "_id" : "shard0002", "host" : "localhost:30002" }
databases:
{ "_id" : "admin", "partitioned" : false, "primary" : "config" }
{ "_id" : "test", "partitioned" : false, "primary" : "shard0001" }
sh is similar to rs, but for sharding: it is a global variable that defines a number of
sharding helper functions. Run sh.help() to see what it defines. As you can see from
the sh.status() output, you have three shards and two databases (admin is created
automatically).
Your test database may have a different primary shard than shown above. A primary
shard is a “home base” shard that is randomly chosen for each database. All of your data
will be on this primary shard. MongoDB cannot automatically distribute your data yet
because it doesn’t know how (or if) you want it to be distributed. You have to tell it, per-
collection, how you want it to distribute data.
A primary shard is different than a replica set primary. A primary shard
refers to the entire replica set composing a shard. A primary in a replica
set is the single server in the set that can take writes.
To shard a particular collection, first enable sharding on the collection’s database. To do
so, run the enableSharding command:
> sh.enableSharding("test")
234 | Chapter 13: Introduction to Sharding
Now sharding is enabled on the test database, which allows you to shard collections
within the database.
When you shard a collection, you choose a shard key. This is a field or two that MongoDB
uses to break up data. For example, if you choose to shard on "username", MongoDB
would break up the data into ranges of usernames: "a1-steak-sauce" through "def
con", "defcon1" through "howie1998", and so on. Choosing a shard key can be thought
of as choosing an ordering for the data in the collection. This is a similar concept to
indexing, and for good reason: the shard key becomes the most important index on
your collection as it gets bigger. To even create a shard key, the field(s) must be indexed.
Before enabling sharding, we have to create an index on the key we want to shard by:
> db.users.ensureIndex({"username" : 1})
Now we’ll shard the collection by "username":
> sh.shardCollection("test.users", {"username" : 1})
Although we are choosing a shard key without much thought here, it is an important
decision that should be carefully considered in a real system. See Chapter 15 for more
advice on choosing a shard key.
If you wait a few minutes and run sh.status() again, you’ll see that there’s a lot more
information displayed than there was before:
--- Sharding Status ---
sharding version: { "_id" : 1, "version" : 3 }
shards:
{ "_id" : "shard0000", "host" : "localhost:30000" }
{ "_id" : "shard0001", "host" : "localhost:30001" }
{ "_id" : "shard0002", "host" : "localhost:30002" }
databases:
{ "_id" : "admin", "partitioned" : false, "primary" : "config" }
{ "_id" : "test", "partitioned" : true, "primary" : "shard0000" }
test.users chunks:
shard0001 4
shard0002 4
shard0000 5
{ "username" : { $minKey : 1 } } -->> { "username" : "user1704" }
on : shard0001
{ "username" : "user1704" } -->> { "username" : "user24083" }
on : shard0002
{ "username" : "user24083" } -->> { "username" : "user31126" }
on : shard0001
{ "username" : "user31126" } -->> { "username" : "user38170" }
on : shard0002
{ "username" : "user38170" } -->> { "username" : "user45213" }
on : shard0001
{ "username" : "user45213" } -->> { "username" : "user52257" }
on : shard0002
{ "username" : "user52257" } -->> { "username" : "user59300" }
A One-Minute Test Setup | 235
on : shard0001
{ "username" : "user59300" } -->> { "username" : "user66344" }
on : shard0002
{ "username" : "user66344" } -->> { "username" : "user73388" }
on : shard0000
{ "username" : "user73388" } -->> { "username" : "user80430" }
on : shard0000
{ "username" : "user80430" } -->> { "username" : "user87475" }
on : shard0000
{ "username" : "user87475" } -->> { "username" : "user94518" }
on : shard0000
{ "username" : "user94518" } -->> { "username" : { $maxKey : 1 } }
on : shard0000
The collection has been split up into a dozen chunks, where each chunk is a subset of
your data. These are listed by shard key range (the {"username" : minValue} -->>
{"username" : maxValue} denotes the range of each chunk). Looking at the "on" :
shard part of the output, you can see that these chunks have been evenly distributed
between the shards.
This process of a collection being split into chunks is shown in Figure 13-3 through
Figure 13-5. Before sharding, the collection is essentially a single chunk. Sharding splits
this into smaller chunks based on the shard key, as shown in Figure 13-4. These chunks
can then be distributed across the cluster, as Figure 13-5 shows.
Figure 13-3. Before a collection is sharded, it can be thought of as a single chunk from
the smallest value of the shard key to the largest
Figure 13-4. Sharding splits the collection into many chunks based on shard key ranges
236 | Chapter 13: Introduction to Sharding
Figure 13-5. Chunks are evenly distributed across the available shards
Notice the keys at the beginning and end of the chunk list: $minKey and $maxKey.
$minKey can be thought of as “negative infinity.” It is smaller than any other value in
MongoDB. Similarly, $maxKey is like “positive infinity.” It is greater than any other value.
Thus, you’ll always see these as the “caps” on your chunk ranges. The values for your
shard key will always be between $minKey and $maxKey. These values are actually BSON
types and should not be used in your application; they are mainly for internal use. If
you wish to refer to them in the shell, use the MinKey and MaxKey constants.
Now that the data is distributed across multiple shards, let’s try doing some queries.
First, try a query on a specific username:
> db.users.find({username: "user12345"})
{
"_id" : ObjectId("50b0451951d30ac5782499e6"),
"username" : "user12345",
"created_at" : ISODate("2012-11-24T03:55:05.636Z")
}
As you can see, querying works normally. However, let’s run an explain to see what
MongoDB is doing under the covers:
> db.users.find({username: "user12345"}}).explain()
{
"clusteredType" : "ParallelSort",
"shards" : {
"localhost:30001" : [
{
"cursor" : "BtreeCursor username_1",
"nscanned" : 1,
"nscannedObjects" : 1,
A One-Minute Test Setup | 237
"n" : 1,
"millis" : 0,
"nYields" : 0,
"nChunkSkips" : 0,
"isMultiKey" : false,
"indexOnly" : false,
"indexBounds" : {
"username" : [
[
"user12345",
"user12345"
]
]
}
}
]
},
"n" : 1,
"nChunkSkips" : 0,
"nYields" : 0,
"nscanned" : 1,
"nscannedObjects" : 1,
"millisTotal" : 0,
"millisAvg" : 0,
"numQueries" : 1,
"numShards" : 1
}
There are two parts to this explain: a somewhat usual-looking explain output nested
inside of another explain’s output. The way to read this is that the outer explain is from
the mongos: this describes what the mongos had to do to process this query. The inner
explain is from any shards that were used in the query, in this case, localhost:30001.
As "username" is the shard key, mongos could send the query directly to the correct
shard. Contrast that with the results for querying for all of the data:
> db.users.find().explain()
{
"clusteredType" : "ParallelSort",
"shards" : {
"localhost:30000" : [
{
"cursor" : "BasicCursor",
"nscanned" : 37393,
"nscannedObjects" : 37393,
"n" : 37393,
"millis" : 38,
"nYields" : 0,
"nChunkSkips" : 0,
"isMultiKey" : false,
"indexOnly" : false,
"indexBounds" : {
238 | Chapter 13: Introduction to Sharding
}
}
],
"localhost:30001" : [
{
"cursor" : "BasicCursor",
"nscanned" : 31303,
"nscannedObjects" : 31303,
"n" : 31303,
"millis" : 37,
"nYields" : 0,
"nChunkSkips" : 0,
"isMultiKey" : false,
"indexOnly" : false,
"indexBounds" : {
}
}
],
"localhost:30002" : [
{
"cursor" : "BasicCursor",
"nscanned" : 31304,
"nscannedObjects" : 31304,
"n" : 31304,
"millis" : 36,
"nYields" : 0,
"nChunkSkips" : 0,
"isMultiKey" : false,
"indexOnly" : false,
"indexBounds" : {
}
}
]
},
"n" : 100000,
"nChunkSkips" : 0,
"nYields" : 0,
"nscanned" : 100000,
"nscannedObjects" : 100000,
"millisTotal" : 111,
"millisAvg" : 37,
"numQueries" : 3,
"numShards" : 3
}
As you can see from this explain, this query has to visit all three shards to find all the
data. In general, if we are not using the shard key in the query, mongos will have to send
the query to every shard.
A One-Minute Test Setup | 239
Queries that contain the shard key and can be sent to a single shard or subset of shards
are called targeted queries. Queries that must be sent to all shards are called scatter-
gather queries: mongos scatters the query to all the shards and then gathers up the results.
Once you are finished experimenting, shut down the set. Switch back to your original
shell and hit Enter a few times to get back to the command line. Then run clus
ter.stop() to cleanly shut down all of the servers:
> cluster.stop()
If you are ever unsure of what an operation will do, it can be helpful to use ShardingT
est to spin up a quick local cluster and try it out.
240 | Chapter 13: Introduction to Sharding
CHAPTER 14
Configuring Sharding
In the previous chapter, you set up a “cluster” on one machine. This chapter covers how
to set up a more realistic cluster and how each piece fits, in particular:
• How to set up config servers, shards, and mongos processes
• How to add capacity to a cluster
•How data is stored and distributed
When to Shard
Deciding when to shard is a balancing act. You generally do not want to shard too early
because it adds operational complexity to your deployment and forces you to make
design decisions that are difficult to change later. On the other hand, you do not want
to wait too long to shard because it is difficult to shard an overloaded system without
downtime.
In general, sharding is used to:
•Increase available RAM
•Increase available disk space
•Reduce load on a server
•Read or write data with greater throughput than a single mongod can handle
Thus, good monitoring is important to decide when sharding will be necessary. Care‐
fully measure each of these metrics. Generally people speed toward one of these bot‐
tlenecks much faster than the others, so figure out which one your deployment will need
to provision for first and make plans well in advance about when and how you plan to
convert your replica set.
241
As you add shards, performance should increase roughly linearly per shard up to hun‐
dreds of shards. However, you will usually experience a performance drop if you move
from a non-sharded system to just a few shards. Due to the overhead of moving data,
maintaining metadata, and routing, small numbers of shards will generally have higher
latency and may even have lower throughput than a non-sharded system. Thus, you
may want to jump directly to three or more shards.
Starting the Servers
The first step in creating a cluster is to start up all of the processes required. As mentioned
in the previous chapter, we need to set up the mongos and the shards. There’s also a third
component, the config servers, which are an important piece. They are normal mon
god servers that store the cluster configuration: who the shards are, what collections are
sharded by, and the chunks.
Config Servers
Config servers are the brains of your cluster: they hold all of the metadata about which
servers hold what data. Thus, they must be set up first and the data they hold is ex‐
tremely important: make sure that they are running with journaling enabled and that
their data is stored on non-ephemeral drives. Each config server should be on a separate
physical machine, preferable geographically distributed.
The config servers must be started before any of the mongos processes, as mongos pulls
its configuration from them. Config servers are standalone mongod processes, so you
can start them up the same way you would a “normal” mongod:
$ # server-config-1
$ mongod --configsvr --dbpath /var/lib/mongodb -f /var/lib/config/mongod.conf
$
$ # server-config-2
$ mongod --configsvr --dbpath /var/lib/mongodb -f /var/lib/config/mongod.conf
$
$ # server-config-3
$ mongod --configsvr --dbpath /var/lib/mongodb -f /var/lib/config/mongod.conf
When you start up config servers, do not use the --replSet option: config servers are
not members of a replica set. mongos writes to all three config servers and does a two-
phase-commit-type operation to ensure that all three servers have the same data, so all
three must be writable (in a replica set, only the primary is writable by clients).
242 | Chapter 14: Configuring Sharding
A common question is why three config servers? The reasoning behind
the choice is that one config server is not enough: you need redundancy.
Conversely, you don’t want too many config servers, since confirming
actions with all of them would be prohibitively time consuming. Also,
if any of them goes down, you cluster’s metadata becomes read-only.
Thus, three was chosen as enough to give redundancy but not have the
downsides of having too many servers. It will probably be made more
flexible in the future.
The --configsvr option indicates to the mongod that you are planning to use it as a
config server. It is not strictly required, as all it does is change the default port mon
god listens on to 27019 and the default data directory to /data/configdb (you can override
either or both of these settings with --port and --dbpath).
It is recommended that you use this option because it makes it easier to tell, operation‐
ally, what these servers are doing. If you start up your config servers without the --
configsvr option, though, it’s not a problem.
In terms of provisioning, config servers do not need much space or many resources. A
generous estimate is 1 KB of config server space per 200 MB of actual data: they really
are just tables of contents. As they don’t use many resources, you can deploy config
servers on machines running other things, like app servers, shard mongods, or mon
gos processes.
If all of your config servers are lost, you must dig through the data on your shards to
figure out which data is where. This is possible, but slow and unpleasant. Take frequent
backups of config server data. Always take a backup of your config servers before per‐
forming any cluster maintenance.
The mongos Processes
Once you have three config servers running, start a mongos process for your application
to connect to. mongos processes need to know where the config servers are, so you must
always start mongos with the --configdb option:
$ mongos --configdb config-1:27019,config-2:27019,config-3:27019 \
> -f /var/lib/mongos.conf
By default, mongos runs on port 27017. Note that it does not need a data directory
(mongos holds no data itself, it loads the cluster configuration from the config servers
on startup). Make sure that you set logpath to save the mongos log somewhere safe.
You can start as many mongos processes as you’d like. A common setup is one mongos
process per application server (running on the same machine as the application server).
Starting the Servers | 243
Each mongos must use the exact same list of config servers, down to the order in which
they are listed.
Adding a Shard from a Replica Set
Finally, you’re ready to add a shard. There are two possibilities: you may have an existing
replica set or you may be starting from scratch. We will cover starting from an existing
set below. If you are starting from scratch, initialize an empty set and follow the steps
below.
If you already have a replica set serving your application, that will become your first
shard. To convert it into a shard, you are going to tell the mongos the replica set name
and give it a seed list of replica set members.
For example, if you have a replica set named spock on server-1, server-2, server-3,
server-4, and server-5, you would connect to the mongos and run:
> sh.addShard("spock/server-1:27017,server-2:27017,server-4:27017")
{
"added" : "spock/server-1:27017,server-2:27017,server-4:27017",
"ok" : true
}
You can specify all the members of the set, but you do not have to. mongos will
automatically detect any members that were not included in the seed list. If you run
sh.status(), you’ll see that MongoDB soon lists the shard as "spock/server-
1:27017,server-2:27017,server-4:27017,server-3:27017,server-5:27017".
The set name, “spock”, is taken on as an identifier for this shard. If we ever want to
remove this shard or migrate data to it, we’ll use “spock” to describe it. This works better
than using a specific server (e.g., server-1), as replica set membership and status can
change over time.
Once you’ve added the replica set as a shard you can convert your application from
connecting to the replica set to connecting to the mongos. When you add the shard,
mongos registers that all the databases in the replica set are “owned” by that shard, so it
will pass through all queries to your new shard. mongos will also automatically handle
failover for your application as your client library would: it will pass the errors through
to you.
Test failing over a shard’s primary in a development environment to ensure that your
application handles the errors received from mongos correctly (they should be identical
to the errors that you receive from talking to the primary directly).
244 | Chapter 14: Configuring Sharding
Once you have added a shard, you must set up all clients to send requests
to the mongos instead of contacting the replica set. Sharding will not
function correctly if some clients are still making requests to the replica
set directly (not through the mongos). Switch all clients to contacting
the mongos immediately after adding the shard and set up a firewall rule
to ensure that they are unable to connect directly to the shard.
There is a --shardsvr option, analogous to the --configsvr option mentioned previ‐
ously. As before, --shardsvr has little practical effect (it changes the default port to
27018) but can be nice to include operationally.
You can also create stand-alone-mongod shards (instead of replica set shards), but this
is not recommend for production (ShardingTest in the previous chapter did this). To
add a single mongod as a shard simply specify the hostname of the standalone server in
the call to addShard:
> sh.addShard("some-server:27017")
Stand-alone-server shards default to being named shard0000, shard0001, and so on. If
you plan to switch to replica sets later, start with one-member replica sets instead of
standalone servers. Switching from a stand-alone-server shard to a replica set requires
downtime (see “Server Administration” on page 285).
Adding Capacity
When you want to add more capacity, you’ll need to add more shards. To add a new,
empty shard, create a replica set. Make sure it has a distinct name from any of your other
shards. Once it is initialized and has a primary, add it to your cluster by running the
addShard command through mongos, specifying the new replica set’s name and its hosts
as seeds.
If you have several existing replica sets that are not shards, you can add all of them as
new shards in your cluster so long as they do not have any database names in common.
For example, if you had one replica set with a “blog” database, one with a “calendar”
database, and one with the “mail”, “tel”, and “music” databases, you could add each
replica set as a shard and end up with a cluster with three shards and five databases.
However, if you had a fourth replica set that also had a database named “tel”, mongos
would refuse to add it to the cluster.
Sharding Data
MongoDB won’t distribute your data automatically until you tell it how to do so. You
must explicitly tell both the database and collection that you want them to be distributed.
Starting the Servers | 245
For example, suppose we want to shard the artists collection in the music database on
the "name" key. First, we enable sharding for the database, music:
> db.enableSharding("music")
Sharding a database is always prerequisite to sharding one of its collections.
Once you’ve enabled sharding on the database level, you can shard a collection by run‐
ning sh.shardCollection:
> sh.shardCollection("music.artists", {"name" : 1})
Now the collection will be sharded by the "name" key. If you are sharding an existing
collection there must be an index on the "name" field; otherwise the shardCollec
tion call will return an error. If you get an error, create the index (mongos will return
the index it suggests as part of the error message) and retry the shardCollection
command.
If the collection you are sharding does not yet exist, mongos will automatically create
the shard key index for you.
The shardCollection command splits the collection into chunks, which are the unit
MongoDB uses to move data around. Once the command returns successfully, Mon‐
goDB will begin balancing the collection across the shards in your cluster. This process
is not instantaneous. For large collections it may take hours to finish this initial
balancing.
How MongoDB Tracks Cluster Data
Each mongos must always know where to find a document, given its shard key. Theo‐
retically, MongoDB could track where each and every document lived, but this becomes
unwieldy for collections with millions or billions of documents. Thus, MongoDB groups
documents into chunks, which are documents in a given range of the shard key. A chunk
always lives on a single shard, so MongoDB can keep a small table of chunks mapped
to shards.
For example, if a user collection’s shard key is {"age" : 1}, one chunk might be all
documents with an "age" field between 3 and 17. If mongos gets a query for {"age" :
5}, it can route the query to the shard where the 3−17 chunk lives.
As writes occur, the number and size of the documents in a chunk might change. Inserts
can make a chunk contain more documents, removes fewer. If we were making a game
for children and preteens, our chunk for ages 3−17 might get larger and larger (one
would hope). Almost all of our users would be in that chunk, and so on a single shard,
somewhat defeating the point of distributing our data. Thus, once a chunk grows to a
certain size, MongoDB automatically splits it into two smaller chunks. In this example,
the chunk might be split into one chunk containing documents with ages 3 through 11
246 | Chapter 14: Configuring Sharding
and the another containing 12 through 17. Note that these two chunks still cover the
entire age range that the original chunk covered: 3−17. As these new chunks grow, they
can be split into still smaller chunks until there is a chunk for each age.
You cannot have chunks with overlapping ranges, like 3−15 and 12−17. If you could,
MongoDB would need to check both chunks when attempting to find an age in the
overlap, like 14. It is more efficient to only have to look in one place, particularly once
chunks begin moving around the cluster.
A document always belongs to one and only one chunk. One consequence to this rule
is that you cannot use an array field as your shard key, since MongoDB creates multiple
index entries for arrays. For example, if a document had [5, 26, 83] in its "age" field, it
would belong in up to three chunks.
A common misconception is that the data in a chunk is physically
grouped on disk. This is incorrect: chunks have no effect on how mon
god stores collection data.
Chunk Ranges
Each chunk is described by the range it contains. A newly sharded collection starts off
with a single chunk and every document lives in this chunk. This chunk’s bounds are
negative infinity to infinity, shown as $minKey and $maxKey in the shell.
As this chunk grows, MongoDB will automatically split it into two chunks, with the
range negative infinity to <some value> and <some value> to infinity. <some value>
is the same for both chunks: the lower chunk contains everything up to (but not in‐
cluding) <some value> and the upper chunk actually contains <some value>.
This may be more intuitive with an example: suppose we were sharding by "age" as
described earlier. All documents with "age" between 3 and 17 are contained on one
chunk: 3 ≤ age < 17. When this is split, we end up with two ranges: 3 ≤ age < 12 on
one chunk and 12 ≤ age < 17 on the other. 12 is called the split point.
Chunk information is stored in the config.chunks collection. If you looked at the contents
of that collection, you’d see documents that looked something like this (some fields have
been omitted for clarity):
> db.chunks.find(criteria, {"min" : 1, "max" : 1})
{
"_id" : "test.users-age_-100.0",
"min" : {"age" : -100},
"max" : {"age" : 23}
}
{
"_id" : "test.users-age_23.0",
How MongoDB Tracks Cluster Data | 247
"min" : {"age" : 23},
"max" : {"age" : 100}
}
{
"_id" : "test.users-age_100.0",
"min" : {"age" : 100},
"max" : {"age" : 1000}
}
Based on the config.chunks documents shown, here are a few examples of where various
documents would live:
{"_id" : 123, "age" : 50}
This document would live in the second chunk, as that chunk contains all docu‐
ments with "age" between 23 and 100.
{"_id" : 456, "age" : 100}
This document would live on the third chunk, as lower bounds are inclusive. The
second chunk contains all documents up to "age" : 100, but not any documents
where "age" equals 100.
{"_id" : 789, "age" : -101}
This document would not be in any of these chunks. It would be in some chunk
with a range lower than the first chunk’s.
With a compound shard key, shard ranges work the same way that sorting by the two
keys would work. For example, suppose that we had a shard key on {"username" : 1,
"age" : 1}. Then we might have chunk ranges such as:
{
"_id" : "test.users-username_MinKeyage_MinKey",
"min" : {
"username" : { "$minKey" : 1 },
"age" : { "$minKey" : 1 }
},
"max" : {
"username" : "user107487",
"age" : 73
}
}
{
"_id" : "test.users-username_\"user107487\"age_73.0",
"min" : {
"username" : "user107487",
"age" : 73
},
"max" : {
"username" : "user114978",
"age" : 119
}
}
248 | Chapter 14: Configuring Sharding
{
"_id" : "test.users-username_\"user114978\"age_119.0",
"min" : {
"username" : "user114978",
"age" : 119
},
"max" : {
"username" : "user122468",
"age" : 68
}
}
Thus, mongos can easily find on which chunk someone with a given username (or a
given username and age) lives. However, given just an age, mongos would have to check
all, or almost all, chunks. If we wanted to be able to target queries on age to the right
chunk, we’d have to use the “opposite” shard key: {"age" : 1, "username" : 1}. This
is often a point of confusion: a range over the second half of a shard key will cut across
multiple chunks.
Splitting Chunks
mongos tracks how much data it inserts per chunk and, once that reaches a certain
threshold, checks if the chunk needs to be split, as shown in Figure 14-1 and
Figure 14-2. If the chunk does need to be split, mongos will update the chunk’s metadata
on the config servers. Chunk splits are just a metadata change (no data is moved). New
chunk documents are created on the config servers and the old chunk’s range ("max")
is modified. Once that process is complete, the mongos resets its tracking for the original
chunk and creates new trackers for the new chunks.
When mongos asks a shard if a chunk needs to be split, the shard makes a rough calcu‐
lation of the chunk size. If it finds that the chunk is getting large, it finds split points and
sends those to the mongos (as shown in Figure 14-3).
A shard may not be able to find any split points though, even for a large chunk, as there
are a limited number of ways to legally split a chunk. Any two documents with the same
shard key must live in the same chunk so chunks can only be split between documents
where the shard key’s value changes. For example, if the shard key was "age", the fol‐
lowing chunk could be split at the points where the shard key changed, as indicated:
{"age" : 13, "username" : "ian"}
{"age" : 13, "username" : "randolph"}
------------ // split point
{"age" : 14, "username" : "randolph"}
{"age" : 14, "username" : "eric"}
{"age" : 14, "username" : "hari"}
{"age" : 14, "username" : "mathias"}
------------ // split point
How MongoDB Tracks Cluster Data | 249
{"age" : 15, "username" : "greg"}
{"age" : 15, "username" : "andrew"}
mongos will not necessarily split a chunk at every split point available, but those are the
possibilities it has to choose from.
For example, if the chunk contained the following documents, it could not be split
(unless the application started inserting fractional ages):
{"age" : 12, "username" : "kevin"}
{"age" : 12, "username" : "spencer"}
{"age" : 12, "username" : "alberto"}
{"age" : 12, "username" : "tad"}
Thus, having a variety of values for your shard key is important. Other important prop‐
erties will be covered in the next chapter.
If one of the config servers is down when a mongos tries to do a split, the mongos won’t
be able to update the metadata (as shown in Figure 14-4). All config servers must be up
and reachable for splits to happen. If the mongos continues to receive write requests for
the chunk, it will keep trying to split the chunk and fail. As long as the config servers
are not healthy, splits will continue not to work and all the split attempts can slow down
the mongos and shard involved (which repeats the process shown in Figure 14-1 through
Figure 14-4 for each incoming write). This process of mongos repeatedly attempting to
split a chunk and being unable to is called a split storm. The only way to prevent split
storms is to ensure that your config servers are up and healthy as much of the time as
possible. You can also restart a mongos to reset its write counter (so that it is no longer
at the split threshold).
Figure 14-1. When a client writes to a chunk, mongos will check its split threshold for
the chunk
250 | Chapter 14: Configuring Sharding
Figure 14-2. If the split threshold has been reached, mongos will send a request for split
points to the shard
Figure 14-3. The shard calculates split points for the chunk and sends them to the
mongos
Figure 14-4. The mongos chooses a split point and attempts to inform the config server
but cannot reach it. Thus, it is still over its split threshold for the chunk and any subse‐
quent writes will trigger this process again.
Another issue is that mongos might never realize that it needs to split a large chunk.
There is no global counter of how big each chunk is. Each mongos simply calculates
How MongoDB Tracks Cluster Data | 251
whether the writes it has received have reached a certain threshold (as shown in
Figure 14-5). This means that if your mongos processes go up and down frequently a
mongos might never receive enough writes to hit the split threshold before it is shut
down again and your chunks will get larger and larger (as shown in Figure 14-6).
Figure 14-5. As mongos processes perform writes, their counters increase toward the
split threshold
252 | Chapter 14: Configuring Sharding
Figure 14-6. If mongos processes are regularly restarted their counters may never hit the
threshold, making chunks grow without bound
The first way to prevent this is to have fewer mongos churn. Leave mongos processes up,
when possible, instead of spinning them up when they are needed and then turning
them off when they are not. However, some deployments may find it too expensive to
run mongos processes that aren’t being used. If you are in this situation, another way of
getting more splits is to make the chunk size smaller than you actually want it to be.
This will prompt splits to happen at a lower threshold.
You can turn off chunk splitting by starting every mongos with --nosplit.
The Balancer
The balancer is responsible for migrating data. It regularly checks for imbalances be‐
tween shards and, if it finds an imbalance, will begin migrating chunks. Although the
balancer is often referred to as a single entity, each mongos plays the part of “the balancer”
occasionally.
The Balancer | 253
Every few seconds, a mongos will attempt to become the balancer. If there are no other
balancers active, the mongos will take a cluster-wide lock from the config servers and
do a balancing round. Balancing doesn’t affect a mongos’s normal routing operations,
so clients using that mongos should be unaffected.
You can see which mongos is the balancer by looking at the the config.locks collection:
> db.locks.findOne({"_id" : "balancer"})
{
"_id" : "balancer",
"process" : "router-23:27017:1355763351:1804289383",
"state" : 0,
"ts" : ObjectId("50cf939c051fcdb8139fc72c"),
"when" : ISODate("2012-12-17T21:50:20.023Z"),
"who" : "router-23:27017:1355763351:1804289383:Balancer:846930886",
"why" : "doing balance round"
}
The config.locks collection keeps track of all cluster-wide locks. The balancer is the
document with the "_id" of "balancer". The lock’s "who" field tells you which mon
gos is—or was—balancing: router-23:27017 in this case. The "state" field indicates
whether the balancer is running; 0 means it is no longer active, 2 means it’s still balancing.
(1 means that the mongos is attempting to take the lock but has not yet acquired it—
you won’t usually see 1.)
Once a mongos has become the balancer, it checks its table of chunks for each collection
to see if any shards have hit the balancing threshold. This is when one shard has signif‐
icantly more chunks than the other shards (the exact threshold varies: larger collections
tolerate larger imbalances than smaller ones). If an imbalance is detected, the balancer
will redistribute chunks until all shards are within one chunk of one another. If no
collections have hit the balancing threshold. The mongos stops being the balancer.
Assuming that some collections have hit the threshold, the balancer will begin migrating
chunks. It chooses a chunk from the overloaded shard and asks the shard if it should
split the chunk before migrating. Once it does any necessary splits, it migrates the chunk
to a machine with fewer chunks.
An application using the cluster does not need be aware that the data is moving: all reads
and writes are routed to the old chunk until the move is complete. Once the metadata
is updated, all mongos processes attempting to access the data in the old location will
get an error. These errors should not be visible to the client: the mongos will silently
handle the error and retry the operation on the new shard.
This is a common cause of errors you might see in mongos logs that are about being
“unable to setShardVersion.” When mongos gets this type of error, it looks up the new
location of the data from the config servers, updates its chunk table, and attempts the
request again. If it successfully retrieves the data from the new location, it will return it
254 | Chapter 14: Configuring Sharding
to the client as though nothing went wrong (but it will print a message in the log that
the error occurred).
If the mongos is unable to retrieve the new chunk location because the config servers
are unavailable, it will return an error to the client. This is another reason why it is
important to always have config servers up and healthy.
The Balancer | 255
CHAPTER 15
Choosing a Shard Key
The most important and difficult task when using sharding is choosing how your data
will be distributed. To make intelligent choices about this, you have to understand how
MongoDB distributes data. This chapter helps you make a good choice of shard key by
covering:
• How to decide among multiple possible shard keys
• Shard keys for several use cases
• What you can’t use as a shard key
• Some alternative strategies if you want to customize how data is distributed
• How to manually shard your data
This chapter assumes that you understand the basic components of sharding as covered
in the previous chapters.
Taking Stock of Your Usage
When you shard a collection you choose a field or two to use to split up the data. This
key (or keys) is called a shard key. Once you have more than a few shards, it’s almost
impossible to change your shard key, so it is important to choose correctly (or at least
notice any issues quickly).
To choose a good shard key, you need to understand your workload and how your shard
key is going to distribute your application’s requests. This can be difficult to picture, so
try to work out some examples or, even better, try it out on a backup data set with sample
traffic. This section has lots of diagrams and explanations, but there is no substitute for
trying it on your own data set.
257
For each collection that you’re planning to shard, start by answering the following
questions:
•How many shards are you planning to grow to? A three-shard cluster has a great
deal more flexibility than a thousand-shard cluster. As a cluster gets larger, you
should not plan to fire off queries that can hit all shards, so almost all queries must
include the shard key.
•Are you sharding to decrease read or write latency? (Latency refers to how long
something takes, e.g., a write takes 20 ms, but we need it to take 10 ms.) Decreasing
write latency usually involves sending requests to geographically closer or more
powerful machines.
•Are you sharding to increase read or write throughput? (Throughput refers to how
many requests the cluster can handle at the same time: the cluster can do 1,000
writes in 20 ms, but we need it to do 5,000 writes in 20 ms.) Increasing throughput
usually involves adding more parallelization and making sure that requests are dis‐
tributed evenly across the cluster.
•Are you sharding to increase system resources (e.g., give MongoDB more RAM per
GB of data)? If so, you want to keep working set size as small possible.
Use these answers to evaluate the following shard key descriptions and decide whether
the shard key you choose would work well in your situation. Does it give you the targeted
queries that you need? Does it change the throughput or latency of your system in the
ways you need? If you need a compact working set, does it provide that?
Picturing Distributions
There are three basic distributions that are the most common ways people choose to
split their data: ascending key, random, and location-based. There are other types of
keys that could be used, but most use cases fall into one of these categories. Each is
discussed in the following sections.
Ascending Shard Keys
Ascending shard keys are generally something like a "date" field or ObjectId—any‐
thing that steadily increases over time. An autoincrementing primary key is another
example of an ascending field, albeit one that doesn’t show up in MongoDB much (unless
you’re importing from another database).
Suppose that we shard on an ascending field, like "_id" on a collection using ObjectIds.
If we shard on "_id", then this will be split into chunks of "_id" ranges, as in
Figure 15-1. These chunks will be distributed across our sharded cluster of, let’s say,
three shards, as shown in Figure 15-2.
