Inside the sql server query optimizer

High Performance SQL Server
978-1-906434-57-1
Inside the SQL Server
Query Optimizer
Benjamin Nevarez

Inside the SQL Server
Query Optimizer
By Benjamin Nevarez
First published by Simple Talk Publishing 2010

Copyright Benjamin Nevarez 2011
ISBN: 978-1-906434-57-1
The right of Benjamin Nevarez to be identified as the author of this work has been asserted by him in
accordance with the Copyright, Designs and Patents Act 1988.
All rights reserved. No part of this publication may be reproduced, stored or introduced into a retrieval
system, or transmitted, in any form, or by any means (electronic, mechanical, photocopying, recording or
otherwise) without the prior written consent of the publisher. Any person who does any unauthorized act in
relation to this publication may be liable to criminal prosecution and civil claims for damages.
This book is sold subject to the condition that it shall not, by way of trade or otherwise, be lent, re-sold,
hired out, or otherwise circulated without the publisher's prior consent in any form other than that in which
it is published and without a similar condition including this condition being imposed on the subsequent
publisher.
Editor: Chris Massey
Technical Review: Grant Fritchey
Cover Image: Andy Martin http://www.this-is-sunderland.co.uk
Typeset & Designed: Matthew Tye & Gower Associates

Table of Contents
Chapter 1: Introduction to Query Optimization........................... 14
How the Query Optimizer Works........................................................................................ 15
Generating candidate execution plans..........................................................................18
Assessing the cost of each plan......................................................................................20
Query execution and plan caching................................................................................20
Hinting................................................................................................................................21
Ongoing Query Optimizer Challenges................................................................................22
A historical perspective....................................................................................................23
Execution Plans.......................................................................................................................24
Join Orders...............................................................................................................................34
Summary...................................................................................................................................41
Chapter 2: The Execution Engine..................................................42
Data Access Operators............................................................................................................43
Scanning............................................................................................................................ 44
Seeking............................................................................................................................... 48
Bookmark lookup.............................................................................................................49
Aggregations............................................................................................................................. 55
Sorting and hashing..........................................................................................................55
Stream Aggregate..............................................................................................................56
Hash Aggregate.................................................................................................................60
Joins.......................................................................................................................................... 66
Nested Loops Join............................................................................................................ 67
Merge Join..........................................................................................................................71
Hash Join............................................................................................................................ 74
Parallelism................................................................................................................................75
Summary...................................................................................................................................78
Chapter 3: Statistics and Cost Estimation...................................80
Statistics....................................................................................................................................81
Creating and updating statistics.....................................................................................81

Inspecting statistics objects............................................................................................ 84
Density............................................................................................................................... 87
Histograms...............................................................................................................................92
Statistics Maintenance.......................................................................................................... 99
Statistics on Computed Columns.......................................................................................104
Filtered Statistics...................................................................................................................107
Cardinality Estimation Errors..............................................................................................111
UPDATE STATISTICS with ROWCOUNT, PAGECOUNT...........................................113
Cost Estimation..................................................................................................................... 118
Summary..................................................................................................................................121
Chapter 4: Index Selection.......................................................... 122
Introduction........................................................................................................................... 122
The Mechanics of Index Selection.....................................................................................124
The Database Engine Tuning Advisor...............................................................................130
The Missing Indexes Feature............................................................................................... 136
Unused Indexes.....................................................................................................................140
Summary.................................................................................................................................144
Chapter 5: The Optimization Process......................................... 145
Overview................................................................................................................................. 145
Peeking at the Query Optimizer.........................................................................................146
Parsing and Binding..............................................................................................................160
Transformation Rules...........................................................................................................162
The Memo..............................................................................................................................170
Optimization Phases.............................................................................................................176
Simplification...................................................................................................................176
Trivial plan........................................................................................................................ 181
Full optimization.............................................................................................................183
Summary.................................................................................................................................192
Chapter 6: Additional Topics....................................................... 193
Updates...................................................................................................................................194
Per-row and per-index plans.........................................................................................197

Halloween protection....................................................................................................200
Data Warehouses.................................................................................................................. 202
Parameter Sniffing................................................................................................................ 208
Optimize for a typical parameter.................................................................................. 211
Optimize on every execution........................................................................................212
Local Variables and OPTIMIZE FOR UNKNOWN.................................................. 213
Auto-parameterization.........................................................................................................214
Forced parameterization......................................................................................................216
Summary.................................................................................................................................218
Chapter 7: Hints............................................................................219
Before You Reach for a Hint…............................................................................................ 220
Types of Hints........................................................................................................................223
Joins.................................................................................................................................. 224
Aggregations.................................................................................................................... 229
FORCE ORDER............................................................................................................... 231
INDEX and FORCESEEK hints....................................................................................237
FAST N.............................................................................................................................240
NOEXPAND, EXPAND VIEWS hints......................................................................... 242
Plan Guides............................................................................................................................ 244
USE PLAN............................................................................................................................. 248
Summary................................................................................................................................. 251

vii
About the Author
Benjamin Nevarez is a database professional based in Los Angeles, California. He has
more than 15 years of experience with relational databases, and has been working with
SQL Server since version 6.5. Benjamin has been the technical editor of the two latest
books by Kalen Delaney, including "SQL Server 2008 Internals." He holds a Master's
Degree in Computer Science and has been a speaker at several technology conferences,
including the PASS Community Summit.
Benjamin's blog is at http://www.benjaminnevarez.com, and he can be reached at
admin@benjaminnevarez.com.
About the Technical Reviewer
Grant Fritchey, Microsoft SQL Server MVP 2009–2010, works for an industry-leading
engineering and insurance company as a principal DBA. Grant has performed the
development of large-scale applications in languages such as Visual Basic, C#, and Java,
and has worked with SQL Server since version 6.0. He spends a lot of time involved in the
SQL Server community, including speaking and blogging, and he is an active participant
in the SQLServerCentral.Com forums. He is the author of several books including
"SQL Server Execution Plans" (Simple Talk Publishing, 2008) and "SQL Server Query
Performance Tuning Distilled" (Apress, 2008).

viii
Acknowledgements
Writing this book was a lot of fun, but also a lot of work; actually, a lot more work
than I originally expected. Fortunately, I got help from several people. First of all, I
would like to thank Chris Massey. Chris helped me as the technical editor of the book,
and guided me through most of the writing process, doing absolutely outstanding work.
Very special thanks also go out to Grant Fritchey who helped us as the technical reviewer,
and went very much beyond just reviewing, as his comments provided invaluable
feedback to improve the quality of this book. Thanks also go to Tony Davis for offering
me this opportunity in the first place, and helping to get the project started.
Outside the Red Gate team, my deepest gratitude goes to Cesar Galindo-Legaria,
Manager of the SQL Server Query Optimization team at Microsoft, for answering my
endless list of questions about the Query Optimizer, and educating me through some
of the information he has published, including numerous research papers and an
excellent chapter of a SQL Server book. I had the opportunity to meet Tony, Cesar, and
Grant at the PASS Summit back in 2009, when I first started this journey, and I hope to
finally meet Chris at the same conference in October 2011.
Although Chris, Cesar, Grant, and Tony have directly helped me to shape the content
of this book, there's also a host of other people who have indirectly influenced the book
through their work, which helped me to learn about, and better understand, the SQL
Server query processor. With that in mind, I would like to thank the authors who have
discussed the query processor in some of the available SQL Server books, Microsoft
white papers and blogs, including Kalen Delaney, Ken Henderson, Lubor Kollar, Craig
Freedman, Conor Cunningham and Eric Hanson.
Research papers have provided me with an unlimited source of information, and helped
me to understand the Query Optimizer at a significantly deeper level than ever before.
So, thanks to all that amazing work by the many people in the research community
including, in addition to Cesar, Goetz Graefe, Surajit Chaudhuri, Yannis Ioannidis,
Vivek Narasayya, Pat Selinger, Florian Waas, and many, many more.

ix
Finally, on the personal side, I would like to thank my parents, Guadalupe and Humberto,
and my family: my wife Rocio, and my three sons, Diego, Benjamin and David. Thanks,
all, for your unconditional support and patience.

x
Preface
The Query Optimizer has always been one of my favorite SQL Server topics, which is why
I started blogging about it, and submitting related presentations to PASS. And so it would
have continued, except that, after several blog posts discussing the Query Optimizer, Red
Gate invited me to write a book about it. This is that book.
I started learning about the Query Optimizer by reading the very few SQL Server books
which discussed the topic, most of which covered it only very briefly. Yet I pressed on,
and later, while trying to learn more about the topic, I found an extremely rich source of
information in the form of the many available research papers. It was hard to fully grasp
them at the beginning, as academic papers can be difficult to read and understand, but I
soon got used to them, and was all the more knowledgeable for it.
Having said that, I feel that I'm in a bit of a minority, and that many people still see
the Query Optimizer just as a black box where a query is submitted and an amazing
execution plan is returned. It is also seen as a very complex component, and rightly so.
It definitely is a very complex component, perhaps the most complex in database
management software, but there is still a lot of great information about the Query
Optimizer that SQL Server professionals can benefit from.
The Query Optimizer is the SQL Server component that tries to give you an optimal
execution plan for your queries and, just as importantly, tries to find that execution plan
as quickly as possible. A better understanding of what the Query Optimizer does behind
the scenes can help you to improve the performance of your databases and applications,
and this book explains the core concepts behind how the SQL Server Query Optimizer
works. With this knowledge, you'll be able to write better queries, provide the Query
Optimizer with the information it needs to produce efficient execution plans, and
troubleshoot the cases when the Query Optimizer is not giving you a good plan.

11
With that in mind, and in case it's not obvious, the content of this book is intended for
SQL Server professionals: database developers and administrators, data architects and,
basically, anybody who submits more than just trivial queries to SQL Server.
Here's a quick overview of what the book covers:
• The first chapter, Introduction to Query Optimization, starts with an overview on
how the SQL Server Query Optimizer works, and introduces the concepts that will
be covered in more detail in the rest of the book. A look into some of the challenges
query optimizers still face today is covered next, along with a section on how to read
and understand execution plans. The chapter closes with a discussion of join ordering,
traditionally one of the most complex problems in query optimization.
• The second chapter talks about the Execution Engine, describing it as a collection of
physical operators that perform the functions of the query processor. It emphasizes
how these operations, implemented by the Execution Engine, define the choices
available to the Query Optimizer when building execution plans. This chapter includes
sections on data access operations, the concepts of sorting and hashing, aggregations,
and joins, to conclude with a brief introduction to parallelism.
• Chapter 3, Statistics and Cost Estimation, shows how the quality of the execution
plans generated by the Query Optimizer is directly related to the accuracy of its
cardinality and cost estimations. The chapter describes Statistics objects in detail,
and includes some sections on how statistics are created and maintained, as well as
how they are used by the Query Optimizer. We'll also take a look at how to detect
cardinality estimation errors, which may cause the Query Optimizer to choose
inefficient plans, together with some recommendations on how to avoid and fix
these problems. Just to round off the subject, the chapter ends with an introduction
to cost estimation.

12
• Chapter 4, Index Selection, shows how SQL Server can speed up your queries and
dramatically improve the performance of your applications, just by using the right
indexes. The chapter shows how SQL Server selects indexes, how you can provide
better indexes, and how you can verify your execution plans to make sure these
indexes are correctly used. We'll talk about the Database Engine Tuning Advisor
and the Missing Indexes feature, which will show how the Query Optimizer itself can
provide you with index tuning recommendations.
• Chapter 5, The Optimization Process, goes right into the internals of the Query
Optimizer and introduces the steps that it performs without you ever knowing. This
covers everything, from the moment a query is submitted to SQL Server, until an
execution plan is generated and ready to be executed, including steps like parsing,
binding, simplification, trivial plan, and full optimization. Important components
which are part of the Query Optimizer architecture, such as transformation rules and
the memo structure, are also introduced.
• Chapter 6, Additional Topics, includes a variety of subjects, starting with the basics
of update operations, and how they also need to be optimized just like any other
query, so that they can be performed as quickly as possible. We'll have an introduction
to Data Warehousing and how SQL Server optimizes star queries, before launching
into a detailed explanation of parameter sniffing, along with some recommendations
on how to avoid some problems presented by this behavior. Continuing with the
topic of parameters, the chapter concludes by discussing auto-parameterization and
forced parameterization.
• Chapter 7 describes hints, and warns that, although hints are a powerful tool which
allow you to take explicit control over the execution plan of a query, they need to be
used with caution, and only as a last resort when no other option is available. The
chapter covers the most used hints, and ends with a couple of sections on plan guides
and the USE PLAN query hint.

13
Before we get started, please bear in mind that this book contains many undocumented
SQL Server statements. These statements are provided only as a way to explore and
understand the Query Optimizer and, as such, should not be used on a production
environment; use them wisely. I hope you enjoy learning about this topic as much as I do.
Benjamin Nevarez

14
Chapter 1: Introduction to Query
Optimization
The SQL Server Query Optimizer is a cost-based optimizer. It analyzes a number of
candidate execution plans for a given query, estimates the cost of each of these plans,
and selects the plan with the lowest cost of the choices considered. Indeed, given that the
Query Optimizer cannot consider every possible plan for every query, it actually has to
find a balance between the optimization time and the quality of the selected plan.
Therefore, it is the SQL Server component that has the biggest impact on the perform-
ance of your databases. After all, selecting the right (or wrong) execution plan could mean
the difference between a query execution time of milliseconds, and one of minutes, or
even hours. Naturally, a better understanding of how the Query Optimizer works can
help both database administrators and developers to write better queries and to provide
the Query Optimizer with the information it needs to produce efficient execution plans.
This book will demonstrate how you can use your newfound knowledge of the Query
Optimizer's inner workings and, in addition, it will give you the knowledge and tools to
troubleshoot the cases when the Query Optimizer is not giving you a good plan.
This first chapter starts with an overview on how the SQL Server Query Optimizer
works, and introduces the concepts that will be covered in more detail in the rest of the
book. We'll also cover some of the background and challenges of query optimization
and, since this book will make extensive use of execution plans, a section on how to
read and understand them is included as well. The chapter closes with a discussion of
join ordering, one of the most complex problems in query optimization, and shows how
joining tables in an efficient order improves the performance of a query but, at the same
time, can exponentially increase the number of execution plans that should be analyzed
by the Query Optimizer.

15
Chapter 1: Introduction to Query Optimization
Note
This book contains a large number of example SQL queries, all of which are based on the Adventure-
Works database, although Chapter 6 additionally uses the AdventureWorksDW database. All code has
been tested on both SQL Server 2008 and SQL Server 2008 R2. Note that these sample databases are not
included in your SQL Server installation by default, but can be downloaded from the CodePlex website.
You need to download the family of sample databases for your version, either SQL Server 2008 or SQL
Server 2008 R2. During installation, you may choose to install all the databases or, at least, Adventure-
Works and AdventureWorksDW.
How the Query Optimizer Works
At the core of the SQL Server Database Engine are two major components: the storage
engine and the query processor, also called the relational engine. The storage engine is
responsible for reading data between the disk and memory in a manner that optimizes
concurrency while maintaining data integrity. The query processor, as the name suggests,
accepts all queries submitted to SQL Server, devises a plan for their optimal execution,
and then executes the plan and delivers the required results.
Queries are submitted to SQL Server using the SQL language (or T-SQL, the Microsoft
SQL Server extension to SQL). Since SQL is a high-level declarative language, it only
defines what data to get from the database, not the steps required to retrieve that data, or
any of the algorithms for processing the request. Thus, for each query it receives, the first
job of the query processor is to devise a plan, as quickly as possible, which describes the
best possible way (or, at the very least, an efficient way) to execute said query. Its second
job is to execute the query according to that plan.
Each of these tasks is delegated to a separate component within the query processor;
the Query Optimizer devises the plan and then passes it along to the execution engine,
which will actually execute the plan and get the results from the database.

16
In order to arrive at what it believes to be the best plan for executing a query, the query
processor performs a number of different steps; the entire query processing process is
shown in Figure 1-1.
Figure 1-1: The query processing process.
We'll look at this whole process in much more detail later in the book, but I'll just run
through the steps briefly now.
• Parsing and binding – the query is parsed and bound. Assuming the query is valid,
the output of this phase is a logical tree, with each node in the tree representing a
logical operation that the query must perform, such as reading a particular table, or
performing an inner join.

17
• Query optimization – the logical tree is then used to run the query optimization
process, which roughly consists of the following two steps:
• generation of possible execution plans – using the logical tree, the Query
Optimizer devises a number of possible ways to execute the query, i.e. a number
of possible execution plans; an execution plan is, in essence, a set of physical
operations (an Index Seek, a Nested Loops Join, and so on), that can be performed
to produce the required result, as described by the logical tree
• cost-assessment of each plan – while the Query Optimizer does not generate every
possible execution plan, it assesses the resource and time cost of each plan it does
generate; the plan that the Query Optimizer deems to have the lowest cost of those
it has assessed is selected, and passed along to the execution engine.
• Query execution, plan caching – the query is executed by the execution engine,
according to the selected plan; the plan may be stored in memory, in the plan cache.
Parsing and binding are the first operations performed when a query is submitted to a
SQL Server instance. Parsing makes sure that the T-SQL query has a valid syntax, and
translates the SQL query into an initial tree representation: specifically, a tree of logical
operators representing the high-level steps required to execute the query in question.
Initially, these logical operators will be closely related to the original syntax of the query,
and will include such logical operations as "get data from the Customer table," "get data
from the Contact table," "perform an inner join," and so on. Different tree representa-
tions of the query will be used throughout the optimization process, and this logical tree
will receive different names until it is finally used to initialize the Memo structure, as will
be discussed later.
Binding is mostly concerned with name resolution. During the binding operation,
SQL Server makes sure that all the object names do exist, and associates every table and
column name on the parse tree with their corresponding object in the system catalog.
The output of this second process is called an algebrized tree, which is then sent to the
Query Optimizer.

18
The next step is the optimization process, which is basically the generation of candidate
execution plans and the selection of the best of these plans according to their cost. As
has already been mentioned, the SQL Server Query Optimizer uses a cost-estimation
model to estimate the cost of each of the candidate plans.
In essence, query optimization is the process of mapping the logical query operations
expressed in the original tree representation to physical operations, which can be carried
out by the execution engine. So, it's actually the functionality of the execution engine that
is being implemented in the execution plans being created by the Query Optimizer, that
is, the execution engine implements a certain number of different algorithms, and it
is from these algorithms that the Query Optimizer must choose, when formulating its
execution plans. It does this by translating the original logical operations into the physical
operations that the execution engine is capable of performing, and execution plans show
both the logical and physical operations. Some logical operations, such as a Sort, translate
to the same physical operation, whereas other logical operations map to several possible
physical operations. For example, a logical join can be mapped to a Nested Loops Join,
Merge Join, or Hash Join physical operator.
Thus, the end product of the query optimization process is an execution plan: a
tree consisting of a number of physical operators, which contain the algorithms
to be performed by the execution engine in order to obtain the desired results from
the database.
Generating candidate execution plans
As stated, the basic purpose of the Query Optimizer is to find an efficient execution plan
for your query. Even for relatively simple queries, there may be a large number of different
ways to access the data to produce the same end result. As such, the Query Optimizer
has to select the best possible plan from what may be a very large number of candidate
execution plans, and it's important that it makes a wise choice, as the time taken to return
the results to the user can vary wildly, depending on which plan is selected.

19
The job of the Query Optimizer is to create and assess as many candidate execution plans
as possible, within certain criteria, in order to arrive at the best possible plan. We define
the search space for a given query as the set of all the possible execution plans for that
query, and any possible plan in this search space returns the same results. Theoretically,
in order to find the optimum execution plan for a query, a cost-based query optimizer
should generate all possible execution plans that exist in that search space, and correctly
estimate the cost of each plan. However, some complex queries may have thousands, or
even millions, of possible execution plans and, while the SQL Server Query Optimizer
can typically consider a large number of candidate execution plans, it cannot perform
an exhaustive search of all the possible plans for every query. If it did, the time taken to
assess all of the plans would be unacceptably long, and could start to have a major impact
on the overall query execution time.
The Query Optimizer must strike a balance between optimization time and plan
quality. For example, if the Query Optimizer spends one second finding a good enough
plan that executes in one minute, then it doesn't make sense to try to find the perfect
or most optimal plan, if this is going to take five minutes of optimization time, plus the
execution time. So SQL Server does not do an exhaustive search, but instead tries to find
a suitably efficient plan as quickly as possible. As the Query Optimizer is working within
a time constraint, there's a chance that the plan selected may be the optimal plan but it is
also likely that it may just be something close to the optimal plan.
In order to explore the search space, the Query Optimizer uses transformation rules and
heuristics. The generation of candidate execution plans is performed inside the Query
Optimizer using transformation rules, and the use of heuristics limits the number of
choices considered, in order to keep the optimization time reasonable. Candidate plans
are stored in memory during the optimization, in a component called the Memo.
Transformation rules, heuristics, and the Memo will be discussed in more detail in
Chapter 5, The Optimization Process.

20
Assessing the cost of each plan
Searching or enumerating candidate plans is just one part of the optimization process.
The Query Optimizer still needs to estimate the cost of these plans and select the least
expensive one. To estimate the cost of a plan, it estimates the cost of each physical
operator in that plan, using costing formulas that consider the use of resources such as
I/O, CPU, and memory. This cost estimation depends mostly on the algorithm used by
the physical operator, as well as the estimated number of records that will need to be
processed; this estimate of the number of records is known as the cardinality estimation.
To help with this cardinality estimation, SQL Server uses and maintains optimizer
statistics, which contain statistical information describing the distribution of values
in one or more columns of a table. Once the cost for each operator is estimated using
estimations of cardinality and resource demands, the Query Optimizer will add up all of
these costs to estimate the cost for the entire plan. Rather than go into more detail here,
statistics and cost estimation will be covered in more detail in Chapter 3, Statistics and
Cost Estimation.
Query execution and plan caching
Once the query is optimized, the resulting plan is used by the execution engine to retrieve
the desired data. The generated execution plan may be stored in memory, in the plan
cache (known as the "procedure cache" in previous versions of SQL Server) in order that
it may be reused if the same query is executed again. If a valid plan is available in the plan
cache, then the optimization process can be skipped and the associated cost of this step,
in terms of optimization time, CPU resources, and so on, can be avoided.
However, reuse of an existing plan may not always be the best solution for a given query.
Depending on the distribution of data within a table, the optimal execution plan for a
given query may differ greatly, depending on the parameters provided in said query, and
a behavior known as parameter sniffing may result in a suboptimal plan being chosen.

21
In fact, given the level of impact which query parameters can have on query performance,
the parameter sniffing behavior (as well as several other parameter-related topics) will be
discussed in plenty of detail in Chapter 6, Additional Topics.
Even when an execution plan is available in the plan cache, some metadata changes, such
as the removal of an index or a constraint, or significant enough changes made to the
contents of the database, may render an existing plan invalid or suboptimal, and thus
cause it to be discarded from the plan cache and a new optimization to be generated. As
a trivial example, removing an index will make a plan invalid if the index is used by that
plan. Likewise, the creation of a new index could make a plan suboptimal, if this index
could be used to create a more efficient alternative plan; and enough changes to the
database contents may trigger an automatic update of statistics, with the same effect on
the existing plan.
Plans may also be removed from the plan cache when SQL Server is under memory
pressure or when certain statements are executed. Changing some configuration options,
for example, max degree of parallelism, will clear the entire plan cache. Alternatively,
some statements, like altering a database with certain ALTER DATABASE options will
clear all the plans associated with that particular database.
Hinting
Most of the time, the Query Optimizer does a great job of choosing highly efficient
execution plans. However, there may be cases when the selected execution plan does
not perform as expected. It is vitally important to differentiate between the occasions
when these cases arise because you are not providing the Query Optimizer with all the
information it needs to do a good job, and the occasions when the problem arises because
of a Query Optimizer limitation. As mentioned earlier, one of the purposes of this book is
to give you the knowledge and tools, both to write better queries, and to troubleshoot the
cases when the Query Optimizer is not giving you a good plan, and your queries are not
performing well.

22
The reality is that query optimizers are highly complex pieces of software which, even
after more than 30 years of research, still face technical challenges, some of which will
be mentioned in the next section. As a result, there may be cases when, even after you've
provided the Query Optimizer with all the information it needs, and there doesn't seem
to be any apparent problem, you are still not getting an efficient plan; in these cases you
may want to resort to hints. However, since hints let you to override the operations of the
Query Optimizer, they need to be used with caution, and only as a last resort when no
other option is available. Hints are instructions that you can send to the Query Optimizer
to influence a particular area of an execution plan. For example, you can use hints to
direct the Query Optimizer to use a particular index or a specific join algorithm. You can
even ask the Query Optimizer to use a specific execution plan, provided that you specify
one in XML format. Hints, and cases where you may need to use them, will be covered in
Chapter 7, Hints.
Ongoing Query Optimizer Challenges
Query optimization is an inherently complex problem, not only in SQL Server, but in
any other relational database system. Despite the fact that query optimization research
dates back to the early seventies, challenges in some fundamental areas are still being
addressed today. The first major impediment to a query optimizer finding an optimal
plan is the fact that, for many queries, it is just not possible to explore the entire search
space. An effect known as combinatorial explosion makes this exhaustive enumeration
impossible, as the number of possible plans grows very rapidly depending on the number
of tables joined in the query. To make the search a manageable process, heuristics are
used to limit the search space (these will be touched upon again in Chapter 5, The
Optimization Process). However, if a query optimizer is not able to explore the entire
search space, there is no way to prove that you can get an absolutely optimal plan, or
even that the best plan is among the candidates being considered. As a result, it is clearly
extremely important that the set of plans which a query optimizer considers contains
plans with low costs.

23
This leads us to another major technical challenge for the Query Optimizer: accurate
cost and cardinality estimation. Since a cost-based optimizer selects the execution plan
with the lowest cost, the quality of the plan selection is only as good as the accuracy of
the optimizer's cost and cardinality estimations. Even supposing that time is not a
concern, and that the query optimizer can analyze the entire search space without a
problem, cardinality and cost estimation errors can still make a query optimizer select
the wrong plan. Cost estimation models are inherently inexact, as they do not consider
all of the hardware conditions, and must necessarily make certain assumptions about
the environment. For example, the costing model assumes that every query starts with a
cold cache (i.e. that its data is read from disk and not from memory) and this assumption
could lead to costing estimation errors in some cases. In addition, cost estimation
relies on cardinality estimation, which is also inexact and has some known limitations,
especially when it comes to the estimation of the intermediate results in a plan. On top of
all that, there are some operations which are not covered by the mathematical model of
the cardinality estimation component, which has to resort to guess logic or heuristics to
deal with these situations. Cardinality and cost estimation will be covered in more detail
in Chapter 3, Statistics and Cost Estimation.
A historical perspective
We've seen some of the challenges query optimizers still face today, but these
imperfections are not for want of time or research. One of these earliest works
describing a cost-based query optimizer was Access Path Selection in a Relational
Database Management System, published in 1979 by Pat Selinger et al., to describe the
query optimizer for an experimental database management system developed in 1975
at what is now the IBM Almaden Research Center. This database management system,
called "System R," advanced the field of query optimization by introducing the use of cost-
based query optimization, the use of statistics, an efficient method of determining join
orders, and the addition of CPU cost to the optimizer's cost estimation formulae.

24
Yet, despite being an enormous influence in the field of query optimization research,
it suffered a major drawback: its framework could not be easily extended to include
additional transformations. This led to the development of more extensible optimiza-
tion architectures, which facilitated the gradual addition of new functionality to query
optimizers. The trailblazers in this field were the Exodus Optimizer Generator, defined by
Goetz Graefe and David DeWitt and, later, the Volcano Optimizer Generator, defined by
Goetz Graefe and William McKenna. Goetz Graefe then went on to define the Cascades
Framework, resolving errors which were present in his previous two endeavors.
While this is interesting, what's most relevant for you and me is that SQL Server
implemented its own cost-based Query Optimizer, based on the Cascades Framework,
in 1999, when its database engine was re-architected for the release of SQL Server 7.0.
The extensible architecture of the Cascades Framework has made it much easier for
new functionality, such as new transformation rules or physical operators, to be
implemented in the Query Optimizer. We will discuss transformation rules in
Chapter 5, The Optimization Process, and physical operators will be discussed in Chapter 2,
The Execution Engine.
Execution Plans
Now that we've got a foundation in the Query Optimizer and how it works its magic, it's
time to consider how we, as users, can interact with it. The primary way we'll interact
with the Query Optimizer is through execution plans which, as I mentioned earlier, are
ultimately trees consisting of a number of physical operators which, in turn, contain
the algorithms to produce the required results from the database. Given that I will make
extensive use of execution plans throughout the book, and because it's very useful to be
familiar with them in any case, in this section I'll show you how to display and read them.
You can request either an actual or an estimated execution plan for a given query, and
either of these two types can be displayed as a graphic, text, or XML plan. The only
difference between these three formats is the level of detail of information displayed.

25
However, when an actual plan is requested, the query needs to be executed, and the plan
is then displayed along with the query results. On the other hand, when an estimated
plan is requested, the query is naturally not executed; the plan displayed is simply the
plan that SQL Server would most probably use if the query were executed (bearing
in mind that a recompile, which we'll discuss later, may generate a different plan at
execution time). Nevertheless, using an estimated plan has several benefits, including
displaying a plan for a long-running query for inspection without actually running the
query, or displaying a plan for update operations without changing the database.
You can display the graphical plans in SQL Server Management Studio by clicking the
Display Estimated Execution Plan or Include Actual Execution Plan buttons from
the SQL Editor toolbar, which is enabled by default. Clicking on Display Estimated
Execution Plan will show the plan immediately, without executing the query whereas,
to request an actual execution plan, you need to click on Include Actual Execution Plan
and then execute the query.
As an example, copy the following query to the Management Studio Query Editor, select
the AdventureWorks database, click the Include Actual Execution Plan button, and
execute the query.
SELECT DISTINCT(City) FROM Person.Address
Listing 1-1.
This displays the plan shown in Figure 1-2.
Figure 1-2: Graphical execution plan.

26
Physical operators, such as the Index Scan and the Hash Aggregate physical operators,
seen in Figure 1-2, are represented as icons in a graphical plan. The first icon is called the
result operator; it represents the SELECT statement, and is usually the root element in
the plan.
Operators implement a basic function or operation of the execution engine; for
example, a logical join operation could be implemented by any of three different
physical join operators: Nested Loops Join, Merge Join or Hash Join. Obviously, there
are many more operators implemented in the execution engine, and all of them are
documented in Books Online, if you're curious about them. The Query Optimizer
builds an execution plan, choosing from these operators, which may read records from
the database, like the Index Scan operator shown in the previous plan, or may read
records from another operator, like the Hash Aggregate, which is reading records from
the Index Scan operator.
After the operator performs some function on the records it has read, the results are
output to its parent. This data flow is represented by arrows between the operators; the
thickness of the arrows corresponds to the relative number of rows. You can hover the
mouse pointer over an arrow to get more information about that data flow, displayed
as a tooltip. For example, if you hover the mouse pointer over the arrow between the
Index Scan and the Hash Aggregate operators (shown in Figure 1-2), you will get the data
flow information between these operators, as shown in Figure 1-3.
Figure 1-3: Data flow between Index Scan and Hash Aggregate operators.

27
By looking at the actual number of rows, you can see that the Index Scan operator is
reading 19,614 rows from the database and sending them to the Hash Aggregate
operator. The Hash Aggregate operator is, in turn, performing some operation on this
data and sending 575 records to its parent, which you can see by placing the mouse
pointer over the arrow between the Hash Aggregate and the SELECT icon.
Basically, in this instance, the Index Scan operator is reading all 19,614 rows from an
index, and the Hash Aggregate is processing these rows to obtain the list of distinct cities,
of which there are 575, which will be displayed in the Results window in Management
Studio. Notice, also, how you can see the estimated, as well as the actual, number of rows;
this is the Query Optimizer's cardinality estimation for this operator. Comparing the
actual and the estimated number of rows can help you to detect cardinality estimation
errors, which can affect the quality of your execution plans, as will be discussed in
Chapter 3, Statistics and Cost Estimation.
To perform their job, physical operators implement at least the following three methods:
Open(), which causes an operator to be initialized, GetRow() to request a row from
the operator, and Close() to shut down the operator once it has performed its role. An
operator can request rows from other operators by calling their GetRow() method. Since
GetRow() produces just one row at a time, the actual number of rows displayed in the
execution plan is also the number of times the method was called on a specific operator,
and an additional call to GetRow() is used by the operator to indicate the end of the result
set. In the previous example, the Hash Aggregate operator calls the Open() method once,
GetRow() 19,615 times and Close() once on the Index Scan operator.
In addition to learning more about the data flow, you can also hover the mouse pointer
over an operator to get more information about it. For example, Figure 1-4 shows infor-
mation about the Index Scan operator; notice that it includes, among other things, data
on estimated costing information like the estimated I/O, CPU, operator and subtree
costs. You can also see the relative cost of each operator in the plan as a percentage of the
overall plan, as shown in Figure 1-2. For example, the cost of the Index Scan is 52% of the
cost of the entire plan.

28
Figure 1-4: Tooltip for the Index Scan operator.
Additional information from an operator or the entire query can be obtained by using
the Properties window. So, for example, choosing the SELECT icon and selecting the
Properties window from the View menu (or pressing F4) will show some properties for
the entire query, as shown in Figure 1-5.

29
Figure 1-5: Properties window for the query.
Once you have displayed a graphical plan, you can also easily display the same plan in
XML format. Simple right-click anywhere on the execution plan window to display a
pop-up window, as shown in Figure 1-6, and select Show Execution Plan XML…; this
will open the XML editor and display the XML plan as shown in Figure 1-7. As you can
see, you can easily switch between a graphical and an XML plan.

30
Figure 1-6: Pop-up window on the execution plan window.
Figure 1-7: XML execution plan.
If needed, graphical plans can be saved to a file by selecting Save Execution Plan As…
from the pop-up window shown in Figure 1-6. The plan, usually saved with a .sqlplan
extension, is actually an XML document containing the XML plan, but can be read by
Management Studio into a graphical plan. You can load this file again, by selecting File >
Open in Management Studio, in order to immediately display it as a graphical plan, which
will behave exactly as before.

31
Table 1-1 shows the different statements you can use to obtain an estimated or actual
execution plan in text, graphic, or XML format. Note that, when you run any of these
statements using the ON clause, it will apply to all subsequent statements until the
option is manually set to OFF again.
Estimated Execution Plan Actual Execution Plan
Text
Plan
SET SHOWPLAN_TEXT ON
SET SHOWPLAN_ALL ON
SET STATISTICS PROFILE ON
Graphic
Plan
Management Studio Management Studio
XML
Plan
SET SHOWPLAN_XML ON SET STATISTICS XML ON
Table 1-1: Statements for displaying query plans.
As you can see in Table 1-1, there are two commands to get estimated text plans; SET
SHOWPLAN_TEXT and SET SHOWPLAN_ALL. Both statements show the estimated
execution plan, but SET SHOWPLAN_ALL also shows some additional information,
including the estimated number of rows, estimated CPU cost, estimated I/O cost, and
estimated operator cost. However, recent versions of Books Online, including that of
SQL Server 2008 R2, indicate that all text versions of execution plans will be deprecated
in a future version of SQL Server.

32
To show an XML plan you can use the following commands.
SET SHOWPLAN_XML ON
GO
GO
SET SHOWPLAN_XML OFF
Listing 1-2.
This will display a link starting with the following text:
<ShowPlanXML xmlns="http://schemas.microsoft.com/sqlserver/2004 ...
Listing 1-3.
Clicking the link will show you a graphical plan, and you can then display the XML plan
using the same procedure as explained earlier. Alternatively, you can use the following
code to display a text execution plan.
GO
GO
SET SHOWPLAN_TEXT OFF
GO
Listing 1-4.
This code will actually display two results sets, the first one returning the text of the
T-SQL statement. In the second result set, you will see the following plan (edited to fit the
page), which shows the same Hash Aggregate and Index Scan operators displayed earlier
in Figure 1-2.

33
|--Hash Match(Aggregate, HASH:([Person].[Address].[City]), RESIDUAL …
|--Index Scan(OBJECT:([AdventureWorks].[Person].[Address]. [IX_Address …
Listing 1-5.
Finally, be aware that there are still other ways to display an execution plan, such as using
SQL trace (for example by using SQL Server Profiler) or the sys.dm_exec_query_plan
dynamic management function (DMF). As mentioned earlier, when a query is optimized,
its execution plan may be stored in the plan cache, and the sys.dm_exec_query_plan
DMF can display such cached plans, as well as any plan which is currently executing.
The following query in Listing 1-6 will show the execution plans for all the queries
currently running in the system. The sys.dm_exec_requests dynamic management
view (DMV), which returns information about each request currently executing, is used
to obtain the plan_handle value, which is needed to find the execution plan using the
sys.dm_exec_query_plan DMF. A plan_handle is a hash value which represents a
specific execution plan, and it is guaranteed to be unique in the system.
SELECT query_plan FROM sys.dm_exec_requests
CROSS APPLY sys.dm_exec_query_plan(plan_handle)
WHERE session_id = 135
Listing 1-6.
The output will be a result set containing links similar to the one shown in Listing 1-3
and, as explained before, clicking the link will show you the graphical execution plan. For
more information about the sys.dm_exec_requests DMV and the sys.dm_exec_
query_plan DMF, you should go to Books Online.
If you're not yet familiar with execution plans in all their myriad forms, this section
should have given you enough background to follow along through the rest of the book.
We'll cover more topics and skills as we go along, but, in the meantime, let's take a look at
one of the most fundamental puzzles of query optimization.

34
Join Orders
Join ordering is one of the most complex problems in query optimization, and one that
has been the subject of extensive research since the seventies. It refers to the process of
calculating the optimal join order, that is, the order in which the necessary tables are
joined, when executing a query. As suggested in the ongoing challenges briefly discussed
earlier, join ordering is directly related to the size of the search space, as the number of
possible plans for a query grows very rapidly, depending on the number of tables joined.
A join combines records from two tables based on some common information, and
the predicate which defines which columns are used to join the tables is called a join
predicate. A join works with only two tables at a time, so a query requesting data from
n tables must be executed as a sequence of n – 1 joins, but it should be noted that the first
join does not have to be completed before the next join can be started. Because the
order of joins is a key factor in controlling the amount of data flowing between each
operator in the execution plan, it's a factor which the Query Optimizer needs to pay
close attention to.
Specifically, the Query Optimizer needs to make two important decisions regarding joins:
• the selection of a join order
• the choice of a join algorithm.
In this section I'll talk about join orders but, since the implementation of join algorithms
is part of the execution engine, selecting a join algorithm will be explained in Chapter 2,
The Execution Engine. Join order is, strictly speaking, a separate concern from the
algorithms provided by the execution engine, so I'll give an overview of the former here.
As mentioned, the order in which the tables are joined determines the cost and
performance of a query. Although the results of the query are the same, regardless
of the join order, the access cost of each different join order can vary dramatically.

35
As a result of the commutative and associative properties of joins, even simple queries
offer many different possible join orders, and this number increases exponentially with
the number of tables that need to be joined. The task of the Query Optimizer is to find
the optimal sequence of joins between the tables used in the query. To clarify this
challenge, let's first clarify the terminology.
The commutative property of a join between tables A and B states that:
A JOIN B is equivalent to B JOIN A.
This defines which table will be accessed first. In a Nested Loops Join, for example, the
first accessed table is called the outer table and the second one the inner table. In a Hash
Join, the first accessed table is the build input and the second one the probe input. As we
will see in the next chapter, correctly defining which table will be the inner and outer
table in a Nested Loops Join, or the build input or probe input in a Hash Join is important
to get right, as it has significant performance and cost implications, and it is a choice
made by the Query Optimizer.
The associative property of a join between tables A, B, and C states that:
(A JOIN B) JOIN C is equivalent to A JOIN (B JOIN C).
This defines the order in which the tables are joined. For example, (A JOIN B) JOIN C
specifies that table A must be joined to table B first, and then the result must be joined to
table C. A JOIN (B JOIN C) means that table B must be joined to table C first and then the
result must be joined to table A. Each possible permutation may have different cost and
performance results depending, for example, on the size of their temporary results.
Costing of the join algorithms will also be explained in the next chapter.
By way of an example, Listing 1-7 shows a query, taken from Books Online, which joins
together three tables in the AdventureWorks database. Click Include Actual Execution
Plan and execute the query.

36
SELECT FirstName, LastName
FROM Person.Contact AS C
JOIN Sales.Individual AS I
ON C.ContactID = I.ContactID
JOIN Sales.Customer AS Cu
ON I.CustomerID = Cu.CustomerID
WHERE Cu.CustomerType = 'I'
Listing 1-7.
By looking at the resultant execution plan, shown on Figure 1-8, you can see that the
Query Optimizer is not using the same join order as that specified in the query; it found a
more efficient one instead. The join order as expressed in the query is (Person.Contact
JOIN Sales.Individual) JOIN Sales.Customer. However, you will see from the
plan shown in Figure 1-8 that the Query Optimizer actually chose the join order (Sales.
Customer JOIN Sales.Individual) JOIN Person.Contact.
Figure 1-8: Execution plan for query joining three tables.
You should also notice that the Query Optimizer chose a Merge Join operator to
implement the join between the first two tables, then a Hash Join operator to join the
result to the Person.Contact table.
Just to experiment, the query shown in Listing 1-8 shows the same query, but this time
using the FORCE ORDER hint to instruct the Query Optimizer to join the tables in the
exact order indicated in the query. Paste this query into the same query window in

37
Management Studio as the one from Listing 1-7, and execute both of them together,
capturing their execution plans.
OPTION (FORCE ORDER)
Listing 1-8.
The result set returned is, of course, exactly the same in each case, but the execution plan
for the FORCE ORDER query (shown in Figure 1-9), indicates that the Query Optimizer
followed the prescribed join order, and this time chose a Hash Match Join operator for
the first join.
Figure 1-9: Execution plan using the FORCE ORDER hint.
This might not seem significant, but if you compare the cost of each query, via the
Query cost (relative to the batch) information at the top of each plan, you will see that
there might be a price to pay for overruling the Query Optimizer, as it has found the

38
hinted query to be more expensive. Specifically, the relative cost of the first query is 38%,
compared to 62% for the FORCE ORDER query.
Estimated subtree costs
You can get the same result by hovering over the SELECT icon of each plan and examining the
Estimated Subtree Cost which, in this case, is the entire tree or query. The first query will show
a cost of 3.2405 and the second one will show 5.3462. Therefore the relative cost of the second query
is 5.3462/(3.2405 + 5.3462)*100 = 62%.
As noted earlier, the number of possible join orders in a query increases exponentially
with the number of tables. In fact, with just a handful of tables, the number of possible
join orders could be numbered in the thousands or even millions, although the exact
number of possible join orders depends on the overall shape of the query tree. Obviously,
it is impossible for the Query Optimizer to look at all those combinations: it would take
far too long. Instead, it uses heuristics, such as considering the shape of the query tree, to
help it narrow down the search space.
As mentioned before, queries are represented as trees in the query processor, and the
shape of the query tree, as dictated by the nature of the join ordering, is so important in
query optimization that some of these trees have names, such as left-deep, right-deep and
bushy trees.
Figure 1-10 shows left-deep and bushy trees for a join of four tables. For example, the
left-deep tree could be: JOIN( JOIN( JOIN(A, B), C), D)
And the bushy tree could be: JOIN(JOIN(A, B), JOIN(C, D))
Left-deep trees are also called linear trees or linear processing trees, and you can see
how their shapes lead to that description. Bushy trees, on the other hand, can take any
arbitrary shape, and so the set of bushy trees actually includes the sets of both left-deep
and right-deep trees.

39
Figure 1-10: Left-deep and bushy trees.
Table 1-2 shows how the number of possible join orders increases as we increase the
number of tables, for both left-deep and bushy trees, and I'll explain how it's calculated
in a moment.
Tables Left-deep trees Bushy trees
1 1 1
2 2 2
3 6 12
4 24 120
5 120 1,680
6 720 30,240
7 5,040 665,280
8 40,320 17,297,280
9 362,880 518,918,400

40
Tables Left-deep trees Bushy trees
10 3,628,800 17,643,225,600
11 39,916,800 670,442,572,800
12 479,001,600 28,158,588,057,600
Table 1-2: Possible join orders for left-deep and bushy trees.
The number of left-deep trees is calculated as n!, or n factorial, where n is the number
of tables in the relation. A factorial is the product of all positive integers less than or
equal to n; so, for example, for a five-table join, the number of possible join orders is
5! = 5 x 4 x 3 x 2 x 1 = 120.
The number of possible join orders for a bushy tree is more complicated, and can be
calculated as (2n–2)!/(n–1)!.
The important point to remember here is that the number of possible join orders grows
very quickly as the number of tables increase, as highlighted by Table 1-2. For example,
in theory, if we had a six-table join, a query optimizer would potentially need to evaluate
30,240 possible join orders.
Of course, we should also bear in mind that this is just the number of permutations for
the join order. On top of this, the Query Optimizer also has to evaluate a number of
possible physical join operators, data access methods (e.g. Table Scan, Index Scan or
Index Seek), as well as optimize other parts of the query, such as aggregations, subqueries
and so on.
So how does the Query Optimizer analyze all these possible plan combinations? The
answer is: it doesn't. Performing an exhaustive evaluation of all possible combinations,
for a complex query, would take too long to be useful, so the Query Optimizer must find
a balance between the optimization time and the quality of the resulting plan. Rather
than exhaustively evaluate every single combination, the Query Optimizer tries to

41
narrow the possibilities down to the most likely candidates, using heuristics (some of
which we've already touched upon) to guide the process, which will be explained in
Chapter 5, The Optimization Process.
Summary
This chapter has covered a lot of ground in a relatively short space, but by now you
should have an understanding (or at least an appreciation) of the concepts we're going to
tackle in more detail in the following chapters.
We've been introduced to the fundamental operations of the SQL Server Query
Optimizer, from parsing the initial query to how the Query Optimizer tries to find
the best possible execution plan for every query submitted to SQL Server. We've also
looked at the complexity of the optimization process, including the challenges it faces in
exploring the potentially vast search space and accurately estimating cardinality and the
cost of candidate execution plans.
As a result of the research that has gone into solving some of those challenges, the Query
Optimizer implemented in SQL Server is based on the extensible Cascades Framework
architecture, which facilitates the addition of new functionality to the query optimizer,
including new operators and transformation rules. Chapters 2 and 5 are going to go into
more detail regarding both of those, and the section on how to read and understand
execution plans will also have given you the basic skills to be able to find information
regarding physical operators. Finally, we touched upon the problem of finding an
efficient join order in a multi-join query, which is still a fundamental challenge in
query optimization.
Now that we've had a first look at the concepts involved, we can start getting into the real
details of how the SQL Server Query Optimizer works.

42
Chapter 2: The Execution Engine
The Execution Engine is, at its heart, a collection of physical operators that perform the
functions of the query processor, which is to execute your query in an efficient way. Or,
to look at it from the other direction, these operations implemented by the Execution
Engine define the choices available to the Query Optimizer when building execution
plans. The Execution Engine and its operators were briefly introduced in the previous
chapter, and now we'll cover some of the most used operators, their algorithms and their
costs. In this chapter, I will focus on operators related to data access, aggregations, joins,
and parallelism, as these ones are the most commonly used in queries, and also the ones
more used in this book. Of course, there are many more operators implemented by the
Execution Engine, and you can find a complete list and description on SQL Server 2008
R2 in Books Online. Since the Query Optimizer is the primary focus of this book, this
chapter will illustrate how it decides between the various choices of operators provided
by the Execution Engine. For example, I will show you how the Query Optimizer reasons
about choosing between a Nested Loops Join or a Hash Join, or between a Stream
Aggregate and a Hash Aggregate operator.
This chapter starts with a look at the data access operations, including the operators
to perform scans, seeks, and bookmark lookups on database structures like heaps and
both clustered and non-clustered indexes. The concepts of sorting and hashing are also
explained, showing how they impact some of the algorithms of both physical joins and
aggregations, which are shown later. The next section focuses on aggregations, and
explains the Stream Aggregate and Hash Aggregate operators in detail. In the same way,
the joins section presents the Nested Loops Join, Merge Join and Hash Join physical
operators. The chapter concludes with an introduction to parallelism and how it can help
to reduce the response time of a query.

43
Data Access Operators
In this section, I will show you the operations that directly access the database, using
either a base table or an index, examples of which include scans and seeks. A scan reads
an entire structure, which could be a heap, a clustered index, or a non-clustered index.
A seek, on the other hand, does not scan an entire structure but, instead, efficiently
retrieves rows from an index. Therefore seeks can only be performed on a clustered or
non-clustered index. Just to make the difference between these structures clear, a heap
contains all of a table's columns, and its data is not stored sorted in any particular order.
Conversely, in a clustered index, the data is stored logically sorted by the clustering key
and, in addition to the clustering key, the clustered index also contains the remaining
columns of the table. On the other hand, a non-clustered index can be defined on a
clustered index or a heap, and usually contains only a subset of the columns of the table.
The operations on these structures are summarized in Table 2-1 below.
Structure Scan Seek
Heap Table Scan
Clustered index Clustered Index Scan Clustered Index Seek
Non-clustered index Index Scan Index Seek
Table 2-1: Data Access operators.

44
Scanning
Let's start with the simplest example, by scanning a heap which, as shown in Table 2-1,
is performed by the Table Scan operator. The following query on the AdventureWorks
database will use a table scan, as shown in Figure 2-1.
SELECT * FROM DatabaseLog
Listing 2-1.
Figure 2-1: A Table Scan operator.
Similarly, the following query will show a Clustered Index Scan operator, as shown in the
plan on Figure 2-2:
SELECT * FROM Person.Address
Listing 2-2.
Figure 2-2: A Clustered Index Scan operator.

45
Both the Table Scan and Clustered Index Scan operations are similar in that they
scan the entire base table, but the former is used for heaps and the second one for
clustered indexes.
Sorting is something to consider when it comes to scans, because even when the data in
a clustered index is stored sorted, using a Clustered Index Scan does not guarantee that
the results will be sorted unless this is explicitly requested. By not automatically sorting
the results, the Storage Engine has the option to find the most efficient way to access this
data without worrying about returning it in an ordered set. Examples of these efficient
methods include an advanced scanning mechanism called "merry-go-round scanning,"
which allows multiple query executions to share full table scans so that each execution
may join the scan at a different location. Alternatively, the Storage Engine may also use an
allocation order scan, based on Index Allocation Map (IAM) pages, to scan the table. I'm
not going to go into more detail regarding these techniques, because what's important
right now is that they exist, and the Storage Engine has the option of implementing them.
If you want to know whether your data has been sorted, the Ordered property can show
if the data was returned in a manner ordered by the Clustered Index Scan operator. So,
for example, the clustering key of the Person.Address table is AddressID, and if you
run the following query and look at the tooltip of the Clustered Index Scan operator, you
will get something similar to what is shown in Figure 2-3.
ORDER BY AddressID
Listing 2-3.

46
Figure 2-3: Properties of the Clustered Index Scan operator.
Notice that the Ordered property shows True. If you run the same query without
the ORDER BY clause, the Ordered property will, unsurprisingly, show False. In some
other cases, SQL Server can benefit from reading the table in the order specified by the
clustered index. One example is shown later in this chapter in Figure 2-15, where a Stream
Aggregate operator can benefit from the fact that a Clustered Index Scan operator can
easily obtain the data already sorted.
Next, I will show you an example of an Index Scan operator. This example uses a
non-clustered index to cover a query; that is, it can solve the entire query without
accessing the base table (bearing in mind that a non-clustered index usually contains only

47
a few of the columns of the table). Run the following query, which will show the plan in
Figure 2-4.
SELECT AddressID, City, StateProvinceID FROM Person.Address
Listing 2-4.
Figure 2-4: An Index Scan operator.
Note that the Query Optimizer was able to solve this query without even accessing
the base table Person.Address, and instead decided to scan the IX_Address_
AddressLine1_AddressLine2_City_StateProvinceID_PostalCode index, which
comprises fewer pages. The index definition includes AddressLine1, AddressLine2,
City, StateProvinceID and PostalCode, so it can clearly cover columns requested in
the query. However, you may wonder where the index is getting the AddressID column
from. When a non-clustered index is created on a table with a clustered index, each
non-clustered index row also includes the table clustering key. This clustering key is used
to find which record from the clustered index is referred to by the non-clustered index
row (a similar approach for non-clustered indexes on a heap will be explained later in this
section). In this case, as I mentioned earlier, AddressID is the clustering key of the table
and it is stored in every row of the non-clustered index, which is why the index was able
to cover this column in the previous query.

48
Seeking
Now let us look at Index Seeks, which can be performed by both the Clustered Index
Seek and the Index Seek operators and which are used against clustered and non-
clustered indexes, respectively. An Index Seek does not scan the entire index, but instead
navigates the B-tree index structure to quickly find one or more records. The next query,
together with the plan on Figure 2-5, shows an example of a Clustered Index Seek. A
benefit of a Clustered Index Seek, compared to a non-clustered Index Seek, is that the
former can cover any column of the table. Of course, since the records of a clustered
index are logically ordered by its clustering key, a table can only have one clustered index.
SELECT AddressID, City, StateProvinceID FROM Person.Address
WHERE AddressID = 12037
Listing 2-5.
Figure 2-5: A Clustered Index Seek operator.
The next query and Figure 2-6 both illustrate a non-clustered Index Seek operator. It is
interesting to note here that the base table was not used at all and it was not even neces-
sary to scan the entire index: there is a non-clustered index on the StateProvinceID
and, as mentioned previously, it also contains the clustering key AddressID.
SELECT AddressID, StateProvinceID FROM Person.Address
WHERE StateProvinceID = 32
Listing 2-6.

49
Figure 2-6: An Index Seek operator.
Bookmark lookup
The question that now comes up is what happens if a non-clustered index is useful
to quickly find one or more records, but does not cover the query? In other words, what
happens if the non-clustered index does not contain all of the columns requested by
the query? In this case, the Query Optimizer has to decide if it is more efficient to both
use the non-clustered index to find these records quickly and also access the base table
to obtain the additional fields, or to just go straight to the base table and scan it. For
example, on the previous query on Listing 2-6, an existing non-clustered index covers
both AddressID and StateProvinceID columns. What about if we also request the
City and ModifiedDate columns on the same query? This is shown in the next query,
which returns one record and produces the plan in Figure 2-7.
SELECT AddressID, City, StateProvinceID, ModifiedDate
FROM Person.Address
Listing 2-7.

50
Figure 2-7: A bookmark lookup example.
As in the previous example, the Query Optimizer is choosing the index IX_Address_
StateProvinceID to find the records quickly. However, because the index does not
cover the additional columns, it also needs to use the base table (in this case the clustered
index) to get that additional information. This operation is called a bookmark lookup,
and it is performed by the Key Lookup operator, which was introduced specifically to
differentiate a bookmark lookup from a regular Clustered Index Seek. Actually, the Key
Lookup operator only appears on a graphical plan (and then only from SQL Server
2005 Service Pack 2 and onwards), although text and XML plans can also show if a
Clustered Index Seek operator is performing a bookmark lookup. For example, run
the following query:
GO
FROM Person.Address
GO
GO
Listing 2-8.

51
The output will show the following text plan including a Clustered Index Seek operator
with the LOOKUP keyword at the end:
|--Nested Loops(Inner Join, OUTER REFERENCES …)
|--Index Seek(OBJECT:([Address].[IX_Address_StateProvinceID]),
SEEK:([Address].[StateProvinceID]=(32)) ORDERED FORWARD)
|--Clustered Index Seek(OBJECT:([Address].[PK_Address_AddressID]),
SEEK:([Address].[AddressID]=[Address].[AddressID]) LOOKUP ORDERED FORWARD)
Listing 2-9.
The XML plan shows the same information in the following way:
<RelOp … PhysicalOp="Clustered Index Seek" …>
…
<IndexScan Lookup="true" …>
Listing 2-10.
Keep in mind that, although SQL Server 2000 implemented a bookmark lookup using a
dedicated operator (called Bookmark Lookup), the operation is basically the same.
Now run the same query but, this time, request StateProvinceID equal to 20. This will
produce the plan shown in Figure 2-8.
FROM Person.Address
Listing 2-11.

52
Figure 2-8: Plan switching to a Clustered Index Scan.
This time, the Query Optimizer has selected a Clustered Index Scan and returned 308
records (compared to just a single record for the StateProvinceID 32). So the Query
Optimizer is producing two different execution plans for the same query, with the only
difference being the value of the StateProvinceID parameter. As I will show in more
detail in the next chapter, in this case, the Query Optimizer uses the value of the query's
StateProvinceID parameter to estimate the cardinality of the predicate as it tries to
produce an efficient plan for that parameter.
This time, the Query Optimizer estimated that more records could be returned than
when StateProvinceID was equal to 32, and it decided that it was cheaper to do a Table
Scan than to do many bookmark lookups. At this stage, you may be wondering at what
point the Query Optimizer decides to change from one method to another. Well, since a
bookmark lookup requires random I/O, which is very expensive, it would not take many
records for the Query Optimizer to switch from a bookmark lookup to a Clustered Index
Scan (or a Table Scan). We already know that, when the query returned one record, for
StateProvinceID 32, the Query Optimizer chose a bookmark lookup. We also saw that,
when we requested the records for StateProvinceID 20, which returned 308 records,
it used a Clustered Index Scan. Logically, we can try requesting somewhere between 1 and
308 records to find this switch-over point, right?
Actually, as you may already suspect, this is a cost-based decision which does not
depend on the actual number of records returned by query, but rather the estimated
number of records. We (or rather, the Query Optimizer) can find these estimates by
analyzing the appropriate statistics object for the IX_Address_StateProvinceID
index, something that will be covered in Chapter 3, Statistics and Cost Estimation.