258 | Chapter 15: Choosing a Shard Key
Figure 15-1. The collection is split into ranges of ObjectIds. Each range is a chunk.
Suppose we create a new document. Which chunk will it be in? The answer is the chunk
with the range ObjectId("5112fae0b4a4b396ff9d0ee5") through $maxKey. This is
called the max chunk, as it is the chunk containing $maxKey.
If we insert another document, it will also be in the max chunk. In fact, every subsequent
insert will be into the max chunk! Every insert’s "_id" field will be closer to infinity than
the previous (because ObjectIds are always ascending), so they will all go to into the
max chunk.
Picturing Distributions | 259
Figure 15-2. Chunks are distributed across shards in a random order
This has a couple of interesting (and often undesirable) properties. First, all of your
writes will be routed to one shard (shard0002, in this case). This chunk will be the only
one growing and splitting, as it is the only one that receives inserts. As you insert data,
new chunks will “fall off” of this chunk’s butt, as shown in Figure 15-3.
260 | Chapter 15: Choosing a Shard Key
Figure 15-3. The max chunk continues growing and being split into multiple chunks
This pattern often makes it more difficult for MongoDB to keep chunks evenly balanced
because all the chunks are being created by one shard. Therefore, MongoDB must con‐
stantly move chunks to other shards instead of correcting small imbalances that might
occur in a more evenly distributed systems.
Randomly Distributed Shard Keys
On the other end of the spectrum are randomly distributed shard keys. Randomly dis‐
tributed keys could be usernames, email addresses, UUIDs, MD5 hashes, or any other
key that has no identifiable pattern in your dataset.
Suppose the shard key is a random number between 0 and 1. We’ll end up with a random
distribution of chunks on the various shards, as shown in Figure 15-4.
Picturing Distributions | 261
Figure 15-4. As in the previous section, chunks are distributed randomly around the
cluster
As more data is inserted, the data’s random nature means that inserts should hit every
chunk fairly evenly. You can prove this to yourself by inserting 10,000 documents and
seeing where they end up:
> var servers = {}
> var findShard = function (id) {
... var explain = db.random.find({_id:id}).explain();
... for (var i in explain.shards) {
... var server = explain.shards[i][0];
262 | Chapter 15: Choosing a Shard Key
... if (server.n == 1) {
... if (server.server in servers) {
... servers[server.server]++;
... } else {
... servers[server.server] = 1;
... }
... }
... }
... }
> for (var i = 0; i < 10000; i++) {
... var id = ObjectId();
... db.random.insert({"_id" : id, "x" : Math.random()});
... findShard(id);
... }
> servers
{
"spock:30001" : 2942,
"spock:30002" : 4332,
"spock:30000" : 2726
}
As writes are randomly distributed, shards should grow at roughly the same rate, lim‐
iting the number of migrates that need to occur.
The only downside to randomly distributed shard keys is that MongoDB isn’t efficient
at randomly accessing data beyond the size of RAM. However, if you have the capacity
or don’t mind the performance hit, random keys nicely distribute load across your
cluster.
Location-Based Shard Keys
Location-based shard keys may be things like a user’s IP, latitude and longitude, or
address. Location shard keys are not necessarily related to a physical location field: the
“location” might be a more abstract way that data should be grouped together. In any
case, it is a key where documents with some similarity fall into a range based on this
field. This can be handy for both putting data close to its users and keeping related data
together on disk.
For example, suppose we have a collection of documents that are sharded on an IP
address. Documents will be organized into chunks based on their addresses and ran‐
domly spread across the cluster, as shown in Figure 15-5.
Picturing Distributions | 263
Figure 15-5. A sample distribution of chunks in the IP address collection
If we wanted certain chunk ranges to be attached to certain shards, we could tag these
shards and then assign chunk ranges to tags. In this example, suppose that we wanted
to keep certain IP blocks on certain shards: say, “56.*.*.*” (the United States Postal
Service’s IP block) on shard0000 and “17.*.*.*” (Apple’s IP block) on either shard0000
or shard0002. We do not care where the other IPs live. We could request that the balancer
do this by setting up tag ranges:
> sh.addShardTag("shard0000", "USPS")
> sh.addShardTag("shard0000", "Apple")
> sh.addShardTag("shard0002", "Apple")
Next, we create the rules:
> sh.addTagRange("test.ips", {"ip" : "056.000.000.000"},
... {"ip" : "057.000.000.000"}, "USPS")
This attaches all IPs greater than or equal to 56.0.0.0 and less than 57.0.0.0 to the shard
tagged “USPS”. Next, we add a rule for Apple:
> sh.addTagRange("test.ips", {"ip" : "017.000.000.000"},
... {"ip" : "018.000.000.000"}, "Apple")
When the balancer moves chunks, it will attempt to move chunks with those ranges to
those shards. Note that this process is not immediate. Chunks that were not covered by
a tag range will be moved around normally. The balancer will continue attempting to
distribute chunks evenly among shards.
Shard Key Strategies
This section presents a number of shard key options for various types of applications.
Hashed Shard Key
For loading data as fast as possible, hashed shard keys are the best option. A hashed
shard key can make any field randomly distributed, so it is a good choice if you’re going
264 | Chapter 15: Choosing a Shard Key
to be using an ascending key a in a lot of queries but want writes to be random dis‐
tributed.
The trade-off is that you can never do a targeted range query with a hashed shard key.
If you will not be doing range queries, though, hashed shard keys are a good option.
To create a hashed shard key, first create a hashed index:
> db.users.ensureIndex({"username" : "hashed"})
Next, shard the collection with:
> sh.shardCollection("app.users", {"username" : "hashed"})
{ "collectionsharded" : "app.users", "ok" : 1 }
If you create a hashed shard key on a nonexistent collection, shardCollection behaves
interestingly: it assumes that you want evenly distributed chunks, so it immediately
creates a bunch of empty chunks and distributes them around your cluster. For example,
suppose our cluster looked like this before creating the hashed shard key:
> sh.status()
--- Sharding Status ---
sharding version: { "_id" : 1, "version" : 3 }
shards:
{ "_id" : "shard0000", "host" : "localhost:30000" }
{ "_id" : "shard0001", "host" : "localhost:30001" }
{ "_id" : "shard0002", "host" : "localhost:30002" }
databases:
{ "_id" : "admin", "partitioned" : false, "primary" : "config" }
{ "_id" : "test", "partitioned" : true, "primary" : "shard0001" }
Immediately after shardCollection returns there are two chunks on each shard, evenly
distributing the key space across the cluster:
> sh.status()
--- Sharding Status ---
sharding version: { "_id" : 1, "version" : 3 }
shards:
{ "_id" : "shard0000", "host" : "localhost:30000" }
{ "_id" : "shard0001", "host" : "localhost:30001" }
{ "_id" : "shard0002", "host" : "localhost:30002" }
databases:
{ "_id" : "admin", "partitioned" : false, "primary" : "config" }
{ "_id" : "test", "partitioned" : true, "primary" : "shard0001" }
test.foo
shard key: { "username" : "hashed" }
chunks:
shard0000 2
shard0001 2
shard0002 2
{ "username" : { "$MinKey" : true } }
-->> { "username" : NumberLong("-6148914691236517204") }
on : shard0000 { "t" : 3000, "i" : 2 }
Shard Key Strategies | 265
{ "username" : NumberLong("-6148914691236517204") }
-->> { "username" : NumberLong("-3074457345618258602") }
on : shard0000 { "t" : 3000, "i" : 3 }
{ "username" : NumberLong("-3074457345618258602") }
-->> { "username" : NumberLong(0) }
on : shard0001 { "t" : 3000, "i" : 4 }
{ "username" : NumberLong(0) }
-->> { "username" : NumberLong("3074457345618258602") }
on : shard0001 { "t" : 3000, "i" : 5 }
{ "username" : NumberLong("3074457345618258602") }
-->> { "username" : NumberLong("6148914691236517204") }
on : shard0002 { "t" : 3000, "i" : 6 }
{ "username" : NumberLong("6148914691236517204") }
-->> { "username" : { "$MaxKey" : true } }
on : shard0002 { "t" : 3000, "i" : 7 }
Note that there are no documents in the collection yet, but when you start inserting
them, writes should be evenly distributed across the shards from the get-go. Ordinarily,
you would have to wait for chunks to grow, split, and move to start writing to other
shards. With this automatic priming, you’ll immediately have chunk ranges on all
shards.
There are some limitations on what your shard key can be if you’re using a hashed shard
key. First, you cannot use the unique option. As with other shard keys, you cannot use
array fields. Finally, be aware of is that floating point values will be rounded to whole
numbers before hashing, so 1 and 1.999999 will both be hashed to the same value.
Hashed Shard Keys for GridFS
Before attempting to shard GridFS collections, make sure that you understand how
GridFS stores data (see Chapter 6 for an explanation).
In the following explanation, the term “chunks” is overloaded since GridFS splits files
into chunks and sharding splits collections into chunks. Thus, the two types of chunks
are referred to as “GridFS chunks” and “sharding chunks” later in the chapter.
GridFS collections are generally excellent candidates for sharding, as they contain mas‐
sive amounts of file data. However, neither of the indexes that are automatically created
on fs.chunks are particularly good shard keys: {"_id" : 1} is an ascending key and
{"files_id" : 1, "n" : 1} picks up fs.files' "_id" field, so it is also an ascending key.
However, if you create a hashed index on the "files_id" field, each file will be randomly
distributed across the cluster. But a file will always be contained in a single chunk. This
is the best of both worlds: writes will go to all shards evenly and reading a file’s data will
only ever have to hit a single shard.
266 | Chapter 15: Choosing a Shard Key
To set this up, you must create a new index on {"files_id" : "hashed"} (as of this
writing, mongos cannot use a subset of the compound index as a shard key). Then shard
the collection on this field:
> db.fs.chunks.ensureIndex({"files_id" : "hashed"})
> sh.shardCollection("test.fs.chunks", {"files_id" : "hashed"})
{ "collectionsharded" : "test.fs.chunks", "ok" : 1 }
As a side note, the fs.files collection may or may not need to be sharded, as it will be
much smaller than fs.chunks. You can shard it if you would like, but it less likely to be
necessary.
The Firehose Strategy
If you have some servers that are more powerful than others, you might want to let them
handle proportionally more load than your less-powerful servers. For example, suppose
you have one shard that is composed of SSDs that can handle 10 times the load of your
other machines (backed by spinning disks). Luckily, you have 10 other shards. You could
force all inserts to go to the SSD, and then allow the balancer to move older chunks to
the other shards. This would give lower-latency writes than the spinning disks would.
To use this strategy, we have to pin the highest chunk to the SSD. First, tag the SSD:
> sh.addShardTag("shard-name", "ssd")
Now, pin the current value of the ascending key through infinity to that shard, so all
new writes go to it:
> sh.addTagRange("dbName.collName", {"_id" : ObjectId()},
... {"_id" : MaxKey}, "ssd")
Now all inserts will be routed to this last chunk, which will always live on the shard
tagged "ssd".
However, ranges from now through infinity will be trapped on this shard unless we
modify the tag range. We could set up a cron job to update the tag range once a day, like
this:
> use config
> var tag = db.tags.findOne({"ns" : "dbName.collName",
... "max" : {"shardKey" : MaxKey}})
> tag.min.shardKey = ObjectId()
> db.tags.save(tag)
Then all of the previous day’s chunks would be able to move to other shards.
Another downside of this strategy is that it requires some changes to scale. If your SSD
can no longer handle the number of writes coming in, there is no trivial way to split the
load between this server and another.
Shard Key Strategies | 267
If you do not have a high-performance server to firehose into or you are not using
tagging, do not use an ascending key as the shard key. If you do, all writes will go to a
single shard.
Multi-Hotspot
Standalone mongod servers are most efficient when doing ascending writes. This con‐
flicts with sharding, in that sharding is most efficient when writes are spread over the
cluster. This technique basically creates multiple hotspots—optimally several on each
shard—so that writes are evenly balanced across the cluster but, within a shard, as‐
cending.
To accomplish this, we use a compound shard key. The first value in the compound key
is a rough, random value with low-ish cardinality. You can picture each value in the first
part of the shard key as a chunk, as shown in Figure 15-6. This will eventually work itself
out as you insert more data, although it will probably never be divided up this neatly
(right on the $minKey lines). However, if you insert enough data, you should eventually
have approximately one chunk per random value. As you continue to insert data, you’ll
end up with multiple chunks with the same random value, which brings us to the second
part of the shard key.
Figure 15-6. A subset of the chunks. Each chunk contains a single state and a range of
_ids
The second part of the shard key is an ascending key. This means that, within a chunk,
values are always increasing, as shown in the sample documents in Figure 15-7. Thus,
268 | Chapter 15: Choosing a Shard Key
if you had one chunk per shard, you’d have the perfect setup: ascending writes on every
shard, as shown in Figure 15-8. Of course having n chunks with n hotspots spread across
n shards isn’t very extensible: add a new shard and it won’t get any writes because there’s
no hot chunk to put on it. Thus, you want a few hotspot chunks per shard (to give you
room to grow). However, you don’t want too many. A few hotspot chunks will keep the
effectiveness of ascending writes. But having, say, a thousand “hotspots” on a shard will
end up being equivalent to random writes.
Figure 15-7. A sample list of inserted documents. Note that all _ids are increasing.
Shard Key Strategies | 269
Figure 15-8. The inserted documents, split into chunks. Note that, within each chunk,
the _ids are increasing.
You can picture this setup as each chunk being a stack of ascending documents. There
are multiple stacks on each shard, each ascending until the chunk is split. Once a chunk
is split, only one of the new chunks will be a hotspot chunk: the other chunk will es‐
sentially be “dead” and never grow again. If the stacks are evenly distributed across the
shards, writes will be evenly distributed.
270 | Chapter 15: Choosing a Shard Key
Shard Key Rules and Guidelines
There are several practical restrictions to be aware of before choosing a shard key.
Determining which key to shard on and creating shard keys should be reminiscent of
indexing because the two concepts are similar. In fact, often your shard key may just be
the index you use most often (or some variation on it).
Shard Key Limitations
Shard keys cannot be arrays. sh.shardCollection() will fail if any key has an array
value and inserting an array into that field is not allowed.
Once inserted, a document’s shard key value cannot be modified. To change a docu‐
ment’s shard key, you must remove the document, change the key, and reinsert it. Thus,
you should choose a field that is unchangeable or changes infrequently.
Most special types of index cannot be used for shard keys. In particular, you cannot
shard on a geospatial index. Using a hashed index for a shard key is allowed, as covered
previously.
Shard Key Cardinality
Whether your shard key jumps around or increases steadily, it is important to choose
a key with values that will vary. As with indexes, sharding performs better on high
cardinality fields. If, for example, we had a "logLevel" key that had only values "DE
BUG", "WARN", or "ERROR", MongoDB wouldn’t be able to break up your data into more
than three chunks (because there would be only three different values for the shard key).
If you have a key with little variation and want to use it as a shard key anyway, you can
do so by creating a compound shard key on that key and a key that varies more, like
"logLevel" and "timestamp". It is important that the combination of keys has high
cardinality.
Controlling Data Distribution
Sometimes, automatic data distribution will not fit your requirements. This section gives
you some options beyond choosing a shard key and allowing MongoDB to do everything
automatically.
As your cluster gets larger or busier, these solutions become less practical. However, for
small clusters, you may want more control.
Shard Key Rules and Guidelines | 271
Using a Cluster for Multiple Databases and Collections
MongoDB evenly distributes collections across every shard your cluster, which works
well if you’re storing homogeneous data. However, if you have a log collection that is
“lower-value” than your other data, you might not want it taking up space on your more
expensive servers. Or, if you have one powerful shard, you might want to use it for only
a realtime collection and not allow other collections to use it. You can set up separate
clusters, but you can also give MongoDB specific directions about where you want it to
put certain data.
To set this up, use the sh.addShardTag() helper in the shell:
> sh.addShardTag("shard0000", "high")
> // shard0001 - no tag
> // shard0002 - no tag
> // shard0003 - no tag
> sh.addShardTag("shard0004", "low")
> sh.addShardTag("shard0005", "low")
Then we can assign different collections to different shards. For instance, our realtime
collection:
> sh.addTagRange("super.important", {"shardKey" : MinKey},
... {"shardKey" : MaxKey}, "high")
This says, “for negative infinity to infinity for this collection, store it on shards tagged
‘high’.” This means that no data from the important collection will be stored on any other
server. Note that this does not effect how other collections are distributed: other col‐
lections will still be evenly distributed between this shard and the others.
We can perform a similar operation to keep the log collection on a low-quality server:
> sh.addTagRange("some.logs", {"shardKey" : MinKey},
... {"shardKey" : MaxKey}, "low")
The log collections will now be split evenly between shard0004 and shard0005.
Assigning a tag range to a collection does not affect it instantly. It is an instruction to
the balancer that, when it runs, these are the viable targets to move the collection to.
Thus, if the entire log collection is on shard0002 or evenly distributed among the shards,
it will take a little while for all of the chunks to be migrated to shard0004 and shard0005.
As another example, perhaps we have a collection where we don’t want it on the shard
tagged “high” but do want it on any other one. We can add a new tag to all of the non-
high-performance shards to create a new grouping. Shards can have as many tags as
you need:
> sh.addShardTag("shard0001", "whatever")
> sh.addShardTag("shard0002", "whatever")
> sh.addShardTag("shard0003", "whatever")
272 | Chapter 15: Choosing a Shard Key
> sh.addShardTag("shard0004", "whatever")
> sh.addShardTag("shard0005", "whatever")
Now we can specify that we want this collection (call it "normal.coll") distributed across
these five shards:
> sh.addTagRange("normal.coll", {"shardKey" : MinKey},
... {"shardKey" : MaxKey}, "whatever")
You cannot assign collections dynamically, i.e., “when a collection is created, randomly
home it to a shard.” However, you could have a cron job that went through and did it
for you.
If you make a mistake or change your mind, you can remove shard tags with sh.remov
eShardTag():
> sh.removeShardTag("shard0005", "whatever")
If you remove all tags described by a tag range (for example, if untagging the shard
marked "high") the balancer won’t distribute the data anywhere because there aren’t
any valid locations listed. All the data will still be readable and writable; it just won’t be
able to migrate until you modify your tags or tag ranges.
There is no helper for removing tag ranges, but you can do so manually. To deal with
tag ranges manually, access the config.tags namespace through the mongos. Similarly,
shard tag information is kept in the config.shards namespace in the "tags" field of each
shard document. If a shard has no "tags" field, then it has no tags.
Manual Sharding
Sometimes, for complex requirements or special situations, you may prefer to have
complete control over which data is distributed where. You can turn off the balancer if
you don’t want data to be automatically distributed and use the moveChunk command
to manually distribute data.
To turn off the balancer, connect to a mongos (any mongos is fine) and update the
config.settings namespace with the following:
> db.settings.update({"_id" : "balancer"}, {"enabled" : false}, true)
Note that this is an upsert: it creates the balancer setting if one does not exist.
If there is currently a migrate in progress, this setting will not take effect until the migrate
has completed. However, once any in-flight migrations have finished, the balancer will
stop moving data around.
Once the balancer is off, you can move data around manually (if necessary). First, find
out which chunks are where by looking at config.chunks:
> db.chunks.find()
Controlling Data Distribution | 273
Now, use the moveChunk command to migrate chunks to other shards. Specify the lower
bound of the chunk-to-be-migrated and give the name of the shard that you want to
move the chunk to:
> sh.moveChunk("test.manual.stuff",
... {user_id: NumberLong("-1844674407370955160")}, "test-rs1")
{ "millis" : 4079, "ok" : 1 }
However, unless you are in an exceptional situation, you should use MongoDB’s auto‐
matic sharding instead of doing it manually. If you end up with a hotspot on a shard
that you weren’t expecting, you might end up with most of your data on that shard.
In particular, do not combine setting up unusual distributions manually with running
the balancer. If the balancer detects an uneven number of chunks it will simply reshuffle
all of your work to get the collection evenly balanced again. If you want uneven distri‐
bution of chunks, use the shard tagging technique discussed in “Using a Cluster for
Multiple Databases and Collections” on page 272.
274 | Chapter 15: Choosing a Shard Key
CHAPTER 16
Sharding Administration
A sharded cluster is the most difficult type of deployment to administer. This chapter
gives advice on performing administrative tasks on all parts of a cluster, including:
• Inspecting what the cluster’s state is: who its members are, where data is held, and
what connections are open
• How to add, remove, and change members of a cluster
•Administering data movement and manually moving data
Seeing the Current State
There are several helpers available to find out what data is where, what the shards are,
and what the cluster is doing.
Getting a Summary with sh.status
sh.status() gives you an overview of your shards, databases, and sharded collections.
If you have a small number of chunks, it will print a breakdown of which chunks are
where as well. Otherwise it will simply give the collection’s shard key and how many
chunks each shard has:
> sh.status()
--- Sharding Status ---
sharding version: { "_id" : 1, "version" : 3 }
shards:
{ "_id" : "shard0000", "host" : "localhost:30000",
"tags" : [ "USPS" , "Apple" ] }
{ "_id" : "shard0001", "host" : "localhost:30001" }
{ "_id" : "shard0002", "host" : "localhost:30002", "tags" : [ "Apple" ] }
databases:
{ "_id" : "admin", "partitioned" : false, "primary" : "config" }
275
{ "_id" : "test", "partitioned" : true, "primary" : "shard0001" }
test.foo
shard key: { "x" : 1, "y" : 1 }
chunks:
shard0000 4
shard0002 4
shard0001 4
{ "x" : { $minKey : 1 }, "y" : { $minKey : 1 } } -->>
{ "x" : 0, "y" : 10000 } on : shard0000
{ "x" : 0, "y" : 10000 } -->> { "x" : 12208, "y" : -2208 }
on : shard0002
{ "x" : 12208, "y" : -2208 } -->> { "x" : 24123, "y" : -14123 }
on : shard0000
{ "x" : 24123, "y" : -14123 } -->> { "x" : 39467, "y" : -29467 }
on : shard0002
{ "x" : 39467, "y" : -29467 } -->> { "x" : 51382, "y" : -41382 }
on : shard0000
{ "x" : 51382, "y" : -41382 } -->> { "x" : 64897, "y" : -54897 }
on : shard0002
{ "x" : 64897, "y" : -54897 } -->> { "x" : 76812, "y" : -66812 }
on : shard0000
{ "x" : 76812, "y" : -66812 } -->> { "x" : 92793, "y" : -82793 }
on : shard0002
{ "x" : 92793, "y" : -82793 } -->> { "x" : 119599, "y" : -109599 }
on : shard0001
{ "x" : 119599, "y" : -109599 } -->> { "x" : 147099, "y" : -137099 }
on : shard0001
{ "x" : 147099, "y" : -137099 } -->> { "x" : 173932, "y" : -163932 }
on : shard0001
{ "x" : 173932, "y" : -163932 } -->>
{ "x" : { $maxKey : 1 }, "y" : { $maxKey : 1 } } on : shard0001
test.ips
shard key: { "ip" : 1 }
chunks:
shard0000 2
shard0002 3
shard0001 3
{ "ip" : { $minKey : 1 } } -->> { "ip" : "002.075.101.096" }
on : shard0000
{ "ip" : "002.075.101.096" } -->> { "ip" : "022.089.076.022" }
on : shard0002
{ "ip" : "022.089.076.022" } -->> { "ip" : "038.041.058.074" }
on : shard0002
{ "ip" : "038.041.058.074" } -->> { "ip" : "055.081.104.118" }
on : shard0002
{ "ip" : "055.081.104.118" } -->> { "ip" : "072.034.009.012" }
on : shard0000
{ "ip" : "072.034.009.012" } -->> { "ip" : "090.118.120.031" }
on : shard0001
{ "ip" : "090.118.120.031" } -->> { "ip" : "127.126.116.125" }
on : shard0001
{ "ip" : "127.126.116.125" } -->> { "ip" : { $maxKey : 1 } }
276 | Chapter 16: Sharding Administration
on : shard0001
tag: Apple { "ip" : "017.000.000.000" } -->> { "ip" : "018.000.000.000" }
tag: USPS { "ip" : "056.000.000.000" } -->> { "ip" : "057.000.000.000" }
{ "_id" : "test2", "partitioned" : false, "primary" : "shard0002" }
Once there are more than a few chunks, sh.status() will summarize the chunk stats
instead of pinting each chunk. To see all chunks, run sh.status(true) (the true tells
sh.status() to be verbose).
All the information sh.status() shows is gathered from your config database.
sh.status() runs a MapReduce to collect this data, so you cannot run sh.status()
when using the --noscripting option.
Seeing Configuration Information
All of the configuration information about your cluster is kept in collections in the config
database on the config servers. You can access it directly, but the shell has several helpers
for exposing this information in a more readable way. However, you can always directly
query the config database for metadata about your cluster.
Never connect directly to your config servers, as you do not want to
take the chance of accidentally changing or removing config server data.
Instead, connect to the mongos and use the config database to see its data,
as you would for any other database:
mongos> use config
If you manipulate config data through mongos (instead of connecting
directly to the config servers), mongos will ensure that all of your config
servers stay in sync and prevent various dangerous actions like acci‐
dentally dropping the config database.
In general, you should not directly change any data in the config database (exceptions
are noted below). If you do change anything, you will generally have to restart all of
your mongos servers to see its effect.
There are several collections in the config database. This section covers what each one
contains and how it can be used.
config.shards
The shards collection keeps track of all the shards in the cluster. A typical document in
the shards collection might looks something like this:
> db.shards.findOne()
{
"_id" : "spock",
"host" : "spock/server-1:27017,server-2:27017,server-3:27017",
Seeing the Current State | 277
"tags" : [
"us-east",
"64gb mem",
"cpu3"
]
}
The shard’s "_id" is picked up from the replica set name, so each replica set in your
cluster must have a unique name.
When you update your replica set configuration (e.g., adding or removing members),
the "host" field will be updated automatically.
config.databases
The databases collection keeps track of all of the databases, sharded and non, that the
cluster knows about:
> db.databases.find()
{ "_id" : "admin", "partitioned" : false, "primary" : "config" }
{ "_id" : "test1", "partitioned" : true, "primary" : "spock" }
{ "_id" : "test2", "partitioned" : false, "primary" : "bones" }
If enableSharding has been run on a database, "partitioned" will be true. The "pri
mary" is the database’s “home base.” By default, all new collections in that database will
be created on the database’s primary shard.
config.collections
The collections collection keeps track of all sharded collections (non-sharded collections
are not shown). A typical document looks something like this:
> db.collections.findOne()
{
"_id" : "test.foo",
"lastmod" : ISODate("1970-01-16T17:53:52.934Z"),
"dropped" : false,
"key" : { "x" : 1, "y" : 1 },
"unique" : true
}
The important fields are:
"_id"
The namespace of the collection.
"key"
The shard key. In this case, it is a compound key on "x" and "y".
"unique"
Indicates that the shard key is a unique index. This field is not displayed unless it
is true (the shard key is unique). By default, the shard key is not unique.
278 | Chapter 16: Sharding Administration
config.chunks
The chunks collection keeps a record of each chunk in all the collections. A typical
document in the chunks collection might look something like this:
{
"_id" : "test.hashy-user_id_-1034308116544453153",
"lastmod" : { "t" : 5000, "i" : 50 },
"lastmodEpoch" : ObjectId("50f5c648866900ccb6ed7c88"),
"ns" : "test.hashy",
"min" : { "user_id" : NumberLong("-1034308116544453153") },
"max" : { "user_id" : NumberLong("-732765964052501510") },
"shard" : "test-rs2"
}
The most useful fields are:
"_id"
The unique identifier for the chunk. Generally this is the namespace, shard key, and
lower chunk boundary.
"ns"
The collection that this chunk is from.
"in"
The smallest value in the chunk’s range (inclusive).
"max"
All values in the chunk are smaller than this value.
"shard"
Which shard the chunk resides on.
The "lastmod" and "lastmodEpoch" fields are used to track chunk versioning. For
example, if a chunk "foo.bar-_id-1" split into two chunks, we’d want a way of distin‐
guishing the new, smaller "foo.bar-_id-1" chunk from its previous incarnation. Thus,
the "t" and "i" fields are the major and minor chunk versions: major versions change
when a chunk is migrated to a new shard and minor versions change when a chunk is
split.
sh.status() uses the config.chunks collection to gather most of its information.
config.changelog
The changelog collection is useful for keeping track of what a cluster is doing, since it
records all of the splits and migrates that have occurred.
Splits are recorded in a document that looks like this:
{
"_id" : "router1-2013-02-09T18:08:12-5116908cab10a03b0cd748c3",
"server" : "spock-01",
Seeing the Current State | 279
"clientAddr" : "10.3.1.71:62813",
"time" : ISODate("2013-02-09T18:08:12.574Z"),
"what" : "split",
"ns" : "test.foo",
"details" : {
"before" : {
"min" : { "x" : { $minKey : 1 }, "y" : { $minKey : 1 } },
"max" : { "x" : { $maxKey : 1 }, "y" : { $maxKey : 1 } },
"lastmod" : Timestamp(1000, 0),
"lastmodEpoch" : ObjectId("000000000000000000000000")
},
"left" : {
"min" : { "x" : { $minKey : 1 }, "y" : { $minKey : 1 } },
"max" : { "x" : 0, "y" : 10000 },
"lastmod" : Timestamp(1000, 1),
"lastmodEpoch" : ObjectId("000000000000000000000000")
},
"right" : {
"min" : { "x" : 0, "y" : 10000 },
"max" : { "x" : { $maxKey : 1 }, "y" : { $maxKey : 1 } },
"lastmod" : Timestamp(1000, 2),
"lastmodEpoch" : ObjectId("000000000000000000000000")
}
}
}
The "details" give information about what the original document looked like and
what it split into.
This output is what the first chunk split of a collection looks like. Note that each new
chunk has its minor version increment: "lastmod" is Timestamp(1000, 1) and Time
stamp(1000, 2), respectively.
Migrates are a bit more complicated and actually create four separate changelog docu‐
ments: one noting the start of the migrate, one for the "from" shard, one for the "to"
shard, and one for the migrate’s commit (when the migration is finalized). The middle
two documents are of interest because these give a breakdown of how long each step in
the process took. This can give you an idea whether it’s the disk, network, or something
else that is causing a bottleneck on migrates.
For example, the document created by the "from" shard looks like this:
{
"_id" : "router1-2013-02-09T18:15:14-5116923271b903e42184211c",
"server" : "spock-01",
"clientAddr" : "10.3.1.71:27017",
"time" : ISODate("2013-02-09T18:15:14.388Z"),
"what" : "moveChunk.to",
"ns" : "test.foo",
"details" : {
"min" : { "x" : 24123, "y" : -14123 },
"max" : { "x" : 39467, "y" : -29467 },
280 | Chapter 16: Sharding Administration
"step1 of 5" : 0,
"step2 of 5" : 0,
"step3 of 5" : 900,
"step4 of 5" : 0,
"step5 of 5" : 142
}
};
Each of the steps listed in "details" is timed and the "stepN of 5" messages show
how long the step took, in milliseconds. When the "from" shard receives a move‐
Chunk command from the mongos, it:
1. Checks the command parameters.
2. Confirms with the config servers that it can acquire a distributed lock for the
migrate.
3. Tries to contact the "to" shard.
4. The data is copied. This is referred to and logged as “the critical section.”
5. Coordinates with the "to" shard and config servers to confirm the migrate.
Note that the "to" and "from" shards must be in close communication starting at "step4
of 5": the shards directly talk to one another and the config server to perform the
migration. If the "from" server has flaky network connectivity during the final steps, it
may end up in a state where it cannot undo the migrate and cannot move forward with
it. In this case, the mongod will shut down.
The "to" shard’s changelog document is similar to the "from" shard’s, but the steps are
a bit different. It looks like:
{
"_id" : "router1-2013-02-09T18:15:14-51169232ab10a03b0cd748e5",
"server" : "spock-01",
"clientAddr" : "10.3.1.71:62813",
"time" : ISODate("2013-02-09T18:15:14.391Z"),
"what" : "moveChunk.from",
"ns" : "test.foo",
"details" : {
"min" : { "x" : 24123, "y" : -14123 },
"max" : { "x" : 39467, "y" : -29467 },
"step1 of 6" : 0,
"step2 of 6" : 2,
"step3 of 6" : 33,
"step4 of 6" : 1032,
"step5 of 6" : 12,
"step6 of 6" : 0
}
}
When the "to" shard receives a command from the "from" shard, it:
Seeing the Current State | 281
1. Migrates indexes. If this shard has never held chunks from the migrated collection
before, it needs to know what fields are indexed. If this isn’t the first time a chunk
from this collection is being moved to this shard, then this should be a no-op.
2. Deletes any existing data in the chunk range. There might be data left over from a
failed migration or restore procedure which we wouldn’t want to interfere with the
current data.
3. Copy all documents in the chunk to the "to" shard.
4. Replay any operations that happened to these document during the copy (on the
"to" shard).
5. Wait for the "to" shard to have replicated the newly migrated data to a majority of
servers.
6. Commit the migrate by changing the chunk’s metadata to say that it lives on the
"to" shard.
config.tags
This collection is created if you configure shard tags for your system. Each tag is asso‐
ciated with a chunk range:
> db.tags.find()
{
"_id" : {
"ns" : "test.ips",
"min" : {"ip" : "056.000.000.000"}
},
"ns" : "test.ips",
"min" : {"ip" : "056.000.000.000"},
"max" : {"ip" : "057.000.000.000"},
"tag" : "USPS"
}
{
"_id" : {
"ns" : "test.ips",
"min" : {"ip" : "017.000.000.000"}
},
"ns" : "test.ips",
"min" : {"ip" : "017.000.000.000"},
"max" : {"ip" : "018.000.000.000"},
"tag" : "Apple"
}
config.settings
This collection contains documents representing the current balancer settings and
chunk size. By changing the documents in this collection, you can turn the balancer on
282 | Chapter 16: Sharding Administration
or off or change the chunk size. Note that you should always connect to mongos, not
the config servers directly, to change values in this collection.
Tracking Network Connections
There are a lot of connections between the components of a cluster. This section covers
some sharding-specific information. See Chapter 23 for more information on
networking.
Getting Connection Statistics
There is a command, connPoolStats, for finding out connection information about
mongoses and mongods. This gives you information about how many connections a
server has open, and to what:
> db.adminCommand({"connPoolStats" : 1})
{
"createdByType": {
"sync": 857,
"set": 4
},
"numDBClientConnection": 35,
"numAScopedConnection": 0,
"hosts": {
"config-01:10005,config-02:10005,config-03:10005": {
"created": 857,
"available": 2
},
"spock/spock-01:10005,spock-02:10005,spock-03:10005": {
"created": 4,
"available": 1
}
},
"totalAvailable": 3,
"totalCreated": 861,
"ok": 1
}
Hosts of the form "host1,host2,host3" are config server connections, also known as
“sync” connections. Hosts that look like "name/host1,host2,...,hostN" are connec‐
tions to shards. The "available" counts are how many connections are currently
available in the connection pools on this instance.
Note that this command only works on mongos processes and mongods that are mem‐
bers of a shard.
You may see connections to other shards in the output of connPoolStats from a shard,
as shards connect to other shards to migrate data. The primary of one shard will connect
directly to the primary of another shard and “suck” its data.
Tracking Network Connections | 283
When a migrate occurs, a shard sets up a ReplicaSetMonitor (a process that monitors
replica set health) to track the health of the shard on the other side of the migrate.
mongod never destroys this monitor, so you may see messages in one replica set’s log
about the members of another replica set. This is totally normal and should have no
effect on your application.
Limiting the Number of Connections
When a client connects to a mongos, mongos creates a connection to at least one shard
to pass along the client’s request. Thus, every client connection into a mongos yields at
least one outgoing connection from mongos to the shards.
If you have many mongos processes, they may create more connections than your shards
can handle: a mongos allows up to 20,000 connections (same as mongod), so if you have
5 mongos processes with 10,000 client connections each, they may be attempting to
create 50,000 connections to a shard!
To prevent this, you can use the maxConns option to your command line configuration
for mongos to limit the number of connections it can create. The following formula can
be used to calculate the maximum number of connections a shard can handle from a
single mongos:
maxConns = 20,000 − (numMongosProcesses × 3) − (numMembersPerRepli‐
caSet × 3) − (other / numMongosProcesses)
Breaking down the pieces of this formula:
(numMongosProcesses × 3)
Each mongos creates three connections per mongod: a connection to forward client
requests, an error-tracking connection (the writeback listener), and a connection
to monitor the replica set’s status.
(numMembersPerReplicaSet × 3)
The primary creates a connection to each secondary and each secondary creates
two connections to the primary, for a total of three connections.
(other / numMongosProcesses)
other is the number of miscellaneous processes that may connect to your mon
gods, such as MMS agents, direct shell connections (for administration), or con‐
nections to other shards for migrations.
Note that maxConns only prevents mongos from creating more than this many connec‐
tions. It doesn’t mean that it does anything particularly helpful when it runs out of
connections: it will block requests, waiting for connections to be “freed.” Thus, you must
284 | Chapter 16: Sharding Administration
prevent your application from using this many connections, especially as your number
of mongos processes grows.
When a MongoDB instance exits cleanly it closes all connections before stopping. The
members who were connected to it will immediately get socket errors on those con‐
nections and be able to refresh them. However, if a MongoDB instance suddenly goes
offline due to a power loss, crash, or network problems, it probably won’t cleanly close
all of its sockets. In this case, other servers in the cluster may be under the impression
that their connection is healthy until they try to perform an operation on it. At that
point, they will get an error and refresh the connection (if the member is up again at
that point).
This is a quick process when there are only a few connections. However, when there are
thousands of connections that must be refreshed one by one you can get a lot of errors
because each connection to the downed member must be tried, determined to be bad,
and re-established. There isn’t a particularly good way of preventing this aside from
restarting processes that get bogged down in a reconnection storm.
Server Administration
As your cluster grows, you’ll need to add capacity or change configurations. This section
covers how to add and remove servers from your cluster.
Adding Servers
You can add new mongos processes at any time. Make sure their --configdb option
specifies the correct set of config servers and they should be immediately available for
clients to connect to.
To add new shards, use the "addShard" command as show in Chapter 14.
Changing Servers in a Shard
As you use your sharded cluster, you may want to change the servers in individual
shards. To change a shard’s membership, connect directly to the shard’s primary (not
through the mongos) and issue a replica set reconfig. The cluster configuration will pick
up the change and update config.shards automatically. Do not modify config.shards by
hand.
The only exception to this is if you started your cluster with standalone servers as shards,
not replica sets.
Server Administration | 285
Changing a shard from a standalone server to replica set
The easiest way to do this is to add a new, empty replica set shard and then remove the
standalone server shard (see “Removing a Shard” on page 286).
If you wish to turn the standalone server into a replica set the process is fairly complex
and involves downtime:
1. Stop requests to the system.
2. Shut down the standalone server (call it server-1) and all mongos processes.
3. Restart the server-1 in replica set mode (with the --replSet option).
4. Connect to server-1 and initiate the set as a one-member replica set.
5. Connect to each config server and replace this shard’s entry in config.shards to have
a setName/server-1:27017 form for the shard name. Make sure all three config
servers have identical information. It is risky to manually edit config servers!
A good way of ensuring that they are identical is to run the dbhash command on
each config server:
> db.runCommand({"dbhash" : 1})
This comes up with an MD5 sum for each collection. Some collections in the config
database will be different on different config servers, but config.shards should not
be.
6. Restart all mongos processes. They will read the shard data off of the config servers
at start up and treat the replica set as a shard.
7. Restart all shards’ primaries to refresh their config data.
8. Send requests to the system again.
9. Add other members to server-1’s set.
This process is complex, error prone, and not recommended. If at all possible, just add
a new shard that’s an empty replica set and let migrations take care of moving your data
to it.
Removing a Shard
In general, shards should not be removed from a cluster. If you are regularly adding and
removing shards, you are putting a lot more stress on the system than necessary. If you
add too many shards it is better to let your system grow into it, not remove them and
add them back later. However, if necessary, you can remove shards.
First make sure that the balancer is on. The balancer will be tasked with moving all the
data on this shard to other shards in a process called draining. To start draining, run
286 | Chapter 16: Sharding Administration
the removeShard command. removeShard takes the shard’s name and drains all the
chunks on a given shard to the other shards:
> db.adminCommand({"removeShard" : "test-rs3"})
{
"msg" : "draining started successfully",
"state" : "started",
"shard" : "test-rs3",
"note" : "you need to drop or movePrimary these databases",
"dbsToMove" : [
"blog",
"music",
"prod"
],
"ok" : 1
}
Draining can take a long time if there are a lot of chunks or large chunks to move. If
you have jumbo chunks (see “Jumbo Chunks” on page 292), you may have to temporarily
raise the chunk size to allow draining to move them.
If you want to keep tabs on how much has been moved, run removeShard again to give
you the current status:
> db.adminCommand({"removeShard" : "test-rs3"})
{
"msg" : "draining ongoing",
"state" : "ongoing",
"remaining" : {
"chunks" : NumberLong(5),
"dbs" : NumberLong(0)
},
"ok" : 1
}
You can run removeShard as many times as you want.
Chunks may have to split to be moved, so you may see the number of chunks increase
in the system during the drain. For example, suppose we have a 5-shard cluster with the
following chunk distributions:
test-rs0 10
test-rs1 10
test-rs2 10
test-rs3 11
test-rs4 11
This cluster has a total of 52 chunks. If we remove test-rs3, we might end up with:
test-rs0 15
test-rs1 15
test-rs2 15
test-rs4 15
Server Administration | 287
The cluster now has 60 chunks, 18 of which came from shard test-rs3 (11 were there to
start and 7 were created from draining splits).
Once all the chunks have been moved, if there are still databases “homed” on the shard,
you’ll need to remove them before the shard can be removed. The output of remove‐
Shard will be something like:
> db.adminCommand({"removeShard" : "test-rs3"})
{
"msg" : "draining ongoing",
"state" : "ongoing",
"remaining" : {
"chunks" : NumberLong(0),
"dbs" : NumberLong(3)
},
"note" : "you need to drop or movePrimary these databases",
"dbsToMove" : [
"blog",
"music",
"prod"
],
"ok" : 1
}
To finish the remove, move the homed databases with the movePrimary command:
> db.adminCommand({"movePrimary" : "blog", "to" : "test-rs4"})
{
"primary " : "test-rs4:test-rs4/ubuntu:31500,ubuntu:31501,ubuntu:31502",
"ok" : 1
}
Once you have moved any databases, run removeShard one more time:
> db.adminCommand({"removeShard" : "test-rs3"})
{
"msg" : "removeshard completed successfully",
"state" : "completed",
"shard" : "test-rs3",
"ok" : 1
}
This is not strictly necessary, but it confirms that you have completed the process. If
there are no databases homed on this shard, you will get this response as soon as all
chunks have been migrated off.
Once you have started a shard draining, there is no built-in way to stop it.
Changing Config Servers
Changing anything about your config servers is difficult, dangerous, and generally in‐
volves downtime. Before doing any maintenance on config servers, take a backup.
288 | Chapter 16: Sharding Administration
All mongos processes need to have the same value for --configdb whenever they are
running. Thus, to change the config servers, you must shut down all your mongos
processes, make sure they are all down (no mongos process can still be running with the
old --configdb argument), and then restart them with the new --configdb argument.
For example, one of the most common tasks is to move from one config server to three.
To accomplish this, shut down your mongos processes, your config server, and all your
shards. Copy the data directory of your config servers to the two new config servers (so
that there is an identical data directory on all three servers). Now, start up all three config
servers and the shards. Then start each of the mongos processes with --configdb
pointing to all three config servers.
Balancing Data
In general, MongoDB automatically takes care of balancing data. This section covers
how to enable and disable this automatic balancing as well as how to intervene in the
balancing process.
The Balancer
Turning off the balancer is a prerequisite to nearly any administrative activity. There is
a shell helper to make this easier:
> sh.setBalancerState(false)
With the balancer off a new balancing round will not begin, but it will not force an
ongoing balancing round to stop immediately: migrations generally cannot stop on a
dime. Thus, you should check the config.locks collection to see whether or not a bal‐
ancing round is still in progress:
> db.locks.find({"_id" : "balancer"})["state"]
0
0 means the balancer is off. See “The Balancer” on page 253 for an explanation of the
balancer states.
Balancing puts load on your system: the destination shard must query the source shard
for all the documents in a chunk, insert them, and then the source shard must delete
them. There are two circumstances in particular where migrations can cause perfor‐
mance problems:
1. Using a hotspot shard key will force constant migrations (as all new chunks will be
created on the hotspot). Your system must have the capacity to handle the flow of
data coming off of your hotspot shard.
Balancing Data | 289
2. Adding a new shard will trigger a stream of migrations as the balancer attempts to
populate it.
If you find that migrations are affecting your application’s performance, you can sched‐
ule a window for balancing in the config.settings collection. Run the following update
to only allow balancing between 1 p.m. and 4 p.m.:
> db.settings.update({"_id" : "balancer"},
... {"$set" : {"activeWindow" : {"start" : "13:00", "stop" : "16:00"}}},
... true )
If you set a balancing window, monitor it closely to ensure that mongos can actually
keep your cluster balanced in the time that you have allotted it.
You must be careful if you plan to combine manual balancing with the automatic bal‐
ancer, since the automatic balancer always determines what to move based on the cur‐
rent state of the set and does not take into account the set’s history. For example, suppose
you have shardA and shardB, each holding 500 chunks. shardA is getting a lot of writes,
so you turn off the balancer and move 30 of the most active chunks to shardB. If you
turn the balancer back on at this point, it will immediately swoop in and move 30 chunks
(possibly a different 30) back from shardB to shardA to balance the chunk counts.
To prevent this, move 30 quiescent chunks from shardB to shardA before starting the
balancer. That way there will be no imbalance between the shards and the balancer will
be happy to leave things as they are. Alternatively, you could perform 30 splits on
shardA’s chunks to even out the chunk counts.
Note that the balancer only uses number of chunks as a metric, not size of data. Thus,
a shard with a few large chunks may end up as the target of a migration from a shard
with many small chunks (but a smaller data size).
Changing Chunk Size
There can be anywhere from zero to millions of documents in a chunk. Generally, the
larger a chunk is, the longer it takes to migrate to another shard. In Chapter 13, we used
a chunk size of 1 MB, so that we could see chunk movement easily and quickly. This is
generally impractical in a live system. MongoDB would be doing a lot of unnecessary
work to keep shards within a few megabytes of each other in size. By default, chunks
are 64 MB, which is generally a good balance between ease of migration and migratory
churn.
Sometimes you may find that migrations are taking too long with 64 MB chunks. To
speed them up, you can decrease your chunk size. To do this, connect to mongos through
the shell and update the config.settings collection:
> db.settings.findOne()
{
290 | Chapter 16: Sharding Administration
"_id" : "chunksize",
"value" : 64
}
> db.settings.save({"_id" : "chunksize", "value" : 32})
The previous update would change your chunk size to 32 MB. Existing chunks would
not be changed immediately, but as splits occurred chunks would trend toward that size.
mongos processes will automatically load the new chunk size value.
Note that this is a cluster-wide setting: it affects all collections and databases. Thus, if
you need a small chunk size for one collection and a large chunk size for another, you
may have to compromise with a chunk size in between the two ideals (or put the col‐
lections in different clusters).
If MongoDB is doing too many migrations or your documents are large, you may want
to increase your chunk size.
Moving Chunks
As mentioned earlier, all the data in a chunk lives on a certain shard. If that shard ends
up with more chunks than the other shards, MongoDB will move some chunks off it.
Moving a chunk is called a migration and is how MongoDB balances data across your
cluster.
You can manually move chunks using the moveChunk shell helper:
> sh.moveChunk("test.users", {"user_id" : NumberLong("1844674407370955160")},
... "spock")
{ "millis" : 4079, "ok" : 1 }
This would move the chunk containing the document with "user_id" of
1844674407370955160 to the shard named "spock". You must use the shard key to find
which chunk to move ("user_id", in this case). Generally, the easiest way to specify a
chunk is by its lower bound, although any value in the chunk will work (the upper bound
will not, as it is not actually in the chunk). This command will move the chunk before
returning, so it may take a while to run. The logs are the best place to see what it is doing
if it takes a long time.
If a chunk is larger than the max chunk size, mongos will refuse to move it:
> sh.moveChunk("test.users", {"user_id" : NumberLong("1844674407370955160")},
... "spock")
{
"cause" : {
"chunkTooBig" : true,
"estimatedChunkSize" : 2214960,
"ok" : 0,
"errmsg" : "chunk too big to move"
},
"ok" : 0,
Balancing Data | 291
"errmsg" : "move failed"
}
In this case, you must manually split the chunk before moving it, using the splitAt
command:
> db.chunks.find({"ns" : "test.users",
... "min.user_id" : NumberLong("1844674407370955160")})
{
"_id" : "test.users-user_id_NumberLong(\"1844674407370955160\")",
"ns" : "test.users",
"min" : { "user_id" : NumberLong("1844674407370955160") },
"max" : { "user_id" : NumberLong("2103288923412120952") },
"shard" : "test-rs2"
}
> sh.splitAt("test.ips", {"user_id" : NumberLong("2000000000000000000")})
{ "ok" : 1 }
> db.chunks.find({"ns" : "test.users",
... "min.user_id" : {"$gt" : NumberLong("1844674407370955160")},
... "max.user_id" : {"$lt" : NumberLong("2103288923412120952")}})
{
"_id" : "test.users-user_id_NumberLong(\"1844674407370955160\")",
"ns" : "test.users",
"min" : { "user_id" : NumberLong("1844674407370955160") },
"max" : { "user_id" : NumberLong("2000000000000000000") },
"shard" : "test-rs2"
}
{
"_id" : "test.users-user_id_NumberLong(\"2000000000000000000\")",
"ns" : "test.users",
"min" : { "user_id" : NumberLong("2000000000000000000") },
"max" : { "user_id" : NumberLong("2103288923412120952") },
"shard" : "test-rs2"
}
Once the chunk has been split into smaller pieces, it should be movable. Alternatively,
you can raise the max chunk size and then move it, but you should break up large chunks
whenever possible. Sometimes, though, they cannot be broken up: these are called
jumbo chunks.
Jumbo Chunks
Suppose you choose the "date" field as your shard key. The "date" field in this collection
is a string that looks like "year/month/day", which means that mongos can create at
most one chunk per day. This works fine for a while, until your application suddenly
goes viral and gets a thousand times its typical traffic for one day.
This day’s chunk is going to be much larger than any other day’s, but it is also completely
unsplittable because every document has the same value for the shard key.
292 | Chapter 16: Sharding Administration
Once a chunk is larger than the max chunk size set in config.settings, the balancer will
not be allowed to move the chunk. These unsplittable, unmovable chunks are called
jumbo chunks and they are inconvenient to deal with.
Let’s take an example. Suppose there are three shards, shard1, shard2, and shard3. If you
use the hotspot shard key pattern described in “Ascending Shard Keys” on page 258, all
your writes will be going to one shard, say shard1. mongos will try to balance the number
of chunks evenly between the shards. But the only chunks that the balancer can move
are the non-jumbo chunks, so it will migrate all the small chunks off the hotspot shard.
Now all the shards will have roughly the same number of chunks, but all of shard2 and
shard3’s chunks will be less than 64 MB in size. And if jumbo chunks are being created,
more and more of shard1’s chunks will be more than 64 MB in size. Thus, shard1 will
fill up a lot faster than the other two shards, even though the number of chunks is
perfectly balanced between the three.
Thus, one of the indicators that you have jumbo chunk problems is that one shard’s size
is growing much faster than the others. You can also look at sh.status() to see if you
have jumbo chunks: they will be marked with a "jumbo" attribute:
> sh.status()
...
{ "x" : -7 } -->> { "x" : 5 } on : shard0001
{ "x" : 5 } -->> { "x" : 6 } on : shard0001 jumbo
{ "x" : 6 } -->> { "x" : 7 } on : shard0001 jumbo
{ "x" : 7 } -->> { "x" : 339 } on : shard0001
...
You can use the dataSize command to check chunk sizes.
First, we use the config.chunks collection to find the chunk ranges:
> use config
> var chunks = db.chunks.find({"ns" : "acme.analytics"}).toArray()
Use these chunk ranges to find possible jumbo chunks:
> use dbName
> db.runCommand({"dataSize" : "dbName.collName",
... "keyPattern" : {"date" : 1}, // shard key
... "min" : chunks[0].min,
... "max" : chunks[0].max})
{ "size" : 11270888, "numObjects" : 128081, "millis" : 100, "ok" : 1 }
Be careful—the dataSize command does have to scan the chunk’s data to figure out how
big it is. If you can, narrow down your search by using your knowledge of your data:
were jumbo chunks created on certain date? For example, if there was a really busy day
on November 1, look for chunks with that day in their shard key range. If you’re using
GridFS and sharding by "files_id", you can look at the fs.files collection to find a file’s
size.
Balancing Data | 293
Distributing jumbo chunks
To fix a cluster thrown off-balance by jumbo chunks, you must evenly distribute them
among the shards.
This is a complex manual process, but should not cause any downtime (it may cause
slowness, as you’ll be migrating a lot of data). In the description below, the shard with
the jumbo chunks is referred to as the “from” shard. The shards that the jumbo chunks
are migrated to are called the “to” shards. Note that you may have multiple “from” shards
that you wish to move chunks off of. Repeat these steps for each:
1. Turn off the balancer. You don’t want to the balancer trying to “help” during this
process:
> sh.setBalancerState(false)
2. MongoDB will not allow you to move chunks larger than the max chunk size, so
temporarily raise the chunk size. Make a note of what your original chunk size is
and then change it to something large, like 10,000. Chunk size is specified in
megabytes:
> use config
> db.settings.findOne({"_id" : "chunksize"})
{
"_id" : "chunksize",
"value" : 64
}
> db.settings.save({"_id" : "chunksize", "value" : 10000})
3. Use the moveChunk command to move jumbo chunks off the “from” shard. If you
are concerned about the impact migrations will have on your application’s perfor‐
mance, use the secondaryThrottle option to prevent them from happening too
quickly:
> db.adminCommand({"moveChunk" : "acme.analytics",
... "find" : {"date" : new Date("10/23/2012")},
... "to" : "shard0002",
... "secondaryThrottle" : true})
secondaryThrottle forces migrations to periodically wait until a majority of sec‐
ondaries have replicated the migration. It only works if you are running with shards
that are replica sets (not standalone servers).
4. Run splitChunk on the remaining chunks on the donor shard until it has a roughly
even number of chunks as the other shards.
5. Set chunk size back to its original value:
> db.settings.save({"_id" : "chunksize", "value" : 64})
6. Turn on the balancer:
294 | Chapter 16: Sharding Administration
> sh.setBalancerState(true)
When the balancer is turned on again it cannot move the jumbo chunks again, as they
are essentially held in place by their size.
Preventing jumbo chunks
As the amount of data you are storing grows, the manual process described in the pre‐
vious section becomes unsustainable. Thus, if you’re having problems with jumbo
chunks, you should make it a priority to prevent them from forming.
To prevent jumbo chunks, modify your shard key to have more granularity. You want
almost every document to have a unique value for the shard key, or at least never have
more than chunksize-worth of data with a single shard key value.
For example, if you were using the year/month/day key described earlier it can quickly
be made finer-grained by adding hours, minutes, and seconds. Similarly, if you’re
sharding on something coarsely-grained key like log level, add a second field to your
shard key with a lot of granularity, such as an MD5 hash or UUID. Then you can always
split a chunk, even if the first field is the same for many documents.
Refreshing Configurations
As a final tip, sometimes mongos will not update its configuration correctly from the
config servers. If you ever get a configuration that you don’t expect or a mongos seems
to be out of date or cannot find data that you know is there, use the flushRouterConfig
command to manually clear all caches:
> db.adminCommand({"flushRouterConfig" : 1})
If flushRouterConfig does not work, restarting all your mongos or mongod processes
clears any possible cache.
Balancing Data | 295
PART V
Application Administration
CHAPTER 17
Seeing What Your Application Is Doing
Once you have an application up and running, how do you know what it’s doing? This
chapter covers how to figure out what kind of queries MongoDB is running, how much
data is being written, and how to investigate what MongoDB is actually doing. You’ll
learn:
• How to find slow operations and kill them
•Getting and interpreting statistics about your collections and databases
•Using command-line tools to give you a picture of what MongoDB is doing
Seeing the Current Operations
An easy way to find slow operations is to see what is running. Anything slow is more
likely to show up and have been running for longer. It’s not guaranteed, but it’s a good
first step to see what might be slowing down an application.
To see the operations that are running, use the db.currentOp() function:
> db.currentOp()
{
"inprog" : [
{
"opid" : 34820,
"active" : true,
"secs_running" : 0,
"op" : "query",
"ns" : "test.users",
"query" : {
"count" : "users",
"query" : {
"username" : "user12345"
},
299
"fields" : {
}
},
"client" : "127.0.0.1:39931",
"desc" : "conn3",
"threadId" : "0x7f12d61c7700",
"connectionId" : 3,
"locks" : {
"^" : "r",
"^test" : "R"
},
"waitingForLock" : false,
"numYields" : 0,
"lockStats" : {
"timeLockedMicros" : {
},
"timeAcquiringMicros" : {
"r" : NumberLong(9),
"w" : NumberLong(0)
}
}
},
...
]
}
This displays a list of operations that the database is performing. Here are some of the
more important fields in the output:
"opid"
This is the operation’s unique identifier. You can use this number to kill an operation
(see “Killing Operations” on page 301).
"active"
If this operation is running. If this field is false, it means the operation has yielded
or is waiting for a lock.
"secs_running"
How long this operation has been running. You can use this to find queries that are
taking too long or sucking up database resources.
"op"
The type of operation. This is generally a query, insert, update, or remove. Note
that database commands are processed as queries.
"desc"
This can be correlated with messages in the logs. Every log message related to this
connection will be prefixed with [conn3], so you can use this to grep the logs for
relevant information.
300 | Chapter 17: Seeing What Your Application Is Doing
"locks"
This describes the types of locks taken by this operation. "^" indicates the global
lock.
"waitingForLock"
Whether this operation is currently blocking, waiting to acquire a lock.
"numYields"
The number of times this operation has yielded, releasing its lock to allow other
operations to go. Generally, any operation that searches for documents (queries,
updates, and removes) can yield. An operation will only yield if there are other
operations enqueued and waiting to take its lock. Basically, if there are no operations
in "waitingForLock" state, the current operations will not yield.
"lockstats.timeAcquiringMicros"
This shows how long it took this operation to acquire the locks it needed.
You can filter currentOp() to only look for operations fulfilling certain criteria, such
as operations on a certain namespace or ones that have been running for a certain length
of time. Filter the results by passing in a query argument:
> db.currentOp({"ns" : "prod.users"})
You can query on any field in currentOp, using all the normal query operators.
Finding Problematic Operations
The most common use for db.currentOp() is looking for slow operations. You can use
the filtering technique described in the previous section to find all queries that take
longer than a certain amount of time, which may suggest a missing index or improper
field filtering.
Sometimes people will find that unexpected queries are running, generally because
there’s an app server running an old or buggy version of software. The "client" field
can help track down where unexpected operations are coming from.
Killing Operations
If you find an operation that you want to stop, you can kill it by passing db.killOp()
its "opid":
> db.killOp(123)
Not all operations can be killed. In general, operations can only be killed when they
yield, so updates, finds, and removes can all be killed. Operations holding or waiting
for a lock usually cannot be killed.
Seeing the Current Operations | 301
Once you have sent a “kill” message to an operation, it will have a "killed" field in the
db.currentOp output. However, it won’t actually be dead until it disappears from list of
current opertations.
False Positives
If you look for slow operations, you may see some long-running internal operations
listed. There are several long-running requests MongoDB may have running, depending
on your setup. The most common are the replication thread (which will continue fetch‐
ing more operations from the sync source for as long as possible) and the writeback
listener for sharding. Any long-running query on local.oplog.rs can be ignored as well
as any writebacklistener commands.
If you kill either of these operations, MongoDB will just restart them. However, you
generally should not do that. Killing the replication thread will briefly halt replication
and killing the writeback listener may cause mongos to miss legitimate write errors.
Preventing Phantom Operations
There is an odd, MongoDB-specific issue that you may run into, particularly if you’re
bulk-loading data into a collection. Suppose you have a job that is firing thousands of
update operations at MongoDB and MongoDB is grinding to a halt. You quickly stop
the job and kill off all the updates that are currently occurring. However, you continue
to see new updates appearing as soon as you kill the old ones, even though the job is no
longer running!
If you are loading data using unacknowledged writes, your application will fire writes
at MongoDB, potentially faster than MongoDB can process them. If MongoDB gets
backed up, these writes will pile up in the operating system’s socket buffer. When you
kill the writes MongoDB is working on, this allows MongoDB to start processing the
writes in the buffer. Even if you stop the client sending the writes, any writes that made
it into the buffer will get processed by MongoDB, since they’ve already been “received”
(just not processed).
The best way to prevent these phantom writes is to do acknowledged writes: make each
write wait until the previous write is complete, not just until the previous write is sitting
in a buffer on the database server.
Using the System Profiler
To find slow operations you can use the system profiler, which records operations in a
special system.profile collection. The profiler can give you tons of information about
operations that are taking a long time, but at a cost: it slows down mongod’s overall
performance. Thus, you may only want to turn on the profiler periodically to capture
302 | Chapter 17: Seeing What Your Application Is Doing
a slice of traffic. If your system is already heavily loaded, you may wish to use another
technique described in this chapter to diagnose issues.
By default, the profiler is off and does not record anything. You can turn it on by running
db.setProfilingLevel() in the shell:
> db.setProfilingLevel(2)
{ "was" : 0, "slowms" : 100, "ok" : 1 }
Level 2 means “profile everything.” Every read and write request received by the database
will be recorded in the system.profile collection of the current database. Profiling is
enabled per-database and incurs a heavy performance penalty: every write has to be
written an extra time and every read has to take a write lock (because it must write an
entry to the system.profile collection). However, it will give you an exhaustive listing of
what your system is doing:
> db.foo.insert({x:1})
> db.foo.update({},{$set:{x:2}})
> db.foo.remove()
> db.system.profile.find().pretty()
{
"ts" : ISODate("2012-11-07T18:32:35.219Z"),
"op" : "insert",
"ns" : "test.foo",
"millis" : 37,
"client" : "127.0.0.1",
"user" : ""
}
{
"ts" : ISODate("2012-11-07T18:32:47.334Z"),
"op" : "update",
"ns" : "test.foo",
"query" : {
},
"updateobj" : {
"$set" : {
"x" : 2
}
},
"nscanned" : 1,
"fastmod" : true,
"millis" : 3,
"client" : "127.0.0.1",
"user" : ""
}
{
"ts" : ISODate("2012-11-07T18:32:50.058Z"),
"op" : "remove",
"ns" : "test.foo",
"query" : {
Using the System Profiler | 303
},
"millis" : 0,
"client" : "127.0.0.1",
"user" : ""
}
You can use the "client" field to see which users are sending which operations to the
database. If we were using authentication, we could see which user was doing each
operation, too.
Often, you do not care about most of the operations that your database is doing, just
the slow ones. For this, you can set the profiling level to 1: profile only slow operations.
By default, level 1 profiles operations that take longer that 100 ms. You can also specify
a second argument, which defines what “slow” means to you. This would record all
operations that took longer than 500 ms:
> db.setProfilingLevel(1, 500)
{ "was" : 2, "slowms" : 100, "ok" : 1 }
To turn profiling off, set the profiling level to 0:
> db.setProfilingLevel(0)
{ "was" : 1, "slowms" : 500, "ok" : 1 }
Generally do not set slowms to a low value. Even with profiling off, slowms has an effect
on mongod: it sets the threshold printing slow operation in the log. Thus, if you set
slowms to 2, every operation that takes longer than 2 ms will show up in the log, even
with profiling off. Thus, if you lower slowms to profile something, you might want to
raise it again before turning off profiling.
You can see the current profiling level with db.getProfilingLevel(). The profiling
level is not persistent: restarting the database clears the level.
There are command-line options for configuring the profiling level, --profile lev
el and --slowms time, but bumping the profiling level is generally a temporary de‐
bugging measure, not something you want to add to your configuration long-term.
If you turn on profiling and the system.profile collection does not already exist, Mon‐
goDB creates a small capped collection for it (a few megabytes in size). If you want to
run the profiler for an extended period of time, this may not be enough space for the
number of operations you need to record. You can make a larger system.profile collection
by turning off profiling, dropping the system.profile collection, and creating a new
system.profile capped collection that is the size you desire. Then enable profiling on the
database.
304 | Chapter 17: Seeing What Your Application Is Doing
Calculating Sizes
In order to provision the correct amount of disk and RAM, it is useful to know how
much space documents, indexes, collections, and databases are taking up. See “Calcu‐
lating the Working Set” on page 348 for information on calculating your working set.
Documents
The easiest way to get the size of a document is to use the shell’s Object.bsonsize()
function. Pass in any document to get the size it would be when stored in MongoDB.
For example, you can see that storing _ids as ObjectIds is more efficient than storing
them as strings:
> Object.bsonsize({_id:ObjectId()})
22
> // ""+ObjectId() converts the ObjectId to a string
> Object.bsonsize({_id:""+ObjectId()})
39
More practically, you can pass in documents directly from your collections:
> Object.bsonsize(db.users.findOne())
This shows you exactly how many bytes a document is taking up on disk. However, this
does not count padding or indexes, which can often be significant factors in the size of
a collection.