53
I performed this exercise and found that the highest estimated number of records to
get a bookmark lookup for this particular example was 62, and the first one to have a
Clustered Index Scan was 106. Let us see both examples here, by running the query
with the StateProvinceID values 163 and 71. You will get the plans on Figure 2-9
and Figure 2-10, respectively.
Figure 2-9: Plan for the StateProvinceID = 163 predicate.
Figure 2-10: Plan for the StateProvinceID = 71 predicate.
By looking at the plans, you can see that, for this specific example, the Query Optimizer
selects a bookmark lookup for an estimated 62 records, and changes to a Clustered Index
Scan when that estimated number of records goes up to 106 (there are no estimated
values between 62 and 106 for this particular statistics object). Although in this case both
the actual and estimated number of rows are the same, keep in mind that the Query
Optimizer makes its decision based on the estimated number of rows. It does not know
the actual number of rows when the execution plan is generated (as the candidate plans

54
are only models and estimations), as the actual number of records is only known when
the plan is executed and the results returned.
Finally, since non-clustered indexes can exist on both heaps and clustered indexes, we can
also have a bookmark lookup on a heap. To follow the next example, create an index on
the DatabaseLog table, which is a heap, by running the following statement:
CREATE INDEX IX_Object ON DatabaseLog(Object)
Listing 2-12.
Then run the following query, which will produce the plan in Figure 2-11:
SELECT * FROM DatabaseLog
WHERE Object = 'City'
Listing 2-13.
Figure 2-11: A RID Lookup.
Note that, instead of the Key Lookup operator shown before, this plan displays a RID
Lookup operator. This is because heaps do not have clustering keys like clustered indexes
do, and instead they have row identifiers (RID). A RID is a row locator that includes
information like the database file, page, and slot numbers to allow a specific record to be

55
easily located. Every row in a non-clustered index created on a heap contains the RID of
the corresponding heap record.
To clean up, simply remove the index you just created:
DROP INDEX DatabaseLog.IX_Object
Listing 2-14.
Aggregations
Aggregations are used in databases to summarize information about some set of data.
The result can be a single value, such as the average salary for a company, or it can be a
per-group value, like the average salary by department. SQL Server has two operators to
implement aggregations, Stream Aggregate and Hash Aggregate, and they can be used to
solve queries with aggregation functions (like SUM, AVG or MAX), the GROUP BY clause, or
the DISTINCT keyword.
Sorting and hashing
Before introducing the remaining operators of this chapter, I would like to add a brief
discussion on sorting and hashing, which play a very important role in some of the
operators and algorithms of the Execution Engine. For example, two of the operators
covered on this chapter, Stream Aggregate and Merge Join, require data to be already
sorted. To provide sorted data, the Query Optimizer may employ an existing index, or it
may explicitly introduce a Sort operator.

56
On the other hand, hashing is used by the Hash Aggregate and Hash Join operators, both
of which work by building a hash table in memory. The Hash Join operator uses memory
only for the smaller of its two inputs, which is defined by the Query Optimizer.
Sorting also uses memory and, similar to hashing, will also use the tempdb database if
there is not enough available memory, which could become a performance problem. Both
sorting and hashing (only during the time the build input is hashed, as explained later) are
blocking or stop-and-go operations; that is, they cannot produce any rows until they have
consumed all their input.
Stream Aggregate
Let us start with the Stream Aggregate operator, using a query with an aggregation
function. Queries using an aggregate function and no GROUP BY clause are called
scalar aggregates, as they return a single value, and are always implemented by the
Stream Aggregate operator. To demonstrate, run the following query, which shows the
plan in Figure 2-12:
SELECT AVG(ListPrice) FROM Production.Product
Listing 2-15.
Figure 2-12: A Stream Aggregate.

57
A text plan can be useful to show more details about both the Stream Aggregate and the
Compute Scalar operators, so you should also run the query in Listing 2-16.
GO
SELECT AVG(ListPrice) FROM Production.Product
GO
GO
Listing 2-16.
The displayed text plan is:
|--Compute Scalar(DEFINE:([Expr1003]=CASE WHEN [Expr1004]=(0) THEN NULL ELSE
[Expr1005]/CONVERT_IMPLICIT(money,[Expr1004],0) END))
|--Stream Aggregate(DEFINE:([Expr1004]=Count(*), [Expr1005]=SUM([Product].
[ListPrice])))
|--Clustered Index Scan(OBJECT:([Product].[PK_Product_ProductID]))
The same information could be obtained from the graphical plan by selecting the
Properties window (by pressing F4) of both the Stream Aggregate and Compute Scalar
operators, and expanding the Defined Values property as shown in Figure 2-13.

58
Figure 2-13: Properties of the Stream Aggregate operator.
Note that, in order to implement the AVG aggregation function, the Stream Aggregate
is computing both a COUNT and a SUM aggregate, the results of which will be stored in
the computed expressions Expr1004 and Expr1005 respectively. The Compute Scalar
verifies that there is no division by zero by using a CASE expression. As you can see in
the text plan, if Expr1004, which is the value for the count, is zero, the Compute Scalar
operator returns NULL, otherwise it calculates and returns the average by dividing the
sum by the count.

59
Now let's see an example of a query using the GROUP BY clause; the following query
produces the plan in Figure 2-14:
SELECT ProductLine, COUNT(*) FROM Production.Product
GROUP BY ProductLine
Listing 2-17.
Figure 2-14: Stream Aggregate using a Sort operator.
A Stream Aggregate operator always requires its input to be sorted by the GROUP BY
clause predicate so, in this case, the Sort operator shown in the plan will provide the data
sorted by the ProductLine column. After receiving its input sorted, the records for
the same group will be next to each other, so the Stream Aggregate operator can count
the records for each group. Note that, although the first example of this section was also
using a Stream Aggregate, it did not require any sorted input: a query without a GROUP BY
clause considers its entire input a single group.
A Stream Aggregate can also use an index to have its input sorted, as in the following
query, which produces the plan on Figure 2-15:
SELECT SalesOrderID, SUM(LineTotal)
FROM Sales.SalesOrderDetail
GROUP BY SalesOrderID
Listing 2-18.

60
Figure 2-15: Stream Aggregate using an existing index.
The Sort operator is not needed in this plan, as the Clustered Index Scan provides
the data already sorted by SalesOrderID, which is part of the clustering key of the
SalesOrderDetail table. As in the previous example, the Stream Aggregate operator
will consume the sorted data, but this time it will calculate the sum of the LineTotal
column for each group.
Since the purpose of the Stream Aggregate operator is to aggregate values based on
groups, its algorithm relies on the fact that its input is already sorted by the GROUP BY
clause, and thus records from the same group are next to each other. Basically, in this
algorithm, the first record read will create the first group, and its aggregate value will
be initialized. Any record read after that will be checked to see if it matches the current
group; if it does match, then the record value will be aggregated to this group. On the
other hand, if the record doesn't match the current group, a new group will be created,
and its own aggregated value initialized. This process will continue until all the records
are processed.
Hash Aggregate
Now let us take a look at the Hash Aggregate operator, shown as Hash Match (Aggregate)
on the execution plans. This chapter describes two hash algorithms, Hash Aggregate and
Hash Join, which work in a similar way and are, in fact, implemented by the same physical
operator: Hash Match. I will cover the Hash Aggregate operator in this section, and the
Hash Join operator in the next one.

61
The Query Optimizer can select a Hash Aggregate for big tables where the data is not
sorted, there is no need to sort it, and its cardinality estimates only a few groups. For
example, the SalesOrderHeader table has no index on the ContactID column, so the
following query will use a Hash Aggregate operator, as shown in Figure 2-16.
SELECT ContactID, COUNT(*)
FROM Sales.SalesOrderHeader
GROUP BY ContactID
Listing 2-19.
Figure 2-16: A Hash Aggregate.
As mentioned earlier in this chapter, a hash operation builds a hash table in memory. The
hash key used for this table is displayed on the Properties window, as the Hash Keys Build
property, as shown in Figure 2-17, which in this case is ContactID. Since this table is not
sorted by the required column, ContactID, every row scanned can belong to any group.
The algorithm for the Hash Aggregate operator is similar to the Stream Aggregate, with
the exceptions that, in this case, the input data is not sorted, a hash table is created in
memory, and a hash value is calculated for each row processed. For each hash value
calculated, the algorithm will check if the corresponding group already exists on the
hash table and, if it does not, it will create a new entry for it. In this way, the values for
each record are aggregated in this entry on the hash table, and only one row for each
group is stored in memory.

62
Figure 2-17: Properties of the Hash Aggregate operator, showing Hash Keys Build property.
Note, again, that a Hash Aggregate helps when the data is not sorted. If you create an
index that can provide sorted data, then the Query Optimizer may select a Stream
Aggregate instead. Run the following statement to create an index, and then execute
the previous query again, to verify that it uses a Stream Aggregate, as shown in the plan
displayed in Figure 2-18.

63
CREATE INDEX IX_ContactID ON Sales.SalesOrderHeader(ContactID)
Listing 2-20.
Figure 2-18: A Stream Aggregate using an index.
To clean up, drop the index using the following DROP INDEX statement:
DROP INDEX Sales.SalesOrderHeader.IX_ContactID
Listing 2-21.
If the input is not sorted and order is explicitly requested in a query, the Query Optimizer
may introduce a Sort operator and a Stream Aggregate as shown previously, or it may
decide to use a Hash Aggregate and then sort the results as in the following query, which
produces the plan on Figure 2-19. The Query Optimizer will estimate which operation
is less expensive: to sort the entire input and use a Stream Aggregate, or to use a Hash
Aggregate and sort the aggregated results.
SELECT ContactID, COUNT(*)
GROUP BY ContactID
ORDER BY ContactID
Listing 2-22.

64
Figure 2-19: A Hash Aggregate followed by a Sort operator.
Finally, a query using the DISTINCT keyword can be implemented by a Stream Aggregate,
a Hash Aggregate or by a Distinct Sort operator. The Distinct Sort operator is used
to both remove duplicates and sort its input. In fact, a query using DISTINCT can be
rewritten as a GROUP BY query, and both can generate the same execution plan. If an
index to provide sorted data is available, the Query Optimizer can use a Stream Aggregate
operator. If no index is available, SQL Server can introduce a Distinct Sort operator or a
Hash Aggregate operator. Let's see all three cases here; the following two queries return
the same data and produce the same execution plan, as shown in Figure 2-20.
SELECT DISTINCT(Title)
FROM HumanResources.Employee
SELECT Title
FROM HumanResources.Employee
GROUP BY Title
Listing 2-23.
Note that the plan is using a Distinct Sort operator. This operator will sort the rows and
eliminate duplicates.
Figure 2-20: A Distinct Sort operator.

65
If we create an index, the Query Optimizer will introduce a Stream Aggregate operator
since the plan can take advantage of the fact that the data is already sorted. To test it,
run this:
CREATE INDEX IX_Title ON HumanResources.Employee(Title)
Listing 2-24.
Then run the previous queries again. Both queries will now produce the plan shown
on Figure 2-21.
Figure 2-21: A Stream Aggregate used by a query with a DISTINCT keyword.
Drop the index before continuing, by using this statement:
DROP INDEX HumanResources.Employee.IX_Title
Listing 2-25.
Finally, for a bigger table without an index to provide order, a Hash Aggregate may be
used, as in the two following examples.

66
SELECT DISTINCT(ContactID)
SELECT ContactID
GROUP BY ContactID
Listing 2-26.
Both queries produce the same results and will use the same execution plan, as shown in
Figure 2-22:
Figure 2-22: A Hash Aggregate used by a query with a DISTINCT keyword.
Joins
I started talking about joins and join orders in Chapter 1, Introduction to Query
Optimization. In this section I will talk about the three join operators that SQL
Server uses to implement logical joins: the Nested Loops Join, the Merge Join and
the Hash Join. It is important to understand that no join algorithm is better than
the others, and that the Query Optimizer will select the best join algorithm depending
on the specific scenario, as I'll explain here.

67
Nested Loops Join
Let's start with a query listing employees who are also sales persons. This creates the plan
in Figure 2-23, which uses a Nested Loops Join:
SELECT e.EmployeeID
FROM HumanResources.Employee AS e
INNER JOIN Sales.SalesPerson AS s
ON e.EmployeeID = s.SalesPersonID
Listing 2-27.
Figure 2-23: A Nested Loops Join.
The input shown at the top in a Nested Loops Join plan is known as the outer input
and the one at the bottom is the inner input. The algorithm for the Nested Loops Join
is very simple: the operator used to access the outer input is executed only once, and the
operator used to access the inner input is executed once for every record that qualifies on
the outer input. Note that, in this example, the plan is scanning a non-clustered index
instead of the base table for the outer input. Since there is no filter on the SalesPerson
table, all of its 17 records are returned and so, as dictated by the algorithm, the inner input
(the Clustered Index Seek) is executed 17 times – once for each row from the outer table.

68
You can validate this information by looking at the operator properties. Figure 2-24 shows
the Index Scan operator properties, where you can find the actual number of executions
(which in this case is 1), and the actual number of rows (in this case, 17). Figure 2-25 shows
the Clustered Index Seek operator properties, which demonstrates that both the actual
number of rows and the number of executions is 17.
Figure 2-24: Properties of the Index Scan operator.

69
Figure 2-25: Properties of the Clustered Index Seek operator.
Let us change the query to add a filter by TerritoryID.
SELECT e.EmployeeID
INNER JOIN Sales.SalesPerson AS s
ON e.EmployeeID = s.SalesPersonID
WHERE TerritoryID = 1
Listing 2-28.
This query produces the plan in Figure 2-26.

70
Figure 2-26: A Nested Loops Join with a filter on the outer table.
Note that the outer input is, again, SalesPerson, but this time it's not using an index;
the new predicate is using the TerritoryID column which is not included in any index,
and so the Query Optimizer decides to do a Clustered Index Scan instead. The filter on
the SalesPerson table is asking for TerritoryID equal to 1, and only three records
qualify this time. As a result, the Clustered Index Seek, which is the operator on the inner
input, is executed only three times. You can verify this information by looking at the
properties of each operator, as we did for the previous query.
To recap briefly, in the Nested Loops Join algorithm, the operator for the outer input will
be executed once, and the operator for the inner input will be executed once for every
row that qualifies on the outer input. The result of this is that the cost of this algorithm
is proportional to the size of the outer input multiplied by the size of the inner input. As
such, the Query Optimizer is more likely to choose a Nested Loops Join when the outer
input is small and the inner input has an index on the join key. This join type can be
especially effective when the inner input is potentially large, as only a few rows, indicated
by the outer input, will be searched.

71
Merge Join
Now let's take a look at a Merge Join example; run the following query, which returns
the name of each customer that is categorized as a store. The execution plan is shown in
Figure 2-27.
SELECT Name
FROM Sales.Store AS S
JOIN Sales.Customer AS C
ON S.CustomerID = C.CustomerID
WHERE C.CustomerType = N'S'
Listing 2-29.
Figure 2-27: A Merge Join example.
One difference between this and a Nested Loops Join is that, in a Merge Join, both input
operators are executed only once. You can verify this by looking at the properties of
both operators, and you'll find that the number of executions is 1. Another difference is
that a Merge Join requires an equality operator and its inputs sorted on the join predicate.
In this example, the join predicate has an equality operator, is using the CustomerID
column, and both clustered indexes are ordered by CustomerID, which is their
clustering key.

72
Taking benefit from the fact that both of its inputs are sorted on the join predicate, a
Merge Join simultaneously reads a row from each input and compares them. If the rows
match, they are returned. If the rows do not match, the smaller value can be discarded
because, since both inputs are sorted, the discarded row will not match any other row on
the other input table.
This process continues until one of the tables is completed. Even if there are still rows on
the other table, they will clearly not match any rows on the fully-scanned table, so there is
no need to continue. Since both tables can potentially be scanned, the maximum cost of a
Merge Join is the sum of both inputs.
If the inputs are not sorted, the Query Optimizer it is not likely to choose a Merge Join,
although you can test this and see what the Query Optimizer does if we force a Merge
Join. If you run the following query, you will notice that it uses a Nested Loops Join, as
SELECT *
INNER JOIN Person.Contact AS c
ON e.ContactID = c.ContactID
Listing 2-30.
Figure 2-28: A Nested Loops Join.

73
In this case, the Contact table is sorted on the join predicate, but Employee is not. If
you're curious, you can force a Merge Join using a hint, as in the following query; the
Query Optimizer will introduce a Sort operator to sort Employee on ContactID, as
SELECT *
INNER JOIN Person.Contact AS c
ON e.ContactID = c.ContactID
OPTION (MERGE JOIN)
Listing 2-31.
Figure 2-29: Plan with a hint to use a Merge Join.
As a summary, given the nature of the Merge Join, the Query Optimizer is more likely to
choose this algorithm when faced with medium to large inputs, where there is an equality
operator on the join predicate, and their inputs are sorted.

74
Hash Join
The third join algorithm used by SQL Server is the Hash Join. Run the following query to
produce the plan displayed in Figure 2-30, and then we'll take a closer look at the Hash
Join operator.
SELECT pv.ProductID, v.VendorID, v.Name
FROM Purchasing.ProductVendor pv JOIN Purchasing.Vendor v
ON (pv.VendorID = v.VendorID)
WHERE StandardPrice > $10
Listing 2-32
Figure 2-30: A Hash Join example.
In the same way as the Merge Join, the Hash Join requires an equality operator on the
join predicate but, unlike the Merge Join, it does not require its inputs to be sorted. In
addition, its operations in both inputs are executed only once, which you can verify by
looking at the operator properties as shown before. However, a Hash Join works by
creating a hash table in memory. The Query Optimizer will use a cardinality estimation
to detect the smaller of the two inputs, called the build input, and will use it to build a
hash table in memory. If there is not enough memory to host the hash table, SQL Server
can use disk space, creating a workfile in tempdb. A Hash Join will also block, but only
during the time the build input is hashed. After the build input is hashed, the second

75
table, called the probe input, will be read and compared to the hash table. If rows are
matched they will be returned. On the execution plan, the table at the top will be used as
the build input, and the table at the bottom as the probe input.
Finally, note that a behavior called "role reversal" may appear. If the Query Optimizer is
not able to correctly estimate which of the two inputs is smaller, the build and probe roles
may be reversed at execution time, and this will not be shown on the execution plan.
In summary, the Query Optimizer can choose a Hash Join for large inputs where there
is an equality operator on the join predicate. Since both tables are scanned, the cost of a
Hash Join is the sum of both inputs.
Parallelism
I will finish this discussion of the Execution Engine operations with a quick introduction
to parallelism. SQL Server can introduce parallelism to help some expensive queries to
execute faster by using several processors simultaneously. However, even when a query
may get better performance by using parallel plans, it may still use more resources than a
similar serial plan.
In order for the Query Optimizer to consider parallel plans, the SQL Server installation
must have access to at least two processors or cores, or a hyper-threaded configuration.
In addition, both the affinity mask and the max degree of parallelism advanced
configuration options must allow the use of at least two processors.
The affinity mask configuration option specifies which processors are eligible to run SQL
Server threads, and the default value of 0 means that all the processors can be used. The
max degree of parallelism configuration option is used to limit the number of processors
that can be used in parallel plans, and its default value of 0 similarly allows all available
processors to be used. As you can see if you have the proper hardware, SQL Server allows
parallel plans by default, with no additional configuration.

76
Parallelism will be considered when the estimated cost of a serial plan is higher than the
value defined in the cost threshold for the parallelism configuration parameter. However,
this doesn't guarantee that parallelism will actually be employed in the final execution
plan, as the final decision to parallelize a query (or not) will be based on cost reasons. That
is, there is no guarantee that the best parallel plan found will have a lower cost than the
best serial plan, so the serial plan may still end up being the better plan. Parallelism is
implemented by the parallelism physical operator, also known as the exchange operator,
which implements the Distribute Streams, Gather Streams, and Repartition Streams
logical operations.
The following query, which lists the names and cities of all the individual customers
ordered by CustomerID, will produce a parallel plan. Since this plan is too big to print in
this book, only a section is displayed in Figure 2-31.
SELECT I.CustomerID, C.FirstName, C.LastName, A.City
JOIN Sales.CustomerAddress AS CA
ON CA.CustomerID = I.CustomerID
JOIN Person.Address AS A
ON A.AddressID = CA.AddressID
ORDER BY I.CustomerID
Listing 2-33.
Figure 2-31: Part of a parallel plan.

77
One benefit of the graphical plans, compared to text and XML plans, is that you can easily
see which operators are being executed in parallel by looking at the parallelism symbol
(a small yellow circle with arrows) included in the operator icon. In this case, it's shown
in Figure 2-31 for the Sort and Hash Join operators.
To see why a parallel plan was considered and selected, you can look at the cost of the
serial plan. One way to do this is by using the MAXDOP hint to force a serial plan, as in the
following query:
SELECT I.CustomerID, C.FirstName, C.LastName, A.City
JOIN Sales.CustomerAddress AS CA
ON CA.CustomerID = I.CustomerID
JOIN Person.Address AS A
ON A.AddressID = CA.AddressID
OPTION (MAXDOP 1)
Listing 2-34.
The forced serial plan has a cost of 5.31282 and, given that the default cost threshold for
parallelism configuration option is 5, this clearly crosses that threshold. An interesting
test you can perform in your own test environment is to change the cost threshold
for parallelism option to 6 by running the following statements:
sp_configure 'cost threshold for parallelism', 6
GO
RECONFIGURE
GO
Listing 2-35.

78
And if you run the same query again, this time without the MAXDOP hint, you will get a
serial plan with the cost of 5.31282. Since the cost threshold for parallelism is now 6, the
Query Optimizer did not even try to find a parallel plan. Do not forget to change the cost
threshold for parallelism configuration option back to the default value of 5 by running
the following statement:
sp_configure 'cost threshold for parallelism', 5
GO
RECONFIGURE
GO
Listing 2-36.
Summary
This chapter described the Execution Engine as a collection of physical operators, which
also defines the choices that are available for the Query Optimizer to build execution
plans with. Some of most commonly used operators of the Execution Engine were
introduced, including their algorithms, relative costs, and the scenarios when the Query
Optimizer is more likely to choose them. In particular, we've looked at operators for data
access, aggregations, joins and parallelism operations.
The concepts of sorting and hashing were also introduced as a mechanism used by
the Execution Engine to match and process data. Data access operations included the
scan of tables and indexes, index seeks and bookmark lookup operations. Aggregation
algorithms like Stream Aggregate and Hash Aggregate were discussed, along with
join algorithms like the Nested Loops Join, Merge Join and Hash Join. An introduction
to parallelism was also presented.

79
Understanding how these operators function, as well as what they are likely to cost, will
give you a much stronger sense of what's actually happening under the hood when you
investigate how your queries are being implemented. This, in turn, will help you to find
potential problems in your execution plans, and to know when to resort to any of the
techniques which I'll describe later in the book.

80
Chapter 3: Statistics and Cost
Estimation
The SQL Server Query Optimizer is a cost-based optimizer, and therefore the quality
of the execution plans it generates is directly related to the accuracy of its cost
estimations. In the same way, the estimated cost of a plan is based on the algorithms
or operators used, and their cardinality estimations. So, to correctly estimate the cost of
an execution plan, the Query Optimizer needs to estimate, as precisely as possible, the
number of records returned by a given query.
During query optimization, SQL Server explores many candidate plans, estimates their
relative costs and selects the most efficient one. As such, incorrect cardinality and cost
estimation may cause the Query Optimizer to choose inefficient plans which can have a
negative impact on the performance of your database.
In this chapter, I'll discuss the statistics used by the Query Optimizer. Statistics contain
three major pieces of information: the histogram, the density information, and the string
statistics, all of which help with different parts of the cardinality estimation process. I
will show you how statistics are created and maintained, and how they are used by the
Query Optimizer. I will also provide you with information on how to detect cardinality
estimation errors that can negatively impact the quality of your execution plans, as well as
recommendations on how to fix them. The chapter ends with an overview of the costing
module, which estimates the I/O and CPU cost for each operator, to finally obtain the
total cost of the plan.

81
Chapter 3: Statistics and Cost Estimation
Statistics
SQL Server creates and maintains statistics to help the Query Optimizer with
cardinality estimation. A cardinality estimate is the estimated number of records
that will be returned by filtering, JOIN predicates or GROUP BY operations. Selectivity is
a concept similar to cardinality estimation, which can be described as the percentage of
rows from an input that satisfy a predicate. A highly selective predicate returns a small
number of rows. Rather than say any more on the subject here, we'll dive into more detail
about these concepts later in this chapter.
Creating and updating statistics
To get started, let's take a look at the various ways statistics can be created and updated.
Statistics are created in several ways: automatically by the Query Optimizer (if the default
option to automatically create statistics, AUTO_CREATE_STATISTICS, is on); when an
index is created; or when they are explicitly created, for example, by using the CREATE
STATISTICS statement. Statistics can be created on one or more columns, and both
the index and explicit creation methods support single- and multi-column statistics.
However, the statistics which are automatically generated by the Query Optimizer are
always single-column statistics. As I've already mentioned briefly, the most important
components of statistics objects are the histogram, the density information, and the
string statistics. Both histograms and string statistics are created only for the first
column of a statistics object, the latter only if the column is of a string data type.
Density information (which I'll discuss in plenty of detail later in this chapter) is
calculated for each set of columns forming a prefix in the statistics object. Filtered
statistics, on the other hand, are not created automatically by the Query Optimizer,
but only when a filtered index is created, or by issuing a CREATE STATISTICS statement
with a WHERE clause. Both filtered indexes and statistics are a new feature introduced in
SQL Server 2008, which we will touch upon later.

82
With the default configuration (if AUTO_UPDATE_STATISTICS is on), the Query
Optimizer automatically updates statistics when they are out of date. As noted, the
Query Optimizer does not automatically create multi-column or filtered statistics,
but once they are created, by using any of the methods described earlier, they can be
automatically updated. Alternatively, index rebuild operations and statements like
UPDATE STATISTICS can also be used to update statistics. Both the auto-create and
auto-update default choices will give you good quality statistics most of the time, and
you naturally have the choice to change this configuration, or use some other statements,
if you need more control over the quality of the statistics.
So, statistics may be automatically created (if non-existent) and updated (if out of date)
as necessary during query optimization. If an execution plan for a specific query already
exists in the plan cache and the statistics used by the plan are out of date, then the plan
is discarded, the statistics are updated, and a new plan is created. In a similar way,
updating statistics, either manually or automatically, invalidates any existing execution
plan that used those statistics, and will cause a new optimization the next time the
query is executed.
When it comes to determining the quality of your statistics, a fact to consider is the size
of the sample used to calculate said statistics. The Query Optimizer always uses a sample
of the target table when it creates or updates statistics, and the minimum sample size is
8 MB, or the size of the table if it's smaller than 8 MB. The sample size will increase for
bigger tables, but it may still only be a small percentage of the table.
If needed, you can explicitly request a bigger sample or scan the entire table to have
better quality statistics. Using the CREATE STATISTICS and UPDATE STATISTICS
statements you can specify a sample size or use the WITH FULLSCAN option to scan the
entire table. Doing either of these can be of benefit, especially with data that is not
randomly distributed throughout the table. Scanning the entire table will naturally give
you the most accurate statistics possible. In fact, given that statistics are always
created alongside a new index, and given that this operation scans the entire table
anyway, index statistics are initially created with the equivalent of the WITH
FULLSCAN option. However, if the Query Optimizer needs to automatically update

83
these index statistics, it can only use a default sample, as it may take too long to scan the
entire table again.
By default, SQL Server needs to wait for the update statistics operation to
complete before optimizing and executing the query; that is, statistics are updated
synchronously. A new database configuration option introduced with SQL Server 2005,
AUTO_UPDATE_STATISTICS_ASYNC, can be used to change this default and let the
statistics be updated asynchronously. As you might have guessed, with asynchronous
statistics update, the Query Optimizer does not wait for the update statistics operation
to complete, and instead just uses the current statistics for the optimization process.
This can help in situations where applications experience timeouts caused by delays
related to the automatic update of statistics. Although the current optimization will use
the out-of-date statistics, they will be updated in the background and will be used by any
later query optimizations.
SQL Server defines when statistics are out of date by using column modification
counters or colmodctrs, which count the number of table modifications, and which are
kept for each table column. Basically, for tables bigger than 500 rows, a statistics object
is considered out of date if the colmodctr value of the leading column has changed by
more than 500 plus 20% of the number of rows in the table. The same formula is used
by filtered statistics but, since they are built only from a subset of the records of the
table, the colmodctr value is first adjusted depending on the selectivity of the filter.
Colmodctrs are usually not exposed by any SQL Server metadata although they can be
accessed by using a dedicated administrator connection and looking at the rcmodified
column of the sys.sysrscols base system table in SQL Server 2008 (same information
can be found on the sysrowset columns for SQL Server 2005).
The density information on multi-column statistics might improve the quality of execu-
tion plans in the case of correlated columns or statistical correlations between columns.
As mentioned previously, density information is kept for all the columns in a statistics
object, in the order that they appear in the statistics definition. By default, SQL Server
assumes columns are independent so, if a relationship or dependency exists between
columns, multi-column statistics can help with cardinality estimation problems in

84
queries which are using these columns. Density information will also help on filters
and GROUP BY operations, as we'll see in the density section later on. Filtered statistics,
which are also explained later in this chapter, can also be used for cardinality estimation
problems with correlated columns.
Inspecting statistics objects
Let us see an example of a statistics object and inspect the data it stores. Existing statistics
for a specific object can be displayed using the sys.stats catalog view, as used in the
following query:
SELECT * FROM sys.stats
WHERE object_id = object_id('Sales.SalesOrderDetail')
Listing 3-1.
An output similar to that in Listing 3-2 (edited to fit the page) will be shown.
object_id name stats_id
--------- --------------------------------------------------- --------
642101328 PK_SalesOrderDetail_SalesOrderID_SalesOrderDetailID 1
642101328 AK_SalesOrderDetail_rowguid 2
642101328 IX_SalesOrderDetail_ProductID 3
Listing 3-2.
One record for each statistics object is shown. You can use the DBCC SHOW_STATISTICS
statement to display the details of a statistics object by specifying the column name or the
name of the statistics object.

85
For example, run the following statement to verify that there are no statistics on the
UnitPrice column of the Sales.SalesOrderDetail table:
DBCC SHOW_STATISTICS ('Sales.SalesOrderDetail', UnitPrice)
Listing 3-3.
If no statistics exists, which is the case for a fresh installation of the AdventureWorks
database, you will receive the following error message:
Msg 2767, Level 16, State 1, Line 2
Could not locate statistics 'UnitPrice' in the system catalogs.
By then running the following query, the Query Optimizer will automatically create
statistics on the UnitPrice column, which is used in the query predicate.
SELECT * FROM Sales.SalesOrderDetail
WHERE UnitPrice = 35
Listing 3-4.
Running the previous DBCC SHOW_STATISTICS statement again will now show a
statistics object similar to the following output (displayed as text and edited to fit
the page).

86
Name Updated Rows Rows Sampled Steps
------------------------- ------------------- ------ ------------ ------
_WA_Sys_00000007_2645B050 Feb 24 2010 2:12PM 121317 110678 200
All density Average Length Columns
----------- -------------- -----------
0.003225806 8 UnitPrice
RANGE_HI_KEY RANGE_ROWS EQ_ROWS DISTINCT_RANGE_ROWS AVG_RANGE_ROWS
------------ ---------- -------- ------------------- ---------------
1.374 0 144.3928 0 1
2.29 34.27779 2779.8 0 1
2.994 429.5555 342.3352 3 1
3.975 34.27779 1 0 18.33333
3.99 34.27779 2064.53 0 1
4.611 146.0489 33.46852 3 1
Listing 3-5.
The output is separated into three result sets called the header, the density vector and
the histogram, all of which you can see above, although the header information has been
truncated to fit onto the page. Let us look at the columns of the header using the previous
statistics object example, bearing in mind that some of the columns I'll describe are not
visible in Listing 3-5.
• Name: _WA_Sys_00000007_2645B050. This is the name of the statistics object, and
will probably be different in your SQL Server instance. All automatically generated
statistics have a name that starts with _WA_Sys. The 00000007 value is the
column_id of the column which these statistics are based on, as can be seen on the
sys.columns catalog, and 2645B050 is the hexadecimal equivalent of the object_id
value of the table (which can be easily verified using the calculator program available
on Windows). Reportedly, WA stands for Washington, the state of the United States
where the SQL Server development team is located.
• Updated: Feb 24 2010 2:12PM. This is the date and time at which the statistics
object was created or last updated.

87
• Rows: 121317. This is the number of rows that existed in the table when the statistics
• Rows Sampled: 111078. This is the number of rows sampled when the statistics
• Steps: 200. This is the number of steps of the histogram, which will be explained in
the next major section.
• Density: 0.07004219. This density value is no longer used by the Query Optimizer
and it is only included for backward compatibility.
• Average key length: 8. This is the average number of bytes for the columns of the
statistics object.
• String Index. NO. This value indicates if the statistics object contains string
statistics and the only choices are YES or NO; SQL Server does not provide additional
details about the string statistics. String statistics contain the data distribution for
string columns, and can help to estimate the cardinality of queries with LIKE
conditions. As indicated before, string statistics are only created for the first column,
and only when the column is of a string data type.
• Filter Expression and Unfiltered Rows. These columns will be explained in the
filtered statistics section, later in the chapter.
Below the header you'll find the density vector, which includes a wealth of potentially
useful density information and will be explained in the next section.
Density
To better explain the density vector, run the statement in Listing 3-6 to inspect the
statistics of the existing index, IX_SalesOrderDetail_ProductID.

88
DBCC SHOW_STATISTICS ('Sales.SalesOrderDetail', IX_SalesOrderDetail_ProductID)
Listing 3-6.
This will display the following density vector, which shows the densities for the
ProductID column, as well as a combination of columns ProductID, SalesOrderID,
and then ProductID, SalesOrderID and SalesOrderDetailID.
All density Average Length Columns
------------ -------------- -------------------------------------------
0.003759399 4 ProductID
8.242868E-06 8 ProductID, SalesOrderID
8.242868E-06 12 ProductID, SalesOrderID, SalesOrderDetailID
Listing 3-7.
Density, which is defined as 1 / "number of distinct values," is listed in the All density
field, and it is calculated for each set of columns, forming a prefix for the columns in
the statistics object. For example, the statistics object in Listing 3-7 was created for the
columns ProductID, SalesOrderID and SalesOrderDetailID, and so the density
vector will show three different density values: one for ProductID, another one for
ProductID and SalesOrderID combined, and a third one for the combination of
ProductID, SalesOrderID, and SalesOrderDetailID. The names of the analyzed
columns will be displayed in the Columns field, and the Average Length column will
show the average number of bytes for each density value. In the previous example, all
the columns were defined using the int data type, so the average lengths for each of the
density values will be 4, 8 and 12 bytes. Now that we've seen how density information is
structured, let's take a look at how it's used.
Density information can be used to improve the Query Optimizer's estimates for GROUP
BY operations, and on equality predicates where a value is unknown, as in the case of
local variables. To see how this is done, let's consider, for example, the number of distinct
values for ProductID on the Sales.SalesOrderDetail table: 266. Density can be

89
calculated, as mentioned earlier, as 1 / "number of distinct values," which in this case
would be 1 / 266, which is 0.003759399 as shown on the first density value on Listing 3-7.
So, the Query Optimizer can use the density information to estimate the cardinality of
GROUP BY queries. GROUP BY queries can benefit from the estimated number of distinct
values, and this information is already available in the density value. If you have this
density information, then all you have to do is to find the estimated number of distinct
values by calculating the reciprocal of the density value. For example, to estimate the
cardinality of the following query using GROUP BY ProductID, we can calculate
the reciprocal of the ProductID density shown in Listing 3-7. In this case, we have
1 / 0.003759399, which gives us 266, which is the estimated number of rows shown on
the plan in Figure 3-1.
SELECT ProductID FROM Sales.SalesOrderDetail
GROUP BY ProductID
Listing 3-8.
Figure 3-1: Cardinality estimation example using a GROUP BY clause.
In a similar way, to test GROUP BY ProductID, SalesOrderID, we would need
1 / 8.242868E-06, which give us 121,317, which you can also verify by obtaining that
query's graphical plan.
Listing 3-9 is an example of how the density can be used to estimate the cardinality of a
query using local variables.

90
DECLARE @ProductID int
SET @ProductID = 921
SELECT ProductID FROM Sales.SalesOrderDetail
WHERE ProductID = @ProductID
Listing 3-9.
In this case, the Query Optimizer does not know the value of the @ProductID local
variable at optimization time, so it is not able to use the histogram (which we'll discuss
shortly) and will use the density information instead. The estimated number of rows is
obtained using the density multiplied by the number of records in the table which, in our
example, is 0.003759399 * 121317, or 456.079 as shown in Figure 3-2.
Figure 3-2: Cardinality estimation example using a local variable.
Actually, since the Query Optimizer does not know the value of @ProductID at
optimization time, the value of 921 in Listing 3-9 does not matter; any other value will
give exactly the same estimated number of rows and execution plan. Finally, run this
query with an inequality operator:
DECLARE @pid int = 897
WHERE ProductID < @pid
Listing 3-10.

91
Just as before, the value 897 does not matter; any other value will give you the same
estimated number of rows and execution plan. However, this time the Query Optimizer
is not able to use the density information and instead it is using the standard guess of 30%
selectivity for inequality comparisons. That means that the estimated number of rows is
always 30% of the total number of records for an inequality operator and, in this case, 30%
of 121,317 is 36,395.1, as shown in Figure 3-3.
Figure 3-3: Cardinality estimation example using a 30% guess.
However, the use of local variables in a query limits the quality of the cardinality estimate
when using the density information with equality operators. Worse, local variables result
in no estimate at all when used with an inequality operator, which results in a guessed
percentage. For this reason, local variables should be avoided in queries, and parameters
or literals should be used instead. When parameters or literals are used, the Query
Optimizer is able to use the histogram, which will provide better quality estimates than
the density information on its own.
As it happens, the last section of the DBCC SHOW_STATISTICS output is the histogram,
which I will now explain.

92
Histograms
In SQL Server, histograms are created only for the first column of a statistics object,
and they compress the information of the distribution of values in those columns by
partitioning that information into subsets called buckets or steps. The maximum number
of steps in a histogram is 200, but even if the input has 200 or more unique values, a
histogram may still have less than 200 steps. To build the histogram, SQL Server finds
the unique values in the column and tries to capture the most frequent ones using a
variation of the maxdiff algorithm, so that the most statistically significant information
is preserved. Maxdiff is one of the available histograms whose purpose is to accurately
represent the distribution of data values in relational databases.
To see how the histogram is used, run the following statement to display the
current statistics of the IX_SalesOrderDetail_ProductID index on the
Sales.SalesOrderDetail table:
DBCC SHOW_STATISTICS ('Sales.SalesOrderDetail', IX_SalesOrderDetail_ProductID)
Listing 3-11.
Both the multi-column index and statistics objects include the columns ProductID,
SalesOrderID, and SalesOrderDetailID, but since the histogram is only for the
first column, this data is only available for the ProductID column.
Next, I will show you some examples of how the histogram may be used to estimate the
cardinality of some simple predicates. Let's take a look at a section of the histogram, as
shown in the output in Listing 3-12.

93
------------ ---------- ------- ------------------- --------------
826 0 305 0 1
831 110 198 3 36.66667
832 0 256 0 1
Listing 3-12.
RANGE_HI_KEY is the upper boundary of a histogram step; the value 826 is the upper
boundary for the first step displayed, and 831 is the upper boundary for the second step
shown. This means that the second step may contain only values from 827 to 831.
With that in mind, and to better understand the rest of the histogram structure and
how the histogram information was aggregated, run the following query to obtain
the real number of records for ProductIDs 827 to 831, and we'll compare them against
the histogram.
SELECT ProductID, COUNT(*) AS Total
WHERE ProductID BETWEEN 827 AND 831
GROUP BY ProductID
Listing 3-13.
This produces the following result:
ProductID Total
--------- -----------
827 31
828 46
830 33
831 198
Listing 3-14.

94
Going back to the histogram, EQ_ROWS is the estimated number of rows whose column
value equals RANGE_HI_KEY. So, in our example, for the RANGE_HI_KEY value of 831,
EQ_ROWS shows 198, which we know is also the actual number of existing records for
ProductID 831.
RANGE_ROWS is the estimated number of rows whose column value falls inside the range
of the step, excluding the upper boundary. In our example, this is the number of records
with values from 827 to 830 (831, the upper boundary or RANGE_HI_KEY, is excluded).
The histogram shows 110 records and we could obtain the same value by getting the
sum of 31 records for ProductID 827, 46 records for ProductID 828, 0 records for
ProductID 829, and 33 records for ProductID 830.
DISTINCT_RANGE_ROWS is the estimated number of rows with a distinct column value
inside this range, once again excluding the upper boundary. In our example, we have
records for three distinct values: 827, 828, and 830, so DISTINCT_RANGE_ROWS is 3.
There are no records for ProductID 829, and 831, which is the upper boundary, is again
excluded.
Finally, AVG_RANGE_ROWS is the average number of rows per distinct value,
excluding the upper boundary, and it is simply calculated as RANGE_ROWS / DISTINCT_
RANGE_ROWS. In our example, we have a total of 110 records for 3 DISTINCT_RANGE_
ROWS, which gives us 110 / 3 = 36.6667, also shown in the second step of the histogram
shown previously.
Now let's see how the histogram is used to estimate the selectivity of some queries.
Let us see the first query:
WHERE ProductID = 831
Listing 3-15.

95
Figure 3-4: Cardinality estimation example using a RANGE_HI_KEY value.
Since 831 is the RANGE_HI_KEY on the second step of the histogram shown in Listing
3-12, the Query Optimizer will use the EQ_ROWS value (the estimated number of rows
whose column value equals RANGE_HI_KEY) directly, and the estimated number of rows
will be 198, as shown on Figure 3-4.
Now run the same query, with the value set to 828. This time, the value is inside the
range of the second step but is not a RANGE_HI_KEY, so the Query Optimizer uses
the value calculated for AVG_RANGE_ROWS (the average number of rows per distinct
value), which is 36.6667 as shown in the histogram. The plan is shown in Figure 3-5 and,
unsurprisingly, we get the same estimated number of rows for any of the other values in
the range (except for the RANGE_HI_KEY, obviously). This also includes 829, even when
there are no records for this ProductID value.
Figure 3-5: Cardinality estimation example using an AVG_RANGE_ROWS value.

96
Let's now use an inequality operator and try to find the records with a ProductID less
than 714. Since this requires all the records, both inside the range of a step and the upper
boundary, we need to calculate the sum of the values of both the RANGE_ROWS and the
EQ_ROWS columns for steps 1 through 7 as shown in the histogram below, which give us a
total of 13,223 rows.
------------ ---------- ------- ------------------- --------------
707 0 3083 0 1
708 0 3007 0 1
709 0 188 0 1
710 0 44 0 1
711 0 3090 0 1
712 0 3382 0 1
713 0 429 0 1
714 0 1218 0 1
715 0 1635 0 1
Listing 3-16.
This is the query in question, and the estimated number of rows is shown on the
execution plan in Figure 3-6.
WHERE ProductID < 714
Listing 3-17.

97
Figure 3-6: Cardinality estimation example using an inequality operator.
Let's now test a query with an AND'ed predicate (this example shows statistics as
estimated in SQL Server 2008; statistics for SQL Server 2008 R2 will have minimal
differences for a default sample).
WHERE ProductID = 870 AND OrderQty = 1
Listing 3-18.
SQL Server will use the histograms of two distinct statistics objects here, one for each
predicate clause. We can use one histogram to obtain the estimated number of records for
ProductID = 870, and the second histogram to obtain the estimated number of records
for OrderQty = 1. Requesting ProductID = 870 AND OrderQty = 1 will return
the intersection between both sets of records, so we should multiply the selectivity of both
clauses to obtain this value.
If you obtain the estimated number of rows for the predicate ProductID = 870
alone, as explained before, you will get 4,688 rows. For a table with 121,317 records, this
corresponds to a selectivity of 4,688 / 121,317, or 0.03864256. In the same way, the
estimated number of rows for the predicate OrderQty = 1 alone is 68,024 rows,
which corresponds to a selectivity of 68,024 / 121,317, or 0.56071284.

98
In order to get the intersection of these sets, we need to multiply the selectivity values of
both predicate clauses, 0.03864256 * 0.56071284 to get 0.0216673795624704. Finally, the
calculated selectivity is multiplied by the number of records to give the estimated number
of records as 0.0216673795624704 * 121,317, or 2,628.62, which is the value shown in the
graphical plan in Figure 3-7.
Figure 3-7: Cardinality estimation example using an AND'ed predicate.
It is also worth noticing that if these two columns, ProductID and OrderQty, were
correlated (which is not the case in this example), then this method to estimate the
cardinality would be incorrect. Two methods to help with correlated columns are
using multi-column statistics, as mentioned before, and filtered statistics, which will be
explained later in this chapter.
Finally, let's test the same query with an OR'ed predicate to see how the information
revealed by the histogram will be helpful.
WHERE ProductID = 870 OR OrderQty = 1
Listing 3-19.
By definition, an OR'ed predicate is the union of the sets of rows of both clauses, without
duplicates. That is, this should be the rows estimated for ProductID = 870 plus the
rows estimated for OrderQty = 1, but if there are any rows that may belong to both

99
sets, then they should be included only once. As indicated in the previous example, the
estimated number of rows for the predicate ProductID = 870 alone, is 4,688 rows, and
the estimated number of rows for the predicate OrderQty = 1 alone is 68,024 rows.
The estimated number of records that belong to both sets is the AND'ed predicate we saw
previously: 2,628.62 rows. So, the estimated number of rows for the OR'ed predicate is
4,688 + 68,024 – 2,628.62, or 70083.4, as shown in the execution plan in Figure 3-8.
Figure 3-8: Cardinality estimation example using an OR'ed predicate.
Statistics Maintenance
As mentioned already, the Query Optimizer will, by default, automatically update
statistics when they are out of date. Statistics can also be updated with the UPDATE
STATISTICS statement which you can schedule to run as a job during your database
maintenance window. Another statement commonly used, sp_updatestats, also runs
UPDATE STATISTICS behind the scenes.
There are two important benefits of updating statistics in a maintenance job. The first
is that your queries will use updated statistics without having to wait for the automatic
update of statistics to be completed, avoiding delays in the optimization of your queries
(although asynchronous statistics updates can also be used to partially help with this
problem). The second benefit is that you can use a bigger sample than the Query
Optimizer will use, or you can even scan the entire table. This can give you better

100
quality statistics for big tables, especially for those where data is not randomly distributed
in their data pages. Manually updating statistics can also be a benefit after operations
such as data loads, that update large amounts of data, are performed.
On the other hand, also note that the update of statistics will cause a recompiling of plans
already in the plan cache which are using these statistics, so you may not want to do this
too often, either.
An additional consideration for manually updating statistics in a maintenance job is how
they relate to index rebuild maintenance jobs, which also update the index statistics. Keep
the following items in mind when combining maintenance jobs for both indexes and
statistics, remembering that there are both index and non-index column statistics, and
that index operations obviously may impact only the first of these.
• Rebuilding an index, for example by using the ALTER INDEX … REBUILD statement,
will also update index statistics by scanning all the rows in the table, which is the
equivalent of using UPDATE STATISTICS WITH FULLSCAN. Rebuilding indexes does
not update any column statistics.
• Reorganizing an index, for example using the ALTER INDEX … REORGANIZE
statement, does not update any statistics, not even index statistics.
• By default, the UPDATE STATISTICS statement updates both index and column
statistics. Using the INDEX option will update index statistics only, and using the
COLUMNS option will update non-indexed column statistics only.
So, depending on your maintenance jobs and scripts, several scenarios can exist. The
simplest maintenance plan is if you want to rebuild all the indexes and update all the
statistics. As mentioned before, if you rebuild all your indexes, then all the index statistics
will also be automatically updated by scanning all the rows on the table. Then you just
need to update your non-indexed column statistics by running UPDATE STATISTICS
WITH FULLSCAN, COLUMNS. Since one job updates only index statistics, and the second
one updates only column statistics, it does not matter which one is executed first.

101
Of course, more complicated maintenance plans can exist, for example, when indexes
are rebuilt or reorganized depending on their fragmentation level. A good starting point
to do this is to use the avg_fragmentation_in_percent column and the index
fragmentation thresholds as defined on the Books Online entry for the sys.dm_db_
index_physical_stats dynamic management function. You should keep in mind
the items mentioned above, so that you can avoid problems like updating the index
statistics twice, as could occur when both index rebuild and update statistics operations
are performed. You could also avoid discarding work previously performed, for example,
when you rebuild the indexes of a table (which also updates statistics by scanning
the entire table), and later running a job updating the statistics with a default or
smaller sample.
Let me show you how these commands work, with some examples. Create a new table
dbo.SalesOrderDetail:
SELECT * INTO dbo.SalesOrderDetail
FROM sales.SalesOrderDetail
Listing 3-20.
The next query uses the sys.stats catalog view to show that there are no statistics
objects for the new table:
SELECT name, auto_created, stats_date(object_id, stats_id) AS update_date
FROM sys.stats
WHERE object_id = object_id('dbo.SalesOrderDetail')
Listing 3-21.

102
Now run the following query:
SELECT * FROM dbo.SalesOrderDetail
WHERE SalesOrderID = 43670 AND OrderQty = 1
Listing 3-22.
Use the previous sys.stats query from Listing 3-21 to verify that two statistics
objects were created, one for the SalesOrderID column, and a second for the OrderQty
column. Now create the following index, and run the sys.stats query again to verify
that a new statistics object for the ProductID column has been created.
CREATE INDEX IX_ProductID ON dbo.SalesOrderDetail(ProductID)
Listing 3-23.
This will be the output of the sys.stats query so far:
name auto_created update_date
-------------------------- ------------ -----------------------
_WA_Sys_00000004_76EBA2E9 1 2010-03-01 14:17:44.610
_WA_Sys_00000001_76EBA2E9 1 2010-03-01 14:17:44.770
IX_ProductID 0 2010-03-01 14:19:00.607
Listing 3-24.
Notice how the value of the auto_created column, which indicates if the statistics were
created by the Query Optimizer, is 0 for the IX_ProductID statistics object. Run the
next command to update just the column statistics:
UPDATE STATISTICS dbo.SalesOrderDetail WITH FULLSCAN, COLUMNS
Listing 3-25.

103
You can validate that only the column statistics were updated, by comparing the
update_date column with the previous output. The update_date column uses the
STATS_DATE function to display the last point in time when the statistics were updated,
as is shown on the following output:
name auto_created update_date
-------------------------- ------------ -----------------------
_WA_Sys_00000004_76EBA2E9 1 2010-03-01 14:21:25.850
_WA_Sys_00000001_76EBA2E9 1 2010-03-01 14:21:25.940
IX_ProductID 0 2010-03-01 14:19:00.607
Listing 3-26.
This command will do the same for just the index statistics:
UPDATE STATISTICS dbo.SalesOrderDetail WITH FULLSCAN, INDEX
Listing 3-27.
… and these commands will update both the index and column statistics:
UPDATE STATISTICS dbo.SalesOrderDetail WITH FULLSCAN
UPDATE STATISTICS dbo.SalesOrderDetail WITH FULLSCAN, ALL
Listing 3-28.
As mentioned earlier, if you run the sys.stats query after each of the next two queries,
you'll see how an ALTER INDEX REBUILD statement only updates index statistics:
ALTER INDEX ix_ProductID ON dbo.SalesOrderDetail REBUILD
Listing 3-29.

104
… and you can verify that reorganizing an index does not update any statistics:
ALTER INDEX ix_ProductID on dbo.SalesOrderDetail REORGANIZE
Listing 3-30.
Finally, for good house-keeping, remove the table you have just created:
DROP TABLE dbo.SalesOrderDetail
Listing 3-31.
Statistics on Computed Columns
Another interesting step performed during query optimization is the automatic
matching of computed columns. Although computed columns have been available in
previous versions of SQL Server, the automatic matching feature was only introduced
with SQL Server 2005. In this section, I will show you how this feature works, and explain
how computed columns can help to improve the performance of your queries.
A problem faced by some queries using scalar expressions is that they usually cannot
benefit from column statistics and, without statistics, the Query Optimizer will use
the 30% selectivity guess on inequality comparisons, which may produce inefficient
execution plans. A solution to this problem is the use of computed columns, as SQL
Server can automatically create and update statistics on these columns. The great
benefit of this solution is that you don't need to specify the name of the computed
column in your queries for SQL Server to use its statistics. The Query Optimizer
automatically matches the computed column definition to an existing scalar
expression in a query, so your applications do not need to be changed.

105
To see an example, run this query, which creates the plan shown in Figure 3-9:
WHERE OrderQty * UnitPrice > 10000
Listing 3-32.
Figure 3-9: Cardinality estimation example using a 30% guess.
The estimated number of rows is 36,395.1, which is 30% of the total number of rows,
121,317, although the query returns only 772 records. SQL Server is obviously using a
selectivity guess, as it cannot estimate the selectivity of the expression
OrderQty * UnitPrice > 10000.
Now create a computed column:
ALTER TABLE Sales.SalesOrderDetail
ADD cc AS OrderQty * UnitPrice
Listing 3-33.
Run the previous query in Listing 3-32 again, and note that, this time, the estimated
number of rows has changed and is close to the actual number of rows returned by the
query, as shown in Figure 3-10 (this plan shows the estimated number of rows as in SQL
Server 2008; statistics for SQL Server 2008 R2 will have minimal differences for a default

106
sample). You can optionally test replacing the 10,000 with some other values, like 10, 100,
1,000, or 5,000, and compare the actual and the estimated number of rows returned.
Figure 3-10: Cardinality estimation example using computed columns.
Note that creating the computed column does not create statistics; these statistics are
created the first time that the query is optimized, and you can run the next query to
display the information about the statistics objects for the Sales.SalesOrderDetail
table:
Listing 3-34.
The newly created statistics object will most likely be at the end of the list. Copy the name
of the object, and use the following command to display the details about the statistics
object (I've used the name of my local object, but you should replace that as appropriate).
You can also use "cc" as the name of the object to get the same results. The "cc" column
should be shown on the Columns field in the density section.
DBCC SHOW_STATISTICS ('Sales.SalesOrderDetail', _WA_Sys_00000013_2645B050)
Listing 3-35.

107
Unfortunately, for automatic matching to work, the expression must be exactly the same
as the computed column definition. So, if I change the query to UnitPrice * OrderQty,
instead of OrderQty * UnitPrice, the execution plan will show an estimated number of
rows of 30% again, as this query will demonstrate:
WHERE UnitPrice * OrderQty > 10000
Listing 3-36.
Finally, drop the created computed column:
ALTER TABLE Sales.SalesOrderDetail
DROP COLUMN cc
Listing 3-37.
Filtered Statistics
Filtered statistics are statistics created on a subset of records in a table. Filtered statistics
are automatically created when filtered indexes are created, but they can also be created
manually by specifying a WHERE clause on the CREATE STATISTICS statement. As you
might imagine, filtered statistics can help on queries accessing specific subsets of data.
They can also be useful in situations like correlated columns, especially when one of
these columns has a small number of unique values, and you can create multiple filtered
statistics for each one of these distinct values. As shown in the histogram section previ-
ously, when using multiple predicates, SQL Server assumes that each clause in a query is
independent and, if the columns used in this query were correlated, then the cardinality
estimation would be incorrect. Filtered statistics may also help on huge tables where a
large number of unique values are not accurately represented in the 200-step limitation
currently enforced on histograms.

108
Next, I will show you how you can use filtered statistics to help in a problem with
correlated columns. Running the following query will correctly estimate the number of
rows to be 93:
WHERE City = 'Los Angeles'
Listing 3-38.
In the same way, running the next query will correctly estimate 4,564 rows:
Listing 3-39.
However, since StateProvinceID 9 corresponds to the state of California (which you
can verify by looking at the Person.StateProvince table) it is possible for somebody to
run this query, which in this case will show a less precise estimate of 21.6403, as shown in
WHERE City = 'Los Angeles' AND StateProvinceID = 9
Listing 3-40.

109
Figure 3-11: Cardinality estimate with the independence assumption.
Because of the assumption of independence, SQL Server will multiply the cardinality of
both predicates, which was explained earlier in this chapter. The calculation, abbreviated
as (93 * 4,564) / 19,614, will give us the value 21.6403 shown in the previous plan (19,614 is
the total number of rows in the table).
However, the assumption of independence is incorrect in this example, as the columns
are statistically correlated. To help with this problem, you can create a filtered statistics
object for the state of California, as shown in the next statement.
CREATE STATISTICS california
ON Person.Address(City)
Listing 3-41.
Clearing the cache and running the previous query again will now give a better estimate,
as shown on the following plan:
DBCC FREEPROCCACHE
GO
WHERE City = 'Los Angeles' AND StateProvinceID = 9
Listing 3-42.

110
Figure 3-12: Cardinality estimate with filtered statistics.
Let us now inspect the filtered statistics object by running the following statement:
DBCC SHOW_STATISTICS('Person.Address', california)
WITH STAT_HEADER
Listing 3-43.
This will show the following output, (edited here to fit the page):
Name Rows Rows Sampled Filter Expression Unfiltered Rows
----------- ----- ------------- ------------------------ ---------------
california 4564 4564 ([StateProvinceID]=(9)) 19614
Listing 3-44.
Notice that the filter definition is shown on the Filter Expression field, and that the
Unfiltered Rows field shows the total number of records on the table when the filtered
statistics were created. Also note that, this time, the Rows column number is less than the
total number of rows in the table, and corresponds to the number of records that satisfied
the filter predicate when the statistics object was created. The filter definition can also be
seen on the filter_definition column of the sys.stats catalog view.

111
Finally, drop the statistics object you have just created, by running the
following statement:
DROP STATISTICS Person.Address.california
Listing 3-45.
Cardinality Estimation Errors
Cardinality estimation errors can lead to the Query Optimizer making poor choices
as to how best to execute a query and, therefore, to badly performing execution plans.
Fortunately, you can easily check if you have cardinality estimation errors by comparing
the estimated against the actual number of rows, as shown on graphical or XML
execution plans, or by using the SET STATISTICS PROFILE statement. In the next query,
I'll show you how to use the SET STATISTICS PROFILE statement with one of our
previous examples where SQL Server is making a blind guess regarding the selectivity of
certain columns:
SET STATISTICS PROFILE ON
GO
WHERE OrderQty * UnitPrice > 10000
GO
SET STATISTICS PROFILE OFF
GO
Listing 3-46.

112
This is the resulting output, with the EstimateRows column manually moved just after
the Rows column, and edited to fit the page:
Rows EstimateRows Executes StmtText
------ ------------ -------- ----------------------------------------
772 36395.1 1 SELECT * FROM [Sales].[SalesOrderDetail]
772 36395.1 1 |--Filter(WHERE:([AdventureWorks] .[Sa
0 121317 0 |--Compute Scalar(DEFINE:([Advent
0 121317 0 |--Compute Scalar(DEFINE:([A
121317 121317 1 |--Clustered Index Scan
Listing 3-47.
Using this output, you can easily compare the actual number of rows, shown on the Rows
column, against the estimated number of records, as shown on the EstimateRows
column, for each operator in the plan.
Because each operator relies on previous operations for its input, cardinality estimation
errors can propagate exponentially throughout the query plan. For example, a cardinality
estimation error on a Filter operator can impact the cardinality estimation of all the other
operators in the plan that consume the data produced by that operator. If your query is
not performing well and you find cardinality estimation errors, check for problems like
missing or out-of-date statistics, very small samples being used, correlation between
columns, use of scalar expressions, guessing selectivity issues, and so on.
Recommendations to help with these issues have been provided throughout this
chapter and include things like using the auto-create and auto-update statistics default
configurations, updating statistics using WITH FULLSCAN, avoiding local variables in
queries, using computed columns, and considering multi-column or filtered statistics,
among other things. That's a fairly long list, but it should help convince you that you are
already armed with pragmatically useful information.