Collections
For seeing information about a whole collection, there is a stats function:
> db.boards.stats()
{
"ns" : "brains.boards",
"count" : 12,
"size" : 32292,
"avgObjSize" : 2691,
"storageSize" : 270336,
"numExtents" : 3,
"nindexes" : 2,
"lastExtentSize" : 212992,
"paddingFactor" : 1.0099999999999825,
"flags" : 1,
"totalIndexSize" : 16352,
"indexSizes" : {
"_id_" : 8176,
"username_1_slug_1" : 8176
},
"ok" : 1
}
Calculating Sizes | 305
stats starts with the namespace ("brains.boards") and then the count of all docu‐
ments in the collection. The next couple of fields have to do with the size of the collection.
"size" is what you’d get if you called Object.bsonsize() on each element in the col‐
lection and added up all the sizes: it’s the actual number of bytes the document in the
collection are taking up. Equivalently, if you take the "avgObjSize" and multiply it by
"count", you’ll get "size".
As mentioned above, a total count of the documents’ bytes leaves out some important
space a collection uses: the padding around each document and the indexes. "storage
Size" not only includes those, but also empty space that has been set aside for the
collection but not yet used. Collections always have empty space at the “end” so that
new documents can be added quickly.
"nindexes" is the number of indexes on the collection. An index is not counted in
"nindexes" until it finishes being built and cannot be used until it appears in this list.
Each index is currently one “bucket” (8 KB), since the collection is so small. In general,
indexes will be a lot larger than the amount of data they store, as there is a lot of free
space to optimize adding new entries. You can minimize this free space by having right-
balanced indexes (as described in “Introduction to Compound Indexes” on page 84).
Indexes that are randomly distributed will generally be approximately 50% free space,
whereas ascending-order indexes will be 10% free space.
As your collections get bigger, it may become difficult to read stats() output with sizes
in the billions of bytes or beyond. Thus, you can pass in a scaling factor: 1024 for kilo‐
bytes, 1024*1024 for megabytes, and so on. For example, this would get the collection
stats in terabytes:
> db.big.stats(1024*1024*1024*1024)
Databases
Databases have a stats function that’s similar to collections’:
> db.stats()
{
"db" : "brains",
"collections" : 11,
"objects" : 1109,
"avgObjSize" : 299.79440937781783,
"dataSize" : 332472,
"storageSize" : 1654784,
"numExtents" : 15,
"indexes" : 11,
"indexSize" : 114464,
"fileSize" : 201326592,
"nsSizeMB" : 16,
"ok" : 1
}
306 | Chapter 17: Seeing What Your Application Is Doing
First, we have the name of the database and the number of collections it contains.
"objects" is the total count of documents across all collections in this database.
The bulk of the document contains information about the size of your data. "file
Size" should always be the largest: that is the total amount allocated to the database
files. This number should be equivalent to adding the sizes of all of the brains.* files in
your data directory.
The next largest field is generally going to be "storageSize", which is the total amount
of space the database is using. This doesn’t match "fileSize" because "fileSize"
includes preallocated files. For example, if you have brains.0, brains.1, and brains.2 in
your data directory, brains.2 will be filled with zeros. As soon as anything is written to
brains.2, brains.3 will be allocated. You should always have an empty (full of zeros) file
for each database: as soon as it’s written to, the next file will be preallocated. Thus, this
empty file (plus anything not being used yet in the previous files) is the difference be‐
tween "fileSize" and "storageSize".
"dataSize" is the amount of space the data in this database is taking up. Note that this
does not include space on the free list, but it does include documents’ padding. Thus,
the difference between this and the "storageSize" should be the size of documents
deleted.
db.stats() can take a scale argument the same way that the collections’ stats() func‐
tion can.
If you call db.stats() on a nonexistent database, the "nsSizeMB" will be 0. This is the
size of the .ns file, which is essentially a database’s table of contents. Any database that
exists needs a .ns file
Keep in mind that listing databases on a system with a high lock percent can be very
slow and block other operations. Avoid doing it, if possible.
Using mongotop and mongostat
MongoDB comes with a few command-line tools that can help you determine what
MongoDB is doing by printing stats every few seconds.
mongotop is similar to the top UNIX utility: it gives you an overview of which collections
are busiest. You can also run mongotop --locks to give you locking statistics for each
database.
mongostat gives server-wide information. By default, mongostat prints out a list of sta‐
tistics once per second, although this is configurable by passing a different number of
seconds on the command line. Each of the fields gives a count of how many times the
activity has happened since the field was last printed.
Using mongotop and mongostat | 307
insert/query/update/delete/getmore/command
These are simple counts of how many of each of these operations there have been.
flushes
How many times mongod has flushed data to disk.
mapped
The amount of memory mongod has mapped. This is generally roughly the size of
your data directory.
vsizeThe amount of virtual memory mongod is using. This is generally twice the size of
your data directory (once for the mapped files, once again for journaling).
res This is the amount of memory mongod is using. This should generally be as close
as possible to all the memory on the machine.
locked db
This shows the database that spent the most time locked in the last timeslice. It
reports the percent of time the database was locked combined with how long the
global lock was held, meaning that this value might be over 100%.
idx miss %
This is the most confusingly-named field in the output. It is how many index ac‐
cesses had to page fault: the index entry (or section of index being searched) was
not in memory, so mongod had to go to disk.
qr|qw
This is the queue size for reads and writes: how many reads and writes are blocking,
waiting to be processed.
ar|aw
How many active clients there are: clients currently performing reads and writes.
netIn
Number of network bytes in, as counted by MongoDB (not necessarily the same as
what the OS would measure).
netOut
Number of network bytes out, as counted by MongoDB.
conn
The number of connections this server has open, both incoming and outgoing.
timeThe time at which these statistics were taken.
308 | Chapter 17: Seeing What Your Application Is Doing
You can run mongostat on a replica set or sharded cluster. If you use the --discover
option, mongostat will try to find all the members of the set or cluster from the member
it initially connects to and will print one line per server per second for each. For a large
cluster, this can get unmanageable fast, but it can be useful for small clusters and tools
that can consume the data and present it in a more readable form.
mongostat is a great way to get a quick snapshot of what your database is doing, but for
long-term monitoring a tool like MMS is preferred (see Chapter 21).
Using mongotop and mongostat | 309
CHAPTER 18
Data Administration
This chapter covers how to administrate your collections and databases. Generally the
things covered in this section are not daily tasks but can be critically important for your
application’s performance, for instance:
• Setting up authentication and user accounts
• Creating indexes on a running system
•“Preheating” a new server to allow it to come online quickly
•Defragmenting data files
•Preallocating new data files manually
Setting Up Authentication
One of the first priorities for systems administrators is to ensure their system is secure.
The best way to handle security with MongoDB is to run it in a trusted environment,
ensuring that only trusted machines are able to connect to the server. That said, Mon‐
goDB supports per-connection authentication, albeit with a fairly coarse-grained per‐
missions scheme.
There are more sophisticated security features in MongoDB Enter‐
prise. See http://docs.mongodb.org/manual/security for the most up-to-
date information about authentication and authorization.
311
Authentication Basics
Each database in a MongoDB instance can have any number of users. When security is
enabled, only authenticated users of a database are able to perform read or write
operations.
There are two special databases: users in the admin and local databases can perform
operations on any database. A user that belongs to either one of these databases can be
thought of as a superuser. After authenticating, admin users are able to read or write
from any database and are able to perform certain admin-only commands, such as
listDatabases or shutdown.
Before starting the database with security turned on, it’s important that at least one
admin user has been added. Let’s run through a quick example, starting from a shell
connected to a server without authentication turned on:
> use admin
switched to db admin
> db.addUser("root", "abcd");
{
"user" : "root",
"readOnly" : false,
"pwd" : "1a0f1c3c3aa1d592f490a2addc559383"
}
> use test
switched to db test
> db.addUser("test_user", "efgh");
{
"user" : "test_user",
"readOnly" : false,
"pwd" : "6076b96fc3fe6002c810268702646eec"
}
> db.addUser("read_user", "ijkl", true);
{
"user" : "read_user",
"readOnly" : true,
"pwd" : "f497e180c9dc0655292fee5893c162f1"
}
Here we’ve added an admin user, root, and two users on the test database. One of those
users, "read_only", has read permissions only and cannot write to the database. From
the shell, a read-only user is created by passing true as the third argument to addUser.
To call addUser, you must have write permissions for the database in question; in this
case we can call addUser on any database because we have not enabled security yet.
312 | Chapter 18: Data Administration
The addUser method is useful for more than just adding new users: it
can be used to change a user’s password or read-only status. Just call
addUser with the username and a new password or read-only setting
for the user.
Now let’s restart the server, this time adding the --auth command-line option to enable
security. After enabling security, we can reconnect from the shell and try it:
> use test
switched to db test
> db.test.find();
error: { "$err" : "unauthorized for db [test] lock type: -1 " }
> db.auth("read_user", "ijkl");
1
> db.test.find();
{ "_id" : ObjectId("4bb007f53e8424663ea6848a"), "x" : 1 }
> db.test.insert({"x" : 2});
unauthorized
> db.auth("test_user", "efgh");
1
> db.test.insert({"x": 2});
> db.test.find();
{ "_id" : ObjectId("4bb007f53e8424663ea6848a"), "x" : 1 }
{ "_id" : ObjectId("4bb0088cbe17157d7b9cac07"), "x" : 2 }
> show dbs
assert: assert failed : listDatabases failed:{
"assertion" : "unauthorized for db [admin] lock type: 1
",
"errmsg" : "db assertion failure",
"ok" : 0
}
> use admin
switched to db admin
> db.auth("root", "abcd");
1
> show dbs
admin
local
test
When we first connect, we are unable to perform any operations (read or write) on the
test database. After authenticating as the read_user user, however, we are able to per‐
form a simple find. When we try to insert data, we are again met with a failure because
of the lack of authorization. test_user, which was not created as read-only, is able to
insert data normally. As a non-admin user, though, test_user is not able to list all the
available databases using the show dbs helper. The final step is to authenticate as an
admin user, root, who is able to perform operations on any database.
Setting Up Authentication | 313
Setting Up Authentication
If authentication is enabled, clients must be logged in to perform any reads or writes.
However, there is one oddity in MongoDB’s authentication scheme: before you create
a user in the admin database, clients that are “local” to the server can perform reads and
writes on the database.
Generally, this is not an issue: create your admin user and use authentication normally.
The only exception is sharding. With sharding, the admin database is kept on the config
servers, so shard mongods have no idea it even exists. Therefore, as far as they know,
they are running with authentication enabled but no admin user. Thus, shards will allow
a local client to read and write from them without authentication.
Hopefully this wouldn’t be an issue: optimally your network will be configured so that
only mongos processes are accessible to clients. However, if you are worried about clients
running locally on shards and connecting directly to them instead of going through the
mongos, you may wish to add admin users to your shards.
Note that you do not want the sharded cluster to know about these admin users: it
already has an admin database. The admin databases you’re creating on the shards are
for your use only. To do this, connect to the primary of each shard and run the addUs
er() function:
> db.addUser("someUser", "theirPassword")
Make sure that the replica sets you create users on are already shards in the cluster. If
you create an admin user and then try to add the mongods as a shard the addShard
command will not work (because the cluster already contains an admin database).
How Authentication Works
Users of a given database are stored as documents in its system.users collection. The
structure of a user document is {"user" : username, "readOnly" : true, "pwd" :
password hash}. The password hash is a hash based on the username and password
chosen.
Knowing where and how user information is stored makes performing some common
administration tasks trivial. For example, to remove a user, simply remove the user
document from the system.users collection:
> db.auth("test_user", "efgh");
1
> db.system.users.remove({"user" : "test_user"});
> db.auth("test_user", "efgh");
0
When a user authenticates, the server keeps track of that authentication by tying it to
the connection used for the authenticate command. This means that if a driver or tool
314 | Chapter 18: Data Administration
is employing connection pooling or fails over to another node, authenticated users will
need to reauthenticate on new connections. This should be handled invisibly by the
driver.
Creating and Deleting Indexes
Chapter 5 covered what commands to run to create an index, but it didn’t go into the
operational aspects of doing so. Creating an index is one of the most resource-intensive
operations you can do on a database, so schedule index creations carefully.
Building an index requires MongoDB to find the indexed field (or lack thereof) in every
document in the collection, then sort all the values found. As you might imagine, this
becomes a very expensive task as your collection gets bigger. Thus, indexing should be
done in a way that affects your production server as little as possible.
Creating an Index on a Standalone Server
On a standalone server, build the index in the background during an off-time. There
isn’t much else you can do to minimize impact. To build an index in the background,
run ensureIndex with the "background" : true option:
> db.foo.ensureIndex({"someField" : 1}, {"background" : true})
Any type of index can be built in the background.
A foreground index build takes less time than a background index build but locks the
database for the duration of the process. Thus, no other operations can read or write to
the database while a foreground index is building. Background indexes yield the write
lock regularly to allow other operations to go. This means that they can take longer—
much longer on write-heavy servers—but the server can serve other clients while build‐
ing an index.
Creating an Index on a Replica Set
The easiest way to create an index on a replica set is to create it on the primary and wait
for it to be replicated to the secondaries. On small collections, this should have minimal
impact.
If you have a large collection, this can lead to the situation where all of your secondaries
start building the index at the same time. Suddenly all of your secondaries will be un‐
available for client reads and may fall behind in replication. Thus, for larger collections,
this is the preferred technique:
1. Shut down a secondary.
2. Start it up as a standalone node, as described in Chapter 6.
Creating and Deleting Indexes | 315
3. Build the index on that server.
4. Reintroduce the member into the replica set.
5. Repeat for each secondary.
When you have finished this process, only the primary should be left without an index.
Now there are two options: you can either build the index on the primary in the back‐
ground (which will put more strain on the primary) or you can step down the primary
and then follow steps 1 through 4 to build the index on it as you did with the other
members of the set. This involves a failover, which may be more or less preferable than
adding load to the primary.
You can also use this isolate-and-build technique to build an index on a member of the
set that isn’t configured to build indexes (one that has the "buildIndexes" : false
option set): start it as a standalone member, build the index, and add it back into the
set.
If you cannot use the rotation method for whatever reason, plan to build new indexes
during an off time (at night, a holiday, a weekend, etc.).
Creating an Index on a Sharded Cluster
To create indexes on a sharded cluster, we want to follow the same procedure as de‐
scribed for replica sets and build the index on one shard at a time.
First, turn off the balancer. Then follow the procedure outlined in the previous section
for each shard, treating it as a individual replica set. Finally, run ensureIndex through
the mongos and turn the balancer back on again.
This procedure is only required for adding an index to existing shards: new shards will
pick up on the index when they start receiving chunks from a collection.
Removing Indexes
If an index is no longer necessary, you can remove it with the dropIndexes command
and the index name. Query the system.indexes collection to figure out what the index
name is, as even the autogenerated names vary somewhat from driver to driver:
> db.runCommand({"dropIndexes" : "foo", "index" : "alphabet"})
You can drop all indexes on a collection by passing in "*" as the value for the "index"
key:
> db.runCommand({"dropIndexes" : "foo", "index" : "*"})
This leaves the "_id" index. The only way to get rid of that one is to drop the whole
collection. Removing all the documents in a collection (with remove) does not affect the
indexes; they will be repopulated when new documents are inserted.
316 | Chapter 18: Data Administration
Beware of the OOM Killer
The Linux out-of-memory killer will kill processes that are using a lot of memory. Be‐
cause of the way MongoDB uses memory, it is not usually an issue, but index creations
are the one time it can be. If you are building an index and your mongod suddenly
disappears, check /var/log/messages for notices from the OOM killer. Running a back‐
ground index build or adding some swap space can prevent this. If you have adminis‐
trative permissions on the machine, you may want to make MongoDB unkillable.
See “The OOM Killer” on page 383 for more information.
Preheating Data
When you restart a machine or bring up a new server, it can take a while for MongoDB
to get all the right data from disk into memory. If your performance constraints require
that data be in RAM, it can be disastrous to bring a new server online and then let your
application hammer it while it gradually pages in the data it needs.
There are a couple of ways to get your data into RAM before officially bringing the
server online, to prevent it from messing up your application.
Restarting MongoDB does change what’s in RAM. RAM is managed by
the OS and the OS won’t evict data from memory until the space is
needed for something else. Thus, if the mongod process needs to be
restarted it should not affect what data is in memory. (However, mon
god will report low resident memory until it has a chance to ask the OS
for the data it needs.)
Moving Databases into RAM
If you need a database in RAM, you can use the UNIX dd tool to load it before starting
the mongod:
$ for file in /data/db/brains.*
> do
> dd if=$file of=/dev/null
> done
Replace brains with the name of the database you want to load.
Replacing /data/db/brains.* with /data/db/* will load the whole data directory (all da‐
tabases) into RAM (assuming there’s enough room for all of them). If you load a database
or set of databases into memory and it takes up more memory than you have, some of
its data will fall back out of memory immediately. In this situation, you may want to use
Preheating Data | 317
one of the techniques outlined in the next section to move more specific parts of your
data into memory.
When you start the mongod, it will ask the operating system for the data files and the
operating system, knowing that the data files are in memory, will be able to quickly
access it.
However, this technique is only helpful if your entire database fits in memory. If it does
not, you can do more fine-grained preheating using the following techniques.
Moving Collections into RAM
MongoDB has a command for preheating data called touch. Start mongod (perhaps on
a different port or firewalled off from your application) and touch a collection to load
it into memory:
> db.runCommand({"touch" : "logs", "data" : true, "index" : true})
This will load all the documents and indexes into memory for the logs collection. You
can specify to only load documents or only load indexes. Once touch completes, you
can allow your application to access MongoDB.
However, an entire collection (even just the indexes) can still be too much data. For
example, your application might only require one index to be in memory, or only a small
fraction of the documents. In that case, you’ll have to custom preheat the data.
Custom-Preheating
When you have more complex preheating needs, you may have to roll your own pre‐
heating scripts. Here are some common preheating requirements and how to deal with
them:
Load a specific index
Let’s say we have an index such as {"friends" : 1, "date" : 1} that must be in
RAM. We can load this index into memory by doing a covered query (see “Using
covered indexes” on page 90):
> db.users.find({}, {"_id" : 0, "friends" : 1, "date" : 1}).
... hint({"friends" : 1, "date" : 1}).explain()
The explain forces mongod to iterate through all of the results for you. You must
specify that you only want to return the indexed fields (the second argument to
find) or the query will end up loading all the documents into memory, too (which
you may want, but it’s good to be aware of). Note that this will always load the index
and documents into memory for indexes that cannot be covered, such as multikey
indexes.
318 | Chapter 18: Data Administration
Load recently updated documents
If you have an index on a date field that you update when you update the document,
you can query on recent dates to load recent documents.
If you don’t have an index on the date field, this query will end up loading all
documents in the collection into memory, so don’t bother. If you don’t have a date
field, you might be able to use the "_id" field if you’re mostly concerned with recent
inserts (see below).
Load recently created documents
If you are using ObjectIds for your "_id" field, you can use the fact that recently
created documents contain a timestamp to query for them. For example, suppose
we wanted to find all documents created in the last week. We could create an "_id"
that was less than every ObjectId created in the last week:
> lastWeek = (new Date(year, month, day)).getTime()/1000
1348113600
Replace year, month, and date appropriately and this gives you the date, in seconds.
Now we need to get an ObjectId from this time. First, convert it into a hex string,
then append 16 zeros to it:
> hexSecs = lastWeek.toString(16)
505a94c0
> minId = ObjectId(hexSecs+"0000000000000000")
ObjectId("505a94c00000000000000000")
Now we just have to query for it:
> db.logs.find({"_id" : {"$gt" : minId}}).explain()
This will load all the docs (and some right-hand branches of the "_id" index) from
the last week.
Replay application usage
MongoDB has a facility for recording and replaying traffic called the diaglog. En‐
abling the diaglog incurs a performance penalty, so it is best to use it temporarily
to gather a “representative” slice of traffic. To gather traffic, run the following com‐
mand in the mongo shell:
> db.adminCommand({"diagLogging" : 2})
The 2 option means “capture reads.” The 1 option will capture writes and the 3
option captures both (0 is the default: off). You probably don’t want to capture writes
because you don’t want extra writes applied to your new member when you replay
the diaglog.
Now let the diaglog record operations by letting mongod run for however long you
want while sending it traffic. Reads will be stored in the diaglog file in the data
directory. Reset diagLogging to 0 when you’re done:
Preheating Data | 319
> db.adminCommand({"diagLogging" : 0})
To use your diaglog files, start up your new server and, from the server where the
diaglog files live, run:
$ nc hostname 27017 < /data/db/diaglog* | hexdump -c
Replace the IP, port, and data directory, if necessary. This sends the recorded op‐
erations to hostname:27017 as a series of normal queries.
Note that the diaglog will capture the command turning on the diaglog, so you’ll
have to log into the server and turn it off when you’re done replaying the diaglog
(you also might want to delete the diaglog files it generates from the replay).
These techniques can be combined: you could load a couple of indexes while replaying
the diaglog, for example. You can also run them all at the same time if you aren’t bot‐
tlenecked on disk IO, either through multiple shells or the startParallelShell com‐
mand (if the shell is local to the mongod):
> p1 = startParallelShell("db.find({}, {x:1}).hint({x:1}).explain()", port)
> p2 = startParallelShell("db.find({}, {y:1}).hint({y:1}).explain()", port)
> p3 = startParallelShell("db.find({}, {z:1}).hint({z:1}).explain()", port)
Replace port with the port on which mongod is running.
Compacting Data
MongoDB uses a lot of disk space. Sometimes, if you have deleted or updated a lot of
data, you’ll end up with collection fragmention. Fragmentation occurs when your data
files have a lot of empty space that MongoDB can’t reuse because the individual chunks
of free space are too small. In this case, you’ll see messages like this in the log:
Fri Oct 7 06:15:03 [conn2] info DFM::findAll(): extent 0:3000 was empty,
skipping ahead. ns:bar.foo
This message is, in and of itself, harmless. However, it means that an entire extent had
no documents in it. To get rid of empty extents and repack collections efficiently, use
the compact command:
> db.runCommand({"compact" : "collName"})
Compaction is very resource-intensive: you should not plan to do a compaction on a
mongod that’s serving clients. Thus, the recommended procedure is similar to that of
building indexes: compact data on each of the secondaries, then step down the primary
and run the final compaction on it.
When you run a compact on a secondary it will drop into recovering state, which means
that it will return errors if sent read requests and it cannot be used as a sync source.
When the compaction is finished, it will return to secondary state.
320 | Chapter 18: Data Administration
Compaction fits documents as closely as it can, as though the padding factor was 1. If
you need a higher padding factor for the collection, you can specify it as an argument
to compact:
> db.runCommand({"compact" : "collName", "paddingFactor" : 1.5})
You can specify a padding factor between 1 and 4. This does not permanently affect the
padding factor, just how MongoDB rearranges documents. Once the compaction is
finished, the padding factor will go back to whatever it was originally.
Compacting does not decrease the amount of disk space a collection uses: it just puts
all of the documents at the “beginning” of a collection, on the assumption that the
collection will expand again to use the available space. Thus, compaction is only a brief
reprieve if you are running out of disk space: it will not decrease the amount of disk
space MongoDB is using, although it may make MongoDB not need to allocate new
space for longer.
You can reclaim disk space by running a repair. Repair makes a full copy of the data so
you must have free space equal to the size of your current data files. This is often an‐
noying, as the most common reason to do a repair is that their machine is running low
on disk space. However, if you can mount another disk, you can specify a repair path,
which is the directory (your newly-mounted drive) that repair should use for the new
copy of the data.
Since it makes a totally clean copy of your data you can interrupt a repair at any time
with no effect on your original data set. If you run into the problem in the middle of a
repair, you can delete the temporary repair files without affecting your actual data files.
To repair, start mongod with the --repair option (and --repairpath, if desired).
You can run repair on a single database in the shell by calling db.repairDatabase().
Moving Collections
You can rename a collection using the renameCollection command. This cannot move
collections between databases, but it can change the collection’s name. This operation
is almost instantaneous, regardless of the size of the collection being renamed. On busy
systems, it can take a few nerve-wracking seconds, but it can be performed in production
with no performance penalty:
> db.sourceColl.renameCollection("newName")
You can optionally pass a second argument: what to do with the newName collection if
it already exists. The options are true (drop it) or false (the default: error out if it
exists).
Moving Collections | 321
To move a collection to another database, you must either dump/restore it or manually
copy the documents (do a find and iterate over the results, inserting them into the new
database).
You can move a collection to a different mongod using the cloneCollection command:
> db.runCommand({"cloneCollection" : "collName", "from" : "hostname:27017"})
You cannot use cloneCollection move a collection within a mongod: it can only move
collections between servers.
Preallocating Data Files
If you know that your mongod will need certain data files, you can run the following
script to preallocate them before your application goes online. This can be especially
helpful for the oplog (which you know will be a certain size) and any databases that you
know will be a given size, at least for a while:
#!/bin/bash
# Make sure db name was passed in
if test $# -lt 2 || test $# -gt 3
then
echo "$0 <db> <number-of-files>"
fi
db=$1
num=$2
for i in {0..$num}
do
echo "Preallocating $db.$i"
head -c 2146435072 /dev/zero > $db.$i
done
Store this in a file (say, preallocate), and make the file executable. Go to your data di‐
rectory and allocate the files that you need:
$ # create test.0-test.100
$ preallocate test 100
$
$ # create local.0-local.4
$ preallocate local 4
Once you start the database and it accesses the datafiles for the first time, you cannot
delete any of the data files. For example, if you allocated data files test.0 through test.
100 and then start the database and realize that you only need test.0 through test.20, you
should not delete test.21-test.100. Once MongoDB is aware of them, it will be unhappy
if they go away.
322 | Chapter 18: Data Administration
CHAPTER 19
Durability
Durability is the guarantee that an operation that is committed will survive permanently.
MongoDB has highly configurable durability settings, from absolutely no guarantees
to completely durable. This section covers:
• How MongoDB guarantees durability
•How to configure your application and server to give you the level of durability you
need
•The implications of running without journaling enabled
•What MongoDB does not guarantee
MongoDB can ensure data integrity after crashes and hard shutdowns, assuming disk
and software are behaving correctly.
Note that relational databases usually use durability to describe transaction persistence.
As MongoDB does not support transactions, it is used a bit differently here.
What Journaling Does
As you perform writes, MongoDB creates a journal that contains the exact disk location
and bytes changed for that write. Thus, if the server suddenly stops, on startup the
journal can be used to replay any writes that were not flushed to disk before the shut‐
down.
The data files are flushed to disk every 60 seconds (by default), so the journal only needs
to hold around 60 seconds of write data. Journaling preallocates a couple of empty files
for this purpose that it puts in /data/db/journal, with the names _j.0, _j.1, and so on.
After you’ve been running MongoDB for a long time, you may look in your journal
directory and see something like _j.6217, _j.6218, and _j.6219. These are the current
323
journal files. The numbers will continue to increase the longer MongoDB is running.
On clean shutdown, the journal files will be removed (as there is no need for them after
a clean shutdown).
If there is a crash (or kill -9), mongod will replay its journal files on startup, spitting
out a lot of lines of checksumming. These lines are verbose and indecipherable, but they
are an indication that everything is working as it should. You may want to try running
kill -9 on mongod in development so you know what to expect on restart if it happens
in production.
Planning Commit Batches
By default, MongoDB writes to the journal every 100 ms once a few megabytes of data
have been written, whichever comes sooner. This means that MongoDB commits
changes in batches: every write isn’t flushed to disk immediately, but with the default
settings you are unlikely to lose more than 100 ms of writes in the event of a crash.
However, this guarantee is not strong enough for some applications, so there are several
ways to get stronger durability guarantees. You can ensure a write has been durably
written by passing the j option to getLastError. getLastError will wait for the previous
write to be written to the journal and journaling will only wait 30 ms (instead of 100
ms) to journal the next batch of writes:
> db.foo.insert({"x" : 1})
> db.runCommand({"getLastError" : 1, "j" : true})
> // The {"x" : 1} document is now safely on disk
Note that this means that, if you use "j" : true as an option on every write, your write
speed will essentially be throttled to 33 writes/sec:
(1 write/30ms) × (1000ms/second) = 33.3 writes/second
It generally doesn’t take this long to flush writes to disk, so you’ll find your write per‐
formance improves if you allow MongoDB to batch most writes instead of committing
after every one. However, the option is always there for important writes.
Committing a write commits all previous writes, as well. Thus, if you have 50 important
writes, you could use “normal” getLastError (without the j option) and then call it with
the j option after the final write. If that succeeds, you know that all 50 writes are flushed
safely to disk.
If you have many connections with incoming writes, you can mitigate the speed penalty
of using j by having many writes happen in parallel. This can increase throughput, even
though latency is high.
324 | Chapter 19: Durability
Setting Commit Intervals
Another option for making journaling less intrusive is that you can shorten (or lengthen)
the amount of time between journal commits. Run the setParameter command to set
the journalCommitInterval to a value between 2 ms and 500 ms. This would commit
to the journal every 10 ms:
> db.adminCommand({"setParameter" : 1, "journalCommitInterval" : 10})
This can also be set as a command line option: --journalCommitInterval.
Regardless of the interval set, calling getLastError with "j" : true will cut the time to
a third of the time set.
If clients try to write faster than the journal can flush, mongod will block writes until
the journal finishes writing to disk. This is the only time that mongod will throttle writes.
Turning Off Journaling
Journaling is recommended for all production deployments, but in some cases you may
wish to turn it off. Journaling impacts the speed with which MongoDB can write, even
without the j option. If value of the data is not worth the speed penalty, you may wish
to disable journaling.
The downside to disabling journaling is that MongoDB has no way of ensuring data
integrity after a crash. After a crash without journaling, you should assume that your
data is corrupt and must either be repaired or replaced. You should not use data after a
crash without journaling unless you don’t care if your database suddenly stops working.
Assuming that you would prefer your database to continue to work after a crash, there
are a few options.
Replacing Data Files
This is your best option. Delete all of the files in the data directory and get new ones:
restore from a backup, take a snapshot of a known-clean member, or initial sync the
server if it’s a member of a replica set. If you have a replica set with a small amount of
data, resyncing is probably your best option: stop the member (if it’s not already down),
delete everything in its data directory, and start it back up again.
Repairing Data Files
If you have no backups, no copies, and no other members of the set, you’ll need to make
due with whatever data can be salvaged. You need to use a “repair” tool on the database,
but repairs are really corruption-ectomies: they’ll remove any corruption, but you may
not have much clean data left.
Turning Off Journaling | 325
There are two repair tools that come with mongod: the repair built into mongod itself
and a more hardcore repair built into mongodump. The mongodump repair may find
more of your data, but it takes a long time to do it (and the built-in repair is not exactly
speedy). Additionally, if you use mongodump’s repair, you’ll still need to restore the data
before you’re ready to start up again.
Thus, you should judge how much time you’re willing to devote to data recovery and
choose accordingly.
To use the repair built into mongod, run mongod with the --repair option:
$ mongod --dbpath /path/to/corrupt/data --repair
MongoDB will not start listening on 27017 when running a repair, but you can watch
the log to see what it is doing. Note that repair makes a complete copy of your existing
data, so if you have 80 GB of data, so you need 80 GB of free disk space. To help with
this (a bit), repair supports a --repairpath option. This lets you mount an “emergency
drive” and repair your data onto that if you don’t have enough space left on your primary
disk. Running with --repairpath looks like:
$ mongod --dbpath /path/to/corrupt/data \
> --repair --repairpath /media/external-hd/data/db
If repair gets killed or errors out (say, runs out of disk space), you won’t be in any worse
of a situation. Repair writes all of its output to new files, not changing your original files
until the very end, so your original data files will be in no worse shape than when you
started the repair.
The other option is using the --repair option on mongodump. This looks like:
$ mongodump --repair
Neither of these are particularly great options, but they should let you get mongod
running with a clean dataset again.
The mongod.lock File
There is a special file called mongod.lock in your data directory that is important when
you’re running without journaling (if you are running with journaling, it should never
come up).
When mongod exits it clears mongod.lock so that mongod knows on startup that it shut
down cleanly. Conversely, if the lock file was not cleaned out, mongod knows that it
existed uncleanly.
If mongod detects that it previously exited uncleanly, it will not allow you to start up
again so that you know you have to get a clean copy of your data. However, some people
have realized that you can get around this check by deleting the lock file before starting
mongod. Please don’t do this. Routinely deleting the lock file on startup means that you
326 | Chapter 19: Durability
don’t know and don’t care whether your data is corrupt. Unless this is the case, respect
the lock file. If it prevents mongod from starting, fix your data, and don’t delete mon
god.lock.