113
Some SQL Server features, such as table variables, do not use statistics, so you might
want to consider using a similar feature like temporary tables if you're having perform-
ance problems related to cardinality estimation errors. In addition, for complex queries
that are not performing well because of cardinality estimation errors, you may want to
consider partitioning the query into several steps while storing the intermediate results
in temporary tables. This will allow SQL Server to create statistics on the intermediate
results, which will help the Query Optimizer to produce a better execution plan.
UPDATE STATISTICS with ROWCOUNT,
PAGECOUNT
In this section I will show you the undocumented ROWCOUNT and PAGECOUNT options of
the UPDATE STATISTICS statement, which can help you in cases where you want to see
which execution plans would be generated for huge tables (with millions of records), but
then test those plans in small, or even empty, tables. As you can imagine, these options
can be helpful for testing in some scenarios where you may not want to spent time or disk
space creating big tables.
By using this method you are essentially tricking the Query Optimizer, as it will generate
execution plans using cardinality estimations which are made as if the table really had
millions of records, even if your table is actually tiny. Note that this option, available
since SQL Server 2005, only helps in creating the execution plan for your queries. Actually
running the query will use the real data in your test table which will, of course, execute
faster than a table with millions of records.
Using these UPDATE STATISTICS options does not change the table statistics, only the
counters for the numbers of rows and pages of a table and, as I will show shortly, the
Query Optimizer uses this information to estimate the cardinality of queries. Finally,
before we look at examples, keep in mind that these are undocumented and unsupported
options, and should not be used in a production environment.

114
So, let's see an example. Run the following query to create a new table on the Adven-
tureWorks database:
SELECT * INTO dbo.Address
FROM Person.Address
Listing 3-48.
Inspect the number of rows by running the following queries; they should show
19,614 rows:
SELECT * FROM sys.partitions
WHERE object_id = object_id('dbo.Address')
Listing 3-49.
SELECT * FROM sys.dm_db_partition_stats
Listing 3-50.
Now run the following query, and inspect the graphical execution plan:
SELECT * FROM dbo.Address
WHERE City = 'London'
Listing 3-51.
Running this query will create a new statistics object for the City column, and will show
the plan in Figure 3-13. Note that the estimated number of rows is 434, and it's using a
simple Table Scan operator.

115
Figure 3-13: Cardinality estimation example using a small table.
We can discover where the Query Optimizer is getting the estimated number of rows by
inspecting the statistics object. Run this query to see the name of the statistics object ...
Listing 3-52.
... and then use the displayed statistics object name in the following statement (the name
may be different in your case).
DBCC SHOW_STATISTICS ('dbo.Address', _WA_Sys_00000004_46136164)
Listing 3-53.
A fragment of the histogram is shown next.
------------- ---------- ------- ------------------- --------------
Lincoln Acres 0 102 0 1
London 32 434 2 16
Long Beach 0 97 0 1
Los Angeles 2 93 2 1
Listing 3-54.

116
By looking at the histogram, you can find the value 434 on EQ_ROWS for the RANGE_HI_
KEY value "London."
Now run the following UPDATE STATISTICS WITH ROWCOUNT, PAGECOUNT statement
(you can specify any other value for ROWCOUNT and PAGECOUNT):
UPDATE STATISTICS dbo.Address WITH ROWCOUNT = 1000000, PAGECOUNT = 100000
Listing 3-55.
If you inspect the number of rows from sys.partitions or sys.dm_db_parti-
tion_stats again, as shown previously, it will now show 1,000,000 rows (sys.dm_db_
partition_stats also shows the new number of pages). Clear the plan cache and run
the query again.
DBCC FREEPROCCACHE
GO
SELECT * FROM dbo.Address
WHERE City = 'London'
Listing 3-56.
Note that the estimated number of rows has changed from 434 to 22,127.1, as shown on
Figure 3-14, and that a different plan was generated using this new cardinality estimation;
this time, the Query Optimizer decided to create a parallel plan.

117
Figure 3-14: Cardinality estimation using ROWCOUNT and PAGECOUNT.
However, if you look at the statistics object again, using DBCC SHOW_STATISTICS as
shown before, you'll see that the histogram has not changed. One way to obtain the
estimated number of rows shown in the new execution plan is by calculating the
percentage (or fraction) of rows for the value "London" from the statistics sample
which, in this case, is 19,614, as will be shown in the header of the statistics object
referred to in Listing 3-54. So the fraction is 434 / 19,614, or 0.022127052. Next,
we apply the same percentage to the new "current" number of rows, which results
in 1,000,000 * 0.022127052, and we get 22,127.1, which is the estimated number of
rows displayed in the plan in Figure 3-14.
Finally, drop the table you just created:
DROP TABLE dbo.Address
Listing 3-57.

118
Cost Estimation
As we've established, the quality of the execution plans the Query Optimizer generates is
directly related to the accuracy of its costing estimates. Even when the Query Optimizer
is able to enumerate low cost plans, an incorrect cost estimation may result in the Query
Optimizer choosing inefficient plans, which can negatively impact the performance of
your database. During query optimization, the Query Optimizer explores many candidate
plans, estimates their cost, and then selects the most efficient one. So, in addition to being
accurate, cost estimation must also be efficient, since it is used multiple times during the
query optimization process.
Costs are estimated for any partial or complete plan; cost computation is done per
operator, and the total plan cost is the sum of the costs of all the operators in that plan.
The cost of each operator depends on its algorithm and the estimated number of records
it returns, and some operators, such as Sort or Hash Join, also consider the available
memory in the system. A high level overview of the cost of the algorithms for some of
the most used operators was included in Chapter 2, The Execution Engine.
So, each operator has an associated CPU cost, and some of them will also have some I/O
cost, and the cost of the operator as a whole is the sum of these costs. An operator like
a Clustered Index Scan has both CPU and I/O costs, whereas some other operators, like
Stream Aggregate, will only have a CPU cost. Since Microsoft does not publish how
these costs are calculated, I will show you a very basic example of how the cost of a plan
is estimated.
To show this in an example, let's look at the largest table in the AdventureWorks
database. Run the following query and look at the estimated CPU and I/O costs for the
Clustered Index Scan operator, as shown in Figure 3-15:
WHERE LineTotal = 35
Listing 3-58.

119
Figure 3-15: Clustered Index Scan operator properties.
Note that, in an older version of SQL Server, the cost used to mean the estimated time
in seconds that a query would take to execute on a specific hardware configuration, but
currently this value is meaningless as an objective unit of measurement, and should not
be interpreted as one.
For a Clustered Index Scan operator, I observed that the CPU cost is 0.0001581 for the
first record, plus 0.0000011 for any additional record after that. Because, in this specific
case, we have an estimated 121,317 records, we can calculate 0.0001581 + 0.0000011 *
(121317 – 1), which comes to 0.133606, which is the value shown as Estimated CPU Cost.
In a similar way, I noticed that the minimum I/O cost is 0.003125 for the first database
page, and then it grows in increments of 0.00074074 for every additional page. Since this
operator scans the entire table, I can use the query in Listing 3-59 to find the number of
database pages (which turns out to be 1,234).

120
SELECT in_row_data_page_count, row_count
FROM sys.dm_db_partition_stats
AND index_id = 1
Listing 3-59.
In this case I have 0.003125 + 0.00074074 * (1234 – 1), which comes to 0.916458, which is
the value shown as Estimated I/O Cost.
Finally, we add both costs, 0.133606 + 0.916458, to get 1.05006, which is the total
estimated cost of the operator. In the same way, adding the cost of all the operators
will give the total cost of the plan. In this case, the cost of the Clustered Index Scan
(1.05006) plus the cost of the first Compute Scalar operator (0.01214), the second
Compute Scalar operator (0.01213), and the cost of the Filter operator (0.0582322), will
give the total cost of the plan: 1.13256, as shown on Figure 3-16.
Figure 3-16: Execution plan displaying total cost.

121
Summary
In this chapter, we have seen how statistics are used by SQL Server to estimate the
cardinality as well as the cost of operators and execution plans. The most important
elements of a statistics object, namely the histogram, the density information, and string
statistics, were introduced and explained. Examples of how to use histograms were
shown, including queries with equality and inequality operators and both AND'ed and
OR'ed predicates. The use of density information was shown in GROUP BY operations, and
in cases when the Query Optimizer is not able to use a histogram, such as in the case of
local variables.
Maintenance of statistics was also explained, with some emphasis on how to proactively
update statistics to avoid delays during query optimization, and how to improve the
quality of statistics by scanning the entire table instead of a default sample. We also
discussed how to detect cardinality estimation errors, which can negatively impact the
quality of your execution plans, and we looked at recommendations on how to fix them.

122
Chapter 4: Index Selection
Index selection is one of the most important techniques used in query optimization. By
using the right indexes, SQL Server can speed up your queries and dramatically improve
the performance of your applications. In this chapter, I will show you how SQL Server
selects indexes, how you can use this knowledge to provide better indexes, and how you
can verify your execution plans to make sure these indexes are correctly used.
This chapter also includes sections about the Database Engine Tuning Advisor and the
Missing Indexes feature, which will show how you can use the Query Optimizer itself to
provide index tuning recommendations. However, it is important to emphasize that, no
matter what index recommendations these tools give, it is ultimately up to the database
administrator or developer to do their own index analysis, and finally decide which of
these recommendations to implement. Also, since we'll be covering these tools mainly
from the point of view of the Query Optimizer, you should use Books Online to obtain
more in-depth information regarding using these features.
Finally, the sys.dm_db_index_usage_stats DMV will be introduced as a tool to
identify existing indexes which your queries may not be using. Indexes that are not being
used will provide no benefit to your databases, but will use valuable disk space and slow
your update operations, and so they should be considered for removal.
Introduction
As mentioned in Chapter 2, The Execution Engine, SQL Server can use indexes to perform
seek and scan operations. Indexes can be used to speed up the execution of a query by
quickly finding records without performing table scans; by delivering all the columns
requested by the query without accessing the base table (i.e. covering the query, which

123
I'll return to in a moment), or by providing sorted order, like in queries with GROUP BY,
DISTINCT or ORDER BY clauses.
Part of the Query Optimizer's job is to determine if an index can be used to evaluate a
predicate in a query. This is basically a comparison between an index key and a constant
or variable. In addition, the Query Optimizer needs to determine if the index covers the
query; that is, if the index contains all the columns required by the query (referred to
as a "covering index"). It needs to confirm this because, as you'll hopefully remember, a
non-clustered index usually contains only a subset of the columns of the table.
SQL Server can also consider using more than one index, and joining them to cover
all the columns required by the query (index intersection). If it's not possible to cover all
of the columns required by the query, it may need to access the base table, which could be
a clustered index or a heap, to obtain the remaining columns. This is called a bookmark
lookup operation (which could be a Key Lookup or a RID Lookup, as explained in Chapter
2). However, since a bookmark lookup requires random I/O, which is a very expensive
operation, its usage can be effective only for a relatively small number of records.
Also keep in mind that, although one or more indexes can be used, it does not mean
that they will finally be selected in an execution plan, as this is always a cost-based
decision. So, after creating an index, make sure you verify that the index is, in fact, used
in a plan (and, of course, that your query is performing better, which is probably the
primary reason why you are defining an index). An index that it is not being used by any
query will just take up valuable disk space, and may negatively impact the performance of
update operations without providing any benefit. It is also possible that an index, which
was useful when it was originally created, is no longer used by any query. This could be
as a result of changes in the database, the data, or even the query itself. To help you avoid
this frustrating situation, the last section in this chapter will show you how you can
identify which indexes are no longer being used by any of your queries.

124
The Mechanics of Index Selection
In a seek operation, SQL Server navigates throughout the B-tree index to quickly find the
required records without the need for an index or table scan. This is similar to using an
index at the end of a book to find a topic quickly, instead of reading the entire book. Once
the first record has been found, SQL Server can then scan the index leaf level forward or
backward to find additional records. Both equality and inequality operators can be used
in a predicate, including =, <, >, <=, >=, <>, !=, !<, !>, BETWEEN, and IN. For example, the
following predicates can be matched to an Index Seek operation if there is an index on
the specified column, or a multi-column index with that column as a leading index key:
• ProductID = 771
• UnitPrice < 3.975
• LastName = 'Allen'
• LastName LIKE 'Brown%'
As an example, look at the next query, which uses an Index Seek operator and produces
SELECT ProductID, SalesOrderID, SalesOrderDetailID
Listing 4-1.
Figure 4-1: Plan with Index Seek.

125
Figure 4-2: Index Seek operator properties.
The SalesOrderDetail table has a multi-column index with ProductID as the leading
column. The Index Seek operator properties, which you can see in Figure 4-2, include the
following Seek predicate on the ProductID column, which shows that SQL Server was
effectively able to use the index to seek on this column.
Seek Keys[1]: Prefix: [AdventureWorks].[Sales]. [SalesOrderDetail].ProductID =
Scalar Operator (CONVERT_IMPLICIT(int,[@1],0))
Listing 4-2.

126
An index cannot be used to seek on some complex expressions, expressions using
functions, or strings with a leading wildcard character, as in the following predicates:
• ABS(ProductID) = 771
• UnitPrice + 1 < 3.975
• LastName LIKE '%Allen'
• UPPER(LastName) = 'Allen'
Compare the following query to the previous example; by adding an ABS function to
the predicate, SQL Server is no longer able to use an Index Seek operator and chooses,
instead, to do an Index Scan as shown on the plan in Figure 4-3.
WHERE ABS(ProductID) = 771
Listing 4-3.
Figure 4-3: Plan with an Index Scan.

127
Figure 4-4: Index Scan operator properties.
Note that, in Figure 4-4, the following predicate is, however, still evaluated on the Index
Scan operator.
abs([AdventureWorks].[Sales].[SalesOrderDetail].
[ProductID]) =CONVERT_IMPLICIT(int,[@1],0)
Listing 4-4.
In the case of a multi-column index, SQL Server can only use the index to seek on the
second column if there is an equality predicate on the first column. So SQL Server can use

128
a multi-column index to seek on both columns in the following cases, supposing that a
multi-column index exists on both columns in the order presented:
• ProductID = 771 AND SalesOrderID > 34000
• LastName = 'Smith' AND FirstName = 'Ian'
That being said, if there is no equality predicate on the first column, or if the predicate
can not be evaluated on the second column, as is the case in a complex expression, then
SQL Server may still only be able to use a multi-column index to seek on just the first
column, as in the following examples:
• ProductID = 771 AND ABS(SalesOrderID) = 34000
• ProductID < 771 AND SalesOrderID = 34000
• LastName > 'Smith' AND FirstName = 'Ian'
However, SQL Server is not able to use a multi-column index for an Index Seek in the
following examples, as it is not even able to search on the first column:
• ABS(ProductID) = 771 AND SalesOrderID = 34000
• LastName LIKE '%Smith' AND FirstName = 'Ian'
Finally, take a look at the following query, and the Index Seek operator properties in
Figure 4-5.
WHERE ProductID = 771 AND ABS(SalesOrderID) = 45233
Listing 4-5.

129
Figure 4-5: Index Seek operator properties.
The seek predicate is using only the ProductID column as shown here:
Seek Keys[1]: Prefix: [AdventureWorks].[Sales].
[SalesOrderDetail].ProductID = Scalar Operator (CONVERT_IMPLICIT(int,[@1],0)
Listing 4-6.

130
An additional predicate on the SalesOrderID column is evaluated like any other scan
predicate, as listed in:
abs([AdventureWorks].[Sales].[SalesOrderDetail]. [SalesOrderID])=[@2]
Listing 4-7.
So, in summary this shows that, as we expected, SQL Server was able to perform a seek
operation on the ProductID column but, because of the use of the ABS function, was
not able to do the same for SalesOrderID. The index was used to navigate directly to
find the rows that satisfy the first predicate, but then had to continue scanning to validate
the second predicate.
The Database Engine Tuning Advisor
Currently, all major commercial database vendors include a physical database design tool
to help with the creation of indexes. However, when these tools were first developed,
there were just two main architectural approaches considered for how the tools should
recommend indexes. The first approach was to build a stand-alone tool with its own
cost model and design rules. The second approach was to build a tool that could use the
Query Optimizer cost model.
A problem with building a stand-alone tool is the requirement for duplicating the cost
module. On top of that, having a tool with its own cost model, even if it's better than the
optimizer's cost model, may not be a good idea because the optimizer still chooses its plan
based on its own model.
The second approach, using the Query Optimizer to help on physical database design,
was proposed in the database research community as far as back as 1988. Since it's the
optimizer which chooses the indexes for an execution plan, it makes sense to use the

131
optimizer itself to help find which missing indexes would benefit existing queries. In this
scenario, the physical design tool would use the optimizer to evaluate the cost of queries
given a set of candidate indexes. An additional benefit of this approach is that, as the
optimizer cost model evolves, any tool using its cost model can automatically benefit
from it.
SQL Server was the first commercial database product to include a physical design tool,
in the shape of the Index Tuning Wizard which shipped with SQL Server 7.0, and
which was later replaced by the Database Engine Tuning Advisor (DTA) in SQL Server
2005. Both tools use the Query Optimizer cost model approach and were created as
part of the AutoAdmin project at Microsoft, the goal of which was to reduce the total
cost of ownership (TCO) of databases by making them self-tuning and self-managing.
In addition to indexes, the DTA can help with the creation of indexed views and
table partitioning.
However, creating real indexes in a DTA tuning session is not feasible; its overhead could
impact operational queries and degrade the performance of your database. So how does
the DTA estimate the cost of using an index that does not yet exist? Actually, even during
a regular query optimization, the Query Optimizer does not use indexes to estimate
the cost of a query. The decision on whether to use an index or not relies only on some
metadata and the statistical information regarding the columns of the index. Index data
itself is not needed during query optimization but will, of course, be required during
query execution if the index is chosen.
So, to avoid creating real indexes during a DTA session, SQL Server uses a special kind of
indexes called hypothetical indexes, which were also used by the Index Tuning Wizard.
As the name implies, hypothetical indexes are not real indexes; they only contain statis-
tics and can be created with the undocumented WITH STATISTICS_ONLY option of the
CREATE INDEX statement. You may not be able to see these indexes during a DTA
session because they are dropped automatically when they are no longer needed.
However, you could see the CREATE INDEX WITH STATISTICS_ONLY and DROP INDEX
statements if you run a SQL Server Profiler session to see what the DTA is doing.

132
Let's take a quick tour to some of these concepts. To get started, create a new table on the
AdventureWorks database:
SELECT *
INTO dbo.SalesOrderDetail
Listing 4-8.
Copy the following query and save it to a file:
Listing 4-9.
Open a new DTA session, and you can optionally run a SQL Server Profiler session
if you want to inspect what the DTA is doing. On the Workload File option, select the
file containing the SQL statement that you just created with Listing 4-9, and specify
AdventureWorks as both the database to tune and the database for workload analysis.
Click the Start Analysis button and, when the DTA analysis finishes, run this query to
inspect the contents of the msdb..DTA_reports_query table:
SELECT * FROM msdb..DTA_reports_query
Listing 4-10.

133
Running that query shows the following output, (edited for space):
StatementString CurrentCost RecommendedCost
------------------------------------------- ------------ ---------------
SELECT * FROM dbo.SalesOrderDetail WHERE... 1.2434 0.00328799
Listing 4-11.
Notice that the query returns information like the query that was tuned, as well as the
current and recommended cost. The current cost, 1.2434, is easy to obtain by directly
requesting an estimated execution plan for the query as shown in Figure 4-6.
Figure 4-6: Plan showing total cost.
Since the DTA analysis was completed, the required hypothetical indexes were
already dropped. To now obtain the indexes recommended by the DTA, click on the
Recommendations tab and look at the Index Recommendations section, where you
can find the code to create any recommended index by then clicking on the Definition
column. In our example, it will show the code in Listing 4-12.

134
CREATE CLUSTERED INDEX [_dta_index_SalesOrderDetail_c_5_1915153868__K5]
ON [dbo].[SalesOrderDetail]
(
[ProductID] ASC
)WITH (SORT_IN_TEMPDB = OFF, IGNORE_DUP_KEY = OFF, DROP_EXISTING = OFF,
ONLINE = OFF) ON [PRIMARY]
Listing 4-12.
In the next statement, and for demonstration purposes only, I will go ahead and create
the index recommended by the DTA but, instead of a regular index, I will create it as a
hypothetical index by adding the WITH STATISTICS_ONLY clause.
CREATE CLUSTERED INDEX cix_ProductID ON dbo.SalesOrderDetail(ProductID)
WITH STATISTICS_ONLY
Listing 4-13.
You can validate that a hypothetical index was created by running the next query:
SELECT * FROM sys.indexes
WHERE object_id = object_id('dbo.SalesOrderDetail')
AND name = 'cix_ProductID'
Listing 4-14.
The output is shown next below; note that the is_hypothetical field shows that this
is, in fact, just a hypothetical index.
object_id name index_id type type_desc is_hypothetical
---------- ------------- -------- ----- --------- ----------------
1915153868 cix_ProductID 3 1 CLUSTERED 1
Listing 4-15.

135
Remove the hypothetical index by running this statement:
DROP INDEX dbo.SalesOrderDetail.cix_ProductID
Listing 4-16.
Finally, implement the DTA recommendation, this time as a regular clustered index:
CREATE CLUSTERED INDEX cix_ProductID ON dbo.SalesOrderDetail(ProductID)
Listing 4-17.
After implementing the recommendation and running the query again, the clustered
index is in fact now being used by the Query Optimizer. This time, the plan shows a
Clustered Index Seek operator and an estimated cost of 0.0033652, which is very close
to the recommended cost listed previously when querying the msdb..DTA_reports_
query table.
Finally, drop the table you just created by running the following statement:
Listing 4-18.

136
The Missing Indexes Feature
SQL Server does provide a second approach that can help you to find useful indexes
for your existing queries. Although not as powerful as the DTA, this option, called the
Missing Indexes feature, does not require the database administrator to decide when
tuning is needed, to explicitly identify what workload represents the load to tune, or to
run any tool. This is a lightweight feature which is always on and, like the DTA, was also
introduced with SQL Server 2005. Let's take a look at what it does.
During optimization, the Query Optimizer defines what the best indexes for a query are
and, if these indexes don't exist, it will make this index information available in the XML
plan for a particular plan (as well as the graphical plan, as of SQL Server Management
Studio 2008). Alternatively, it will aggregate this information for queries optimized since
the instance was started, and make it all available on the sys.dm_db_missing_index
DMVs. Note that, just by displaying this information, the Query Optimizer is not only
warning you that it might not be selecting an efficient plan; it is also showing you which
indexes may help to improve the performance of your query. In addition, database
administrators and developers should be aware of the limitations of this feature, as
described on the Books Online entry, Limitations of the Missing Indexes Feature.
So, with all that in mind, let's take a quick look to see how this feature works. Create
the dbo.SalesOrderDetail table on the AdventureWorks database by running the
SELECT *
INTO dbo.SalesOrderDetail
FROM sales.SalesOrderDetail
Listing 4-19.

137
Run this query and request a graphical or XML execution plan:
WHERE SalesOrderID = 43670 AND SalesOrderDetailID > 112
Listing 4-20.
This query could benefit from an index on the SalesOrderID and SalesOrder-
DetailID columns, but no missing indexes information is shown this time. One
limitation of the Missing Indexes feature which this example has revealed is that it
does not work with a trivial plan optimization. You can verify that this is a trivial plan
by looking at the graphical plan properties, shown as Optimization Level TRIVIAL, or
by looking at the XML plan, where the StatementOptmLevel is shown as TRIVIAL.
You can avoid the trivial plan optimization in several ways, as I'll explain in Chapter 5,
The Optimization Process (for now, you'll just have to take it on faith). In our case, we're
just going to create a non-related index by running the following statement:
CREATE INDEX IX_ProductID ON dbo.SalesOrderDetail(ProductID)
Listing 4-21.
What is significant about this is that, although the index created will not be used by our
previous query, the query no longer qualifies for a trivial plan. Run the query again, and
this time the XML plan will contain the following entry:
<MissingIndexes>
<MissingIndexGroup Impact="99.7137">
<MissingIndex Database="[AdventureWorks]" Schema="[dbo]"
Table="[SalesOrderDetail]">
<ColumnGroup Usage="EQUALITY">
<Column Name="[SalesOrderID]" ColumnId="1" />
</ColumnGroup>

138
<ColumnGroup Usage="INEQUALITY">
<Column Name="[SalesOrderDetailID]" ColumnId="2"/>
</ColumnGroup>
</MissingIndex>
</MissingIndexGroup>
</MissingIndexes>
Listing 4-22.
The MissingIndexes entry in the XML plan can show up to three groups: equality,
inequality, and included; and the first two are shown in this example using the
ColumnGroup attribute. The information contained in these groups can be used to
create the missing index; the key of the index can be built by using the equality columns,
followed by the inequality columns, and the included columns can be added using the
INCLUDE clause of the CREATE INDEX statement. SQL Server 2008 Management Studio
can build the CREATE INDEX statement for you and, in fact, if you look at the graphical
plan, you can see a Missing Index warning at the top, including a CREATE INDEX
command, as shown in Figure 4-7:
Figure 4-7: Plan with a Missing Index warning.
Notice the impact value of 99.7137 – Impact is a number between 0 and 100 which gives
you an estimate of the average percentage benefit that the query could obtain if the
proposed index were available.
You can right-click on the graphical plan and select Missing Index Details to see the
CREATE INDEX command that can be used to create this desired index, as shown
in Listing 4-23.

139
/*
Missing Index Details from SQLQuery1.sql – The Query Processor estimates that
implementing the following index could improve the query cost by 99.7137%.
*/
/*
USE [AdventureWorks]
GO
CREATE NONCLUSTERED INDEX [<Name of Missing Index, sysname,>]
ON [dbo].[SalesOrderDetail] ([SalesOrderID], [SalesOrderDetailID])
GO
*/
Listing 4-23.
Create the recommended index, after you provide a name for it, by running the
CREATE NONCLUSTERED INDEX IX_SalesOrderID_SalesOrderDetailID
ON [dbo].[SalesOrderDetail]([SalesOrderID], [SalesOrderDetailID])
Listing 4-24.
If you run the query in Listing 4-20 again and look at the execution plan, this time you'll
see an Index Seek operator using the index you've just created, and both the Missing
Index warning and the MissingIndex element of the XML plan are gone, as shown in
Figure 4-8.

140
Figure 4-8: Plan without the Missing Index warning.
Finally, remove the dbo.SalesOrderDetail table you've just created by running the
Listing 4-25.
Unused Indexes
I'll end this chapter on indexes by introducing the functionality of the sys.dm_db_
index_usage_stats DMV, which you can use to learn about the operations performed
by your indexes, and which is especially helpful in discovering indexes that are not used
by any query, or are only minimally used. As we've already discussed, indexes that are not
being used will provide no benefit to your databases, but will use valuable disk space, slow
your update operations, and should be considered for removal.
The sys.dm_db_index_usage_stats DMV stores the number of seek, scan, lookup,
and update operations performed by both user and system queries, including the last
time each type of operation was performed. Keep in mind that this DMV, in addition to

141
non-clustered indexes, will also include heaps, listed as index_id equal to 0, and
clustered indexes, listed as index_id equal to 1. For the purposes of this section, you
may want to just focus on non-clustered indexes, which include index_id values 2 or
greater. Since heaps and clustered indexes contain the table's data, they may not even be
candidates for removal in the first place.
By inspecting the user_seeks, user_scans, and user_lookup values of your
non-clustered indexes you can see how your indexes are being used, and you can inspect
the user_updates values to see the amount of updates performed on the index. All of
this information will help to give you a sense as to how useful an index actually is. Bear in
mind that all I'll be demonstrating is how to call up information from this DMV, and what
sort of situations will trigger different updates to the information it returns. How you
react to the information it returns is a task I leave to you.
As an example, run the following code to create a new table with a non-clustered index:
SELECT * INTO dbo.SalesOrderDetail
CREATE NONCLUSTERED INDEX IX_ProductID ON dbo.SalesOrderDetail(ProductID)
Listing 4-26.
If you want to keep track of the values for this example follow these steps carefully, as
every query execution may change the index usage statistics. When you run the following
query, it will initially contain only one record, which was created because of table access
performed when the index on Listing 4-26 was created.
SELECT DB_NAME(database_id) as database_name,
OBJECT_NAME(s.object_id) as object_name, i.name, s.*
FROM sys.dm_db_index_usage_stats s join sys.indexes i
ON s.object_id = i.object_id AND s.index_id = i.index_id
and s.object_id = object_id('dbo.SalesOrderDetail')
Listing 4-27.

142
However, the values that we will be inspecting in this exercise, user_seeks,
user_scans, user_lookups, and user_updates are all set to 0.
Now run the following query, let's say, three times:
Listing 4-28.
This query is using a Table Scan operator, so, if you rerun the code in Listing 4-27,
the DMV will show the value 3 on the user_scans column. Note that the column
index_id is 0, denoting a heap, and the name of the table is also listed (as a heap is
just a table with no clustered index).
Run the next query, which uses an Index Seek, twice. After the query is executed, a new
record will be added for the non-clustered index, and the user_seeks counter will show
a value of 2.
SELECT ProductID FROM dbo.SalesOrderDetail
Listing 4-29.
Now, run the following query four times, and it will use both Index Seek and RID Lookup
operators. Since the user_seeks for the non-clustered index had a value of 2, it will be
updated to 6, and the user_lookups value for the heap will be updated to 4.
Listing 4-30.

143
Finally, run the following query once:
UPDATE dbo.SalesOrderDetail
SET ProductID = 666
Listing 4-31.
Note that the UPDATE statement is doing an Index Seek and a Table Update, so
user_seek will be updated for the index, and user_updates will be updated
once for both the non-clustered index and the heap. This is the final output of the
query in Listing 4-27 (edited for space):
name index_id user_seeks user_scans user_lookups user_updates
------------ -------- ---------- ---------- ------------ ------------
NULL 0 0 3 4 1
IX_ProductID 2 7 0 0 1
Listing 4-32.
Finally, drop the table you just created:
Listing 4-33.

144
Summary
This chapter explained how you can define the key of your indexes so that they are likely
to be considered for seek operations, which can improve the performance of your queries
by finding records more quickly. Predicates were analyzed in the contexts of both single
and multi-column indexes, and we also covered how to verify an execution plan to
validate that indexes were selected and properly used by SQL Server.
The Database Engine Tuning Advisor and the Missing Indexes feature, both introduced
with SQL Server 2005, were presented to show how the Query Optimizer itself can be
used to provide index tuning recommendations.
Finally, the sys.dm_db_index_usage_stats DMV was introduced, together with
its ability to provide valuable information regarding your non-clustered indexes usage.
While we didn't have time to discuss all the practicalities of using this DMV, we covered
enough for you to be able to easily find non-clustered indexes that are not being used by
your SQL Server instance.

145
Chapter 5: The Optimization Process
In this chapter, I'll go into the internals of the Query Optimizer and introduce the
steps that it performs in the background, and which we don't see. This covers everything,
from the time a query is submitted to SQL Server until an execution plan is generated
and is ready to be executed, and includes steps like parsing, binding, simplification, trivial
plan optimization and full optimization. Important components which are part of the
Query Optimizer architecture, such as transformation rules and the memo structure, are
also introduced.
The purpose of the Query Optimizer, as we're all aware, is to provide an optimum
execution plan and, in order to do so, it generates possible alternative execution plans
through the use of transformation rules. These alternative plans are stored for the
duration of the optimization process in a structure called the memo. Given that finding
the optimum plan for some queries would take an unacceptably long optimization time,
some heuristics are used to limit the number of alternative plans considered, instead of
using the entire search space – remember that the goal is to find a good enough plan as
quickly as possible. Heuristics help the Query Optimizer to cope with the combinatorial
explosion which occurs in the search space as queries get progressively more complex.
However, the use of transformation rules and heuristics does not necessarily reduce
the cost of the available alternatives, so each candidate plan is also costed, and the best
alternative is chosen based on those costs.
Overview
The query optimization and execution process were introduced in Chapter 1, Introduction
to Query Optimization, and will be explained in more detail throughout the rest of this
chapter. However, before we get started, I'll very briefly describe the inner workings of the

146
query optimization process, which extends both before and after the Query Optimizer
itself. So, if I mention terminology or concepts you've not seen before, don't panic – I'll go
into much more detail and explain everything as we go through the chapter.
Parsing and binding are the first operations performed when a query is submitted to a
SQL Server instance. They produce a tree representation of the query, which is then sent
to the Query Optimizer to perform the optimization process. At the beginning of this
optimization process, this logical tree will be simplified, and the Query Optimizer will
check if the query qualifies for a trivial plan. If it does, then a trivial execution plan is
returned and the optimization process immediately ends. The parsing, binding, simplifi-
cation and trivial plan processes do not depend on the contents of the database (such as
the statistics and the data itself), but only on the database schema and query definition.
These processes also don't use statistics, cost estimation or cost-based decisions, all of
which are only employed during the full optimization process.
If the query does not qualify for a trivial plan, then the Query Optimizer will run the
full optimization process, which is executed in up to three stages, and a plan may be
produced at the end of any of these stages. In addition, to consider all of the information
gathered in the previous phases, like the query definition and database schema, the full
optimization process will also use statistics and cost estimation, and will select the best
execution plan (within the available time) based solely on that plan's cost.
Peeking at the Query Optimizer
In this section I will show you two DMVs which you can use to gain additional insight
into the work being performed by the Query Optimizer. The first one, sys.dm_exec_
query_optimizer_info, which is only partially documented, provides information
regarding the optimizations performed on the SQL Server instance. The second one,
sys.dm_exec_query_transformation_stats, which is also undocumented,
provides information regarding how the Query Optimizer is using the defined
transformation rules. Although both DMVs contain cumulative statistics, recorded

147
since the given SQL Server instance was started, they can also be used to get optimization
information for a specific query or workload, as we'll see in a moment.
Let us look at sys.dm_exec_query_optimizer_info first; as mentioned, you can
use this DMV to obtain statistics regarding the operation of the Query Optimizer, such as
how queries have been optimized, and how many of them have been optimized since the
instance started. This DMV returns three columns:
• Counter – the name of the optimizer event
• Occurrence – the number of occurrences of the optimization event for this counter
• Value – the average value per event occurrence.
38 counters were defined for SQL Server 2005, and a new one, called merge stmt, was
added in SQL Server 2008, giving a total of 39.
To view the statistics for all the Query Optimizer events since the SQL Server instance
was started, we can just run:
SELECT * FROM sys.dm_exec_query_optimizer_info
Listing 5-1.
Table 5-1 shows some example output from one of my SQL Server instances. It shows that
there have been 691,473 optimizations since the instance was started, that the average
elapsed time for each optimization was 0.0078 seconds, and that the average estimated
cost of each optimization, in internal cost units, was about 1.398. This particular example
shows optimizations of inexpensive queries, typical of an OLTP system.

148
Counter Occurrence Value
optimizations 691473 1
elapsed time 691465 0.007806012
final cost 691465 1.398120739
trivial plan 29476 1
tasks 661989 332.5988816
no plan 0 NULL
search 0 26724 1
search 0 time 31420 0.01646922
search 0 tasks 31420 1198.811617
Table 5-1.
The query shown in Listing 5-2 displays the percentage of optimizations in the system
that include hints. This information could be useful to show how extensive the use of
hints in your application is, which, in turn, can show that your code may be less flexible
than anticipated, and may require additional maintenance. Hints are explained in detail
in Chapter 7.
SELECT ( SELECT occurrence
FROM sys.dm_exec_query_optimizer_info
WHERE counter = 'hints'
) * 100.0 / ( SELECT occurrence
FROM
sys.dm_exec_query_optimizer_info
WHERE counter = 'optimizations'
)
Listing 5-2.

149
Although the sys.dm_exec_query_optimizer_info DMV was completely
documented in the original version of SQL Server 2005 Books Online, more recent
versions omit descriptions of nearly half (18 out of 39) of the counters, and instead label
them as "Internal only."
Therefore, in Table 5-2, I am including the current Books Online documentation
plus descriptions of the 18 undocumented counters, according to their original
documentation, which is still valid for SQL Server 2008 R2. The additional descriptions
are shown in italics.
optimizations
Total number of
optimizations.
Not applicable.
elapsed time
Total number of
optimizations.
Average elapsed time per optimization of an
individual statement (query), in seconds.
final cost
Total number of
optimizations.
Average estimated cost for an optimized plan,
in internal cost units.
trivial plan
Total number of trivial plans
(used as final plan).
Not applicable.
tasks
Number of optimizations
that applied tasks (explo-
ration, implementation,
property derivation).
Average number of tasks executed.
no plan
for which no plan was found
after a full optimization
was run, and where no other
errors were issued during
query compilation.
Not applicable.

150
search 0
Total number of final plans
found in search 0 phase.
Not applicable.
search 0 time
that entered search 0.
Average time spent in search 0, in seconds.
search 0 tasks
Average number of tasks run in search 0.
search 1
Not applicable.
search 1 time
Average time spent in search 1, in seconds.
search 1 tasks
search 2
Not applicable.
search 2 time
Average time spent in search 2.
search 2 tasks
gain stage 0 to
stage 1
Number of times search 1 was
run after search 0.
Average gain from stage 0 to stage 1
as (MinimumPlanCost(search 0)
– MinimumPlanCost(search 1)) /
MinimumPlanCost(search 0).
gain stage 1 to
stage 2
Number of times search 2 was
run after search 1.
Average gain from stage 1 to stage 2
as (MinimumPlanCost(search 1)
– MinimumPlanCost(search 2)) /
MinimumPlanCost(search 1).

151
timeout
for which internal timeout
occurred.
Not applicable.
memory limit
exceeded
Number of optimizations for
which an internal memory
limit was exceeded.
Not applicable.
insert stmt
that are for INSERT
statements.
Not applicable.
delete stmt
that are for DELETE
statements.
Not applicable.
update stmt
Number of optimiza
tions that are for UPDATE
statements.
Not applicable.
merge stmt
Number of optimiza-
tions that are for MERGE
statements.
Not applicable.
contains
subquery
for a query that contains at
least one subquery.
Not applicable.
unnest failed
Number of times where
subquery unnesting could not
remove the subquery.
Not applicable.
tables
Total number of
optimizations.
Average number of tables referenced per query
optimized.

152
hints
Number of times some
hint was specified. Hints
counted include: JOIN,
GROUP, UNION and FORCE
ORDER query hints, FORCE
PLAN set option, and
join hints.
Not applicable.
order hint
Number of times a force
order hint was specified.
Not applicable.
join hint
Number of times the join
algorithm was forced by a
join hint.
Not applicable.
view reference
Number of times a view has
been referenced in a query.
Not applicable.
remote query
where the query referenced
at least one remote data
source, such as a table with
a four-part name or an
OPENROWSET result.
Not applicable.
maximum DOP
Total number of
optimizations.
Average effective MAXDOP value for an
optimized plan. By default, effective MAXDOP
is determined by the max degree of parallelism
server configuration option, and may be
overridden for a specific query by the value
of the MAXDOP query hint.

153
maximum
recursion level
in which a MAXRECUR-
SION level greater than 0
has been specified with the
query hint.
Average MAXRECURSION level in optimizations
where a maximum recursion level is specified
with the query hint.
indexed views
loaded
Number of queries for which
one or more indexed views
were loaded for consideration
for matching.
Average number of views loaded.
indexed views
matched
where one or more indexed
views have been matched.
Average number of views matched.
indexed views
used
where one or more indexed
views are used in the output
plan after being matched.
Average number of views used.
indexed views
updated
of a DML statement that
produce a plan which
maintains one or more
indexed views.
Average number of views maintained.
dynamic cursor
request
in which a dynamic cursor
request has been specified.
Not applicable.
fast-forward
cursor request
Number of optimizations in
which a fast-forward cursor
request has been specified.
Not applicable.
Table 5-2: Books Online documentation, with undocumented counters.

154
As mentioned previously, you can use this DMV in two different ways: you can use it
to get information regarding the history of accumulated optimizations on the system
since the instance was started or, rather more usefully, you can use it to get optimization
information for a particular query or a workload.
In order to capture data on the latter, you need to take two snapshots of the DMV – one
before optimizing your query, and another one after the query has been optimized – and
manually find the difference between them. Unfortunately, there is no way to initialize
the values of this DMV.
There are several issues to consider when capturing this information. Firstly, you need
to eliminate the effects of system-generated queries, or queries executed by other users,
which may be running at the same time as your sample query. Try to isolate the query
or workload on your own instance, and make sure that the number of optimizations
reported is the same as the number of optimizations you are requesting. If the former is
greater, then the data probably includes some of those queries submitted by the system or
other users. Of course, it's also possible that your own query against the sys.dm_exec_
query_optimizer_info DMV may count as an optimization.
Secondly, you need to make sure that a query optimization is actually taking place. For
example, if you run the same query more than once, then the Query Optimizer may
simply use an existing plan from the plan cache. You can force an optimization by using
the RECOMPILE hint, as shown later, or by clearing the plan cache. For instance, as of
SQL Server 2008, the DBCC FREEPROCCACHE statement can be used to remove a specific
plan, all the plans related to a specific resource pool, or the entire plan cache. But of
course, you should never clear the plan cache of a production environment.
With all of this in mind, the script shown in Listing 5-3 will display the optimization
information for a specific query, while avoiding all of the aforementioned issues. The
script is based on an original idea by Lubor Kollar, and has a section to include the query
which you want to get optimization information about.

155
- – optimize these queries now
- – so they do not skew the collected results
GO
SELECT *
INTO after_query_optimizer_info
GO
SELECT *
INTO before_query_optimizer_info
GO
DROP TABLE before_query_optimizer_info
DROP TABLE after_query_optimizer_info
GO
- – real execution starts
GO
SELECT *
INTO before_query_optimizer_info
GO
- – insert your query here
SELECT *
FROM Person.Address
- – keep this to force a new optimization
OPTION (RECOMPILE)
GO
SELECT *
INTO after_query_optimizer_info
GO
SELECT a.counter,
(a.occurrence – b.occurrence) AS occurrence,
(a.occurrence * a.value – b.occurrence *
b.value) AS value
FROM before_query_optimizer_info b
JOIN after_query_optimizer_info a
ON b.counter = a.counter
WHERE b.occurrence <> a.occurrence
DROP TABLE before_query_optimizer_info
DROP TABLE after_query_optimizer_info
Listing 5-3.

156
Note that some queries are listed twice in the code. The purpose of this is to optimize
them the first time that they are executed, so that their plan can be available in the plan
cache for all the executions after that. In this way, we aim as far as possible to isolate
the optimization information from the queries we are trying to analyze. Care must be
taken that both queries are exactly the same, including case, comments, and so on, and
separated in their own batch for the GO statements.
If you run this script against the AdventureWorks database, the output should look like
what's shown in Table 5-3. Note that the times shown obviously may be different from the
ones you get in your system, (for both this and other examples in this chapter). This table
indicates, among other things, that there was one optimization, referencing one table,
with a cost of 0.230042585.
elapsed time 1 0
final cost 1 0.230042585
maximum DOP 1 0
optimizations 1 1
tables 1 1
trivial plan 1 1
Table 5-3.
Certainly, for this simple query, we could find the same information in some other places,
such as in an execution plan. However, as I will show later in this chapter, this DMV can
provide optimization information that is not available anywhere else.
The second DMV, sys.dm_exec_query_transformation_stats, provides informa-
tion about the existing transformation rules and how they are being used by the Query

157
Optimizer. Similar to the sys.dm_exec_query_optimizer_info DMV, you can also
use it to get optimization information for a particular query or workload by taking two
snapshots of the DMV (before and after optimizing your query), and manually finding the
difference between them.
To start looking at this DMV, run the following query:
SELECT * FROM sys.dm_exec_query_transformation_stats
Listing 5-4.
Table 5-4 contains a sample output in my test system using SQL Server 2008 R2, showing
the first 10 records out of 377, and edited to fit the page.
Name promise_avg Promised built_substitute Succeeded
JNtoNL 49 2 0 0
LOJNtoNL 456.1428571 7 7 7
LSJNtoNL 0 0 0 0
LASJNtoNL 0 0 0 0
JNtoSM 454 2 2 2
FOJNtoSM 0 0 0 0
LOJNtoSM 454 7 7 0
ROJNtoSM 454 7 7 0
LSJNtoSM 0 0 0 0
RSJNtoSM 0 0 0 0
Table 5-4.

158
The sys.dm_exec_query_transformation_stats DMV returns the transformation
rules currently defined in the system, of which there are 377 for the current release of SQL
Server 2008 R2, and includes what is known as their promise information, which tells
the Query Optimizer how useful a given transformation rule might be. The first field in
the results output is the name of the rule; for example, the first three listed are JNtoNL
(Join to Nested Loops Join), LOJNtoNL (Left Outer Join to Nested Loops Join), and
JNtoSM (Join to Sort Merge Join), which is the academic name of the SQL Server Merge
Join operator.
The same issues shown for the sys.dm_exec_query_optimizer_info DMV
regarding collecting data also apply to the sys.dm_exec_query_transformation_
stats DMV, so the following query can help you to isolate the optimization information
for a specific query, while avoiding data from related queries as much as possible. The
query is based on the succeeded column, which keeps track of the number of times a
transformation rule was used and successfully produced a result.
- – optimize these queries now
- – so they do not skew the collected results
GO
SELECT *
INTO before_query_transformation_stats
FROM sys.dm_exec_query_transformation_stats
GO
SELECT *
INTO after_query_transformation_stats
GO
DROP TABLE after_query_transformation_stats
DROP TABLE before_query_transformation_stats
- – real execution starts
GO
SELECT *
INTO before_query_transformation_stats
GO
- – insert your query here
SELECT * FROM dbo.DatabaseLog

159
- – keep this to force a new optimization
OPTION (RECOMPILE)
GO
SELECT *
INTO after_query_transformation_stats
GO
SELECT a.name, (a.promised – b.promised) as promised
FROM before_query_transformation_stats b
JOIN after_query_transformation_stats a
ON b.name = a.name
WHERE b.succeeded <> a.succeeded
DROP TABLE before_query_transformation_stats
DROP TABLE after_query_transformation_stats
Listing 5-5.
For example, testing with a very simple AdventureWorks query like the following, which
is already included in the code in Listing 5-5 ...
Listing 5-6.
... will show that the following transformation rules are being used.
name promised
GetIdxToRng 1
GetToTrivialScan 1
Table 5-5.

160
We will be using these two DMVs in several sections later in this chapter, and you should,
hopefully, come to see why they are very useful in providing additional insight into the
work being performed by the Query Optimizer.
Parsing and Binding
Parsing and binding are the first operations that SQL Server executes when you submit a
query to a database and, in the current version, they are performed by a component called
the Algebrizer. Parsing first makes sure that the T-SQL query has a valid syntax, and then
uses the query information to build a tree of relational operators. By that, I mean the
parser translates the SQL query into an algebra tree representation of logical operators,
which is called a parse tree.
Parsing only checks for valid T-SQL syntax, not for valid table or column names, which
are verified in the next phase: binding.
Parsing is similar to the parse functionality available in Management Studio (by clicking
the Parse button on the default toolbar) or the SET PARSEONLY statement. For example,
the following query will successfully parse on the AdventureWorks database, even when
the listed columns and table do not exist in said database.
SELECT lname, fname FROM authors
Listing 5-7.
However, if you incorrectly write the SELECT or FROM keywords, SQL Server will return
an error message complaining about the incorrect syntax.

161
Once the parse tree has been constructed, the Algebrizer performs the binding
operation, which is mostly concerned with name resolution. During this operation,
the Algebrizer makes sure that all of the objects named in the query do actually exist,
confirms that the requested operations between them are valid, and verifies that the
objects are visible to the user running the query. It also associates every table and column
name on the parse tree with their corresponding object in the system catalog. Name
resolution for views includes the process of view substitution, where a view reference is
expanded to include the view definition; for example, to directly include the tables used
in the view. The output of the binding operation, which is called an algebrized tree, is
then sent to the Query Optimizer for (as you'll have guessed) optimization.
Originally, this tree will be represented as a series of logical operations which are closely
related to the original syntax of the query. These will include such logical operations as
"get data from the Customer table," "get data from the Contact table," "perform an inner
join," and so on. Different tree representations of the query will be used throughout the
optimization process, and this logical tree will receive different names, until it is finally
used to initialize the memo structure, as we'll discuss later.
For example, the following query will have a tree representation as shown in Figure 5-1.
SELECT c.CustomerID, COUNT(*)
FROM Sales.Customer c JOIN Sales.SalesOrderHeader o
ON c.CustomerID = o.CustomerID
WHERE c.TerritoryID = 4
GROUP BY c.CustomerID
Listing 5-8.

162
Figure 5-1: Query tree representation.
Transformation Rules
The SQL Server Query Optimizer uses transformation rules to explore the search space;
that is, to explore the set of possible execution plans for a specific query. Transformation
rules are based on relational algebra, taking a relational operator tree and generating
equivalent alternatives, in the form of equivalent relational operator trees. At the
most fundamental level, a query consists of logical expressions, and applying these
transformation rules will generate equivalent logical and physical alternatives, which
are stored in memory, in a structure called the memo, for the entire duration of the
optimization process. As already mentioned, and explained later in this chapter, the
Query Optimizer uses three optimization stages, and different transformation rules are
applied in each stage.
Each transformation rule has a pattern and a substitute. The pattern is the expression
to be analyzed and matched, and the substitute is the equivalent expression that it is

163
generated as an output. For example, for the commutativity rule, which is explained later,
a transformation rule can be defined as: Expr1 join Expr2 – > Expr2 join Expr1.
SQL Server will match the pattern Expr1 join Expr2, like in Individual join Customer,
and will produce the equivalent expression, Customer join Individual. The two
expressions are equivalent because both return exactly the same results.
Initially, the query tree contains only logical expressions, and transformation rules are
applied to these logical expressions to generate either logical or physical expressions. As
an example, a logical expression can be the definition of a logical join, whereas a physical
expression could be an actual join implementation, like a Merge Join or a Hash Join. Bear
in mind that transformation rules cannot be applied to physical expressions.
The main types of transformation rules include simplification, exploration and imple-
mentation rules. Simplification rules produce simpler logical trees as their outputs,
and are mostly used during the simplification phase, before the full optimization.
Exploration rules, also called logical transformation rules, generate logical equivalent
alternatives; and implementation rules, or physical transformation rules, are used to
obtain physical alternatives. Both exploration and implementation rules are executed
during the full optimization phase.
Examples of exploration rules include the commutativity and associativity rules, which
are used in join optimization. Commutativity and associativity rules are defined as A join
B – > B join A and (A join B) join C – > A join (B join C) respectively. The commutativity
rule, A join B – > B join A, means that A join B is equivalent to B join A, and joining
the tables A and B in any order will return the same results. Also note that applying
the commutativity rule twice will generate the original expression again; that is, if you
initially apply this transformation to obtain B join A, and then later apply the same
transformation, you can obtain A join B again. However, the Query Optimizer can
handle this problem in order to avoid duplicated expressions. In the same way, the
associativity rule shows that (A join B) join C is equivalent to A join (B join C) as they
also both produce the same results.

164
An example of an implementation rule would be selecting a physical algorithm for a
logical join, such as a Merge Join or a Hash Join.
So the Query Optimizer is using sets of transformation rules to generate and examine
possible alternative execution plans. However, it's important to remember that applying
transformations does not necessarily reduce the cost of the generated alternatives, and
the costing component still needs to estimate their costs. Although both logical and
physical alternatives are kept in the memo structure, only the physical alternatives are
costed. It's important, then, to bear in mind that, although these alternatives may be
equivalent and return the same results, their physical implementations may have very
different costs. The final selection, as is hopefully clear now, will be the best (or, if you
like, the "cheapest") physical alternative stored in the memo.
For example, implementing A join B may have different costs depending on whether
a Nested Loops Join or a Hash Join is selected. In addition, for the same physical join,
implementing the A join B expression may have different performance from B join A.
As explained in Chapter 2, The Execution Engine, the performance of a join is different
depending on which table is chosen as the inner or outer table in a Nested Loops Join,
or the build and the probe inputs in a Hash Join. If you want to find out why the Query
Optimizer might not choose a specific join algorithm, you can use a hint to force a
specific physical join and compare the cost of both the hinted and the original plans.
Those are the foundation principles of transformation rules and, as we saw briefly earlier
in this chapter, according to the sys.dm_exec_query_transformation_stats
DMV, SQL Server currently has 377 transformation rules, and more can be added in
future versions of the product. Just as a point of reference, a quick look at the first
Community Technology Preview (CTP) of the next version of SQL Server, code-named
"Denali," already shows 382 transformation rules included in this DMV!
So, let's go back to the sys.dm_exec_query_transformation_stats DMV defined
earlier and see a few examples of transformation rules used by the query processor.
Include the following query into the code in Listing 5-5 to explore the transformation
rules it uses:

165
Listing 5-9.
As shown in the following output, 17 transformation rules were exercised during the
optimization process.
Name Promised
AppIdxToApp 0
EnforceSort 23
GbAggBeforeJoin 4
GbAggToHS 8
GbAggToStrm 8
GenLGAgg 2
GetIdxToRng 0
GetToIdxScan 4
GetToScan 4
ImplRestrRemap 3
JNtoHS 6
JNtoIdxLookup 6
JNtoSM 6
JoinCommute 2

166
ProjectToComputeScalar 2
SelIdxToRng 6
SELonJN 1
Table 5-6.
Now, (as I will explain in more detail in Chapter 7, Hints) hints may disable some of these
transformation rules for the duration of a query in order to obtain a specific desired
behavior. As a way of experimenting with the effects of these rules, you can also use
the undocumented statements DBCC RULEON and DBCC RULEOFF to enable or disable
transformation rules, and thereby get additional insight into how the Query Optimizer
works. However, before you do that, first be warned that, since these statements impact
the entire optimization process performed by the Query Optimizer, they should be used
only in a test system for experimentation purposes.
To demonstrate the effects of these statements, the query in Listing 5-9 shows the plan
seen below in Figure 5-2.
Figure 5-2: Original execution plan.
Here you can see, among other things, that SQL Server is pushing an aggregate below
the join (a Stream Aggregate before the Merge Join). The Query Optimizer can push
aggregations that significantly reduce cardinality estimation as early in the plan as
possible. This is performed by the transformation rule GbAggBeforeJoin (or Group
By Aggregate Before Join), which is included in the output of Table 5-6. This specific

167
transformation rule is used only if certain requirements are met; for example, when
the GROUP BY clause includes the joining columns, which is the case in our example.
Run the following statement to temporarily disable the use of the GbAggBeforeJoin
transformation rule for the current session:
DBCC RULEOFF('GbAggBeforeJoin')
Listing 5-10.
After disabling this transformation rule and running the query again, the plan, shown
in Figure 5-3, will now show the aggregate after the join, which, according to the Query
Optimizer, is a more expensive plan. You can verify this by looking at their estimated
costs: 0.285331 and 0.312394, respectively. (These are not shown on the figures, but you
can see them by hovering the mouse over the SELECT icon and examining the Estimated
Subtree Cost value, as explained before.) Note that, for this exercise, an optimization may
need to be forced to see the new plan, perhaps using the OPTION (RECOMPILE) hint or
one of the methods which we've discussed to remove the plan from the cache, like DBCC
FREEPROCCACHE.
Figure 5-3: Plan with GbAggBeforeJoin rule disabled.
In addition, there are a couple of additional undocumented statements to show which
transformation rules are enabled and disabled, and these are DBCC SHOWONRULES and
DBCC SHOWOFFRULES. By default, DBCC SHOWONRULES will list all the 377 transformation
rules listed by the sys.dm_exec_query_transformation_stats DMV. To test it, try
running the code in Listing 5-11.

168
DBCC TRACEON (3604)
DBCC SHOWONRULES
Listing 5-11.
We start this exercise with the DBCC TRACEON (3604) command, which enables trace
flag 3604, and instructs SQL Server to send the results to the client, in this case, your
Management Studio session. That means that, in this exercise, the output of the DBCC
SHOWONRULES, and later DBCC SHOWOFFRULES, DBCC RULEON and DBCC RULEOFF
statements will be conveniently available to us. An output similar to Listing 5-12 will be
displayed in this first instance (only 12 rules out of 376 rules are shown here, to preserve
space). The previously disabled rule will not be shown in this output.
DBCC execution completed. If DBCC printed error messages, contact your system administrator.
Rules that are on globally:
JNtoNL
LOJNtoNL
LSJNtoNL
LASJNtoNL
JNtoSM
FOJNtoSM
LOJNtoSM
ROJNtoSM
LSJNtoSM
RSJNtoSM
LASJNtoSM
RASJNtoSM
…
Listing 5-12.
In the same way, the following code will show the rules that are disabled:
DBCC SHOWOFFRULES
Listing 5-13.

169
In our case, it will show that only one rule has been disabled:
Rules that are off globally:
GbAggBeforeJoin
Listing 5-14.
To continue with our example of the effects of the transformation rules, we can disable
the use of a Merge Join by disabling the rule JNtoSM (Join to Sort Merge Join) by running
the following code:
DBCC RULEOFF('JNtoSM')
Listing 5-15.
If you have followed the example, this time DBCC RULEOFF will show some output
indicating that the rule is off for some specific SPID. Running the sample query again will
give us this totally new plan, using both a Hash Join and a Hash Aggregate, as shown in
Figure 5-4.
Figure 5-4: Plan with JNtoSM rule disabled.
In Chapter 7, Hints, you will learn how to obtain this same behavior in your queries using
(unsurprisingly) hints.

170
Finally, before we finish, don't forget to re-enable the GbAggBeforeJoin and JNtoSM
transformation rules by running the following commands ...
DBCC RULEON('JNtoSM')
DBCC RULEON('GbAggBeforeJoin')
Listing 5-16.
... and verify that no transformation rules are still disabled, by running:
DBCC SHOWOFFRULES
Listing 5-17.
You may also want to clear your plan cache to make sure none of these experiment plans
were left in memory, by once again running:
DBCC FREEPROCCACHE
Listing 5-18.
The Memo
The memo structure was originally defined in The Volcano Optimizer Generator by Goetz
Graefe and William McKenna in 1993. As discussed in Chapter 1, Introduction to Query
Optimization, the SQL Server Query Optimizer is based on the Cascades Framework,
which was, in fact, a descendent of the Volcano optimizer.
The memo is a search data structure that is used to store the alternatives which are
generated and analyzed by the Query Optimizer. These alternatives can be logical or

171
physical operators, and are organized into groups of equivalent alternatives, such that
each alternative in the same group produces the same results. Alternatives in the same
group also share the same logical properties and, in the same way that operators can
reference other operators on a relational tree, groups can also reference other groups in
the memo structure.
A new memo structure is created for each optimization. The Query Optimizer first
copies the original query tree's logical expressions into the memo structure, placing
each operator from the query tree in its own group, and then triggers the entire optimi-
zation process. During this process, transformation rules are applied to generate all the
alternatives, starting with these initial logical expressions.
As the transformation rules produce new alternatives, these are added to their equivalent
groups. Transformation rules may also produce a new expression which is not
equivalent to any existing group, and which causes a new group to be created. As I
mentioned, each alternative in a group is a simple logical or physical expression, like
a join or a scan, and a plan will be built using a combination of these alternatives. The
number of these alternatives, and even groups, in a memo structure can be huge.
Although there is the possibility that different combinations of transformation rules may
end up producing the same expressions, the memo structure is designed to avoid both
the duplication of these alternatives and redundant optimizations. By doing this, it saves
memory and is more efficient, as it does not have to search the same plan alternatives
more than once.
Although both logical and physical alternatives are kept in the memo structure, only the
physical alternatives are costed. Thus, at the end of the optimization process, the memo
contains all of the alternatives considered by the Query Optimizer, but only one plan is
selected, based on its cost.
Now, I will show a simplified example of how the memo structure is built for a simple
query, using listing 5-19.