Sneaky Unclean Shutdowns
One important reason to not delete the lock file is that you may not even notice a hard
crash. Suppose that you reboot a machine for some routine maintenance. The init scripts
should take care of shutting down mongod before the server shuts down. However, init
systems will generally try shutting down a process gently and, if it doesn’t shut down
after a couple of seconds, hard-kill it. On a busy system, MongoDB can easily take 30
seconds to shut down: it is the unusual init script that will wait for it. Thus, you may
have more hard shutdowns than you’re aware of.
What MongoDB Does Not Guarantee
There are a couple of situations where MongoDB cannot guarantee durability, such as
if there are hardware issues or filesystem bugs. In particular, if a hard disk is corrupt,
there is nothing MongoDB can do to protect your data.
Also, different varieties of hardware and software may have different durability guar‐
antees. For example, some cheaper or older hard disks report a write’s success while the
write is queued up to be written, not when it has actually been written. MongoDB cannot
defend against misreporting at this level: if the system crashes, data may be lost.
Basically, MongoDB is only as safe as the underlying system: if the hardware or filesys‐
tem destroys the data, MongoDB cannot prevent it. Use replication to defend against
system issues. If one machine fails, hopefully another will still be functioning correctly.
Checking for Corruption
The validate command can be used to check a collection for corruption. To run validate
on the foo collection, do:
> db.foo.validate()
{
"ns" : "test.foo",
"firstExtent" : "0:2000 ns:test.foo",
"lastExtent" : "1:3eae000 ns:test.foo",
"extentCount" : 11,
"datasize" : 75960008,
"nrecords" : 1000000,
"lastExtentSize" : 37625856,
"padding" : 1,
"firstExtentDetails" : {
"loc" : "0:2000",
What MongoDB Does Not Guarantee | 327
"xnext" : "0:f000",
"xprev" : "null",
"nsdiag" : "test.foo",
"size" : 8192,
"firstRecord" : "0:20b0",
"lastRecord" : "0:3fa0"
},
"deletedCount" : 9,
"deletedSize" : 31974824,
"nIndexes" : 2,
"keysPerIndex" : {
"test.foo.$_id_" : 1000000,
"test.foo.$str_1" : 1000000
},
"valid" : true,
"errors" : [ ],
"warning" : "Some checks omitted for speed. use {full:true}
option to do more thorough scan.",
"ok" : 1
}
The main field you’re looking for is "valid", near the end, which will hopefully be true.
If it is not, validate will give some details about the corruption it found.
Most of the output from validate describes internal structures of the collection, which
are not particularly useful for debugging. See Appendix B for more information on
collection internals.
firstExtent
The disk offset of first extent in this collection. It is located in the file test.0 at byte
offset 0x2000.
lastExtent
The offset of the last extent in this collection. It is in test.1 at byte offset 0x3eae000.
extentCount
The number of extents in the collection.
lastExtentSize
The size in bytes of the most-recently-allocated extent. Extents get larger as a col‐
lection expands, growing up to 2 GB in size.
firstExtentDetails
A subobject describing the first extent in the collection. It contains pointers to the
previous and next extents ("xnext" and "xprev"), the extent’s size (note how much
smaller it is than the last extent in the collection: the first extent will generally be
quite small), and pointers to the first and last record in the extent. The records are
the structures that actually hold documents.
328 | Chapter 19: Durability
deletedCount
The number of documents that have been removed from this collection during its
lifetime.
deletedSize
The size of the free list (all of the free space available) for this collection. This is not
just documents that have been deleted but all the space that has been preallocated,
as well.
You can only run validate on collections, not indexes, so you generally cannot tell if an
index is corrupt unless you walk it. Walk an index by running a query for every docu‐
ment in the collection hinting the desired index. Then traverse all results.
If you get an assertion about invalid BSONObj, this is usually corruption. The worst
errors are those that mention pdfile. pdfile is basically the core of MongoDB’s data stor‐
age: an assertion originating from pdfile almost guarantees that your data files are
corrupt.
If you are hitting corruption, you may see something like this in your logs:
Tue Dec 20 01:12:09 [initandlisten] Assertion: 10334:
Invalid BSONObj size: 285213831 (0x87040011)
first element: _id: ObjectId('4e5efa454b4ae20fa6000013')
If the first element shown is garbage, there isn’t much you can do. If the first element is
viewable (as the ObjectId in this example is), you may be able to remove the corrupt
document. You can try by running:
> db.remove({_id: ObjectId('4e5efa454b4ae20fa6000013')})
Replace the "_id" with the "_id" from your log’s assertion. Note that this technique
may not work if the corruption is not limited to that document: you may still have to
repair.
Durability with Replication
Due to the majority issues discussed in Chapter 10, a write to a replica set may be rolled
back until it has been written to a majority of the set. To put together the options here
with the journaling ones above, you can say:
> db.runCommand({"getLastError" : 1, "j" : true, "w" : "majority"})
As of this writing, this only guarantees that the write has been written durably to the
primary and has been written (not necessarily durably) to the secondaries. Theoretically,
it is possible for a majority of the servers to crash in the 100 ms between the write being
written and it being journaled, in which case it would be rolled back on the current
primary. This is an edge case but is obviously sub-optimal. Unfortunately it is not trivial
to fix, but there are a few open bugs to change this behavior.
Durability with Replication | 329
PART VI
Server Administration
CHAPTER 20
Starting and Stopping MongoDB
In Chapter 2, we covered the basics of starting MongoDB. This chapter will go into more
detail about which options are important for setting up MongoDB in production,
including:
• Commonly used options
• Starting up and shutting down MongoDB
•Security-related options
•Logging considerations
Starting from the Command Line
The MongoDB server is started with the mongod executable. mongod has many config‐
urable startup options; to view all of them, run mongod --help from the command line.
A couple of the options are widely used and important to be aware of:
--dbpath
Specify an alternate directory to use as the data directory; the default is /data/db/
(or, on Windows, \data\db\ on the MongoDB binary’s volume). Each mongod pro‐
cess on a machine needs its own data directory, so if you are running three instances
of mongod on one machine, you’ll need three separate data directories. When mon
god starts up, it creates a mongod.lock file in its data directory, which prevents any
other mongod process from using that directory. If you attempt to start another
MongoDB server using the same data directory, it will give an error:
"Unable to acquire lock for lockfilepath: /data/db/mongod.lock."
333
--port
Specify the port number for the server to listen on. By default, mongod uses port
27017, which is unlikely to be used by another process (besides other mongod pro‐
cesses). If you would like to run more than one mongod process on a single machine,
you’ll need to specify different ports for each one. If you try to start mongod on a
port that is already being used, it will give an error:
"Address already in use for socket: 0.0.0.0:27017"
--fork
Fork the server process, running MongoDB as a daemon.
If you are starting up mongod for the first time (with an empty data directory), it
can take the filesystem a few minutes to allocate database files. The parent process
will not return from forking until the preallocation is done and mongod is ready
to start accepting connections. Thus, fork may appear to hang. You can tail the log
to see what it is doing. You must use --logpath if you specify --fork.
--logpath
This option sends all output to the specified file rather than outputting on the
command line. This will create the file if it does not exist, assuming you have write
permissions to the directory. It will also overwrite the log file if it already exists,
erasing any older log entries. If you’d like to keep old logs around, use the --
logappend option in addition to --logpath (highly recommended).
--directoryperdb
This puts each database in its own directory. This allows you to mount different
databases on different disks, if necessary or desired. Common uses for this are
putting a local database (replication) on its own disk or moving a database to a
different disk if the original one fills up. You could also put databases that handle
more load on faster disks and lower on slower. It basically gives you more flexibility
to move things around later.
--config
Use a configuration file for additional options not specified on the command line.
This is typically used to make sure options are the same between restarts. See “File-
Based Configuration” on page 336 for details.
For example, to start the server as a daemon listening on port 5586 and sending all
output to mongodb.log, we could run this:
$ ./mongod --port 5586 --fork --logpath mongodb.log --logappend
forked process: 45082
all output going to: mongodb.log
Note that mongod may decide to preallocate journal files before it considers itself “start‐
ed.” If it does, fork will not return to the command prompt until the preallocation has
334 | Chapter 20: Starting and Stopping MongoDB
finished. You can tail mongodb.log (or wherever you redirected the log file) to watch its
progress.
When you first install and start MongoDB, it is a good idea to look at the log. This might
be an easy thing to miss, especially if MongoDB is being started from an init script, but
the log often contains important warnings that prevent later errors from occurring. If
you don’t see any warnings in the MongoDB log on startup, then you are all set. (Startup
warnings will also appear on shell startup.)
If there are any warnings in the startup banner, take note of them. MongoDB will warn
you about a variety of issues: that you’re running on a 32-bit machine (which MongoDB
is not designed for), that you have NUMA enabled (which can slow your application to
a crawl), or that your system does not allow enough open file descriptors (MongoDB
uses a lot of file descriptors).
The log preamble won’t change when you restart the database, so feel free to run
MongoDB from an init script and ignore the logs, once you know what they say. How‐
ever, its a good idea to check again each time you do an install, upgrade, or recover from
a crash, just to make sure MongoDB and your system are on the same page.
When you start the database, MongoDB will write a document to the local.startup_log
collection that describes the version of MongoDB, underlying system, and flags used:
> db.startup_log.findOne()
{
"_id" : "spock-1360621972547",
"hostname" : "spock",
"startTime" : ISODate("2013-02-11T22:32:52Z"),
"startTimeLocal" : "Mon Feb 11 17:32:52.547",
"cmdLine" : {
},
"pid" : 28243,
"buildinfo" : {
"version" : "2.4.0-rc1-pre-",
...
"versionArray" : [
2,
4,
0,
-9
],
"javascriptEngine" : "V8",
"bits" : 64,
"debug" : false,
"maxBsonObjectSize" : 16777216
}
}
This collection can be useful for tracking upgrades and changes in behavior.
Starting from the Command Line | 335
File-Based Configuration
MongoDB supports reading configuration information from a file. This can be useful
if you have a large set of options you want to use or are automating the task of starting up
MongoDB. To tell the server to get options from a configuration file, use the -f or --
config flags. For example, run mongod --config ~/.mongodb.conf to use ~/.mon
godb.conf as a configuration file.
The options supported in a configuration file are exactly the same as those accepted at
the command line. Here’s an example configuration file:
# Start MongoDB as a daemon on port 5586
port = 5586
fork = true # daemonize it!
logpath = /var/log/mongodb.log
logappend = true
This configuration file specifies the same options we used earlier when starting with
regular command-line arguments. It also highlights most of the interesting aspects of
MongoDB configuration files:
• Any text on a line that follows the # character is ignored as a comment
•The syntax for specifying options is option = value, where option is case-sensitive
•For command-line switches like --fork, the value true should be used
Stopping MongoDB
Being able to safely stop a running MongoDB server is at least as important as being
able to start one. There are a couple of different options for doing this effectively.
The cleanest way to shut down a running server is to use the shutdown command,
{"shutdown" : 1}. This is an admin command and must be run on the admin database.
The shell features a helper function to make this easier:
> use admin
switched to db admin
> db.shutdownServer()
server should be down...
The shutdown command, when run on a primary, steps down the primary and waits
for a secondary to catch up before shutting down the server. This minimizes the chance
of rollback, but the shutdown isn’t guaranteed to succeed. If there is no secondary avail‐
able that can catch up within a few seconds, the shutdown command will fail and the
(former) primary will not shut down:
336 | Chapter 20: Starting and Stopping MongoDB
> db.shutdownServer()
{
"closest" : NumberLong(1349465327),
"difference" : NumberLong(20),
"errmsg" : "no secondaries within 10 seconds of my optime",
"ok" : 0
}
You can force the shutdown command to shutdown a primary by using the force option:
db.adminCommand({"shutdown" : 1, "force" : true})
This is equivalent to sending a SIGINT or SIGTERM signal (all three of these options
result in a clean shutdown, but there may be unreplicated data). If the server is running
as the foreground process in a terminal, a SIGINT can be sent by pressing Ctrl-C.
Otherwise, a command like kill can be used to send the signal. If mongod has 10014 as
its PID, the command would be kill -2 10014 (SIGINT) or kill 10014 (SIGTERM).
When mongod receives a SIGINT or SIGTERM, it will do a clean shutdown. This means
it will wait for any running operations or file preallocations to finish (this could take a
moment), close all open connections, flush all data to disk, and halt.
Security
Do not set up publicly addressable MongoDB servers. You should restrict access as
tightly as possible between the outside world and MongoDB. The best way to do this is
to set up firewalls and only allow MongoDB to be reachable on internal network ad‐
dresses. Chapter 23 covers what connections are necessary to allow between MongoDB
servers and clients.
Beyond firewalls, there are a few options you can add to your config file to make it more
secure:
--bind_ip
Specify the interfaces that you want MongoDB to listen on. Generally you want this
to be an internal IP: something application servers and other members of your
cluster can access but is inaccessible to the outside world. localhost is fine for mon
gos processes if you’re running the application server on the same machine. For
config servers and shards, they’ll need to be addressable from other machines, so
stick with non-localhost addresses.
--nohttpinterface
By default, MongoDB starts a tiny HTTP server on a port 1000 above wherever you
started MongoDB. This gives you some information about your system, but nothing
you can’t get elsewhere and is somewhat useless on a machine you probably only
access via SSH and exposes information that should be inaccessable to the outside
world.
Security | 337
Unless you’re in development, this should be turned off.
--nounixsocket
If you’re not planning to connect via file system socket, you might as well disable
this option. You would only connect via file system socket on a machine that is also
running an application server: you must be local to use a file system socket.
--noscripting
This entirely disallows server-side JavaScript execution. Most security issues that
have been reported with MongoDB have been JavaScript-related and it is generally
safer to disallow it, if your application allows.
Several shell helpers assume that JavaScript is available on the server, notably
sh.status(). You will see errors if you attempt to run any of these helpers with
JavaScript disabled.
Do not enable the REST interface. It is disabled by default and allows running many
commands on the server. It is not intended for production use.
Data Encryption
As of this writing, MongoDB provides no built-in mechanism for encrypting data stor‐
ed. If you require data to be encrypted, use filesystem encryption. Another possibility
is manually encrypting certain fields (although MongoDB has no special ability to query
for encrypted values).
SSL Connections
By default, connections to MongoDB transfer data unencrypted. However, SSL con‐
nection support is available. Due to licensing issues the default builds do not have SSL,
but you can download a subscriber build at http://www.10gen.com, which supports SSL.
You can also compile MongoDB from source to enable SSL support. Consult your driv‐
er’s documentation on how to create SSL connections using your language.
Logging
By default, mongod sends its logs to stdout. Most init scripts use the --logpath option
to send logs to a file. If you have multiple MongoDB instances on a single machine (say,
a mongod and a mongos), make sure that their logs are stored in separate files. Make
sure that you know where the logs are and have read access to the files.
MongoDB spits out a lot of log messages, but please do not run with the --quiet option
(which suppresses some of them). Leaving the log level at the default is usually perfect:
there is enough info for basic debugging (why is this slow, why isn’t this starting up,
338 | Chapter 20: Starting and Stopping MongoDB
etc.), but the log does not take up too much space. If you are debugging a specific issue
with your application, there are a couple options for getting more info from the logs.
First, you can change the log level, either by restarting MongoDB with more v’s or
running the setParameter command:
> db.adminCommand({"setParameter" : 1, "logLevel" : 3})
Remember to turn log level back down to 0, or your logs may be needlessly noisy. You
can turn log level up to 5, at which point mongod will print out almost every action it
takes, including the contents of every request handled. This can cause a lot of IO as
mongod writes everything to the log file, which can slow down a busy system. Turning
on profiling is a better option if you need to see every operation as it’s happening.
By default, MongoDB logs information about queries that take longer than 100 ms to
run. If 100 ms it too short or too long for your application, you can change the threshold
with setProfilingLevel:
> // Only log queries that take longer than 500ms
> db.setProfilingLevel(1, 500)
{ "was" : 0, "slowms" : 100, "ok" : 1 }
> db.setProfilingLevel(0)
{ "was" : 1, "slowms" : 500, "ok" : 1 }
The second line will turn off profiling, but the value in milliseconds given in the first
line will continue to be used as a threshold for the log (across all databases). You can
also set this parameter by restarting MongoDB with the --slowms option.
Finally, set up a cron job that rotates your log every day or week. If MongoDB was started
with --logpath, sending the process a SIGUSR1 signal will make it rotate the log. There
is also a logRotate command that does the same thing:
> db.adminCommand({"logRotate" : 1})
You cannot rotate logs if MongoDB was not started with --logpath.
Logging | 339
CHAPTER 21
Monitoring MongoDB
Before you deploy, it is important to set up some type of monitoring. Monitoring should
allow you to track what your server is doing and alert you if something goes wrong.
This chapter will cover:
• How to track MongoDB’s memory usage
• How to track application performance metrics
•How to diagnose replication issues
Examples use chapters from the Mongo Monitoring Service (MMS) to demonstrate what
to look for when monitoring. There are installation instructions for MMS at https://
mms.10gen.com. If you do not want to use MMS, please use some type of monitoring.
It will help you detect potential issues before they cause problems and let you diagnose
issues when they occur.
Monitoring Memory Usage
Accessing data in memory is fast and accessing data on disk is slow. Unfortunately,
memory is expensive (and disk is cheap) and typically MongoDB uses up memory
before any other resource. This section covers how to monitor MongoDB’s interactions
with disk and memory, and what to watch for.
Introduction to Computer Memory
Computers tend to have a small amount of fast-to-access memory and a large amount
of slow-to-access disk. When you request a page of data that is stored on disk (and not
yet in memory), your system page faults and copies the page from disk into memory. It
can then access the page in memory extremely quickly. If your program stops regularly
341
using the page and your memory fills up with other pages, the old page will be evicted
from memory and only live on disk again.
Copying a page from disk into memory takes a lot longer than reading a page from
memory. Thus, the less MongoDB has to copy data from disk, the better. If MongoDB
can operate almost entirely in memory, it will be able to access data much faster. Thus,
MongoDB’s memory usage is one of the most important stats to track.
Tracking Memory Usage
There are several “types” of memory MongoDB reports using. First is resident memo‐
ry: this is the memory that MongoDB explicitly owns in RAM. For example, if we query
for a document and it is paged into memory, that page is added to MongoDB’s resident
memory.
MongoDB is given an address for that page. This address isn’t the literal address of the
page in RAM. It’s a virtual address. MongoDB can pass it to the kernel and the kernel
will look up where the page really lives. This way, if the kernel needs to evict the page
from memory, MongoDB can still use the address to access it. MongoDB will request
the memory from the kernel, the kernel will look at its page cache, see that the page is
not there, page fault to copy the page into memory, and return it to MongoDB. The
pages of data MongoDB has addresses for is how MongoDB’s mapped memory is cal‐
culated: it includes all of the data MongoDB has ever accessed. It will usually be about
the size of your data set.
MongoDB keeps an extra virtual address for each page of mapped memory for jour‐
naling to use (see Chapter 19). This doesn’t mean that there are two copies of the data
in memory, just two addresses. Thus, the total virtual memory MongoDB uses will be
approximately twice your mapped memory size (or twice your data size). If journaling
is disabled, mapped and virtual memory sizes will be approximately equal.
Note that both virtual memory and mapped memory are not “real” memory allocations:
they do not tell you anything about how much RAM is being used. They are just map‐
pings that MongoDB is keeping. Theoretically, MongoDB could have a petabyte of
memory mapped and only a couple of gigabytes in RAM. Thus, you do not have to
worry if mapped or virtual memory sizes exceed RAM.
Figure 21-1 shows the MMS graph for memory information, which describes how much
resident, virtual, and mapped memory MongoDB is using. On a box dedicated to Mon‐
goDB, resident should be a little less than the total memory size (assuming your working
set is as large or larger than memory). Resident memory is that only statistic that actually
tracks how much data is in physical RAM, but by itself this stat does not tell you much
about how MongoDB is using memory.
342 | Chapter 21: Monitoring MongoDB
Figure 21-1. From the top line to the bottom: virtual, mapped, and resident memory
If your data fits entirely in memory, resident should be approximately the size of your
data. When we talk about data being “in memory,” we’re always talking about the data
being in RAM.
Tracking Page Faults
As you can see from Figure 21-1, memory metrics tend to be fairly steady, but as your
data set grows virtual and mapped will grow with it. Resident will grow to the size of
your available RAM and then hold steady.
You can use other statistics to find out how MongoDB is using memory, not just how
much of each type it has. One useful stat is number of page faults, which tells you how
often the data MongoDB is looking for is not in RAM. Figure 21-2 and Figure 21-3 are
graphs page faults over time. Figure 21-3 is page faulting less than Figure 21-2, but by
itself this information is not very useful. If the disk in Figure 21-2 can handle that many
faults and the application can handle the delay of the disk seeks, there is no particular
problem with having so many faults (or more). On the other hand, if your application
cannot handle the increased latency of reading data from disk, you have no choice but
to store all of your data in memory (or use SSDs).
Monitoring Memory Usage | 343
Figure 21-2. A system that is page faulting hundreds of times a minute
Figure 21-3. A system that is page faulting a few times a minute
Regardless of how forgiving the application is, page faults become a problem when the
disk is overloaded. The amount of load a disk can handle isn’t linear: once a disk begins
getting overloaded, each operation must queue for a longer and longer period of time,
creating a chain reaction. There is usually a tipping point where disk performance begins
degrading quickly. Thus, it is a good idea to stay away from the maximum load that your
disk can handle.
Track your page fault numbers over time. If your application is behaving well with a
certain number of page faults, you have a baseline for how many page faults the system
can handle. If page faults begin to creep up and performance deteriorates, you have a
threshold to alert on.
You can see page fault stats per-database by looking at "recordStats" field of server‐
Status’s output:
> db.adminCommand({"serverStatus" : 1})["recordStats"]
{
"accessesNotInMemory": 200632,
"test": {
344 | Chapter 21: Monitoring MongoDB
"accessesNotInMemory": 1,
"pageFaultExceptionsThrown": 0
},
"pageFaultExceptionsThrown": 6633,
"admin": {
"accessesNotInMemory": 1247,
"pageFaultExceptionsThrown": 1
},
"bat": {
"accessesNotInMemory": 199373,
"pageFaultExceptionsThrown": 6632
},
"config": {
"accessesNotInMemory": 0,
"pageFaultExceptionsThrown": 0
},
"local": {
"accessesNotInMemory": 2,
"pageFaultExceptionsThrown": 0
}
},
"accessesNotInMemory" gives you a count of how many times MongoDB has had to go
to disk (since startup).
Minimizing Btree Misses
Accessing index entries that are not in memory is particularly inefficient, as it often
causes two page faults. There is one fault to load the index entry into memory and then
another to load the document into memory. When an index lookup causes a page fault
it’s called a btree miss. MongoDB also tracks btree hits: when an index access does not
have to go to disk. Both are shown in Figure 21-4.
Indexes are so frequently used that they are generally in memory, but if there is too little
memory, a lot of indexes, or an unusual access pattern (e.g., a lot of table scans), btree
misses may be higher. They should generally be low so if you’re seeing a lot of them,
track down the cause.
Monitoring Memory Usage | 345
Figure 21-4. A chart showing btree stats
IO Wait
Page faults in general are closely tied to how long the CPU is idling waiting for the disk,
called IO wait. Some IO wait is normal (MongoDB has to go to disk sometimes and,
although it tries not to block anything when it does, cannot completely avoid it). The
important thing is that IO wait is not increasing or near 100%, as shown in
Figure 21-5. This indicates that the disk is getting overloaded.
Figure 21-5. IO wait hovering around 100%
MMS can track CPU information if you install the munin plug-in. See their website for
installation instructions.
Tracking Background Flush Averages
One other disk metric to watch is how long it takes MongoDB to write its dirty pages
to disk, also known as the background flush average. This is a good canary-in-the-coal-
mine stat. If the background flush average starts creeping up, you know that your disk
is having trouble keeping up with requests.
346 | Chapter 21: Monitoring MongoDB
At least once a minute (by default), MongoDB will flush all writes that have happened
to disk. (Depending on the operating system, MongoDB may flush writes more fre‐
quently if there are a lot of dirty pages.) You can also configure the interval by passing
a number of seconds to the --syncdelay option when starting mongod. More frequent
syncs will make the amount of data to be synced smaller, but can also be less efficient.
A common misconception is that syncdelay has something to do with
data durability. It has absolutely no effect on durability. To ensure du‐
rability, use journaling. syncdelay is only for tuning disk performance.
Generally, you want to see background flush averages of less than a second. On a slow
disk or a busy system, this can creep up, taking longer and longer as the disk gets over‐
loaded. At some point, the disk will be so overloaded that flushes will take longer than
60 seconds, meaning MongoDB will be trying to flush constantly (which puts even more
load on the disk). Occasional spikes in disk flush times are expected. What you don’t
want to see is a trend towards tens of seconds.
Figure 21-6 shows a graph of background flush averages over time. This system’s hard
drive is working hard: it always takes it more than 5 seconds to write the preceding
minute’s data to disk. This is a bit slow, especially with the regular spikes of nearly 20
seconds, so it might be worth turning syncdelay down a bit, say to 40 seconds, and
seeing if writing less data per flush helped.
Figure 21-6. Background flush averages on a somewhat overloaded system
If background flush average creeps up beyond what is reasonable for your disks (prob‐
ably a few seconds) over a broad time period, start thinking about how you’re going to
lighten the load on your disks.
MongoDB only has to flush dirty data (that is, data that’s changed) so background flush
average will generally reflect write load. Thus, if you have a low write load, background
Monitoring Memory Usage | 347
flush average may not show that your disk is straining. You should always track IO wait
and page faults in addition to background flush average.
Calculating the Working Set
In general, the more of your data that is in memory, the faster MongoDB will perform.
Thus, in order from fastest to slowest, an application could have:
1. The entire data set in memory. This is nice to have but is often too expensive or
infeasible. It may be necessary for applications that depend on fast response times.
2. The working set in memory. This is the most common choice.
Your working set is the data and indexes that your application uses. This may be
everything, but generally there’s a core data set (for example, the users collection
and the last month of activity) that covers 90% of requests. If this working set fits
in RAM, MongoDB will generally be fast: it only has to go to disk for a few “unusual”
requests.
3. The indexes in memory.
4. The working set of indexes in memory. This generally requires right-balanced in‐
dexes (see Chapter 5).
5. No useful subset of data in memory. If possible, avoid this. It will be slow.
You must know what your working set is (and how large it is) to know if you can keep
it in memory. The best way to calculate working set is to track common operations to
find out how much your application is reading and writing. For example, suppose your
application creates 2 GB of new data per week and 800 MB of that data it is regularly
accessed. Users tend to access data up to a month old and data that’s older than that is
mostly unused. Your working set size is probably about 3.2 GB (800 MB/week × 4 weeks),
plus a fudge factor for indexes, so call it 5 GB.
One way to think about this is to track data accessed over time, as shown in
Figure 21-7. If you choose a cutoff where 90% of your request fall, then the data (and
indexes) generated in that period of time are your working set, like Figure 21-8. You
can measure for that amount of time to figure out how much your data set grows. Note
that this example uses time, but it’s possible that there’s another access pattern that makes
more sense for your application (time being the most common one).
348 | Chapter 21: Monitoring MongoDB
Figure 21-7. A plot of data accesses by age of data
Figure 21-8. The working set is data used in the requests before the cutoff of “frequent
requests”
You can also use MongoDB’s stats to estimate the working set. MongoDB keeps a map
of what it thinks is its memory, which you can see by passing in the "workingSet" :
1 option to serverStatus:
> db.adminCommand({"serverStatus" : 1, "workingSet" : 1})
{
...
"workingSet" : {
"note" : "thisIsAnEstimate",
"pagesInMemory" : 18,
"computationTimeMicros" : 3685,
"overSeconds" : 2363
},
...
}
"pagesInMemory" is how many pages MongoDB thinks are currently in memory. Mon‐
goDB does not actually know how many pages are in memory, but it should be close. If
Calculating the Working Set | 349
the number of pages returns equals the size of your RAM this it isn’t very helpful; but
if it is smaller it is probably about how large your working set is.
The "workingSet" field is not included in serverStatus’s output by default.
Some Working Set Examples
Suppose that you have a 40 GB working set. A total of 90% of requests hit the working
set, and 10% hit other data. If you have 500 GB of data and 50 GB of RAM, your working
set fits entirely in RAM. Once your application is has accessed the data it usually accesses
(a process called preheating), it should never have to go to disk again for the working
set. It had 10 GB of space for the 460 GB of less-frequently-accessed data. Obviously,
MongoDB will almost always have to go to disk for the nonworking set data.
On the other hand, suppose our working set does not fit in RAM. Say we have only 35
GB of RAM. Then the working set will generally take up most of RAM. The working
set has a higher probability of staying in RAM because it’s accessed more frequently, but
at some point the less-frequently-accessed data will have to be paged in, evicting working
set (or other less-frequently-accessed data). Thus, there is a constant churn back and
forth from disk: accessing the working set does not have predictable performance
anymore.
Tracking Performance
Performance of queries is often important to track and keep consistent. There are several
ways to track if MongoDB is having trouble with the current request load.
CPU is generally IO bound with MongoDB (IO wait is high, and the other metrics are
negligible). However, if user or system time is approaching 100% (or 100% multiplied
by the number of CPUs you have) the most common cause is that you’re missing an
index on a frequently-used query. The other possibility is that you are running a lot of
MapReduces or other server-side JavaScript. It is a good idea to track CPU (particularly
after deploying a new version of your application) to ensure that all your queries are
behaving as they should.
Note that the graph shown in Figure 21-9 is fine: if there is a low number of page faults,
IO wait may be dwarfed by other CPU activities. It is only when the other activities creep
up that bad indexes may be a culprit.
350 | Chapter 21: Monitoring MongoDB
Figure 21-9. A CPU with minimal IO wait. The top line is user and the lower line is
system. The other stats are very close to 0%.
A similar metric is queuing: how many request are waiting to be processed by MongoDB.
A request is considered queued when it is waiting for the lock it needs to do a read or
a write. Figure 21-10 shows a graph of read and write queues over time. No queues are
preferred (basically an empty graph), but this graph is nothing to be alarmed about. In
a busy system, it isn’t unusual for an operation to have to wait a bit for the correct lock
to be available.
Figure 21-10. Read and write queues over time
You can see if requests are piling up by looking at the number of requests enqueued.
Generally, queue size should be low. A large and ever-present queue is an indication
that mongod cannot keep up with its load. You should decrease the load on that server
as fast as possible.
You can correlate statistics about queuing with lock percentage: the amount of time
MongoDB spends locked. Often disk IO will throttle writes more than locking but
locking is still important to track, especially for systems with fast disks or many se‐
quential writes. Again, one of the most common causes of high lock percentage is that
you are missing an index. As lock percentage increases, operations on average have to
Tracking Performance | 351
wait longer and longer for a lock. Thus, there is an unfortunate cascading nature to high
lock percentages making everything slower, causing requests to build up, causing more
load on the system and even higher lock percentages. Figure 21-11 shows an alarmingly
high lock percentage, which should be dealt with as soon as possible.
Lock percentage is often spiky depending on traffic levels; but if it trends upwards over
time it’s a good indication that your system is under stress and that something needs to
change. Thus, you should alert on lock percentage over a long time (so that a sudden
spike in traffic won’t trigger it).
On the other hand, you may want to also trigger an alert if lock percentage suddenly
spikes, say 25% over its normal value. This might be an indication that your system
cannot handle load spikes and that you may have to add capacity.
Figure 21-11. A lock percentage hovering worryingly near 100%
In addition to the global lock percentage, MongoDB tracks locking per database, so you
can see if you have a particular database with a lot of contention.
Tracking Free Space
One other metric that is basic but important to monitor is disk usage: track free disk
space. Sometimes users wait until their disk runs out of space before they think about
how they want to handle it. By monitoring your disk usage, you can predict how long
your current drive will be sufficient and plan in advance what to do when it is not.
As you run out of space, there are several options:
• If you are using sharding, add another shard.
•Shut down each member of a replica set (one at a time) and copy its data to a larger
disk, which can then be mounted. Restart the member and proceed to the next.
352 | Chapter 21: Monitoring MongoDB
•Replace members of your replica set with members with a larger drive: remove an
old member and add a new member, and allow that one to catch up with the rest
of the set. Repeat for each member of the set.
•If you are using the directoryperdb option and you have a particularly fast-
growing database, move it to its own drive. Then mount the volume as a directory
in your data directory. This way the rest of your data doesn’t have to be moved.
Regardless of the technique you choose, plan ahead to minimize the impact on your
application. You need time to take backups, modify each member of your set in turn,
and copy your data from place to place.
Monitoring Replication
Replication lag and oplog length are important to track.