172
SELECT FirstName, LastName, CustomerType
Listing 5-19.
After a logical tree is created for this query, as explained before, each operator is placed in
its own group, as shown in Table 5.7.
Group 6 Join 3 & 4
Group 5
Group 4 Scan Customer
Group 3 Join 1 & 2
Group 2 Scan Individual
Group 1 Scan Contact
Table 5-7: Initial memo structure.
Notice how both joins reference the other groups instead of the operators. We call Group
6 the root because it is the root operator of the initial plan, that is, it is the root node of
the original query tree. I just left Group 5 blank so it is easier to introduce new groups and
visualize the root group at the top. During optimization, several transformation rules will
be executed, and they will create new alternatives. For example, if we apply either of the
two following associativity rules …

173
(Contact join Individual) join Customer – > Contact join(Individual join Customer)
Listing 5-20.
or
(1 join 2) join 4 – > 1 join(2 join 4)
Listing 5-21.
... we obtain two new operators that are not yet present in the memo structure. The first
one is Join 2 & 4. Since we do not have an equivalent group for this new operator, we
place it in a new group, Group 5 in this case. The second operator joins Group 1 and the
new operator we just added to the memo structure. This new operator would be Join 1 &
5, and since Join 1 & 5 is equivalent to Join 3 & 4, we place them in the same group, which
is also the root group. These two changes are shown in Table 5-8.
Group 6 Join 3 & 4 Join 1 & 5
Group 5 Join 2 & 4
Group 3 Join 1 & 2
Table 5-8: Memo structure after associativity rule has been applied.

174
We can show an example of the commutativity rule like this:
1 Join 5 – > 5 Join 1
Listing 5-22.
Since the new alternative is equivalent to the original operator, it is placed in the same
group, as shown in Table 5-9, along with two other example operators which were added
using the same commutativity rule.
Group 6 Join 3 & 4 Join 1 & 5 Join 5 & 1
Table 5-9: Memo structure after commutativity rule.
Given that this is a simplified example, and I imagine you've got the picture by now,
I'll stop generating logical alternatives at this point. However, rest assured that the
Query Optimizer has many other transformation rules in its arsenal with which to
generate alternatives.
Towards the end of the process, after some implementation rules are applied, physical
operators will be added to the memo structure. A few of them have been added to Table
5-10, and they include data access operators like Clustered Index Scan, and physical join
operators like Nested Loops Join, together with Merge Join and Hash Join.

175
Group 6 Join 3 & 4 Join 1 & 5 Join 5 & 1 Nested Loops 5 & 1 Hash Join 5 & 1
Group 5 Join 2 & 4 Join 4 & 2 Nested Loops 2 & 4 Merge Join 4 & 2
Group 4
Scan
Customer
Clustered Index
Scan
Group 3 Join 1 & 2 Join 2 & 1 Nested Loops 1 & 2
Group 2
Scan
Individual
Clustered Index
Scan
Group 1
Scan
Contact
Clustered Index
Scan
Table 5-10: Memo structure with physical operators.
After the cost is estimated for each physical operator, the Query Optimizer will look for
the cheapest way to assemble a plan using the alternatives available. In our example, it
would select the plan operators shaded in gray in Table 5-10. As a result, the execution
plan selected by SQL Server is the plan shown in Figure 5-5, and you should notice that
the Query Optimizer did not select the join order that was explicitly requested in the
query text and shown in Group 1's tree representation. Instead, the Query Optimizer
found a better join order with a lower cost.
Figure 5-5: Selected execution plan.

176
Optimization Phases
The Query Optimizer has several optimization phases designed to try to optimize queries
as quickly and simply as possible, and to not use more expensive and sophisticated
options unless absolutely necessary. These phases are called the simplification, trivial
plan optimization and full optimization stages. In the same way, the full optimization
phase itself consists of three stages simply called search 0, search 1 and search 2.
Plans can be produced in any of these phases except for the simplification one, which I'll
discuss in a moment. In this section, I'll use the sys.dm_exec_query_optimizer_
info DMV, introduced earlier in this chapter, to show additional information about
these optimization phases.
Simplification
Query rewrites or, more exactly, tree rewrites are performed on this stage to reduce the
query tree into a simpler form in order to make the optimization process easier. Some of
these simplifications include those below.
• Subqueries are converted into joins, but since a subquery does not always translate
directly to an inner join, outer join and group by operations may be added as necessary.
• Redundant inner and outer joins may be removed. A typical example is the Foreign Key
Join elimination which occurs when SQL Server can detect that some joins may not
be needed, as foreign key constraints are available and only columns of the referencing
table are requested. An example of Foreign Key Join elimination is shown later.
• Filters in WHERE clauses are pushed down in the query tree in order to enable early
data filtering, and potentially better matching of indexes and computed columns later
in the optimization process (this simplification is known as predicate pushdown).

177
• Contradictions are detected and removed. Since these parts of the query are not
executed at all, SQL Server saves resources like I/O, locks, memory and CPU, making
the query to be executed faster. For example, the Query Optimizer may know that no
records can satisfy a predicate even before touching any page of data. A contradiction
may be related to a check constraint, or may be related to the way the query is written.
Both scenarios will be shown in an example later in this section.
The output of the simplification process is a simplified logical operator tree.
Let's see a couple of examples of the simplification process, starting with the Foreign Key
Join elimination. The query we used on Listing 5-19 joins three tables and shows
the execution plan seen in Figure 5-5. Let's see what happens if we comment out the
CustomerType column, as shown in Listing 5-23.
SELECT FirstName, LastName – -, CustomerType
Listing 5-23.
If you run the query again, this time only two tables are joined, and the Customer table
has been removed, as can be seen in the execution plan in Figure 5-6.
Figure 5-6: Foreign Key Join elimination example.

178
There are two reasons for this change. First, since the CustomerType column is no
longer required, there are no columns requested from the Customer table. However,
it seems like the Customer table is still needed, as it is required as part of the equality
operation on a join condition. That is, SQL Server needs to make sure that a Customer
record exists for each related record on the Individual table.
Actually this validation is performed by the existing foreign key constraint, so the Query
Optimizer realizes that there is no need to use the Customer table at all. As a test,
temporarily disable the foreign key by running the following statement:
ALTER TABLE Sales.Individual NOCHECK CONSTRAINT FK_Individual_Customer_CustomerID
Listing 5-24.
Now run the Listing 5-23 query again. Without the foreign key constraint, SQL Server
has no choice but to perform the join in order to make sure that the join condition is
executed. As a result, it will use a plan joining all three tables, similar to the one shown
previously in Figure 5-5. Finally, don't forget to re-enable the foreign key by running the
statement in Listing 5-25.
ALTER TABLE Sales.Individual WITH CHECK CHECK CONSTRAINT
FK_Individual_Customer_CustomerID
Listing 5-25.
Now for an example of contradiction detection; first, I need a table with a
check constraint and, handily, the Employee table has the following check
constraint definition:
([VacationHours]>=(-40) AND [VacationHours]<=(240))
Listing 5-26.

179
This check constraint makes sure that the number of vacation hours is a number
between –40 and 240, so if I request:
SELECT * FROM HumanResources.Employee
WHERE VacationHours > 80
Listing 5-27.
… SQL Server will use a Clustered Index Scan operator, as shown in Figure 5-7.
Figure 5-7: Plan without contradiction detection.
However, if I request all of the employees with more than 300 vacation hours then,
because of this check constraint, the Query Optimizer must immediately know that no
records qualify for predicate. Run the query in Listing 5-28.
WHERE VacationHours > 300
Listing 5-28.
As expected, the query will return no records, but this time it will show the execution
plan shown in Figure 5-8.

180
Figure 5-8. Contradiction detection example.
Note that, this time, instead of a Clustered Index Scan, SQL Server is using a Constant
Scan operator. Since there is no need to access the table at all, SQL Server saves resources
like I/O, locks, memory and CPU, making the query to be executed faster.
Now, let's see what happens if I disable the check constraint:
ALTER TABLE HumanResources.Employee NOCHECK CONSTRAINT CK_Employee_VacationHours
Listing 5-29.
This time, running the Listing 5-28 query once again uses a Clustered Index Scan
operator, as the Query Optimizer can no longer use the check constraint to guide its
decisions. Don't forget to enable the constraint again by running the following statement:
ALTER TABLE HumanResources.Employee WITH CHECK CHECK CONSTRAINT CK_Employee_
VacationHours
Listing 5-30.
The second type of contradiction case is when the query itself explicitly contains a
contradiction. Take a look at the query in Listing 5-31.
WHERE ManagerID > 10 AND ManagerID < 5
Listing 5-31.

181
In this case there is no check constraint involved; both predicates are valid and each
will individually return records, but they contradict each other when they are run
together. As a result, the query returns no records and the plan shows a Constant Scan
operator similar to the plan in Figure 5-8. This may just look like a badly written query,
but remember that some predicates may already be included in, for example, view
definitions, and the developer of the query may be unaware of those. For example, in
Listing 5-31, a view may include the predicate ManagerID > 10 and a developer may
call the view using the predicate ManagerID < 5. Since both predicates contradict each
other a Constant Scan operator will be used again instead.
Trivial plan
The optimization process may be expensive to initialize and run for very simple queries
that don't require any cost estimation. To avoid this expensive operation for simple
queries, SQL Server uses the trivial plan optimization. In short, if there's only one way,
or one obvious best way to execute the query, depending on the query definition and
available metadata, a lot of work can be avoided. For example, the following Adventure-
Works query will produce a trivial plan:
Listing 5-32.
The execution plan will show if a trivial plan optimization was performed; the
Optimization Level entry in the Properties window of a graphical plan will show
TRIVIAL, as shown in Figure 5-9. In the same way, an XML plan will show the
StatementOptmLevel attribute as TRIVIAL, as shown in the XML fragment
in Listing 5-33.

182
<StmtSimple StatementOptmLevel="TRIVIAL" StatementSubTreeCost="0.471671"
StatementText="SELECT * FROM dbo.DatabaseLog;" StatementType="SELECT">
Listing 5-33.
Figure 5-9: Trivial plan properties.
As I mentioned at the start of this chapter, additional information regarding the
optimization process can be shown using the sys.dm_exec_query_optimizer_info
DMV, which will produce an output similar to Table 5-1 for this query.
elapsed time 1 0
final cost 1 0.471671267
maximum DOP 1 0

183
optimizations 1 1
tables 1 1
trivial plan 1 1
Table 5-11: Optimization information for a trivial plan.
Table 5-11 shows that this was in fact a trivial plan optimization, using one table and a
maximum DOP of 0, and it also displays the elapsed time and final cost. This same query
was also used earlier to demonstrate the sys.dm_exec_query_transformation_
stats DMV, which illustrated the transformation rules used by the Query Optimizer in
this query, as shown previously in Listing 5-5.
The other possible value for the Optimization Level or StatementOptLevel properties
is FULL, which obviously means that the query did not qualify for a trivial plan, and
a full optimization was performed instead. Full optimization is used for more
complicated queries or queries using more complex features, which will require
comparisons of candidate plans' costs in order to guide decisions; this will be explained
in the next section.
Full optimization
If a query does not qualify for a trivial plan, SQL Server will run the cost-based optimiza-
tion process, which uses transformation rules to generate alternative plans, stores these
alternatives in the memo structure, and uses cost estimation to select the best plan. This
optimization process is executed in three stages, with different transformation rules being
applied at each stage.

184
Since some queries may have a huge number of possible query plans, it's sometimes not
feasible to explore their entire search space, as it would take too long. So, in addition to
applying transformation rules, a number of heuristics are used by the Query Optimizer
to control the search strategy and to limit the number of alternatives generated, in order
to quickly find a good plan. The Query Optimizer needs to balance the optimization time
and the quality of the selected plan. For example, as explained in Chapter 1, Introduction
to Query Optimization, optimizing join orders can create a huge number of possible
alternatives. So, a common heuristic used by SQL Server to reduce the size of the search
space is to avoid considering bushy trees.
In addition, the optimization process can immediately finish if a good enough plan
(relative to the Query Optimizer's internal thresholds) is found at the end of any of these
three phases. However if, at the end of any given phase, the best plan is still very expen-
sive, then the Query Optimizer will run the next phase, which will run an additional set
of (usually more complex) transformation rules. These phases are shown as search 0,
search 1 and search 2 on the sys.dm_exec_query_optimizer_info DMV.
Search 0
Similar to the concept of the trivial plan, the first phase, search 0, will aim to find a
plan as quickly as possible without trying sophisticated transformations. Search 0,
called the transaction processing phase, is ideal for the small queries typically found on
transaction processing systems and it is used for queries with at least three tables. Before
the full optimization process is started, the Query Optimizer generates the initial set of
join orders based on heuristics. These heuristics begin by first joining the smallest tables
or the tables that achieve the largest filtering based on their selectivity. Those are the
only join orders considered on search 0. At the end of this phase, the Query Optimizer
compares the cost of the best generated plan to an internal cost threshold and, if the plan
is still very expensive, SQL Server will run the next phase.

185
Search 1
The next phase, search 1, also called Quick Plan, uses additional transformation rules,
limited join reordering, and is appropriate for more complex queries. At the end of this
phase, SQL Server compares the cost of the cheapest plan to a second internal cost
threshold and, if the best plan is cheap enough, then the plan is selected. If the query is
still expensive and the system can run parallel queries (as described in the Parallelism
section in Chapter 2, The Execution Engine), this phase is executed again to find a good
parallel plan, but no plan is selected for execution after this point. At the end of this
phase, the costs of the best serial and parallel plans are compared, and the cheapest one is
used in the following phase, search 2, which we'll come to in just a moment.
As an example, the following query does not qualify for search 0 and will go directly to
search 1:
WHERE ManagerID = 12
Listing 5-34.
Using the sys.dm_exec_query_optimizer_info DMV as shown in Listing 5-2, you
can display its optimization information seen in Table 5-12, which shows that only the
search 1 phase was executed.
elapsed time 1 0.004
final cost 1 0.00657038
maximum DOP 1 0
optimizations 1 1

186
search 1 1 1
search 1 tasks 1 81
search 1 time 1 0
tables 1 1
tasks 1 81
Table 5-12: Optimization information for search 1 phase.
The sys.dm_exec_query_optimizer_info DMV includes a counter named
"gain stage 0 to stage 1" which shows the number of times search 1 was executed after
search 0, and includes the average decrease in cost from one stage to the other, as
defined by:
(MinimumPlanCost(search 0) – MinimumPlanCost(search 1)) / MinimumPlanCost(search 0)
Listing 5-35.
For example, the query we have been using before in this chapter:
SELECT FirstName, LastName, CustomerType
Listing 5-36.
... will show the optimization information seen in Table 5-13.

187
final cost 1 3.239871842
maximum DOP 1 0
optimizations 1 1
search 0 tasks 1 230
search 0 time 1 0.001
search 1 1 1
gain stage 0 to stage 1 1 0.490795403
tables 1 3
tasks 1 607
Table 5-13: Optimization information for search 0 and 1 phases.
The output shows that the optimization process went through both the search 0 and
search 1 stages and that a plan was found on the latter. It also shows a cost improvement
of almost 50% by going from the search 0 to the search 1 stage.
Search 2
The last phase, search 2, is called full optimization, and is used for queries ranging from
complex to very complex. A larger set of the potential transformation rules, parallel
operators and other advanced optimization strategies are considered in this phase and,

188
since this is the last phase, an execution plan must be found here (perhaps with the
exception of the timeout event, as explained later).
The sys.dm_exec_query_optimizer_info DMV includes another useful counter,
named "gain stage 1 to stage 2," to show the number of times search 2 was executed after
search 1, together with the average decrease in cost from one stage to the other,
as defined by:
(MinimumPlanCost(search 1) – MinimumPlanCost(search 2)) / MinimumPlanCost(search 1)
Listing 5-37.
For example, the following query, as taken from Books Online, will create the
optimization information shown in Table 5-14.1
SELECT I.CustomerID, C.FirstName, C.LastName, A.AddressLine1, A.City,
SP.Name AS State, CR.Name AS CountryRegion
JOIN Sales.Individual AS I ON C.ContactID = I.ContactID
JOIN Sales.CustomerAddress AS CA ON CA.CustomerID = I.CustomerID
JOIN Person.Address AS A ON A.AddressID = CA.AddressID
JOIN Person.StateProvince SP ON
SP.StateProvinceID = A.StateProvinceID
JOIN Person.CountryRegion CR ON
CR.CountryRegionCode = SP.CountryRegionCode
Listing 5-38.
1 Output for SQL Server 2008 is shown; number of tasks will vary for SQL Server 2008 R2.

189
final cost 1 5.8466425
indexed views matched 1 1
maximum DOP 1 0
optimizations 1 1
search 2 1 1
tables 1 6
tasks 1 7100
Table 5-14: Optimization information for search 2 phase.
The optimization information shows that this query went throughout all the three stages
of optimization (as show in Figure 5-10) and, among other things, also includes both of
the stage–to-stage gain counters described earlier.

190
Figure 5-10: The optimization process.
As we've touched upon previously, the Query Optimizer has to find the best plan possible
within the shortest amount of time. More to the point, it must eventually return a plan,
even if that plan is not as efficient as it would like. To that end, the optimization process
also includes the concept of a timeout value. This timeout is not a fixed amount of time,
but is, instead, calculated based on the number of transformations applied together with
the elapsed time.

191
When a timeout is found, the Query Optimizer stops the optimization process and
returns the least expensive plan it has found so far. The best plan found so far could
be a plan found during the current optimization stage, but most likely it would be the
best plan found in the previous one. This obviously means that a timeout can happen
only on the search 1 and search 2 stages; no timeouts can occur on stage 0, as the Query
Optimizer needs to finish at least one optimization stage in order to find a good (or at
least viable) initial plan (the same applies when stage 1 is chosen as the first optimization
phase for a query). This timeout event is shown in the properties of a graphical plan as
Reason For Early Termination Of Statement Optimization, as shown in Figure 5-11, or
in an XML plan as StatementOptmEarlyAbortReason. This event is also shown as the
timeout counter on the sys.dm_exec_query_optimizer_info DMV.
Figure 5-11: Timeout example.
Finally, at the end of the optimization process, the chosen plan will be sent to the
execution engine to be run, and the results will be sent back to the client.

192
Summary
This chapter showed the internals of the Query Optimizer, how your query goes from
a SQL statement submitted to SQL Server, all the way to the selected execution plan,
including parsing, binding, simplification, trivial plan, and the full optimization stages.
Important components which are part of the Query Optimizer architecture, such as
transformation rules and the memo structure, were also introduced.
The Query Optimizer generates a solution space and selects the best possible execu-
tion plan from it, based on the plan cost. Transformation rules are used to generate
these alternatives, which are stored in a memory structure called the memo. Instead
of exploring the search space exhaustively, heuristics are also introduced to limit the
number of possible solutions. Finally, the alternatives stored in the memo are costed,
and the best solution is returned for execution.

193
Chapter 6: Additional Topics
So far we have focused on how the Query Optimizer solves SELECT queries with
mostly joins and aggregations. In this chapter, we'll see some additional SQL features
that traditionally are not covered in query optimization topics, including updates and
data warehouse queries, plus some other topics related to query parameters including
parameter sniffing, auto-parameterization, and forced parameterization.
Update operations are an intrinsic part of database operations, and they also need to be
optimized so that they can be performed as quickly as possible. Just to be clear, when I say
"updates," in truth I'm referring to any operation performed by the INSERT, DELETE or
UPDATE statements, as well as the MERGE statement, which is new in SQL Server 2008.
In this chapter, I'll explain the basics of update operations, and how they can quickly
become complicated, as they need to update existing indexes, access multiple tables and
enforce existing constraints. I will show how the Query Optimizer can select per-row
and per-index plans to optimize UPDATE statements, and I will describe the Halloween
protection problem, as well as how SQL Server avoids it.
Data warehouses are becoming increasingly popular as decision support systems for
organizations of all sizes. Not only are more organizations building data warehouses,
those databases are also growing in size, and multi-terabyte environments are very
common today. In this chapter, I will introduce data warehousing and briefly compare
it to online transaction processing systems. I will explain the basic concepts of data
warehousing, including fact and dimension tables, as well as star and snowflake schemas.
After that, we'll focus on how SQL Server optimizes star join queries, as well as how it
can automatically detect star and snowflake schemas, and reliably identify fact and
dimension tables.
We'll end the chapter with two sections discussing the use of query parameters. In the
first section, I will talk about the parameter sniffing behavior and explain that, although
looking at the parameters of a query helps the Query Optimizer to produce better

194
execution plans, the occasional reuse of some of these plans can also be a performance
problem. I will demonstrate how you can identify problems related to the parameter
sniffing behavior, and I will provide a few recommendations on how to avoid
them, including the use of the OPTIMIZE FOR or RECOMPILE hints. Finally, the auto-
parameterization behavior is explained, along with forced parameterization, a more
drastic choice to parameterize queries, which was introduced with SQL Server 2005.
Updates
Even when performing an update involves some other areas of SQL Server, such as
transactions, concurrency control or locking, update processing is still totally integrated
within the SQL Server query processor framework. Update operations are also optimized
so they can be performed as quickly as possible. So, in this section I will talk about
updates from the point of view of the Query Optimizer. As mentioned earlier, for the
purposes of this section, I'll refer to any operation performed by the INSERT, DELETE,
UPDATE, or MERGE statements as updates.
Update plans can be complicated, as they need to update existing indexes alongside
data and, because of objects like check constraints, referential integrity constraints
and triggers, those plans may also have to access multiple tables and enforce existing
constraints. Updates may also require the updating of multiple tables when cascading
referential integrity constraints or triggers are defined. Some of these operations, such
as updating indexes, can have a big impact on the performance of the entire update
operation, and we'll take a deeper look at that now.
Update operations are performed in two steps, which can be summarized as a read
section followed by the update section. The first step provides the details of the changes
to apply and which records will be updated. For INSERT operations, this includes the
values to be inserted and, for DELETE operations, it includes obtaining the keys of the
records to be deleted, which could be the clustering keys for clustered indexes or the RIDs
for heaps. Just to keep you on your toes, for update operations, a combination of both the

195
keys of the records to be updated and the data to be inserted is needed. In this first step,
SQL Server may read the table to be updated just like in any other SELECT statement.
In the second step, the update operations are performed, including updating indexes,
validating constraints and executing triggers. The update operation will fail and roll back
if it violates a constraint.
Let me start with an example of a very simple update operation. Inserting a new record
on the Person.CountryRegion table using the next query creates a very simple plan, as
INSERT INTO Person.CountryRegion (CountryRegionCode, Name) VALUES ('ZZ', 'New
Country')
Listing 6-1.
Figure 6-1: An insert example.
However, the same scenario gets complicated very quickly when you try to delete the
same record by running the next statement, as shown on the plan in Figure 6-2.
DELETE FROM Person.CountryRegion
WHERE CountryRegionCode = 'ZZ'
Listing 6-2.

196
Figure 6-2: A delete example.
As you can see in this plan, in addition to CountryRegion, two additional tables
(StateProvince and CountryRegionCurrency) are accessed. The reason behind
this is that these two tables have a foreign key referencing CountryRegion, and so SQL
Server needs to validate that no records exist on these tables for this specific value of
CountryRegionCode. So, the tables are accessed, and an Assert operator is included at
the end of the plan to perform this validation. If a record with the CountryRegionCode
to be deleted exists in any of these tables, the Assert operator will throw an exception and
SQL Server will roll back the transaction, returning the following error message:
The DELETE statement conflicted with the REFERENCE constraint "FK_StateProvince_CountryRegion_
CountryRegionCode." The conflict occurred in database "AdventureWorks," table "Person.
StateProvince," column 'CountryRegionCode'.
Listing 6-3.
So as you can see, the previous example showed how update operations can access some
other tables not included in the original query, in this case, because of the definition of
referential integrity constraints. The updating of non-clustered indexes is covered in the
next section.

197
Per-row and per-index plans
An important operation performed by updates is the modifying and updating of existing
non-clustered indexes, which is done by using per-row or per-index maintenance plans
(also called narrow and wide plans, respectively). In a per-row maintenance plan, the
updates to the base table and the existing indexes are performed by a single operator, one
row at a time. On the other hand, in a per-index maintenance plan, the base table and
each non-clustered index are updated in separated operations.
Except for a few cases where per-index plans are mandatory, the Query Optimizer can
choose between a per-row and per-index plan based on performance reasons, and on an
index-by-index basis. Although factors like the structure and size of the table, as well as
the other operations performed by the UPDATE statement, are all considered, choosing
between per-index and per-row plans will mostly depend on the number of records
being updated. The Query Optimizer is more likely to choose a per-row plan when a
small number of records are being updated, and a per-index plan when the number of
records to be updated increases, as this choice scales better. A drawback with the per-row
approach is that the storage engine updates the non-clustered index rows using the
clustered index key order, which is not efficient when a large number of records need to
be updated.
The query in Listing 6-4 will create a per-row plan, which is shown in Figure 6-3 (two
additional queries may be shown on the plan due to the execution of an existing trigger).
Note
The following two queries delete data from the AdventureWorks database, so perhaps you should request
an estimated plan if you don't want the records to be deleted. Alternatively, you could perform a database
backup before running these queries, so that you will be able to restore the database later.

198
DELETE FROM Sales.SalesOrderDetail
WHERE SalesOrderDetailID = 61130
Listing 6-4.
Figure 6-3: A per-row plan.
In addition to updating the clustered index, this delete operation will update two existing
non-clustered indexes, IX_SalesOrderDetail_ProductID and AK_SalesOrder-
Detail_rowguid, which can be seen listed on the Object property in the Properties
window of the Clustered Index Delete operator, as shown in Figure 6-4.
Figure 6-4: Properties of the Clustered Index Delete operator.

199
When a large number of records are being updated, the Query Optimizer may choose a
per-index plan, which the following query will demonstrate, by creating the per-index
plan shown in Figures 6-5 and 6-6.
DELETE FROM Sales.SalesOrderDetail
WHERE SalesOrderDetailID < 43740
Listing 6-5.
In a per-index update, the base table is updated first, which is shown by the Clustered
Index Delete operator in Figure 6-5.
In the second part of the plan, which is shown in Figure 6-6, a Table Spool operator is
used to read the data of the key values of the indexes to be updated, and then a Sort
operator sorts the data in the order of the index. Later, an Index Delete operator updates
a specific non-clustered index in one operation (the name of which you can see on the
graphical plan). Although, the table spool is listed twice in the plan, it is actually the same
operator being reused. Finally, the Sequence operator makes sure that each Index Delete
operation is performed in sequence, as shown from top to bottom.
Figure 6-5: Right part of the per-index plan.
Figure 6-6: Left part of the per-index plan.

200
In summary, keep in mind that, except for a few cases where per-index plans are
mandatory, the Query Optimizer can choose between a per-row and per-index plan
on an index-by-index basis, so it is even possible to have both maintenance choices in
the same execution plan.
Halloween protection
Halloween protection refers to a problem which appears in certain update operations,
and which was found more than thirty years ago by researchers working on the System R
project (mentioned in Chapter 1, Introduction to Query Optimization) at the IBM Almaden
Research Center. The System R team was testing a query optimizer when they ran a query
to update the salary column on an Employee table. The query was supposed to give a
10% raise to every employee with a salary of less than $25,000 but, to their surprise, no
employee had a salary under $25,000 after the update query was completed. They noticed
that the query optimizer had selected the salary index and had updated some records
multiple times, until they reached the $25,000 salary. Since the salary index was used to
scan the records, when the salary column was updated, some records were moved within
the index and were then scanned again later, updating those records more than once. The
problem was called Halloween problem simply because it was discovered on Halloween
around 1976 or 1977.
As I mentioned at the beginning of this section, update operations have a read section
followed by an update section, and that is a crucial distinction to bear in mind at this
stage. To avoid the Halloween problem, the read and update sections must be completely
separated; the read section must be completed in its entirety before the write section is
run. I'll show you how SQL Server avoids the Halloween problem in the next example.
Run the statement in Listing 6-6 to create a new table.