Lag is when the secondaries cannot keep up with the primary. Lag is calculated by
subtracting the time of the last op applied on a secondary from the time of the last op
on the primary. For example, if a secondary just applied an op with the timestamp
3:26:00 p.m. and the primary just applied an op with the timestamp 3:29:45 p.m., the
secondary is lagging by 3 minutes and 45 seconds. You want lag to be as close to 0 as
possible, and it is generally on the order of milliseconds. If a secondary is keeping up
with the primary, the replication lag should look something like the graph shown in
Figure 21-12: basically 0 all the time.
Figure 21-12. A replica set with no lag. This is what you want to see.
If a secondary cannot replicate writes as fast as the primary can write, you’ll start seeing
nonzero lag. The most extreme case of this is when replication is stuck: it cannot apply
any more operations for some reason. At this point, lag will grow by one second per
second, creating the steep slope shown in Figure 21-13. This could be caused by network
issues or a missing "_id" index, which is required on every collection for replication to
function properly.
Monitoring Replication | 353
If a collection is missing an "_id" index, take the server out of the replica set, start it as
a standalone server, and build the "_id" index. Make sure you create the "_id" index
as a unique index. Once created, the "_id" index cannot be dropped or changed (other
than by dropping the whole collection).
Figure 21-13. Replication getting stuck and, just before February 10, beginning to re‐
cover. The red lines are server restarts.
If a system is overloaded, a secondary may gradually fall behind. But you generally won’t
see the characteristic “one second per second” slope in the graph, but some replication
will still be happening. Still, it is important to be aware if the secondaries cannot keep
up with peak traffic or are gradually falling further behind.
Primaries do not throttle writes to “help” secondaries catch up, so it common for sec‐
ondaries to fall behind on overloaded systems (particularly as MongoDB tends to pri‐
oritize writes over reads, which means replication can be starved on the primary). You
can force throttling of the primary to some extent by using “w” with your write concern.
You also might want to try removing load from the secondary by routing any requests
it was handling to another member.
If you are on an extremely underloaded system you may see another interesting pattern:
sudden spikes in replication lag, as shown in Figure 21-14. The spikes shown are not
actually lag—they are caused by variations in sampling. The mongod is processing one
write every couple of minutes. Because lag is measured as the difference between time‐
stamps on the primary and secondary, measuring the timestamp of the secondary right
before a write on the primary makes it look minutes behind. If you increase the write
rate, these spikes should disappear.
354 | Chapter 21: Monitoring MongoDB
Figure 21-14. A low-write system can cause “phantom” lag
The other important metric to track is the length of each member’s oplog. Every member
that might become primary should have an oplog longer than a day. If a member may
be a sync source for another member, it should have an oplog longer than the time an
initial sync takes to complete. Figure 21-15 shows what a standard oplog-length graph
looks like. This oplog has an excellent length: 1,111 hours is over a month of data! In
general, oplogs should be as long as you can afford the disk space to make them. Given
the way they’re used, they take up basically no memory and a long oplog can mean the
difference between a painful ops experience and an easy one.
Figure 21-15. A typical oplog-length graph
Figure 21-16 shows a slightly unusual variation caused by a fairly short oplog and vari‐
able traffic. This is still healthy, but the oplog on that machine is probably too short
(between 6 and 11 hours of maintenance window). The administrator may want to make
the oplog longer when she gets a chance.
Monitoring Replication | 355
Figure 21-16. An oplog length of an application with daily traffic peaks
356 | Chapter 21: Monitoring MongoDB
CHAPTER 22
Making Backups
It is important to take regular backups of your system. Backups are good protection
against most types of failure, and very little can’t be solved by restoring from a clean
backup. This chapter covers the common options for taking backups:
• Single-server backups
• Special considerations for backing up replica sets
•How to back up a sharded cluster
Backups are only useful if you are confident about deploying them in an emergency.
Thus, for any backup technique you choose, be sure to practice both taking backups
and restoring from backups until you are comfortable with the restore procedure.
Backing Up a Server
There are a variety of ways of taking backups. Regardless of method, taking a backup
can cause strain on a system: it generally requires reading all your data into memory.
Thus, backups should generally be done on replica set secondaries (as opposed to the
primary) or, for standalone servers, at an off time.
The techniques in this section apply to any mongod, whether a standalone or a member
of a replica set, unless otherwise noted.
Filesystem Snapshot
The simplest way to make a backup is to take a filesystem snapshot. However, this
requires your filesystem to support snapshotting and you must be running mongod with
journaling enabled. If your system fulfills these two prerequisites, this method requires
no preparation: simply take a snapshot at any time.
357
To restore, ensure that mongod is not running. The exact command for restoring from
a snapshot varies by filesystem, but basically you restore the snapshot and then start
mongod. As you took a snapshot on a live system, the snapshot is essentially what the
data files would look like if mongod had been kill -9-ed at the time the snapshot was
taken. Thus, on startup, mongod will replay the journal files and then begin running
normally.
Copying Data Files
Another way of creating backups is to make a copy of everything in the data directory.
Because you cannot copy all of the files at the same moment without filesystem support,
you must prevent the data files from changing while you are making the copy. This can
be accomplished with a command called fsynclock:
> db.fsyncLock()
This command locks the database against any further writes and then flushes all dirty
data to disk (fsync), ensuring that the files in the data directory have the latest consistent
information and are not changing.
Once this command has been run, mongod will enqueue all incoming writes. It will not
process any further writes until it has been unlocked. Note that this command stops
writes to all databases (not just the one db is connected to).
Once the fsynclock command returns, copy all of the files in your data directory to a
backup location. On Linux, this can be done with a command such as:
$ cp -R /data/db/* /mnt/external-drive/backup
Make sure that you copy absolutely every file and folder from the data directory to the
backup location. Excluding files or directories may make the backup unusable or
corrupt.
Once you have finished copying the data, unlock the database to allow it to take writes
again:
> db.fsyncUnlock()
Your database will begin handling writes again normally.
Note that there are some locking issues with authentication and fsynclock. If you are
using authentication, do not close the shell between calling fsyncLock() and fsyncUn
lock(). If you disconnect, you may be unable to reconnect and have to restart mon
god. The fsyncLock() setting does not persist between restarts, mongod will always start
up unlocked.
358 | Chapter 22: Making Backups
As an alternative to fsynclocking, you can instead shut down mongod, copy the files,
and then start mongod back up again. Shutting down mongod effectively flushes all
changes to disk and prevents new writes from occurring during the backup.
To restore from data directory copies, ensure that mongod is not running and that the
data directory you want to restore into is empty. Copy the backed-up data files to the
data directory, and then start mongod. For example, the following command would
restore the files backed up with the command shown earlier:
$ cp -R /mnt/external-drive/backup/* /data/db/
$ mongod -f mongod.conf
Despite the warnings about partial data directory copies, you can use this method to
backup individual databases if you know what to copy. To back up an individual database
(called, say, "myDB"), copy all of the myDB.* files (including the .ns file) to backup. If
you are using the --directoryperdb option, copy the entire myDB directory.
You can restore specific databases by copying just the files with the correct database
name into your data directory. You must be starting from a clean shutdown to restore
piecemeal like this. If you had a crash or a hard shutdown, do not attempt to restore a
single database from backup: replace the entire directory from backup and start the
mongod to allow the journal files to be replayed.
Never use fsyncLock in conjunction with mongodump. Depending on
what else your database is doing, mongodump may hang forever if the
database is locked.
Using mongodump
The final way of taking a backup is to use mongodump. mongodump is mentioned last
because it has some downsides. It is slower (both to get the backup and restore from it)
and it has some issues with replica sets which are discussed in “Backing Up a Replica
Set” on page 361. However, it also has some benefits: it is a good way to backup individual
databases, collections, and even subsets of collections.
mongodump has a variety of options that you can see by running mongodump --help.
Here, we will focus on the most useful ones to use for backup.
To backup all databases, simply run mongodump. If you are running mongodump on
the same machine as the mongod, you can simply specify the port mongod is running
on:
$ mongodump -p 31000
mongodump will create a dump directory in the current directory, which contains a
dump of all of your data. This dump directory is organized by database and collection
Backing Up a Server | 359
into folders and subfolders. The actual data is stored in .bson files, which merely contain
every document in a collection in BSON, concatenated together. You can exam‐
ine .bson files using the bsondump tool, which comes with MongoDB.
You do not even need to have a server running to use mongodump: you can use the --
dbpath option to specify your data directory and mongodump will use the data files to
copy data:
$ mongodump --dbpath /data/db
You should not use --dbpath if mongod is running.
One issue with mongodump is that it is not an instantaneous backup: the system may
be taking writes while the backup occurs. Thus, someone might begin a backup that
causes mongodump to dump the database A. While mongodump is dumping B, someone
drops A. However, mongodump has already dumped it, so you’ll end up with a snapshot
of the data in a state it never existed in on the original server.
To avoid this, if you are running mongod with --replSet, you can use mongodump’s
--oplog option. This will keep track of all operations that occur on the server while the
dump is taking place, so these operations can be replayed when the backup is restored.
This gives you a consistent point-in-time snapshot of data from the source server.
If you pass mongodump a replica set connection string (e.g., "setName/
seed1,seed2,seed3"), it will automatically choose a secondary to dump from, if one is
available.
To restore from a mongodump backup, use the mongorestore tool:
$ mongorestore -p 31000 --oplogReplay dump/
If you used the --oplog option to dump the database, you must use the --
oplogReplay option with mongorestore to get the point-in-time snapshot.
If you are replacing data on a running server, you may (or may not) wish to use the --
drop option, which drops a collection before restoring it.
The behavior of mongodump and mongorestore has changed over time. To prevent
compatibility issues, try to use the same version of both utilities (you can see their
versions by running mongodump --version and mongorestore --version).
Moving collections and databases with mongodump and mongorestore
You can restore into an entirely different database and collection than you dumped from.
This can be useful if different environments use different database names (say, dev and
prod) but the same collection names.
To restore a .bson file into a specific database and collection, specify the targets on the
command line:
360 | Chapter 22: Making Backups
$ mongorestore --db newDb --collection someOtherColl dump/oldDB/oldColl.bson
Administrative complications with unique indexes
If you have a unique index (other than "_id") on any of your collections, you should
consider using a different type of backup than mongodump/mongorestore. Specifically,
unique indexes require that the data does not change in ways that would violate the
unique index constraint during the copy. The safest way to do this is to choose a method
that “freezes” the data, then take a backup as described in either of the previous two
sections.
If you are determined to use mongodump/mongorestore, you may need to preprocess
your data when you restore from backup.
Backing Up a Replica Set
Generally, you should take backups from a secondary: this keeps load off of the primary
and you can lock a secondary without affecting your application (so long as your ap‐
plication isn’t sending it read requests). You can use any of the three methods outlined
previously to backup a replica set member, but file system snapshot or data file copy are
recommended. Either of these techniques can be applied to replica set secondaries with
no modification.
mongodump is not quite as simple to use when replication is enabled. First, if you are
using mongodump, you must take your backups using the --oplog option to get a point-
in-time snapshot; otherwise the backup’s state won’t match the state of anyone else in
the cluster. You must also create an oplog when you restore from a mongodump backup,
or the restored member will not know where it was synced to.
To restore a replica set member from a mongodump backup, start the target replica set
member as a standalone server with an empty data directory. First, run mongorestore
(as described in the previous section) with the --oplogReplay option. Now it should
have a complete copy of the data, but it still needs an oplog. Create an oplog using the
createCollection command:
> use local
> db.createCollection("oplog.rs", {"capped" : true, "size" : 10000000})
Specify the size of the collection in bytes. See “Resizing the Oplog” on page 220 for
advice on oplog sizing.
Now we need to populate the oplog. The easiest way to do this is to restore the oplog.bson
backup file from the dump into the local.oplog.rs collection:
$ mongorestore -d local -c oplog.rs dump/oplog.bson
Backing Up a Replica Set | 361
Note that this is not a dump of the oplog (dump/local/oplog.rs.bson) but rather the oplog
operations that occurred during the dump. Once this mongorestore is complete, you can
restart this server as a replica set member.
Backing Up a Sharded Cluster
Sharded clusters are impossible to “perfectly” back up while active: you can’t get a snap‐
shot of the entire state of the cluster at a point in time. However, this limitation is
generally sidestepped by the fact that as your cluster gets bigger, it becomes less and less
likely that you’d ever have to restore the whole thing from backup. Thus, when dealing
with a sharded cluster, we focus on backing up pieces: the config servers and the replica
sets individually.
Turn off the balancer before performing any of these operations on a sharded cluster
(either backup or restore). You cannot get a consistent snapshot of the world with chunks
flying around. See “Balancing Data” on page 289 for instructions on turning the balancer
on and off.
Backing Up and Restoring an Entire Cluster
When a cluster is very small or in development, you may want to actually dump and
restore the entire thing. You can accomplish this by turning off the balancer and then
running mongodump through the mongos. This creates a backup of all of the shards on
whatever machine mongodump is running on.
To restore from this type of backup, run mongorestore connected to a mongos.
After turning off the balancer, you can alternatively take filesystem or data directory
backups of each shard and the config servers. However, you will inevitably get copies
from each at slightly different times, which may or may not be a problem. Also, as soon
as you turn on the balancer and a migrate occurs, some of the data you backed up from
one shard will no longer be there.
Backing Up and Restoring a Single Shard
Most often, you’ll only need to restore a single shard in a cluster. If you are not too picky,
you can restore from a backup of that shard taken using one of the single-server methods
just described.
There is one important issue to be aware of: suppose you take a backup of your cluster
on Monday. On Thursday, your disk melts down and you have to restore from backup.
However, in the intervening days, new chunks may have moved to this shard. Your
backup of the shard from Monday will not contain these new chunks. You may be able
to use a config server backup to figure out where the disappearing chunks lived on
362 | Chapter 22: Making Backups
Monday, but it is a lot more difficult than simply restoring the shard. In most cases,
restoring the shard and losing the data in those chunks is the preferable route.
You can connect directly to a shard to restore from backup (instead of going through
mongos).
Creating Incremental Backups with mongooplog
All of the backup methods outlined must make a full copy of the data, even if very little
of it has changed since the last backup. If you have data that is very large relative to the
amount that is being written, you may want to look into incremental backups.
Instead of making full copies of the data every day or week, you take one backup and
then use the oplog to back up all operations that have happened since the backup. This
technique is much more complex than the ones described above, so prefer them unless
incremental backups are absolutely necessary.
This technique requires two machines, A and B, running mongod. A is your main
machine (probably a secondary) and B is your backup machine:
1. Make a note of the latest optime in A’s oplog:
> op = db.oplog.rs.find().sort({$natural: -1}).limit(1).next();
> start = op['ts']['t']/1000
Keep this somewhere safe—you’ll need it for a later step.
2. Take a backup of your data, using one of the techniques above to get a point-in-
time backup. Restore this backup to the data directory on B.
3. Periodically add any operations that have happened on A to B’s copy of the data.
There is a special tool that comes with MongoDB distributions that makes this easy:
mongooplog (pronounced mon-goop-log) which copies data from the oplog of one
server and applies it to the data set on another. On B, run:
$ mongooplog --from A --seconds 1234567
--seconds should be passed the number of seconds between the start variable
calculated in step 1 and the current time, then add a bit (better to replay operations
a second time than miss them).
This keeps your backup relatively up-to-date with your data. This technique is sort of
like keeping a secondary up-to-date manually, so you may just want to use a slave-
delayed secondary instead.
Creating Incremental Backups with mongooplog | 363
CHAPTER 23
Deploying MongoDB
This chapter gives recommendations for setting up a server to go into production. In
particular, it covers:
• Choosing what hardware to buy and how to set it up
• Using virtualized environments
•Important kernel and disk IO settings
•Network setup: who needs to connect to whom
Designing the System
You generally want to optimize for data safety and as fast access as you can afford. This
section discusses the best way to accomplish these goals when choosing disks, RAID
configuration, CPU, and other hardware and low-level software components.
Choosing a Storage Medium
In order of preference, we would like to store and retrieve data from:
1. RAM
2. SSD
3. Spinning disk
Unfortunately, most people have limited budgets or enough data that storing everything
in RAM is impractical and SSDs are too expensive. Thus, the typical deployment is a
small amount of RAM (relative to total data size) and a lot of space on a spinning disk.
If you are in this camp, the important thing is that your working set is smaller than RAM
and you should be ready to scale out if the working set gets bigger.
365
If you are able to spend what you like on hardware, buy a lot of RAM and/or SSDs.
Reading data from RAM takes a few nanoseconds (say, 100). Conversely, reading from
disk takes a few milliseconds (say, 10). It can be hard to picture the difference between
these two numbers, so suppose we scale them up to more relatable numbers: if accessing
RAM took 1 second, accessing disk would take over a day!
100 nanoseconds × 10,000,000 = 1 second
10 milliseconds × 10,000,000 = 1.16 days
These are very back-of-the-envelope calculations (your disk might be a bit faster or your
RAM a bit slower), but the magnitude of this difference doesn’t change much. Thus, we
want to access disk as seldom as possible.
Fast spinning disks do not cut down disk access time that much, so don’t spend lots of
money on them. It is more productive to get more memory or SSDs.
An example from the wild
The graphs in Figure 23-1 through Figure 23-6 illustrate some of the advantages of SSDs.
These graphs are from a user bringing a new shard online midday on August 8. The
user’s deployment was previously just spinning disks. He added a new shard backed by
SSDs and, in the graphs, was concurrently running one shard with SSDs and one with
spinning disks.
As shown in Figure 23-1, the spinning disk peaks at nearly 5,000 queries per second,
but generally only handed a few hundred queries per second.
Figure 23-1. Queries on a spinning disk
In comparison, the chart in Figure 23-2 plots queries on the SSD drive. The SSD handles
5,000 queries per second consistently and peaks to 30,000 queries per second! This new
shard could essentially handle the entire cluster’s load by itself.
366 | Chapter 23: Deploying MongoDB
Figure 23-2. Queries on an SSD
The other interesting point to note about SSDs versus spinning disks is the amount of
stress these relative loads placed on the system. If we take a look at hardware monitoring
on the spinning disk server (Figure 23-3), we can see that the disk is quite busy. The
main visible line on the chart is IO wait: the percent of time that the CPU was waiting
for disk IO. You can see that the CPU is waiting for disk IO at least 10% of the time and
this often spikes to above 50% of the time. This means that the user’s workload was
essentially being throttled by his disk (which was why he was adding SSDs).
Figure 23-3. CPU usage during the queries
In contrast, Figure 23-4 shows the CPU usage on the SSD machine. IO wait actually
isn’t even present on this graph: the two visible lines represent system time and user
time (the higher and lower lines, respectively). Thus, the limiting factor on this machine
is how fast the CPU can run. As the numbers are greater than 100%, this graph also
shows that multiple processors were being utilized. Contrast that to the chart in figure
Figure 23-3, where not even a single core could be fully utilized due to the lag from disk
IO.
Designing the System | 367
Figure 23-4. CPU usage during the queries
Finally, you can see some of the effect this has on MongoDB by looking at the graphs
of lock times in Figure 23-5. On the spinning disk, the database spent between 10% and
25% of the time locked and sometimes spiked to locked 100% of the time.
Figure 23-5. MongoDB lock percentage during the queries
Contrast this with the lock percentage on the SSD machine in Figure 23-6. MongoDB
is essentially unlocked the entire time. (The bump at the beginning of the chart was
from a data-loading operation that he performed before bringing the SSD online.)
368 | Chapter 23: Deploying MongoDB
Figure 23-6. Lock percentage on the SSD machine
As you can see, SSDs can shoulder a lot more load than spinning disks, but unfortunately
they aren’t an option for a lot of deployments. If you can use them, do so. Even if it’s not
possible to use SSDs for your entire cluster, consider deploying as many as possible and
then using the forced hot spot data pattern in Chapter 15 to take advantage of them.
Note that generally you cannot add an SSD to an existing replica set (where the other
members have spinning disks). If the SSD machine becomes primary and handles any‐
thing close to the load it is capable of, the other members will not be able to replicate
from it quickly enough and will fall behind. Thus, it is a good idea to add a new shard
to your cluster if you are introducing SSDs.
Note that SSDs are excellent for normal data usage patterns, but spinning disks actually
work very well for the journal. Putting the journal on a spinning disk and your data on
SSDs can save you SSD space and should have no impact on performance.
Recommended RAID Configurations
RAID is hardware or software that lets you treat multiple disks as though they were a
single disk. It can be used for reliability, performance, or both. A set of disks using RAID
is referred to as a RAID array (somewhat unfortunately, as RAID stands for redundant
array of inexpensive disks... array).
There are a number of ways to configure RAID depending on the features you’re looking
for, generally some combination of speed and fault-tolerance. These are the most com‐
mon varieties:
RAID0
Striping disks for improved performance. Each disk holds part of the data, similar
to MongoDB’s sharding. Because there are multiple underlying disks, lots of data
can be written to disk at the same time. This improves throughput on writes. How‐
ever, if a disk fails and the data is lost, there are no copies of it. It also can cause slow
reads (we’ve particularly seen this on Amazon’s Elastic Block Store), as some data
volumes may be slower than others.
Designing the System | 369
RAID1
Mirroring for improved reliability. An identical copy of the data is written to each
member of the array. This has lower performance than RAID0, as a single member
with a slow disk can slow down all writes. However, if a disk fails, you will still have
a copy of the data on another member of the array.
RAID5
Striping disks, plus keeping an extra piece of data about the other data that’s been
stored to prevent data loss on server failure. Basically, RAID5 can handle one disk
going down and hide that failure from the user. However, to do this, it is slower
than any of the other varieties listed here because it needs to calculate this extra
piece of information whenever data is written. This is particularly expensive with
MongoDB, as a typical workload does many small writes.
RAID10
A combination of RAID0 and RAID1: data is stripped for speed and mirrored for
reliability.
We recommend using RAID10: it is safer than RAID0 and can smooth out performance
issues that can occur with RAID1. Some people feel that RAID1 on top of replica sets
is overkill and opt for RAID0. It is a matter of personal preference: how much risk are
you willing to trade for performance?
Do not use RAID5: it is very, very slow.
CPU
MongoDB tends to be very light on CPU (note Figure 23-3 and Figure 23-4: 10,000
queries per second only uses about two CPU’s-worth of processing power). If you have
a choice between investing in memory and investing in CPU, go with memory every
time. Theoretically, you could max out multiple cores on reads or in-memory sorts, but
in practice that’s rare. CPU is heavily used for index builds and MapReduces, but as of
this writing adding more cores doesn’t help either of these cases.
When choosing between speed and number of cores, go with speed. MongoDB is better
at taking advantage of more cycles on a single processor than increased parallelization.
Choosing an Operating System
64-bit Linux is the operating system MongoDB runs best on. If possible use some flavor
of that. CentOS and RedHat Enterprise Linux are probably the most popular choices,
but any flavor should work (Ubuntu and Amazon Linux are also common). Use the
most recent stable versions of operating systems because old, buggy packages or kernels
can sometimes cause issues.
64-bit Windows is also well supported.
370 | Chapter 23: Deploying MongoDB
Other flavors of Unix are not as well supported: proceed with caution if you’re using
Solaris or one of the BSD variants. Builds for these systems have, at least historically,
had a lot of issues.
One important note on cross-compatibility: MongoDB uses the same wire protocol and
lays out data files identically on all systems, so you can deploy on a combination of
operating systems. For example, you could have a mongos process running on Windows
and the mongods that are its shards running on Linux. You can also copy data files from
Windows to Linux or visa versa with no compatibility issues.
Do not deploy any data-bearing server on a 32-bit system because this limits you to
about 2 GB of data (due to MongoDB using memory-mapped files). Arbiters and mon
gos processes can be run on 32-bit machines. Do not run any other type of MongoDB
server on a 32-bit machine.
MongoDB only works with little-endian architectures. Most drivers support both little-
and big- endian systems, so you can run clients on either. However, the server must
always be run on a little-endian machine.
Swap Space
You should allocate a small amount of swap in case memory limits are reached to prevent
the kernel from killing MongoDB. However, MongoDB does not usually use any swap
space.
The majority of memory MongoDB uses is “slippery”: it’ll be flushed to disk and replaced
with other memory as soon as the system requests the space for something else. There‐
fore, database data should never be written to swap space: it’ll be flushed back to disk
first.
However, occasionally MongoDB will use swap for operations that require ordering
data: either building indexes or sorting. It attempts not to use too much memory for
these types of operations, but by performing many of them at the same time you may
be able to force swapping.
If your application is managing to make MongoDB use swap space, you should look
into redesigning your application or reducing load on the swapping server.
Filesystem
For Linux, the ext4 filesystem or XFS are recommended for your data volumes. It is nice
to have a filesystem that can do filesystem snapshots for backups, but that’s a matter of
preference.
ext3 is not recommended, as it takes a long time to preallocate data files. MongoDB has
to regularly allocate and zero-fill 2 GB data files, which can freeze for minutes at a time
Designing the System | 371
on ext3. If ext3 is necessary, there are some hacks around this. However, try to use
something else if at all possible.
On Windows, either NTFS or FAT are fine.
Do not use NFS directly mounted for MongoDB storage. Some client
versions lie about flushing, randomly remount and flush the page cache,
and do not support exclusive file locking. Using NFS can cause journal
corruption and should be avoided at all costs.
Virtualization
Virtualization is a great way to get cheap hardware and be able to expand fast. However,
there are some downsides, particularly unpredictable network and disk IO. This section
covers virtualization-specific issues.
Turn Off Memory Overcommitting
The memory overcommit setting controls what happens when processes request too
much memory from the operating system. Depending on this setting, the kernel may
give memory to processes even if that memory is not actually available (in the hopes
that it’ll become available by the time the process needs it). That’s called overcommitting:
the kernel promises memory that isn’t actually there. This does not work well with
MongoDB.
The possible values for vm.overcommit_memory are 0 (the kernel guesses about how
much to overcommit), 1 (memory allocation always succeeds), or 2 (don’t commit more
virtual address space than swap space plus a fraction of the overcommit ratio). The value
2 is complicated, but it’s the best option available. To set this, run:
$ echo 2 > /proc/sys/vm/overcommit_memory
You do not need to restart MongoDB after changing this setting.
Mystery Memory
Sometimes the virtualization layer does not handle memory provisioning correctly.
Thus, you may have a virtual machine that claims to have 100 GB of RAM available but
only ever allows you to access 60 GB of it. Conversely, we’ve seen people that were
supposed to have 20 GB of memory end up being able to fit an entire 100 GB data set
into RAM!
Assuming you don’t end up on the lucky side, there isn’t much you can do. If your
readahead is set appropriately and your virtual machine just won’t use all the memory
it should, you may just have to switch virtual machines.
372 | Chapter 23: Deploying MongoDB
Handling Network Disk IO Issues
One of the biggest problems with using virtualized hardware is that you are generally
sharing a disk with other tenants, which exacerbates the disk slowness mentioned pre‐
vious because everyone is competing for disk IO. Thus, virtualized disks can have very
unpredictable performance: they can work fine while your neighbors aren’t busy and
suddenly slow down to a crawl if someone else starts hammering the disk.
The other issue is that this storage is often not physically attached to the machine Mon‐
goDB is running on, so even when you have a disk all to yourself disk IO will be slower
than it would be with a local disk. There is also the unlikely-but-possible scenario of
your MongoDB server losing its network connection to your data.
Amazon has what is probably the most widely-used networked block store, called Elastic
Block Store (EBS). EBS volumes can be connected to Elastic Cloud (EC2) instances and
can give a machine almost any amount of disk immediately. On the plus side, this makes
backups very easy (take a snapshot from a secondary, mount the EBS drive on another
instance, and start up mongod). On the downside, you may encounter very variable
performance.
If you require more predictable performance, there are a couple of options. The most
straightforward way to guarantee the performance you expect is to not host MongoDB
in the cloud. Host it on your own servers and you know no one else is slowing things
down. However, that’s not an option for a lot of people, so the next-best thing is to get
an instance that guarantees a certain number of IOPS (IO Operations Per Second). See
http://docs.mongodb.org for up-to-date recommendations on hosted offerings.
If you can’t pursue either of these options and you need more disk IO than an overloaded
EBS volume can sustain, there is a way to hack around it.
Basically, what you can do is to keep monitoring against the volume MongoDB is using.
If and when that volume slows down, immediately kill that instance and bring up a new
one with a different data volume.
There are a couple of statistics to watch for:
•Spiking IO utilization (“IO wait” on MMS), for obvious reasons.
•Page faults rates spiking. Note that changes in application behavior could also cause
working set changes: you should disable this assassination script before deploying
new versions of your application.
•The number of lost TCP packets go up (Amazon is particularly bad about this: when
performance starts to fall, it drops TCP packets all over the place).
•MongoDB’s read and write queues spiking (this can be seen on MMS or in mongo
stat’s qr/qw column).
Virtualization | 373
If your load varies over the day or week, make sure your script takes that into account:
you don’t want a rogue cron job killing off all of your instances because of an unusually-
heavy Monday morning rush.
This hack relies on you having recent backups or relatively quick-to-sync data sets. If
you have each instance holding terabytes of data, you might want to pursue an alter‐
native approach. Also, this is only likely to work: if your new volume is also being ham‐
mered, it will be just as slow as the old one.
Using Non-Networked Disks
This particular section uses Amazon-specific vocabulary. However, it may apply to other
providers.
Ephemeral drives are the actual disks attached to the physical machine your VM is
running on, so they don’t have a lot of the problems networked storage does. Local disks
can still be overloaded by other users on the same box, but with a large box you can be
reasonably sure you’re not sharing disks with too many others. Even with a smaller
instance, often the ephemeral drive will give better performance than a networked drive
so long as the other tenants aren’t doing tons of IOPS.
The downside is in the name: these disks are ephemeral. If your EC2 instance goes down,
there’s no guarantee you’ll end up on the same box when you restart the instance and
then your data will be gone.
Thus, ephemeral drives should be used with care. You should make sure that you do
not store any important or unreplicated data on these disks. In particular, do not put
the journal on these ephemeral drives or your database on network storage. In general,
think of ephemeral drives as a slow cache rather than a fast disk and use them
accordingly.
Configuring System Settings
There are several system settings which can help MongoDB run more smoothly and
which are mostly related to disk and memory access. This section covers each of these
options and how you should tweak it.
Turning Off NUMA
When machines had a single CPU, all RAM was basically the same in terms of access
time. As machines started to have more processors, engineers realized that having all
memory be equally far from each CPU (as shown in Figure 23-7) was less efficient than
having each CPU have some memory that is especially close to it and fast for that par‐
ticular CPU to access.
374 | Chapter 23: Deploying MongoDB
This architecture where each CPU has its own “local” memory is called non-uniform
memory architecture (NUMA), shown in Figure 23-8.
Figure 23-7. Uniform memory architecture: all memory has the same access cost for
each CPU
Figure 23-8. Non-Uniform Memory Architecture: certain memory is attached to a
CPU, giving the CPU faster access to that particular memory. CPUs can still access oth‐
er CPUs’ memory, but it is more expensive than accessing their own.
For lots of applications, NUMA works well: the processors often need different data
because they’re running different programs. However, this works terribly for databases
in general and MongoDB in particular because databases have such different memory
access patterns than other types of applications. MongoDB uses a massive amount of
memory and needs to be able to access memory that is “local” to other CPUs. However,
the default NUMA settings on many systems makes this difficult.
CPUs favor using the memory that is attached to them and processes tend to favor one
CPU over the others. This means that memory often fills up unevenly, leaving you with
one processor using 100% of its local memory and the other processors using only a
fraction of their memory, as shown in Figure 23-9.
Configuring System Settings | 375
Figure 23-9. Sample memory usage in a NUMA system
In the situation from Figure 23-9, suppose CPU1 needs some data that isn’t in memory
yet. It must use its local memory for data that doesn’t have a “home” yet, but its local
memory is full. Thus, it has to evict some of the data in its local memory to make room
for the new data, even though there’s plenty of space left on the memory attached to
CPU2! This process tends to cause MongoDB to run much slower than expected, as it
only has a fraction of the memory available that it should have. MongoDB vastly prefers
semi-efficient access to more data than extremely efficient access to less data.
Turning off NUMA is one of the magic “go faster” buttons that you definitely want to
make sure you’ve pressed. Like using SSDs, disabling NUMA just makes everything
work better.
If possible, disable NUMA on your BIOS. For example, if you’re using grub, you can
add the numa=off option to grub.cfg:
kernel /boot/vmlinuz-2.6.38-8-generic root=/dev/sda ro quiet numa=off
If your system cannot turn it off in BIOS, you’ll have to start mongod with:
$ numactl --interleave=all mongod [options]
Add this to any init scripts you use.
Also, turn off zone_reclaim_mode. This nasty setting can be thought of as “super NU‐
MA.” If it’s enabled, whenever a page of memory is accessed by a CPU, it will moved to
that CPU’s local memory. Thus, if you have a threadA on one CPU and threadB on
another and they’re both hitting a page of memory, that page will be copied from one
CPU’s local memory to the other’s on every single access. This is very, very slow.
To disable zone_reclaim_mode, run:
$ echo 0 > /proc/sys/vm/zone_reclaim_mode
You do not have to restart mongod for zone_reclaim_mode changes to take effect.
If you have NUMA enabled, your hosts will show up in yellow on MMS, as shown in
Figure 23-10. You can see the actual warning that’s causing it to be yellow by going to
the “Last Ping” tab. Figure 23-11 shows the warning you’ll see if NUMA is enabled.
376 | Chapter 23: Deploying MongoDB
Figure 23-10. A host with startup warnings in MMS
Figure 23-11. Startup warnings about NUMA
Once NUMA has been disabled, MMS will display the host in blue again. (There are
several other reasons a host may appear in yellow. Check for other startup warnings, as
well.)
Setting a Sane Readahead
Readahead is an optimization where the operating system reads more data from disk
than was actually requested. This is useful because most workloads computers handle
are sequential: if you load the first 20 MB of a video, you are probably going to want the
next couple of megabytes of it. Thus, the system will read more from disk than you
actually request and store it in memory, just in case you need it soon.
However, MongoDB is not a typical workload and readahead is a frequent issue on
MongoDB systems. MongoDB tends to read many small pieces of data from random
places on the disk, so the default system settings do not work very well. If readahead is
high, memory gradually fills up with data that MongoDB didn’t request, forcing Mon‐
goDB to go to disk more often.
For example, if you wanted to read one sector (512 bytes) from disk, the disk controller
might actually be instructed to read 256 sectors, on the assumption that you will request
them soon, anyway. However, if you are accessing data fairly randomly across the disk,
all of those prefetched sectors would be wasted. If memory was contained your working
set, 255 sectors of your working set would have to be evicted to make room for these
sectors that aren’t going to be used. 256 sectors is actually a small readahead, too: some
systems default to thousands of sectors of readahead.
Fortunately, there’s a fairly easy way to see if your readahead setting is actively hurting
you: check the resident set size of MongoDB and compare it to the system’s total RAM.
Configuring System Settings | 377
Assuming that your RAM is smaller than your data size, MongoDB’s resident set size
should be a little lower than your total size of RAM (for example, if you have 50 GB of
RAM, MongoDB should be using at least 46 GB). If it is much lower, then your readahead
is probably too high.
This technique of comparing resident set size with total memory size works because
data that’s been “read ahead” from disk is in memory but MongoDB didn’t request it,
so it isn’t included in the calculation of MongoDB’s resident memory size.
To see your current readahead settings, use the blockdev command:
$ sudo blockdev --report
RO RA SSZ BSZ StartSec Size Device
rw 256 512 4096 0 80026361856 /dev/sda
rw 256 512 4096 2048 80025223168 /dev/sda1
rw 256 512 4096 0 2000398934016 /dev/sdb
rw 256 512 1024 2048 98566144 /dev/sdb1
rw 256 512 4096 194560 7999586304 /dev/sdb2
rw 256 512 4096 15818752 19999490048 /dev/sdb3
rw 256 512 4096 54880256 1972300152832 /dev/sdb4
This shows you the settings for each of your block devices. Readahead (the RA column)
is measured in 512-byte sectors. Thus, this system’s readahead is set to 128 KB for each
device (512 bytes/sector * 256 sectors).
You can change this setting for a device by running this command with the --setra
option:
$ sudo blockdev --setra 16 /dev/sdb3
So, what’s a good setting for readahead? Between 16 and 256 is recommended. You don’t
want to set readahead too low, either: you don’t want to have to go to disk multiple times
to fetch a single document. If you have large documents (greater than a megabyte in
size), consider a higher readahead. If your documents are small, stick with something
low, like 32. Do not go below 16 even if your documents are tiny, as that will make
fetching indexes less efficient (index buckets are 8 KB).
When using RAID, you have to set readahead on everything: the RAID controller and
the individual volumes.
Unintuitively, you must restart MongoDB for readahead settings to take effect. You’d
think you’re setting a disk property, so it should apply to all running programs. Un‐
fortunately, processes make a copy of the readahead value when they start and continue
to use that value until terminated.
Disabling Hugepages
Hugepages cause similar issues to high readahead. Do not use hugepages unless:
378 | Chapter 23: Deploying MongoDB
• All of your data fits into memory and
• You have no plans for it to ever grow beyond memory.
MongoDB needs to page in lots of tiny pieces of memory, so using hugepages can result
in more disk IO.
Systems move data from disk to memory and back by the page. Pages are generally a
couple of kilobytes (x86 defaults to 4096-byte pages). If a machine has many gigabytes
of memory, keeping track of each of these (relatively tiny) pages can be slower than just
tracking a few larger-granularity pages, so hugepages allows you to have pages that are
up to 256 MB (on ia64 architectures). However, using hugepages means that you are
keeping megabytes of data from one section of disk in memory. If your data does not
fit in RAM, then swapping in larger pieces from disk will just fill up your memory
quickly with data that will need to be swapped out again. Also, flushing any changes to
disk will be slower, as the disk must write megabytes of “dirty” data, instead of a few
kilobytes.
Note that on Windows this is called Large Pages, not hugepages. Some versions of Win‐
dows have it enabled by default and some do not, so check and make sure it is turned
off.
Hugepages were actually developed to benefit databases, so this may be surprising to
experienced database admins. However, MongoDB tends to do a lot less sequential disk
access than relational databases do.
Choosing a Disk Scheduling Algorithm
The disk controller receives requests from the operating system and processes them in
an order determined by a scheduling algorithm. Sometimes changing this algorithm
can improve disk performance. For other hardware and workloads, it may not make a
difference. The best way to decide is to test them out yourself on your workload. Dead‐
line and completely fair queueing (CFQ) both tend to be good choices.
There are a couple of situations where the noop scheduler (a contraction of “no-op” is
the best choice): if you’re in a virtualized environment, use the noop scheduler. The
noop scheduler basically passes the operations through to the underlying disk controller
as quickly as possible. It is fastest to do this and let the real disk controller handle any
reordering that needs to happen.
Similarly, on SSDs, the noop scheduler is generally the best choice. SSDs don’t have the
same locality issues that spinning disks do.
Finally, if you’re using a RAID controller with caching, use noop. The cache behaves
like an SSD and will take care of propagating the writes to the disk efficiently.
Configuring System Settings | 379
You can change the scheduling algorithm by setting the --elevator option in your boot
configuration.
The option is called elevator because the scheduler behaves like an el‐
evator, picking up people (IO requests) from different floors (processes/
times) and dropping them off where they want to go in an arguable-
optimal way.
Often all of the algorithms work pretty well; you may not see much of a difference
between them.
Don’t Track Access Time
By default, the system tracks when files were last accessed. As the data files used by
MongoDB are very high-traffic, you can get a performance boost by disabling this
tracking. You can do this on Linux by changing atime to noatime in /etc/fstab:
/dev/sda7 /data ext4 rw,noatime 1 2
You must remount the device for the changes to take effect.
atime is more of an issue on older kernels (e.g., ext3), as newer ones use relatime as a
default which is less aggressively updated. Also, be aware that setting noatime can affect
other programs using the partition, such as mutt or backup tools.
Similarly, on Windows you should set the disablelastaccess option. To turn off last
access time recording, run:
C:\> fsutil behavior set disablelastaccess 1
You must reboot for this setting to take effect. Setting this may affect the Remote Storage
service, but you probably shouldn’t be using a service that automatically moves your
data to other disks anyway.
Modifying Limits
There are two limits that MongoDB tends to blow by: the number of threads a process
is allowed to spawn and the number of file descriptors a process is allowed to open. Both
of these should generally be set to unlimited.
Whenever a MongoDB server accepts a connection, it spawns a thread to handle all
activity on that connection. Therefore, if you have 3,000 connections to the database,
the database will have 3,000 threads running (plus a few other threads for non-client-
related tasks). Depending on your application server configuration, your client may
spawn anywhere from a dozen to thousands of connections to MongoDB.
380 | Chapter 23: Deploying MongoDB
If your client will dynamically spawn more child processes as traffic increases (most
application servers will do this), it is important to make sure that these child processes
are not so numerous that they can max out MongoDB’s limits. For example, if you have
20 application servers, each one of which is allowed to spawn 100 child processes and
each child process can spawn 10 threads that all connect to MongoDB, that could spawn
20 × 100 × 10 = 20,000 connections at peak traffic. MongoDB is probably not going to
be very happy about spawning tens of thousands of threads and, if you run out of threads
per process, will simply start refusing new connections.
The other limit to modify is the number of file descriptors MongoDB is allowed to open.
Every incoming and outgoing connection uses a file descriptor, so having the client
connection storm above would create 20,000 open filehandles (incidentally, the maxi‐
mum number MongoDB will allow).
mongos in particular tends to create connections to many shards. When a client connects
to a mongos and makes a request, the mongos opens connections to any and all shards
necessary to fulfill that request. Thus, if a cluster had 100 shards and a client connects
to a mongos and tries to query for all of their data, the mongos must open 100 connec‐
tions: one connection to each shard. This can quickly mount in number of connections,
as you can imagine from the previous example. Suppose a liberally configured app server
made a hundred connections to a mongos process. Then this could get translated to 100
inbound connections × 100 shards = 10,000 connections to shards! (This assumes a
non-targeted query on each connection, which would be a bad design, so this is a
somewhat extreme example).
Thus, there are a couple adjustments to make: many people purposefully configure
mongos processes to only allow a certain number of incoming connections by using the
maxConns option. This is a good way to enforce that your client is behaving well.
You should also increase the limit on the number of file descriptors, as the default (gen‐
erally 1024) is simply too low. Set the max number of file descriptors to unlimited or, if
you’re nervous about that, 20,000. Each system has a different way of changing these
limits, but in general, make sure that you change both the hard and soft limits. A hard
limit is enforced by the kernel and can only be changed by an administrator, versus a
soft limit, which is user-configurable.
If the number of connections is left at 1024, MMS will warn you by displaying the host
in yellow on the host list (as shown in the NUMA example above). If low limits are the
issue, the “Last Ping” tab should display a message similar to that shown in Figure 23-12.
Configuring System Settings | 381
Figure 23-12. MMS low ulimit warning
Even if you have a non-sharded setup and an application that only uses a small number
of connections, increase the hard and soft limits to at least 4096. That will stop MongoDB
from warning you about them and give you some breathing room, just in case.
Configuring Your Network
This section covers which servers should have connectivity to which other servers.
Often, for reasons of network security (and sensibility), you may want to limit the con‐
nectivity of MongoDB servers. Note that multiserver MongoDB deployments should
handle networks being partitioned or down, but it isn’t recommended as a general de‐
ployment strategy.
For a standalone server, clients must be able to make connections to the mongod.
Members of a replica set must be able to make connections to every other member.
Clients must be able to connect to all nonhidden, nonarbiter members. Depending on
network configuration, members may also attempt to connect to themselves, so you
should allow mongods to create connections to themselves.
Sharding is a bit more complicated. There are four components: mongos servers, shards,
config servers, and clients. Connectivity can be summarized in the following three
points:
•A client must be able to connect to a mongos.
•A mongos must be able to connect to the shards and config servers.
•A shard must be able to connect to the other shards and the config servers.
The full connectivity chart is described in Table 23-1.
Table 23-1. Sharding connectivity
Connectivity from server type
to server type mongos Shard Config server Client
mongos Not required Not required Not required Required
Shard Required Required Not required Not recommended
Config server Required Required Not required Not recommended
Client Not required Not required Not required Not MongoDB-related
382 | Chapter 23: Deploying MongoDB
There are three possible values in the table: “Required” means that connectivity between
these two components is required for sharding to work as designed. MongoDB will
attempt to degrade gracefully if it loses these connections due to network issues, but
you shouldn’t purposely configure it that way.
“Not required” means that these two elements never talk in the direction specified, so
no connectivity is needed.
“Not recommended” means that these two elements never talk, but due to user error
they could. For example, it is recommended that clients only make connections to the
mongos, not the shards, so that clients do not inadvertently make requests directly to
shards. Similarly, clients should not be able to directly access config servers so that they
cannot accidentally modify config data.
Note that mongos processes and shards talk to config servers, but config servers don’t
make connections to anyone, even one another.
Shards must communicate during migrates: shards connect to one another directly to
transfer data.
As mentioned earlier, replica set members that compose shards should be able to con‐
nect to themselves.
System Housekeeping
This section covers some common issues you should be aware of before deploying.
Synchronizing Clocks
In general, it’s safest to have your systems’ clocks within a second of each other. Replica
sets should be able to handle nearly any clock skew. Sharding can handle some skew (if
it gets beyond a few minutes, you’ll start seeing warnings in the logs), but it’s best to
minimize. Having in-sync clocks also makes figuring out what’s happening from logs
easier.
You can keep clocks synchronized using the w32tm tool on Windows and the ntp
daemon on Linux.
The OOM Killer
Very occasionally, MongoDB will allocate enough memory that it will be targeted by
the OOM killer (out-of-memory killer). This particularly tends to happen during index
builds, as that is one of the only times when MongoDB’s resident memory should put
any strain on the system.
System Housekeeping | 383
If your MongoDB process suddenly dies with no errors or exit messages in the logs,
check /var/log/messages (or wherever your kernel logs such things) to see if it has any
messages about terminating mongod.
If the kernel has killed MongoDB for memory overuse, you should see something like
this in the kernel log:
kernel: Killed process 2771 (mongod)
kernel: init invoked oom-killer: gfp_mask=0x201d2, order=0, oomkilladj=0
If you were running with journaling, you can simply restart mongod at this point. If you
were not, restore from a backup or resync the data from a replica.
The OOM killer gets particularly nervous if you have no swap space and start running
low on memory, so a good way to prevent it from going on a spree is to configure a
modest amount of swap. MongoDB should never use it, but it makes the OOM killer
happy.
If the OOM killer kills a mongos, you can simply restart it.
Turn Off Periodic Tasks
Check that there aren’t any cron jobs or daemons that might periodically pop to life and
steal resources. One culprit we’ve seen is package managers’ automatic update. These
programs will come to life, consume a ton of RAM and CPU, and then disappear. This
is not something that you want running on your production server.
384 | Chapter 23: Deploying MongoDB
APPENDIX A
Installing MongoDB
MongoDB binaries are available for Linux, Mac OS X, Windows, and Solaris. This means
that, on most platforms, you can download an archive from http://www.mongodb.org/
downloads, inflate it, and run the binary.
The MongoDB server requires a directory it can write database files to and a port it can
listen for connections on. This section covers the entire install on the two variants of
system: Windows and everything else (Linux, Max, Solaris).
When we speak of “installing MongoDB,” generally what we are talking about is setting
up mongod, the core database server. mongod can be used as a standalone server or as a
member of a replica set. Most of the time, this will be the MongoDB process you are
using.
Choosing a Version
MongoDB uses a fairly simple versioning scheme: even-point releases are stable, and
odd-point releases are development versions. For example, anything starting with 2.4
is a stable release, such as 2.4.0, 2.4.1, and 2.4.15. Anything starting with 2.5 is a devel‐
opment release, such as 2.5.0, 2.5.2, or 2.5.10. Let’s take the 2.4/2.5 release as a sample
case to demonstrate how the versioning timeline works:
1. MongoDB 2.4.0 is released. This is a major release and will have an extensive
changelog.
2. After the developers start working on the milestones for 2.6 (the next major stable
release), they release 2.5.0. This is the new development branch that is fairly similar
to 2.4.0 but probably with an extra feature or two and maybe some bugs.
3. As the developers continue to add features, they will release 2.5.1, 2.5.2, and so on.
These releases should not be used in production.
385
4. Some minor bug fixes may be backported to the 2.4 branch, which will cause releases
of 2.4.1, 2.4.2, and so on. Developers are conservative about what is backported;
few new features are ever added to a stable release. Generally, only bug fixes are
ported.
5. After all of the major milestones have been reached for 2.6.0, 2.5.7 (or whatever the
latest development release is) will be turned into 2.6.0-rc0.
6. After extensive testing of 2.6.0-rc0, usually there are a couple minor bugs that need
to be fixed. Developers fix these bugs and release 2.6.0-rc1.
7. Developers repeat step 6 until no new bugs are apparent, and then 2.6.0-rc2 (or
whatever the latest release ended up being) is renamed 2.6.0.
8. Start over from step 1, incrementing all versions by 0.2.
You can see how close a production release is by browsing the core server roadmap on
the MongoDB bug tracker.
If you are running in production, you should use a stable release. If you are planning
to use a development release in production, ask about it first on the mailing list or IRC
to get the developers’ advice.
If you are just starting development on a project, using a development release may be
a better choice. By the time you deploy to production, there will probably be a stable
release with the features you’re using (MongoDB attempts to stick to a regular cycle of
stable releases every six months). However, you must balance this against the possibility
that you would run into server bugs, which can be discouraging to a new user.
Windows Install
To install MongoDB on Windows, download the Windows zip from the MongoDB
downloads page. Use the advice in the previous section to choose the correct version of
MongoDB. There are 32-bit and 64-bit releases for Windows, so select whichever ver‐
sion you’re running. When you click the link, it will download the .zip. Use your favorite
extraction tool to unzip the archive.
Now you need to make a directory in which MongoDB can write database files. By
default, MongoDB tries to use the \data\db directory on the current drive as its data
directory (for example, if you’re running mongod on C:, it’ll use C:\data\db). You can
create this directory or any other empty directory anywhere on the filesystem. If you
chose to use a directory other than \data\db, you’ll need to specify the path when you
start MongoDB, which is covered in a moment.
Now that you have a data directory, open the command prompt (cmd.exe). Navigate to
the directory where you unzipped the MongoDB binaries and run the following:
$ bin\mongod.exe
386 | Appendix A: Installing MongoDB
If you chose a directory other than C:\data\db, you’ll have to specify it here, with the
--dbpath argument:
$ bin\mongod.exe --dbpath C:\Documents and Settings\Username\My Documents\db
See Chapter 20 for more common options, or run mongod.exe --help to see all options.
Installing as a Service
MongoDB can also be installed as a service on Windows. To install, simply run with the
full path, escape any spaces, and use the --install option. For example:
$ C:\mongodb-windows-32bit-1.6.0\bin\mongod.exe
--dbpath "\"C:\Documents and Settings\Username\My Documents\db\"" --install
It can then be started and stopped from the Control Panel.
POSIX (Linux, Mac OS X, and Solaris) Install
Choose a version of MongoDB, based on the advice in the section “Choosing a Ver‐
sion” on page 385. Go to the MongoDB downloads page, and select the correct version
for your OS.
If you are using a Mac, check whether you’re running 32-bit or 64-bit.
Macs are especially picky that you choose the correct build and will
refuse to start MongoDB and give confusing error messages if you
choose the wrong build. You can check what you’re running by clicking
the apple in the upper-left corner and selecting the About This Mac
option.
You must create a directory for the database to put its files. By default, the database will
use /data/db, although you can specify any other directory. If you create the default
directory, make sure it has the correct write permissions. You can create the directory
and set the permissions by running the following:
$ mkdir -p /data/db
$ chown -R $USER:$USER /data/db
mkdir -p creates the directory and all its parents, if necessary (i.e., if the /data directory
didn’t exist, it will create the /data directory and then the /data/db directory). chown
changes the ownership of /data/db so that your user can write to it. Of course, you can
also just create a directory in your home folder and specify that MongoDB should use
that when you start the database, to avoid any permissions issues.
Decompress the .tar.gz file you downloaded from http://www.mongodb.org:
POSIX (Linux, Mac OS X, and Solaris) Install | 387
$ tar zxf mongodb-linux-i686-1.6.0.tar.gz
$ cd mongodb-linux-i686-1.6.0
Now you can start the database:
$ bin/mongod
Or if you’d like to use an alternate database path, specify it with the --dbpath option:
$ bin/mongod --dbpath ~/db
See section TODO for a summary of the most common options, or run mongod with
--help to see all the possible options.
Installing from a Package Manager
On these systems, there are many package managers that can also be used to install
MongoDB. If you prefer using one of these, there are official packages for RedHat,
Debian, and Ubuntu as well as unofficial packages for many other systems. If you use
an unofficial version, make sure it installs a relatively recent version.
On OS X, there are unofficial packages for Homebrew and MacPorts. If you go for the
MacPorts version, be forewarned: it takes hours to compile all the Boost libraries, which
are MongoDB prerequisites. Start the download and leave it overnight.
Regardless of the package manager you use, it is a good idea to figure out where it is
putting the MongoDB log files before you have a problem and need to find them. It’s
important to make sure they’re being saved properly in advance of any possible issues.
388 | Appendix A: Installing MongoDB
APPENDIX B
MongoDB Internals
It is not necessary to understand MongoDB’s internals to use it effectively, but it may
be of interest to developers who wish to work on tools, contribute, or simply understand
what’s happening under the hood. This appendix covers some of the basics. The Mon‐
goDB source code is available at https://github.com/mongodb/mongo.
BSON
Documents in MongoDB are an abstract concept—the concrete representation of a
document varies depending on the driver/language being used. Because documents are
used extensively for communication in MongoDB, there also needs to be a representa‐
tion of documents that is shared by all drivers, tools, and processes in the MongoDB
ecosystem. That representation is called Binary JSON, or BSON (no one knows where
the J went).
BSON is a lightweight binary format capable of representing any MongoDB document
as a string of bytes. The database understands BSON, and BSON is the format in which
documents are saved to disk.
When a driver is given a document to insert, use as a query, and so on, it will encode
that document to BSON before sending it to the server. Likewise, documents being
returned to the client from the server are sent as BSON strings. This BSON data is
decoded by the driver to its native document representation before being returned to
the client.
The BSON format has three primary goals:
Efficiency
BSON is designed to represent data efficiently, without using much extra space. In
the worst case BSON is slightly less efficient than JSON; and in the best case (e.g.,
when storing binary data or large numerics), it is much more efficient.
389
Traversability
In some cases, BSON does sacrifice space efficiency to make the format easier to
traverse. For example, string values are prefixed with a length rather than relying
on a terminator to signify the end of a string. This traversability is useful when the
MongoDB server needs to introspect documents.
Performance
Finally, BSON is designed to be fast to encode to and decode from. It uses C-style
representations for types, which are fast to work with in most programming
languages.
For the exact BSON specification, see http://www.bsonspec.org.
Wire Protocol
Drivers access the MongoDB server using a lightweight TCP/IP wire protocol. The
protocol is documented on the MongoDB wiki but basically consists of a thin wrapper
around BSON data. For example, an insert message consists of 20 bytes of header data
(which includes a code telling the server to perform an insert and the message length),
the collection name to insert into, and a list of BSON documents to insert.
Data Files
Inside of the MongoDB data directory, which is /data/db/ by default, there are separate
files for each database. Each database has a single .ns file and several data files, which
have monotonically increasing numeric extensions. So, the database foo would be stored
in the files foo.ns, foo.0, foo.1, foo.2, and so on.
The numeric data files for a database will double in size for each new file, up to a max‐
imum file size of 2 GB. This behavior allows small databases to not waste too much space
on disk, while keeping large databases in mostly contiguous regions on disk.
MongoDB also preallocates data files to ensure consistent performance. (This behavior
can be disabled using the --noprealloc option.) Preallocation happens in the back‐
ground and is initiated every time that a data file is filled. This means that the MongoDB
server will always attempt to keep an extra, empty data file for each database to avoid
blocking on file allocation.
Namespaces and Extents
Within its data files, each database is organized into namespaces, each storing a specific
collection’s data. The documents for each collection have their own namespace, as does
each index. Metadata for namespaces is stored in the database’s .ns file.
390 | Appendix B: MongoDB Internals
The data for each namespace is grouped on disk into sections of the data files, called
extents. In Figure B-1 the foo database has three data files, the third of which has been
preallocated and is empty. The first two data files have been divided up into extents
belonging to several different namespaces.
Figure B-1. Namespaces and extents
Figure B-1 shows us several interesting things about namespaces and extents. Each
namespace can have several different extents, which are not (necessarily) contiguous
on disk. Like data files for a database, extents for a namespace grow in size with each
new allocation. This is done to balance wasted space used by a namespace versus the
desire to keep data for a namespace mostly contiguous on disk. The figure also shows
a special namespace, $freelist, which keeps track of extents that are no longer in use (e.g.,
extents from a dropped collection or index). When a namespace allocates a new extent,
it will first search the freelist to see whether an appropriately sized extent is available.
Memory-Mapped Storage Engine
The default storage engine (and only supported storage engine at the time of this writ‐
ing) for MongoDB is a memory-mapped engine. When the server starts up, it memory
maps all its data files. It is then the responsibility of the operating system to manage
flushing data to disk and paging data in and out. This storage engine has several im‐
portant properties:
Memory-Mapped Storage Engine | 391
•MongoDB’s code for managing memory is small and clean because most of that
work is pushed to the operating system.
•The virtual size of a MongoDB server process is often very large, exceeding the size
of the entire data set. This is OK because the operating system will handle keeping
the amount of data resident in memory contained.
•32-bit MongoDB servers are limited to a total of about 2 GB of data per mongod.
This is because all of the data must be addressable using only 32 bits.
392 | Appendix B: MongoDB Internals
We’d like to hear your suggestions for improving our indexes. Send email to index@oreilly.com.
Index
Symbols
$ (dollar sign)
$**, creating index on, 117
$- operators (see query operators)
position operator, 41
reserved character, 8
- (minus sign), using to not include a string in
full-text index queries, 118
. (dot), reserved character, 8
2D indexes, 122
2dsphere type, 120
using with ensureIndex to create geospatial
indexes, 120
32-bit systems, 371
64-bit systems, 370
A
access time, not tracking, 380
accessNotInMemory field, serverStatus com‐
mand, 345
acknowledged writes, 51
add function, 211
addShardTag function, 272
$addToSet operator, 39, 137
using with $each, 40
addUser function, 312
admin database, 11
adding a user, 312
commands requiring administrator access,
77
sharding and, 314
user privileges, 312
adminCommand function, 77
administration, 209–227
data administration, 311–322
authentication, 311–315
compacting data, 320
indexes, 315–317
moving collections, 321
preallocating data files, 322
preheating data, 317–320
manipulating replica set member state, 213–
214
preventing elections, 213
turning primaries into secondaries, 213
using maintenance mode, 213
master-slave setup, 225–227
converting to replica set, 226
mimicking behavior with replica sets, 226
monitoring replication, 214–225
building indexes, 222
calculating lag, 219
disabling chaining, 219
getting status, 214
lower-cost replication, 223
replication loops, 218
resizing the oplog, 220
393
restoring from delayed secondary, 221
tracking of lag by primary, 224
visualizing replication graph, 216–218
of sharding, 275–295
balancing data, 289–295
refreshing configurations, 295
seeing configuration information, 277–
283
server administration, 285–289
tracking network connections, 283–285
replica set configuration, 210–212
changing set members, 211
creating a replica set, 210
creating larger sets, 211
forcing reconfiguration, 212
starting replica set members in standalone
mode, 209
aggregate function, 128
aggregation, 4, 127–151
aggregation framework, 127
commands, 146–151
count, 146
distinct, 147
group, 147–151
MapReduce, 140–146
categorizing web pages (example), 143
finding all keys in collection (example),
141–143
MongoDB and, 143–146
pipeline operations, 129–140
$group, 135–137
$limit, 139
$match, 129
$project, 130–135
$skip, 139
$sort, 139
$unwind, 137–139
using pipelines, 140
$all operator, 59
allowChaining setting, 219
Amazon, networked block store (EBS), 373
$and operator, 57, 133
application design, 153–165
managing consistency, 163
migrating schemas, 164
normalization versus denormalization, 153–
160
cardinality, 157
considerations for social graph data, 158–
160
examples of data representations, 154–
157
optimizations for data manipulation, 160
planning databases and collections, 162
when not to use MongoDB, 165
application operations, checking on, 299–309
calculating sizes, 305–307
of collections, 305
of databases, 306
of documents, 305
seeing current operations, 299–302
false positives, 302
finding problematic operations, 301
killing operations, 301
preventing phantom operations, 302
using mongotop and mongostat, 307
using system profiler, 303–304
arbiters
creating election arbiters, 182
downside to using, 183
using at most one, 182
arithmetic operators, 135
array modifiers
$ position operator, 41
$pop, 41
$pull, 41
$push, 37
$slice, 39
$sort, 39
performance and, 44
using arrays as sets, 39
$addToSet modifier, 39
array operators, 137
array type, 17
arrays, 18
indexing, 96
multikey indexes and, 97
querying, 59–63
$all operator, 59
$size operator, 60
$slice operator, 61
interactions of array and range queries,
62
returning matching element with $ (posi‐
tion) operator, 62
update modifiers for, 37
ascending shard keys, 258–261
394 | Index
authentication, 311–315
basics of, 312
how it works, 314
locking issues with fsyncLock, 358
setting up, 314
autoIndexId option, createCollection command,
114
automatic failover, 172
autosharding, 231
$avg operator, 136
B
background flush averages, 346
backups, 357–363
backing up a replica set, 361
backing up a server, 357–361
copying data files, 358
filesystem snapshot, 357
using mongodump, 359
backing up sharded cluster, 362
incremental, with mongooplog, 363
balancers, 253, 289
turning off, 273
balancing data, 289–295
balancer, 289
changing chunk size, 290
jumbo chunks, 292–295
distributing, 294
preventing, 295
batch inserts, 29
big-endian systems, 371
binary data type, 18
blockdev command, 378
boolean expressions, 133
boolean type, 17
$box operator, querying all points in a rectangle,
122
BSD variants, 370
BSON, 389
dump directory .bson files, 360
invalid BSONObj, 329
size, checking for documents, 30
bsondump utility, 360
btree misses, minimizing, 345
buildIndexes setting, 185, 316
C
capped collections, 109, 162
(see also collections)
cardinality, 98, 157
shard key, 271
case, string operations affecting, 133
$center operator, 123
chaining of secondaries, disabling, 219
changelog collection, 279
from shard’s changelog, 280
to shard’s changelog, 281
chunks
breaking large file into, for storage, 125
changing size of, 290
config.chunks collection, 279
jumbo, 292–295
moving, 291
ranges, 247
splitting, 249–253
turning off, 253
splitting collection into, 246
for sharding, 236
circles, finding all points within, using $center
operator, 123
client-to-replica-set connection behavior, 199
clients, connectivity in sharded clusters, 382
clocks, synchronizing, 383
cloneCollection command, 322
cloud, MongoDB hosting and, 373
cluster.stop function, 240
clusters
components of, 232
creating, 232
getting overall view of with sh.status func‐
tion, 234
tracking network connections in, 283–285
using cluster for multiple databases and col‐
lections, 272
$cmd collection, 76
$cmp operator, 133
code type, 18
collections, 8
accessing from db variable, 14
capped, 109–114
creating, 111
natural sorting, 112
no-_id collections, 114
tailable cursors, 113
config.collections collection, 278
Index | 395
dynamic schemas, 8
inconvenient or invalid names, dealing with,
27
keeping MapReduce output collections, 144
moving, 321
moving into RAM, 318
moving with mongodump and mongores‐
tore, 360
naming, 9
planning in application design, 162
sharding, 246
special types, 4
using a cluster for multiple collections, 272
viewing information about, 305
collMod command
expireAfterSecs option, 115
setting usePowerOf2Sizes option, 45
command line, starting MongoDB from, 333–
336
file-based configuration, 336
mongod startup options, 333
commit batches, planning, 324
compacting data, 320
comparisons
comparison expressions, 133
comparison operators, 55
comparison order for types, 69
compound indexes
geospatial, 121
introduction to, 84–89
unique, 104
using, 89
choosing key directions, 90
covered indexes, 91
implicit indexes, 91
$concat operator, 133
$cond operator, 134
conditionals, 55
semantics, 57
config database, 11
config function, 211
--config option, mongodb, 334
config servers, 242
changing, 288
connectivity, 382
config.changelog collection, 279
config.chunks collection, 247, 279
config.collections collection, 278
config.databases collection, 278
config.locks collection, 254, 289
config.settings collection, 283
config.shards collection, 277
updates to, 285
config.tags collection, 282
--configdb option, mongos, 289
configuration files, reading from at startup, 336
configurations, refreshing, 295
connection pooling, 164
connectTo function, 24
connPoolStats command, 283
consistency
embedding versus references, 156
managing, 163
control statements, 134
convertToCapped command, 112
corruption
checking for, 327–329
removing through repairs, 325
count command, 146
covered indexes, 91
CPU usage, 350
SSDs versus spinning disks, 367
CPUs, 370
crashes
with journaling, 324
without journaling, 325
create operations, mongo shell, 14
createCollection command, 111
autoIndexId option set to false, 114
cron job to rotate logs, 339
Ctrl-C, sending SIGINT with, 337
currentOp function, 299–302
cursors, 67–75
advanced query options, 71
avoiding large skips, 70–71
creating with mongo shell, 67
getting consistent results, 72–74
immortal, 75
in explain function output field, 100
iterating through results with next and has‐
Next methods, 67
limits, skips, and sorts, 68
tailable, 113
D
data administration, 311–322
compacting data, 320
396 | Index
creating and deleting indexes, 315–317
creating index on replica set, 315
creating index on sharded cluster, 316
creating index on stand-alone server, 315
OOM (out-of-memory) killer, 317
removing indexes, 316
moving collections, 321
preallocating data files, 322
preheating data, 317–320
custom preheating, 318
moving collections into RAM, 318
moving databases into RAM, 317
setting up authentication, 311–315
basics of authentication, 312
how authentication works, 314
data directory, 12
data distribution, controlling, 271
manual sharding, 273
using cluster for multiple databases and col‐
lections, 272
data types, 16–21
arrays, 18
basic, 16–18
comparison order in MongoDB, 69
dates, 18
embedded documents, 19
_id and ObjectIds, 20
type-specific queries, 58–65
embedded documents, 63
null type, 58
querying arrays, 59–63
regular expressions, 58
database commands, 75–77
how they work, 76
databases
calculating sizes with stats function, 306
config.databases collection, 278
directoryperdb option, 353
in MongoDB, 10
moving into RAM, 317
moving with mongodump and mongores‐
tore, 360
planning in application design, 162
repairing a single database, 321
sharding, 246
using a cluster for multiple databases, 272
dataSize command, 293
dates, 18
date expressions in aggregation pipeline, 132
date type, 17
range queries for, 55
$dayOfMonth operator, 132
$dayOfWeek operator, 132
$dayOfYear operator, 132
db global variable, 14
scripts’ access to, 23
db.addUser function, 312
db.boards.stats function, 305
db.currentOp function, 299–302
db.enableSharding function, 246
db.getProfilingLevel function, 304
db.help command, 22
db.killOp function, 301
db.printReplicationInfo function, 219
db.printSlaveReplicationInfo function, 219
db.repairDatabase function, 321
db.setProfilingLevel function, 303, 339
db.stats function, 306
dbhash command, 286
--dbpath option, mongod, 333
dd tool, 317
delete operations, mongo shell, 16
denormalization, 153
(see also normalization and denormaliza‐
tion)
deploying MongoDB, 365–384
configuring system settings, 374–382
disabling hugepages, 378
disk scheduling algorithm, 379
modifying limits, 380
not tracking access time, 380, 380
setting sane readahead, 377
turning off NUMA, 375–377
configuring your network, 382
designing the system, 365–384
choosing operating system, 370
choosing storage medium, 365–372
CPU, 370
filesystem, 371
RAID configurations, 369
swap space, 371
system housekeeping, 383
virtualization, 372–374
diaglog, 319
--directoryperdb option, mongod, 334
directoryperdb option, 353
disk scheduling algorithm, 379
disk usage, monitoring, 352
Index | 397
distinct command, 147
document-oriented database, 3
documents, 29–52
adding to MongoDB collection, 14
embedded, 17, 19
getting size of, 305
in MongoDB, 7
inserting and saving, 29
loading recently created documents into
RAM, 319
optimizing for document growth, 160
removing, 31
setting a write concern, 51
updating, 32–51
by document replacement, 32
multiple documents, 47
returning updated documents, 48–50
upserts, 45
using modifiers, 34
draining, 287
drop function, 31
dropIndex command, 108
dropIndexes command, 107, 316
dump directory, 360
duplicates
dropping for unique indexes, 105
filtering out of unique indexes, 104
durability, 323–329
checking for corruption, 327
journaling, 323
situations without guarantee of, 327
turning off journaling, 325
mongod.lock file, 326
repairing data files, 325
replacing data files, 325
sneaky unclean shutdowns, 327
with replication, 329
dynamic schemas, collections, 8
E
$each modifier, 38
using with $addToSet, 40
EBS (Elastic Block Store), 369, 373
elections, 181, 193
creating election arbiters, 182
how they work, 181
preventing, 213
$elemMatch operator, 63
grouping query criteria without specifying
every key, 65
embedded document type, 17
embedded documents, 19
changing using $set update modifier, 35
indexing, 96
querying on, 63
embedding versus references, 156
cardinality and, 157
enableSharding command, 234
enableSharding function, 246
encrypting data, 338
endianness, 371
ensureIndex command, 83, 107
creating index on stand-alone server, 315
expireAfterSecs option, 114
identifying indexes, 108
using with 2dsphere type to create geospatial
indexes, 120
ephemeral drives, 374
$eq operator, 133
errors
failures of database commands, 76
using unacknowledged writes, 51
exact phrase searches, 118
explain function, 82
for sharding test setup, 237
using on indexed and non-indexed queries,
98–102
using on indexed and nonindexed queries
output fields, 99
expressions
using in $project pipeline, 131
date expressions, 132
logical expressions, 133
mathematical expressions, 131
string expressions, 132
extents, 391
extreme operators for data set edges, 136
F
features in MongoDB, 4
field-order sensitivity, database commands, 77
fields, manipulating with $project operator, 130
file storage, 4
file system sockets, connection via, 338
file-based configuration, 336
398 | Index
files
storing with GridFS, 123–126
how GridFS works, 125
mongofiles utility, 124
filesystems, 371
filesystem snapshots, 357
filters, adding to run MapReduce on subset of
documents, 145
finalize function, 144
using with group command, 150
find method, 53–55
chaining limit function onto call to find, 68
chaining skip function onto call to find, 68
chaining sort function onto call to find, 69
limitations on, 55
querying a collection, 15
specifying which keys to return, 54
findAndModify command, 49–50
fields, 50
findOne method
querying a collection, 15
specifying which keys to return, 54
firehose strategy for shard keys, 267
firewalls, 337
$first operator, 136
floats, 16
flushRouterConfig command, 295
forcing reconfiguration of replica set, 212
--fork option, mongod, 334
fragmentation, 45
free disk space, tracking, 352
freeze function, 213
fs.chunks collection, 125
fs.files collection, 126
fsyncLock command, 358
fsyncUnlock command, 358
full-text indexes, 115–120
optimizing full text searches, 119
search syntax, 118
searching in other languages, 119
functions
grouping function, using as a key, 151
scope in JavaScript, 67
scope in MapReduce operations, 145
G
$geoIntersects operator, 121
GeoJSON objects, 120
$geometry operator, 120
geospatial indexing, 120–123
2D indexes, 122
compound geospatial indexes, 121
types of geospatial indexes, 120
getCollection function, 27
getLastError command
custom replication guarantees, 202
displaying number of documents updated in
multiupdate, 48
getting information about what was updated,
48
j option, 324
waiting for replication writes, 200–202
getProfilingLevel function, 304
global variables, setting up, using .mongorc.js
file, 25
GridFS
hashed shard keys for collections, 266
sharding by files_id, 293
storing files with, 123–126
how GridFS works, 125
mongofiles utility, 124
working with GridFS from MongoDB
drivers, 124
$group operator, 128, 135
arithmetic operators, 135, 135
array operators, 137
extreme operators, 136, 136
grouping behavior, 137
group command, 147–151
component keys, 148
using a finalizer, 150
using a function as a key, 151
$gt operator, 55, 133
use of indexes, ranges and, 93
$gte operator, 55, 133
H
hashed shard keys, 264–267
for GridFS collections, 266
hasNext method, 67
heartbeats, 191
replica set member states, 191
help command, 22
helper functions for sharding, 234
hidden option, replica set members, 184
hint function
forcing table scan, 103
forcing use of certain index, 88
Index | 399
using, 102
hotspots, multi-hotspot strategy, 268–270
HTTP server, 337
hugepages, disabling, 378
I
_id keys, 20
autogeneration of, 21
collection missing _id index, 354
default return by queries, modifying, 54
in document replacement updates, 33
no-_id collections, 114
storing as ObjectIds versus strings, 305
unique index on, 104
update modifiers and, 34
$ifNull operator, 134
immortal cursors, 75
implicit indexes, 91
$in operator, 56
order of documents returned, 95
$inc modifier, 34, 36, 57
using $ (position operator) with, 42
incremental backups with mongooplog, 363
incrementing and decrementing with $inc
modifier, 36
indexing, 4, 81–108
administrative complications with unique
indexes, 361
arrays, 96
buildIndexes setting on replica set members,
185
building indexes, 222
checking index build progress, 83
collection missing an _id index, 354
compound indexes
introduction to, 84–89
using, 89
creating and deleting indexes, 315–317
creating index on replica set, 315
creating index on sharded cluster, 316
creating index on stand-alone server, 315
OOM (out-of-memory) killer, 317
removing indexes, 316
creating index on key for sharding, 235
full-text indexes, 115–120
geospatial, 120–123
in initial sync process, 189
index administration, 107
changing indexes, 108
identifying indexes, 108
index cardinality, 98
limits on indexes per collection, 84
loading a specific index into RAM, 318
no-_id index for collections, 114
objects, 96
selection of indexes by query optimizer and,
102
size of indexes, 306
TTL (time-to-live) indexes, 114
types of indexes, 104–106
sparse indexes, 106
unique indexes, 104
use of indexes by $-operators, 91–95
inefficient operators, 91
OR queries, 94
using explain and hint functions, 98
when not to index, 103
initial syncing, 188
initiate function, 210
insert method, 14
inserts
adding a document to a collection, 29
batch inserts, 29
validation of, 30
installing MongoDB, 385–388
integers, 16
NumberInt or NumberLong classes, 17
intersection geospatial queries, 121
IO issues, virtualized network disk, 373
IO wait, 346
CPU with minimal, 350
IOPS (IO Operations Per Second), 373
isMaster command, 171
running on secondary to see if it has become
new primary, 172
J
JavaScript
code type in MongoDB, 18
disallowing server side execution of scripts,
338
documents in MongoDB as “JSON-like”, 16
equivalents to shell helpers, 23
executing as part of a query, 65
function scope, 67
MongoDB shell, 13
mongorc.js file for frequenly loaded scripts,
25
400 | Index
property names, 27
running scripts with mongo shell, 23
server-side scripting, security risks, 66
using cursor in forEach loop, 67
joins
no joining facilities in MongoDB, 153
relational databases versus MongoDB, 165
journalCommitInterval, 325
journaling, 323
commit batches and, 324
setting commit intervals, 325
preallocation of journal files before mongod
starts, 335
supporting snapshotting, 357
turning off, 325
JSON
documents in MongoDB, resemblance to, 16
GeoJSON format for points, lines, and poly‐
gons, 120
K
key/value pairs
finding all keys in a collection, using MapRe‐
duce, 141–143
in MongoDB documents, 7
specifying which to return with find method,
54
using a grouping function as a key, 151
using update modifiers with keys, 35
_id key for documents, 20
keys
choosing key directions in compound index
sorting, 89
mandated by GridFS, 126
shard key, 257
(see also sharding)
kill command, 337
kill -9, 324
killOp function, 301
L
lag calculating, 353
calculating for replication, 219
low-write system causing phantom lag, 354
tracking by primary, 224
languages, searching in other languages, 119
$last operator, 136
latency, 258
$limit operator, 128, 139
limiting number of query results, 68
lines, specifying in GeoJSON format, 120
Linux, 370
access time, not tracking, 380
installing MongoDB, 387
OOM (out-of-memory) killer, 383
listCommands command, 76
little-endian systems, 371
load method, 23
local database, 11
user privileges, 312
local.me collection, 188, 224
local.oplog.rs collection, 361
local.slaves collection, 224
local.startup_log collection, 335
local.system.replset collection, 210, 212
location-based shard keys, 263–264
lock percentage, 352
spinning disk versus SSD, 368
logging, 338
setting log level, 339
logical expressions, 133
--logpath option, mongod, 334
logRotate command, 339
$lt operator, 55, 133
use of indexes, ranges and, 93
$lte operator, 55, 133
M
Mac OS X, installing MongoDB, 387
maintenance mode, 213
majority keyword, passing to w parameter, 200
majority, replication set members, 178
manual sharding, 231, 273
many-to-many relationships, 157
map function, 141
mapped memory, 342
MapReduce, 140–146
categorizing web pages (example), 143
finding all keys in a collection (example),
141–143
MongoDB and, 143
finalize function, 144
getting more output, 146
keeping output collections, 144
running MapReduce on subset of docu‐
ments, 145
Index | 401
using a scope, 145
master-slave setup, 225–227
converting to a replica set, 226
mimicking behavior with replica sets, 226
replica sets versus, 225
setting up, 225
$match operator, 129
mathematical expressions, 131
$max operator, 136
max function, 63
maxConns option, mongos, 284, 381
$maxKey, 237
MaxKey constant, 237
md5 key, 126
memory
memory overcommitting, 372
NUMA (non-uniform memory architec‐
ture), 375–377
OOM (out-of-memory) killer, 383
virtualization and mystery memory, 372
memory usage, monitoring, 341–348
IO wait, 346
minimizing btree misses, 345
tracking background flush averages, 346
tracking memory usage, 342
tracking page faults, 343
memory-mapped storage engine, 391
metadata, files stored by GridFS, 126
$min operator, 136
min function, 63
$minKey, 237
MinKey constant, 237
MMS (MongoDB Monitoring Service), 341
statistics on network disk IO issues, 373
mongo (shell), 13–16, 21–28
--nodb option, 170
basic operations, 14
create, 14
read, 15
update, 15
connecting to MongoDB instance, 21
connecting to mongos, 233
creating a cursor, 67
creating mongorc.js file for frequently load‐
ed scripts, 25
customizing the prompt, 26
dealing with inconvenient collection names,
27
editing complex variables, 27
help or tips for using, 22
MongoDB client, 14
running, 13
running scripts with, 23
starting without connecting to mongod, 22
write concerns and, 51
MongoClient connections, 52
mongod, 11
--syncdelay option, 347
-f or --config flags, 336
adding single mongod as a shard, 245
config servers, starting up, 242
converting from master-slave to replica set,
226
master-slave setup, 225
receiving SIGINT or SIGTERM signal, 337
refreshing configuration, 295
repairs, 321, 326
shards, 232
shutdowns after crashes, 327
starting replica set members in standalone
mode, 209
starting with --replSet option, 174
startup options, 333
turning off JavaScript execution with --
noscripting option, 66
mongod.lock file, 326
MongoDB
advantages offered by, 3–5
getting and starting, 11
when not to use, 165
MongoDB Monitoring Service (MMS), 341
statistics on network disk IO issues, 373
MongoDB wiki, 164
mongodb.log file, 334
mongodump utility
backing up entire sharded cluster, 362
fsyncLock and, 359
repair, 326
using for replication set backups, 361
using for single-server backup, 359
mongofiles utility, 124
mongoimport tool, 30
mongooplog utility, 363
mongorestore utility, 360, 361
applying rolled back operations to current
primary, 196
moving collections and databases, 360
restoring entire sharded cluster, 362
402 | Index
mongos, 232
--configdb option, 289
adding shard from a replica set, 244
balancers, 254
connecting to cluster’s mongos, 232
connectivity, 382
explain function output for processing of
query, 238
maxConns option, 284, 381
refreshing configuration, 295
splitting chunks, 249–253
starting processes, 243
turning off balancer associated with, 273
turning off chunk splitting, 253
mongostat utility, 307
mongotop utility, 307
monitoring MongoDB, 341–355
memory usage, 341–348
calculating working set, 348–350
IO wait, 346
minimizing btree misses, 345
tracking background flush averages, 346
tracking memory usage, 342
tracking page faults, 343
replication, 353–355
tracking performance, 350–353
$month operator, 132
moveChunk command, 274, 281, 291
secondaryThrottle option, 294
movePrimary command, 288
moves, 43
relocation of enlarged documents, 73
speed difference between in place updates
and, 44
multi-hotspot strategy, 268–270
multi-value queries, 86
multikey indexes, 97
multiupdates, 47
mystery memory, 372
N
namespaces, 11, 390
naming
collections, 9
databases, 10
$natural operator, 103
natural order, 84
natural sorts, 112
$ne (not equal) operator, 55, 133
use of indexes, 92
using with $push modifier, 39
$near operator
2D index queries, 122
using in geospatial queries, 121
networking
configuring network for MongoDB, 382
considerations for replica sets, 176
handling network disk IO issues in virtual‐
ized disks, 373
tracking network connections in clusters,
283–285
next method, 67
$nin operator, 56
use of table scans instead of indexes, 92
non-networked disks, 374
non-uniform memory architecture (NUMA),
375–377
$nor operator, 57
normalization and denormalization, 153–160
cardinality, 157
examples of data representations, 154–157
social media linking people, friends, follow‐
ers, etc., 158–160
$not operator, 57, 134
use of indexes, 92
NOT queries, search syntax for full-text indexes,
118
null type, 17
querying on, 58
NUMA (non-uniform memory architecture),
375–377
number type, 17
NumberInt class, 17
NumberLong class, 17
O
object id type, 18
Object.bsonsize function, 305
calling on collection elements, 306
ObjectIds, 20
storing _ids as, 305
objects, indexing, 96
one-to-many relationships, 157
one-to-one relationships, 157
OOM (out-of-memory) killer, 317, 383
operating systems, 370
Index | 403
operators
aggregation framework, 127
position operator ($), 41, 42
query operators, 91–95
oplogs, 187
creating in replica set member restore, 361
getting summary of, 219
mongodump’s --oplog option, 360
mongooplog utility, 363
mongorestore’s --oplogReplay option, 360
resizing, 220
tracking length for each member, 355
optimizations
for data manipulation, 160
otpimizing for document growth, 160
removing old data, 162
full-text search, 119
$or operator, 56, 57, 134
OR queries, 56
search syntax for full-text indexes, 118
use of indexes by $or operator, 94
P
package manager, installing MongoDB from,
388
padding factor, 43
page faults, 342
IO wait and, 346
tracking, 343
pagesInMemory, working set, 350
paginating query results without skip, 70
partitioning, 231
(see also sharding)
of full-text indexes, 119
passive members, 183
password hash, 314
pdfile, 329
performance, 5
indexes and, 84
monitoring, 350–353
tracking free space, 352
remove speed, 31
speed of update modifiers, 42
periodic tasks, turning off, 384
Perl Compatible Regular Expression (PCRE) li‐
brary, 59
phantom operations, preventing, 302
PHP, tailable cursor, using, 113
plain queries, 71
points
indexing in 2D indexes, 122
specifying in GeoJSON format, 120
polygons
specifying as array of points in 2D queries,
123
specifying in GeoJSON format, 120
$pop modifier, 41
--port option, mongod, 334
position operator ($), 41
returning a matching array element, 62
using to modify individual key/value pairs,
36
POSIX (Linux, Mac OS X, and Solaris) install,
387
preallocating data files, 322
primaries, 170
automatic failover, 172
demoting to secondaries, 213
election of, 178
how elections work, 181
manually promoting new primary, 227
only one primary in MongoDB, 180
shutdown command, 336
tracking of lag, 224
print method, 23
printReplicationInfo function, 219
printSlaveReplicationInfo function, 219
priority, replication set members, 183, 223, 227
problematic operations, finding, 301
profiling, turning on, 339
$project operator, 127, 130–135
pipeline expressions, 131
date expressions, 132
logical expressions, 133
mathematical expressions, 131
projection example, 134
string expressions, 132
projection, 130
example of, 134
prompt, customizing for mongo shell, 26
publication-subscription systems, 158
$push modifier, 37
using $each with, 38
using $slice with, 39
using with $ne, 39
$push operator, 137
PyMongo, working with GridFS from, 124
404 | Index
Python
PyMongo driver for MongoDB, 124
scope, 66
Q
query operators
use of indexes, 91–95
inefficient operators, 91
OR queries, 94
ranges, 93
query optimizer, 102
querying, 53–77
$where queries, 65
find method, 53–55
query criteria, 55–58
conditional semantics, 57
conditionals, 55
OR queries, 56
OR queries/$not, 57
returning results using cursors, 67–75
advanced query options, 71
avoiding large skips, 70–71
getting consistent results, 72–74
immortal cursors, 75
limits, skips, and sorts, 68
server-side scripting, 66
type-specific queries, 58–65
null type, 58
querying arrays, 59–63
querying on embedded documents, 63
regular expressions, 58
using database commands, 75–77
queuing, 351
R
RAID configurations, 369
RAM
moving data into, 317–320
storing data in, 365
randomly distributed shard keys, 261–263
ranges
chunk, 247
range queries for dates, 55
range query interaction with array queries,
62
use of indexes by $-operators for ranges, 93
read operations, mongo shell, 15
readahead settings, 377
reconfig function, 211
force option, 212
reconfiguration of replica set
changing member’s settings, 211
forcing, 212
RECOVERING state, 213
reduce function, 141
references, embedding versus, 156
regular expressions, 17
querying with, 58
relational databases, features not available in
MongoDB, 4
remove function, 16, 211
removes, 31
removing old data, optimization for, 162
speed of, 31
removeShard command, 287–288
removeShardTag function, 273
renameCollection command, 163, 321
repairs, 321
repairing data files, 325
replaying application usage, 319
replica sets, 169–185
adding a shard from, 244
backing up, 361
changing configuration, 176–178
changing shard from stand-alone server to,
286
config servers not members of, 242
configuration, 210–212
configuring, 174
networking considerations, 176
rs helper functions, 176
connecting to from your application, 199–
207
client-to-replica-set connection behavior,
199
custom replication guarantees, 202–205
waiting for replication on writes, 200–
202
converting from master-slave setup to, 226
creating an index on, 315
defined, 169
designing, 178–181
how elections work, 181
elections, 193
heartbeats, 191
member states, 191
manipulating member state, 213–214
Index | 405
master-slave versus, 225
member configuration options, 181–185
building indexes, 185
creating election arbiters, 182
hidden, 184
priority, 183
slave delay, 185
mimicking master-slave behavior, 226
rollbacks, 193–197
failure of rollbacks, 197
starting members in standalone mode, 209
syncing, 187–191
handling staleness, 190
initial sync, 188
test setup, 170–173
replication
custom gurarantees, 202–205
durability with, 329
introduction to, 169
monitoring, 214–225, 353–355
building indexes, 222
calculating lag, 219
disabling chaining, 219
getting status of replica set members, 214
lower-cost replication, 223
replication loops, 218
resizing the oplog, 220
restoring from delayed secondary, 221
tracking of lag by primary, 224
visualizing replication graph, 216–218
waiting for replication writes, 200–202
problems with replication set, 201
replSetGetStatus command, 214
replSetMaintenance command, 192
replSetMaintenanceMode command, 214
replSetReconfig command, 210
replSetSyncFrom command, 218
reserved database names, 11
resident memory, 342
restores
from data directory copies, 359
from filesystem snapshots, 358
from mongodump backup, using mongores‐
tore, 360, 361
of a single shard in a cluster, 363
of an entire sharded cluster, 362
restoring from delayed secondary, 221
right-balanced indexes, 89
rollbacks, 193–197, 202
defined, 194
failure of, 197
preventing, 196
ROLLBACK state, 192
rs helper functions, 176
rs.add function, 176, 178, 211
rs.addArb function, 182
rs.config function, 177, 184, 211
rs.freeze function, 213
rs.initiate function, 210
rs.isMaster function, 184
rs.reconfig function, 177, 211
force option, 212
rs.remove function, 178, 211
rs.status function, 184, 214
syncingTo field, 216, 218
useful fields, 215
rs.stepDown function, 213
rs.syncFrom function, 218
runCommand function, 76
S
save shell helper, 47
scalability of MongoDB, 3
scatter-gather queries, 240
schemas
dynamic schemas for collections, 8
migrating, 164
no predefined schemas for MongoDB, 3
scope, 66, 67
using in MapReduce operations, 145
secondaries, 171
attempts to write to, 172
becoming primary in automatic failover, 172
delayed, using slaveDelay setting, 185
election to primary, 181
lower-cost server for, 223
restoring from delayed secondary, 221
syncing from another secondary, 219
security, 337
authentication (see authentication)
config file options, 337
data encryption, 338
risks with server-side scripting, 66
SSL connections, 338
seeds, 199
server-side scripting, 66, 338
406 | Index
servers
administration in sharded cluster, 285–289
adding servers, 285
changing config servers, 288
changing from stand-alone server to rep‐
lica set, 286
changing servers in a shard, 285
removing servers, 286–288
single-server backups, 357–361
stand-alone, creating index on, 315
serverStatus command
recordStats field, 344
workingSet option, 349
$set modifier, 34
using $ (position operator) with, 42
$setOnInsert modifier, 46, 161
setParameter command
journalCommitInterval, 325
logLevel option, 339
textSearchEnabled=true option, 115
setProfilingLevel function, 303, 339
sets, using arrays as, 39
settings collection, 283
sh global variable, 234
sh.addShard function, 244
sh.addShardTag function, 272
sh.help function, 234
sh.moveChunk function, 291
sh.removeShardTag function, 273
sh.status function, 234, 235
getting summary of current state, 275
shard key, 235, 257
(see also sharding)
shardCollection command, 246
sharding
admin database and authentication, 314
administration, 275–295
balancing data, 289–295
getting summary with sh.status, 275
refreshing configurations, 295
seeing configuration information, 277–
283
server administration, 285–289
tracking network connections, 283–285
backing up sharded cluster, 362
balancer, 253
choosing a shard key, 257–274
ascending shard keys, 258–261
cardinality of shard keys, 271
controlling data distribution, 271
firehose strategy, 267
hashed shard key, 264–267
limitations of shard keys, 271
location-based shard keys, 263–264
multi-hotspot strategy, 268–270
randomly distributed shard keys, 261–
263
taking stock of usage, 257
components of a cluster, 232
connectivity, 382
creating index on sharded cluster, 316
deciding when to shard, 241
defined, 231
enabling on a collection, 234
helper functions, 234
starting the servers, 242
adding capacity, 245
adding shard from a replica set, 244
config servers, 242
mongos processes, 243
sharding data, 246
test setup, 232–240
tracking of cluster data by MongoDB, 246
chunk ranges, 247
splitting chunks, 249–253
ShardingTest class, 232
shards, 232, 285
(see also servers)
adding from a replica set, 244
config.shards collection, 277
connectivity, 382
shell, 7
(see also mongo)
multiple shells, 320
shell helpers, 23, 76
shutdown command, 336
force option, 337
shutdowns, hard, 327
SIGINT or SIGTERM signal, 337
$size operator, 60
$skip operator, 139
skipping query results, 68
avoiding large skips, 70
finding a random document, 70
paginating results without skip, 70
slaveDelay setting, 185
$slice modifier, 39
$slice operator, 61
Index | 407
slowms field, db.setProfilingLevel function, 304
snapshots, filesystem, 357
snapshotting a query, 74
Solaris, 370
installing MongoDB, 387
solid-state drives (see SSDs)
$sort modifier, 39
$sort operator, 128, 139
sorting
find method results, 69
in compound indexes, 84
results of query using an index, 86
sparse indexes, 106
spinning disks, 365
performance, SSDs versus, 366–369
split point, 247
split storm, 250
splitAt command, 292
SSDs (solid-state drives)
choosing as storage medium, 365
advantages over spinning disks, 366–369
SSL connections, 338
staleness, handling for secondaries, 190
standalone mode, starting members in, 209
starting and shutting down MongoDB, 333–337
starting from command line, 333–336
file-based configuration, 336
stopping MongoDB, 336
startParallelShell command, 320
stats function
for collections, 305
for databases, 306
status function, 214
useful fields, 215
stemming words, languages and, 119
stepDown function, 213
storage medium, choosing, 365–369
$strcasecmp operator, 133
strings
creating full-text index on all string fields
with $**, 117
string expressions, 132
string type, 17
subcollections, 10
$substr operator, 132
$subtract operator, 131
$sum operator, 135
swap space, 371
syncdelay option, 347
syncingTo field, 216, 218
system housekeeping, 383
system profiler, using, 303–304
system settings, 374–382
disabling hugepages, 378
disk scheduling algorithm, 379
modifying limits, 380
not tracking access time, 380
sane readahead, 377
turning off NUMA, 374–377
system.indexes collection, 107
system.profile collection, 303–304
system.users collection, 314
T
table scans, 81
forcing, 103
indexing versus, situations for use, 103
queries using $not and $nin operators, 92
tags (shard), config.tags collection, 282
tailable cursors, 113
TCP/IP, lightweight wire protocol, 390
test database on MongoDB server, 14
text command, 116
textSearchEnabled option, setParameter com‐
mand, 115
throughput, 258
$toLower operator, 133
$toUpper operator, 133
TTL (time-to-live) collections, using ro remove
old data, 162
TTL (time-to-live) indexes, 114
two-phase-commit-type operation, config
servers, 242
U
unacknowledged writes, 51
underloaded systems, 354
unique indexes, 104
administrative complications with, 361
compound, 104
dropping duplicates, 105
sparse indexes, 106
$unset modifier, 35
$unset operator, removing the garbage field, 161
$unwind operator, 137–139
update modifiers, 34–45
$inc modifier, 36
408 | Index
$set modifier, 34
$setOnInsert, 46
array modifiers, 37
$each, 38
$pop, 41
$pull, 41
$push, 37
$slice, 39
$sort, 39
inability to apply multiple modifiers to sin‐
gle key, 57
positional array modifications with $, 41
speed of, 42–45
using arrays as sets
$addToSet modifier, 39
$ne modifier, 39
update operations, mongo shell, 15
updates, 32–51
fully replacing a document, 32
multiple documents, 47
relocation of enlarged documents and con‐
sistency of query results, 73
returning updated documents, 48
speed difference between in place updates
and moves, 44
upserts, 45
using modifiers, 34–45
upserts, 45
usePowerOf2Sizes option, 45, 159
automatically enabled with full-text indexes,
119
users
of databases, stored in system.users, 314
setting up user accounts and authentication,
312
V
validate command, 327
variables
editing commplex variables using the shell,
27
injecting into the shell, using scripts, 24
verbose output for MapReduce, 146
versions of MongoDB, 385
virtual memory, 342
virtualization, 372–374
mystery memory, 372
network disk IO issues, 373
turning off memory overcommitting, 372
using non-networked disks, 374
votes for replica set members, 227
W
w parameter, getLastError command, 200
other options for w, 202
web pages, categorizing (MapReduce example),
143
$week operator, 132
weight, setting for each field for full text index,
117
Wheaton, Wil, 159
$where clauses in queries, 65
$where operator, inability to use indexes, 91
Windows, 370
installing MongoDB, 386
as as service, 387
not tracking access time, 380
wire protocol, 390
$within operator
2D index queries, 122
using for geospatial queries, 121
working set, calculating, 348–350
wrapped queries, 71
write concern, setting, 51
writebacklistener commands, 302
wtimeout option, getLastError command, 201
Y
$year operator, 132
Z
zone_reclaim_mode, 376
Index | 409
About the Author
Kristina Chodorow is a software engineer who worked on the MongoDB core for five
years. She led MongoDB’s replica set development as well as writing the PHP and Perl
drivers. She has given talks on MongoDB at meetups and conferences around the world
and maintains a blog on technical topics at http://www.kchodorow.com. She currently
works at Google.
Colophon
The animal on the cover of MongoDB: The Definitive Guide, Second Edition is a mon‐
goose lemur, a member of a highly diverse group of primates endemic to Madagascar.
Ancestral lemurs are believed to have inadvertently traveled to Madagascar from Africa
(a trip of at least 350 miles) by raft some 65 million years ago. Freed from competition
with other African species (such as monkeys and squirrels), lemurs adapted to fill a wide
variety of ecological niches, branching into the almost 100 species known today. These
animals’ otherworldly calls, nocturnal activity, and glowing eyes earned them their
name, which comes from the lemures (specters) of Roman myth. Malagasy culture also
associates lemurs with the supernatural, variously considering them the souls of ances‐
tors, the source of taboo, or spirits bent on revenge. Some villages identify a particular
species of lemur as the ancestor of their group.
Mongoose lemurs (Eulemur mongoz) are medium-sized lemurs, about 12 to 18 inches
long and 3 to 4 pounds. The bushy tail adds an additional 16 to 25 inches. Females and
young lemurs have white beards, while males have red beards and cheeks. Mongoose
lemurs eat fruit and flowers and they act as pollinators for some plants; they are par‐
ticularly fond of the nectar of the kapok tree. They may also eat leaves and insects.
Mongoose lemurs inhabit the dry forests of northwestern Madagascar. One of the two
species of lemur found outside of Madagascar, they also live in the Comoros Islands
(where they are believed to have been introduced by humans). They have the unusual
quality of being cathemeral (alternately wakeful during the day and at night), changing
their activity patterns to suit the wet and dry seasons. Mongoose lemurs are threatened
by habitat loss and they are classified as a vulnerable species.
The cover image is from Lydekker’s Royal Natural History. The cover font is Adobe ITC
Garamond. The text font is Adobe Minion Pro; the heading font is Adobe Myriad Con‐
densed; and the code font is Dalton Maag’s Ubuntu Mono.

Navigation menu