201
SELECT *
INTO dbo.Product
FROM Production.Product
Listing 6-6.
Run the following UPDATE statement, which produces the execution plan on Figure 6-7.
UPDATE dbo.Product
SET ListPrice = ListPrice * 1.2
Listing 6-7.
Figure 6-7: An update without Halloween protection.
No Halloween protection is needed in this case, as the statement updates the
ListPrice column, which is not part of any index, and so updating the data does
not move any rows around. Now, to demonstrate the problem, I'll create a clustered
index on ListPrice column.
CREATE CLUSTERED INDEX cix ON dbo.Product(ListPrice)
Listing 6-8.
Run the same UPDATE statement from Listing 6-7 again. The query will show a similar
plan, but this time including a Table Spool operator, which is a blocking operator,
separating the read section from the write section. A blocking operator has to read all of
the relevant rows before producing any output rows to the next operator. In this example,

202
the table spool separates the Clustered Index Scan from the Clustered Index Update, as
The spool operator scans the original data and saves a copy of it in a hidden spool
table in tempdb before it is updated. A Table Spool operator is usually used to avoid
the Halloween problem as it is a cheap operator. However, if the plan already includes
another operator that can be used, such as a Sort, then the Table Spool operator is not
needed, and the Sort can perform the same blocking job instead.
Figure 6-8: An update with Halloween protection.
Finally, drop the table you have just created.
DROP TABLE dbo.Product
Listing 6-9.
Data Warehouses
A data warehouse is a decision support system for business decision making, designed to
execute queries from users as well as running reporting and analytical applications. It is
also structurally different from an online transaction processing (OLTP) system, which
focuses on operational transaction processing (we'll look at some of these differences in
a just a moment). Because of these different purposes, both systems also have different
workloads: a data warehouse will usually have to support complex and large queries,
compared to the typically small transactions of an OLTP system.

203
Another main difference between OLTP databases and data warehouses is the degree of
normalization found in them. An OLTP system uses normalized databases, usually at a
third normal form, while a data warehouse uses a denormalized dimensional model. An
OLTP normalized model helps to remove data redundancies,and focus on data integrity;
it benefits update operations as data needs to be updated in one place only. On the other
hand, a data warehouse dimensional model is more appropriate for ad hoc complex
queries, and will usually have fewer tables and require fewer joins.
Dimensional data modeling on data warehouses relies on the use of fact and dimension
tables. Fact tables contain facts or numerical measures of the business, which can
participate in calculations, while dimension tables are the attributes or descriptions of
the facts. Fact tables also usually have foreign keys to link them to the primary keys of the
dimension tables.
Data warehouses also usually follow star and snowflake schema structures. A star
schema contains a fact table and a single table for each dimension. Snowflake schemas
are similar to star schemas to the extent that they also have a fact table but, in addition,
dimension tables can also be normalized, and each dimension can have more than one
table. Fact tables are typically huge and can store millions or billions of rows, compared
to dimension tables, which are significantly smaller. The size of data warehouse databases
tends to range from hundreds of gigabytes to terabytes.
SQL Server sample databases includes AdventureWorksDW, a data warehouse database
whose purpose is to demonstrate the SQL Server business intelligence features. The
AdventureWorksDW database will be used for the example in this section.
Queries that join a fact table to dimension tables are called star join queries. SQL Server
includes special optimizations for star join queries (which we'll look at shortly), can
automatically detect star and snowflake schemas, and can reliably identify fact and
dimension tables. This is significant because sometimes, in order to avoid the overhead of
constraint enforcement during updates, data warehouse implementations don't explicitly
define foreign key constraints. In these cases, the Query Optimizer may need to rely on
heuristics to detect star schemas.

204
One such heuristic is to consider the largest table of the star join query as the fact table
(which, in addition, must have a specified minimum size, currently defined as 100 pages).
The second heuristic requires that all the joins in a star join query need to be inner joins,
and use equality predicates on a single column. It should also be noticed that even in the
rare case where a dimension table is incorrectly chosen as a fact table through the use of
these heuristics, the Query Optimizer will still select a valid plan which will return the
correct data, although it may not be an efficient one.
Regarding optimizations for star join queries, it is interesting to consider the use of
Cartesian (or Cross) products of the dimension tables with multi-column index lookups
on a fact table. Although Cross products are avoided during the regular optimization
process because they can generate huge intermediate results, they can be used for data
warehouse queries involving small dimension tables. As the rows of the Cross product are
being generated, they are immediately used to look up on a multi-column index without
requiring a lot of memory for the intermediate results.
In Optimizing Star Join Queries for Data Warehousing in Microsoft SQL Server,1
Cesar
Galindo-Legaria et al. define three different approaches to optimizing star join queries
based on the selectivity of the fact table, as shown next. As mentioned in Chapter 3,
Statistics and Cost Estimation, selectivity is a measure of the number of records that
are estimated to be returned by a query and, slightly anti-intuitively, smaller numbers
represent higher selectivity (i.e. fewer rows).
For highly selective queries which return up to 10% of the rows in the fact table, the
Query Optimizer may produce a plan with Nested Loops Joins, Index Seeks and
bookmark lookups. For medium selectivity queries, which return anywhere from 10 to
75% of the records in the fact table, SQL Server may recommend Hash Joins with bitmap
filters in combination with fact table scans or fact table range scans. Finally, for the least
selective queries, processing more than 75% of the fact table, the Query Optimizer mostly
will recommend regular Hash Joins with fact table scans. The choice of these operators
1 Published in the Proceedings of the 2008 IEEE 24th International Conference on Data Engineering.

205
and plans is not surprising for the highly and least selective queries, as it is their standard
usage as explained in Chapter 2, The Execution Engine. What is new is the choice of Hash
Joins and bitmap filtering for medium selectivity queries, so that's what we'll look at next.
Bitmap filtering is an optimization for star join queries that was introduced with
SQL Server 2008 and it is only available on the Enterprise, Developer and Evaluation
editions. It is referred to as optimized bitmap filtering in order to differentiate it from
the standard bitmap filtering which was already available in previous versions of SQL
Server. Optimized bitmap filtering improves the performance of star join queries by
removing unnecessary rows from processing early in the query plan, so that subsequent
operators have fewer rows to process. In our case, it filters rows from the fact table to
avoid additional join processing.
This strategy is called "semi-join reduction" and relies on the fact that only the records
from the second table that qualify for the join with the first table are processed. SQL
Server bitmap filters are based on bloom filters, originally conceived by Burton Bloom in
1970. Other semi-join reduction technologies like bitmap indexes have been used by other
database vendors.
Optimized bitmap filtering works with Hash Joins which (as we saw in Chapter 2) use
two inputs, the smaller of which (the build table) is being completely read into memory.
Optimized bitmap filtering takes advantage of the fact that a Hash Join has to process
the build input anyway so, as SQL Server is processing the build table, it creates a bitmap
representation of the join key values found. Since SQL Server can reliably detect fact and
dimension tables, and the latter are almost always the smaller of the two, the build input
upon which the bitmap is based will be a dimension table. This bitmap representation
of the dimension table will be used to filter the second input of the Hash Join, the probe
input, which in this case will be the fact table. This basically means that only the rows in
the fact table that qualify for the join to the dimension table will be processed.

206
Next, let's see an example of optimized bitmap filtering. Run the query in Listing 6-10.
USE AdventureWorksDW
GO
SELECT *
FROM dbo.FactInternetSales AS f
JOIN dbo.DimProduct AS p ON f.ProductKey = p.ProductKey
JOIN dbo.DimCustomer AS c ON f.CustomerKey = c.CustomerKey
WHERE p.ListPrice > 50 AND c.Gender = 'M'
Listing 6-10.
Figure 6-9: A bitmap filtering example.
Note
You may not get the plan shown earlier on a test system with a limited number of logical processors, but
you can simulate that your SQL Server instance has (for example) 8 processors by using the –P startup
parameter. In order to do that, open Configuration Manager, right-click on your SQL Server service,
select Properties, select the Advanced tab, and edit the Startup Parameters entry by adding ";-P8" at
the end of the line. Click OK and restart the instance, remembering to remove this entry when you finish
your testing.
Since this plan is too big to show here, only a section is included in Figure 6-9 (this
plan was created with SQL Server 2008 R2, so the one for SQL Server 2008 may vary a
little). This part of the plan shows one of the two available Bitmap operators, in this case
processing the rows from the DimCustomer table, which is the build input of the Hash

207
Join shown on the left. This Bitmap operator is identified as Opt_Bitmap1007, as you can
verify in the operator's Properties window, and the Opt_ prefix indicates that optimized
bitmap filtering is, in fact, being used. The second Bitmap operator, not shown on this
part of the plan, is identified as Opt_Bitmap1008. The probe input of the Hash Join is
a Table Scan operator on the FactInternetSales fact table, which is also shown in
Figure 6-9. Finally, the predicate section of this Table Scan operator's properties, as
shown in Figure 6-10, shows that both the Opt_Bitmap1007 and Opt_Bitmap1008
bitmap filters are being applied to the fact table to remove non-qualifying rows before
the Hash Join.
Figure 6-10: Fact table Table Scan operator properties.
Finally, bitmap filtering can significantly improve the performance of data warehouse
queries by pushing the filters down into the scan of the fact table early in the query plan,
so that subsequent operators have fewer rows to process.

208
Parameter Sniffing
As we saw in Chapter 3, Statistics and Cost Estimation, SQL Server can use the histograms
of statistics objects to estimate the cardinality of a query, and then use this information to
try to produce an optimal execution plan. The Query Optimizer accomplishes this by first
inspecting the values of the query parameters.
This behavior is called parameter sniffing, and it is a very good thing: getting an execution
plan tailored to the current parameters of a query naturally improves the performance
of your applications. We also know that the plan cache can store these execution plans
so that they can be reused the next time the same query needs to be executed. This saves
optimization time and CPU resources, as the query does not need to be optimized again.
However, although the Query Optimizer and the plan cache work well together most
of the time, some performance problems can occasionally appear. Given that the Query
Optimizer can produce different execution plans for syntactically identical queries,
depending on their parameters, caching and reusing only one of these plans may create
a performance issue for alternative instances of this query which would benefit from a
better plan. This is a known problem with queries using explicit parameterization, such as
stored procedures, for example. Next, I'll show you an example of this problem, together
with a few recommendations on how to fix it.
Let's write a simple stored procedure using the Sales.SalesOrderDetail table on the
AdventureWorks database:
CREATE PROCEDURE test (@pid int)
AS
WHERE ProductID = @pid
Listing 6-11.

209
Run the following statement to execute the stored procedure, and request to display the
execution plan:
EXEC test @pid = 897
Listing 6-12.
The Query Optimizer estimates that only a few records will be returned by this query,
and produces the execution plan shown in Figure 6-11, which uses an Index Seek operator
to quickly find the records on an existing non-clustered index, and a Key Lookup operator
to search on the base table for the remaining columns requested by the query.
Figure 6-11: Plan using Index Seek and Key Lookup operators.
This combination of Index Seek and Key Lookup operators was a good choice because,
although it's a relatively expensive combination, the query was highly selective. However,
what if a different parameter is used, producing a less selective predicate? For example,
try the following query, including a SET STATISTICS IO ON statement to display the
amount of disk activity generated by the query:
SET STATISTICS IO ON
GO
GO
Listing 6-13.

210
The Messages tab will show an output similar to the one in Listing 6-14.
Table 'SalesOrderDetail'. Scan count 1, logical reads 15615, physical reads 87, read-ahead
reads 150, …
Listing 6-14.
As you can see, on this execution alone, SQL Server is performing 15,615 logical reads
when the base table only has 1,244 pages; so it's using over 12 times more I/Os than just
simply scanning the entire table. As we have seen before, performing Index Seeks plus Key
Lookups to the base table, which uses random I/Os, is a very expensive operation.
Now clear the plan cache to remove the execution plan currently held in memory, and
run the stored procedure again, using the same parameter as in Listing 6-13.
DBCC FREEPROCCACHE
GO
GO
Listing 6-15.
This time, you'll get a totally different execution plan. The I/O information now will show
that only around 1,240 pages were read, and the execution plan will include a Clustered
Index Scan as shown in Figure 6-12. Since, this time, there was no optimized version of
the stored procedure stored in the plan cache, SQL Server optimized it from scratch using
the new parameter, and created a new optimal execution plan.
Figure 6-12: Plan using a Clustered Index Scan.

211
Of course, this doesn't mean that you're not supposed to trust your stored procedures
any more, or that maybe all your code is incorrect. This is just a problem that you need
to be aware of and research, especially if you have queries where performance changes
dramatically when different parameters are introduced. If you happen to have this
problem, there are a few choices available, which we'll explore now.
Optimize for a typical parameter
There might be cases when most of the executions of a query use the same execution
plan and you want to avoid an ongoing optimization cost. In these cases you can use the
OPTIMIZE FOR hint, which helps when an optimal plan is generated for the majority of
values used in a specific parameter. As a result, only the few executions using an atypical
parameter will not have an optimal plan.
Suppose that almost all of the executions of our stored procedure would benefit from the
previous plan using an Index Seek and a Key Lookup operator. To take advantage of that,
you could write the stored procedure as in Listing 6-16.
ALTER PROCEDURE test (@pid int)
AS
OPTION (OPTIMIZE FOR (@pid = 897))
Listing 6-16.
When you run the stored procedure for the first time, it will be optimized for the value
897, no matter what parameter value was actually specified for the execution. If you want
check, test the case in Listing 6-17.

212
Listing 6-17.
You can find the following entry close to the end of the XML plan.
<ParameterList>
<ColumnReference Column="@pid" ParameterCompiledValue="(897)"
ParameterRuntimeValue="(870)" />
</ParameterList>
Listing 6-18.
This entry clearly shows which parameter value was used during optimization, and which
one was used during execution. In this case, the stored procedure is optimized only once,
and the plan is stored in the plan cache and reused as many times as needed. The benefit
of using this hint, in addition to avoiding optimization cost, is that you have total control
over which plan is stored in the plan cache.
Optimize on every execution
If you want the best performance for every query, the solution might be to optimize for
every execution. You will get the best possible plan on every execution but will end up
paying for the optimization cost, so you'll need to decide if that's a worthwhile trade-off.
To do this, use the RECOMPILE hint as shown in Listing 6-19.
AS
OPTION (RECOMPILE)
Listing 6-19.

213
The XML plan for this execution:
Listing 6-20.
… will show:
<ParameterList>
<ColumnReference Column="@pid" ParameterCompiledValue="(897)"
ParameterRuntimeValue="(897)" />
</ParameterList>
Listing 6-21.
Local Variables and OPTIMIZE FOR UNKNOWN
Another solution that has been traditionally implemented in the past is the use of local
variables instead of parameters. As mentioned in Chapter 3, Statistics and Cost Estimation,
SQL Server is not able to see the values of local variables at optimization time, as these
values are only known at execution time. However, by using local variables you are not
only disabling parameter sniffing, you're also disabling the Query Optimizer's option of
using the statistics histogram to find an optimal plan for the query. Instead, it will rely on
just the density information of the statistics object, as explained in Chapter 3.
This solution will simply ignore the parameter values and use the same execution plan
for all the executions, but at least you're getting a consistent plan every time. A varia-
tion of the OPTIMIZE FOR hint shown previously, OPTIMIZE FOR UNKNOWN, which was
introduced with SQL Server 2008, has the same effect.

214
Running the following two versions of our stored procedure will have equivalent
outcomes, and will produce the same execution plan. The first version uses local
variables, and the second one uses the new OPTIMIZE FOR UNKNOWN hint.
AS
DECLARE @p int = @pid
WHERE ProductID = @p
Listing 6-22.
AS
OPTION (OPTIMIZE FOR UNKNOWN)
Listing 6-23.
In this case, the Query Optimizer will create the plan using the Clustered Index Scan
shown previously.
Auto-parameterization
The Query Optimizer might decide to parameterize queries in those cases where the
value of a specific parameter does not impact the choice of an execution plan. That is, in
the cases where it does not matter which parameter value is used, the plan returned will
be the same.
This is a very conservative policy and SQL Server will only use it when it is safe to do
so, and the performance of the queries will not be negatively impacted. In this case, the

215
parameterized plan can be reused by similar queries which differ only in the value of their
parameters. This feature, which helps to avoid optimization time and cost, is called
auto-parameterization, and was introduced with SQL Server 7.0.
For example, the next two SQL statements, which were introduced as part of a stored
procedure in Listing 6-22, will produce different execution plans and will not be param-
eterized, even when the queries are syntactically identical and only the parameters are
different. In this case, the Query Optimizer decides that it isn't safe to auto-parameterize
them (and thereby reuse an existing execution plan).
Listing 6-24.
Listing 6-25.
On the other hand, the following query will be auto-parameterized:
SELECT * FROM Sales.SalesOrderHeader
WHERE SalesOrderID = 43669
Listing 6-26.
In this case, the SalesOrderID column is the primary key of the SalesOrderHeader
table, so it is guaranteed to be unique. In addition, the query predicate is using an equality
operator, so there will always be a maximum of one record returned. Given these factors,
SQL Server decides that it is safe to parameterize this plan by using a Clustered Index
Seek operator. You can verify if your query is using a parameterized plan by inspecting the
plan cache, as in the query shown in Listing 6-27.

216
SELECT text
FROM sys.dm_exec_cached_plans
CROSS APPLY sys.dm_exec_sql_text(plan_handle)
WHERE text LIKE '%SalesOrderID%'
Listing 6-27.
The output will include the following auto-parameterized query which will show
placeholders like @1 for the parameter values:
(@1 int)SELECT * FROM [Sales].[SalesOrderHeader] WHERE [SalesOrderID]=@1
Listing 6-28.
Forced parameterization
Finally, a new feature, called forced parameterization, was introduced in SQL Server
2005 to parameterize queries more aggressively. This feature is disabled by default and
can be enabled at the database level, or it can be used on an individual query by using the
PARAMETERIZATION FORCED query hint.
By enabling forced parameterization you can reduce the frequency of query optimiza-
tions, but you may also introduce suboptimal plans for some instances of those queries,
so you should do extensive analysis and testing of your application to verify that your
performance is, in fact, being improved. To differentiate it from forced parameterization,
auto-parameterization is also referred to as simple parameterization. For more
information about forced parameterization please consult Books Online.
To show how forced parameterization works, execute the statement in Listing 6-29 to
enable forced parameterization at the database level.

217
ALTER DATABASE AdventureWorks SET PARAMETERIZATION FORCED
Listing 6-29.
With this new configuration, the two queries in Listings 6-24 and 6-25, which returned
two distinct execution plans, will now be parameterized and produce only one plan. Run
the following query again:
Listing 6-30.
Because of the parameter used on this instance of the query, SQL Server will create a
plan using an Index Seek and a Key Lookup, which may be used by any similar query.
Of course, the risk for this specific example is that the first query will get a plan which
is tailored to its parameters, but which may be suboptimal for some other instances
of the same query with different parameters (like the ProductID 870 query used in
Listing 6-25). Run the following query to verify that the plan was, in fact, parameterized:
SELECT text
FROM sys.dm_exec_cached_plans
CROSS APPLY sys.dm_exec_sql_text(plan_handle)
WHERE text LIKE '%Sales%'
Listing 6-31.
It will show an output similar to this:
(@0 int)select * from Sales . SalesOrderDetail where ProductID = @0
Listing 6-32.

218
Finally, do not forget to set parameterization back to its default value by running the
ALTER DATABASE AdventureWorks SET PARAMETERIZATION SIMPLE
Listing 6-33.
Summary
The focus of the book so far has been on optimizing SELECT queries with joins and
aggregates. This chapter is the first time we've considered additional optimization topics
related to updates, data warehousing, parameter sniffing and auto-parameterization.
We've seen how the Query Optimizer decides between per-row and per-index plans
to optimize UPDATE statements, and we've examined how updates need to perform
additional operations like updating existing indexes, accessing additional tables, and
enforcing existing constraints.
Basic data warehousing concepts, including fact and dimension tables as well as star and
snowflake schemas, were introduced, with the focus being on how SQL Server optimizes
star join queries. Some optimizations, such as Cross products of dimension tables with
multi-column index lookups, and bitmap filtering, were also explained.
We've ended the chapter with topics related to the use of query parameters, and how
they affect both the query optimization process and the reuse of plans by the plan cache.
Topics like parameter sniffing, auto-parameterization and forced parameterization have
also been introduced.

219
Chapter 7: Hints
SQL is a declarative language; it only defines what data to retrieve from the database. It
doesn't describe the manner in which the data should be fetched. That, as we know, is
the job of the Query Optimizer, which analyzes a number of candidate execution plans
for a given query, estimates the cost of each of these plans, and selects an efficient plan by
choosing the cheapest of the choices considered.
But there may be cases when the execution plan selected is not performing as you have
expected and, as part of your query troubleshooting process, you may try to find a better
plan yourself. Before doing this, keep in mind that, just because your query does not
perform as you have expected, that does not mean that a better plan is always possible.
Your plan may be an efficient one, but this is, in fact, probably an expensive query to
perform, or your system may be having performance bottlenecks which are impacting the
query execution.
However, although the Query Optimizer does an excellent job most of the time, it does
occasionally fail to produce an efficient plan, as we've seen throughout this book. That
being said, even in the cases when you're not getting an efficient plan, you should still try
to distinguish between the times when the problems arise because you're not providing
the Query Optimizer with all the information it needs to do a good job, and those when
the problems are a result of a Query Optimizer limitation. Part of the focus of this book
so far has been to help you to provide the Query Optimizer with the information it needs
to produce an efficient execution plan, such as the right indexes or good quality statis-
tics, and also how to troubleshoot the cases when you are not getting a good plan. This
chapter will cover what to do if you hit a Query Optimizer limitation.
Having said that, there might be cases when the Query Optimizer just gets it wrong
and, in such cases, we may be forced to resort to the use of hints. These are essentially
optimizer directives which allow us to take explicit control over the execution plan for a
given query, with the goal of improving its performance. In reaching for a hint, however,

220
Chapter 7: Hints
we are going against the declarative property of the SQL language and, instead, giving
direct instructions to the Query Optimizer. Overriding the Query Optimizer is a risky
business; hints need to be used with caution, and only as a last resort when no other
option is available to produce a viable plan.
With this warning in mind, this chapter will review some of the hints that SQL Server
provides, should the need arise, as well as how and when they might be used. It does not
attempt to provide comprehensive coverage; indeed, we'll focus only on those hints I've
most often seen provide positive performance benefits in certain circumstances, and we'll
look at those in a few pages. Some other query hints, like OPTIMIZE FOR, OPTIMIZE FOR
UNKNOWN and RECOMPILE have already been covered in the Parameter Sniffing section of
Chapter 6, and will not be touched upon again in this chapter.
Before You Reach for a Hint…
Hints are a powerful means by which we can cause our decisions to overrule those of the
Query Optimizer. However, we should only do so with extreme caution, because hints
restrict the choices available to the Query Optimizer, will make your code less flexible,
and will require additional maintenance. A hint should only be employed once you're
certain that you have no alternative options. As a minimum, before you reach for a hint,
you should explore the potential issues below.
• Check for system problems – You need to make sure that your performance problem
is not linked to other system-related issues, such as blocking, or bottlenecks in server
resources such as I/O, memory, or CPU.
• Check for cardinality estimation errors – The Query Optimizer often misses the
correct plan because of cardinality estimation errors. Cardinality estimation errors
can be detected using the SET STATISTICS PROFILE ON statement, and can often be
fixed using solutions like updating statistics, using a bigger sample for your statistics
(or scanning the entire table), using computed columns, or filtered statistics, etc. There

221
Chapter 7: Hints
might be cases where the cardinality estimation errors are caused by the use of features
in which statistics are not supported at all, such as table variables or multi-statement
table-valued functions. In these particular instances you may consider using standard
or temporary tables if you are not getting an efficient plan. Statistics and cardinality
estimation errors are covered in more detail in Chapter 3, Statistics and Cost Estimation.
• Additional troubleshooting – You may need to perform additional troubleshooting
before considering the use of hints. One of the obvious choices for improving the
performance of your queries is providing the Query Optimizer with the right indexes.
How to make sure that your indexes are selected by the Query Optimizer is covered
in Chapter 4, Index Selection. You might also consider some other, less obvious trouble-
shooting procedures, like partitioning your query into steps or smaller pieces and
storing any intermediate results in temporary tables. Temporary tables can give you
the benefit of additional statistics which can help the Query Optimizer to produce
more efficient plans. You can use this method just as a troubleshooting procedure, for
example, to find out which part of the original query is expensive, so you can focus
on it. Alternatively, you can keep it as the final version of your query if these changes
alone give you better performance.
As discussed in this book's introduction, query optimizers have improved radically after
more than 30 years of research, but still face some technical challenges. The SQL Server
Query Optimizer will give you an efficient execution plan for most of your queries, but
will be increasingly challenged as the complexity of the query grows with more tables
joined, plus the use of aggregations, and other SQL features.
If, after investigating the troubleshooting options and recommendations described above
and throughout this book, you still find that the Query Optimizer is not finding a good
execution plan for your query, then you may need to consider using hints to direct the
Query Optimizer toward what you feel is the optimal execution path.
Always remember that, by applying a hint, you effectively disable some of the available
transformation rules to which the Query Optimizer usually has access, and so restrict the
available search space. Only transformation rules that help to achieve the requested plan

222
Chapter 7: Hints
will be executed. For example, if you use hints to force a particular join order, the Query
Optimizer will disable rules that reorder joins. Always try to use the least restrictive hint,
as this will retain as much flexibility as possible in your query, and make maintenance
somewhat easier. In addition, hints can not be used to generate an invalid plan or a plan
that the Query Optimizer normally would not consider during query optimization.
Furthermore, a hint that initially does a great job might actively hinder performance at
a later point in time when some conditions change; for example, as a result of schema
updates, service packs, new versions of SQL Server, or even enough data changes. The
hints may prevent the Query Optimizer from modifying the execution plan accord-
ingly, and thus result in degraded performance. It is your responsibility to monitor and
maintain your hinted queries to make sure that they continue to perform well after such
system changes or, even better, to remove them if they are no longer needed.
Plan guides
Plan guides can help in this scenario as they allow you to apply hints without changing the text of the
query directly. They separate the hint specification from the query itself, and so are an excellent choice
for applying a hint, or even specifying an entire plan, that can then be easily removed in the future. This
makes them particularly useful when dealing with third-party application code, or if you simply want
to apply hints in a more easily maintainable way. There is a whole section dedicated to exploring plan
guides, at the end of this chapter.
Remember, also, that if you decide to use a hint to change a single section or physical
operator of a plan, then after applying the hint, the Query Optimizer will perform a
completely new optimization. The Query Optimizer will obey your hint during the
optimization process, but it still has the flexibility to change everything else in the
plan, so the end result of your tweaking may be unintended changes to other sections
of the plan.
Finally, note that the Query Optimizer cannot perform miracles. The fact that your query
is not performing as you hoped does not always mean that the Query Optimizer is not

223
Chapter 7: Hints
giving you the best possible execution plan. If the operation you are performing is simply
expensive and resource intensive, then it's possible that no amount of tuning or hinting
will help you achieve the performance you'd like.
Types of Hints
SQL Server provides a wide range of hints which can be classified as follows:
• query hints tell the optimizer to apply "this hint" throughout the entire query and are
specified using the OPTION clause, which is included at the end of the query
• join hints apply to a specific join in a query, and can be specified by using the ANSI-
style join hints
• table hints apply to a single table and are usually included using the WITH keyword on
the FROM clause.
Another useful classification is dividing hints into physical operator and goal oriented
hints. Physical operator hints, as the name suggests, request the use of a specific physical
operator, join order or aggregation placement. On the other hand, a goal oriented hint
does not specify how to build the plan, but instead specifies a goal to achieve, leaving the
Query Optimizer to find the best physical operators to achieve that goal. Goal oriented
hints are usually safer and require less knowledge about the internal workings of the
Query Optimizer. Examples of goal oriented hints include the OPTIMIZER FOR or FAST N
hints. Almost all the remaining hints covered in this chapter are physical hints.
Locking hints do not affect plan selection, so they will not be covered here. Plan guides,
which allow you to apply a hint to a query without changing the code in your application,
and the USE PLAN query hint, which allows you to force the Query Optimizer to use a
specified execution plan for a query, are covered separately, later in the chapter.

224
Chapter 7: Hints
In the next few sections, I will discuss hints affecting joins, join order, aggregations,
index scans or seeks, views, and so on. Note that, with a very simple database like
AdventureWorks, the Query Optimizer will most likely give you an optimal plan
for all of the examples in this chapter, so I am just looking for alternate plans for
demonstration purposes.
Joins
We can explicitly ask the Query Optimizer to use any of the available join algorithms:
Nested Loops, Merge and Hash Join. We could do this at the query level, in which case
all the existing joins in the query will be affected, or we can specify it at the join level,
impacting only that join. However, this last choice will also impact the join order, as will
be explained in the FORCE ORDER section.
Let's focus on join hints at the query level first; in this case, the join algorithm is speci-
fied using the OPTION clause. You can also specify two of the three available joins, which
basically asks the Query Optimizer to exclude the third physical join operator from
consideration. The decision between which of the remaining two joins to use will be
cost-based. For example, the following unhinted query will produce the plan in Figure 7-1,
which uses a Hash Join.
Listing 7-1.

225
Chapter 7: Hints
Figure 7-1: Execution plan using a Hash Join.
On the other hand, the following query will request the Query Optimizer to exclude
a Hash Join by requesting either a Nested Loops or Merge Join. In this case, the Query
Optimizer chooses a Merge Join, as shown in the plan in Figure 7-2.
OPTION (LOOP JOIN, MERGE JOIN)
Listing 7-2.
Figure 7-2: Execution plan excluding a Hash Join.

226
Chapter 7: Hints
Join hints can not only force the joins we explicitly specify in our query text, but can
also impact most of the joins introduced by the Query Optimizer, such as foreign key
validation or cascading actions. Other joins, like the Nested Loops used in a bookmark
lookup, cannot be changed, as it would defeat the purpose of using the bookmark lookup
in the first place. For example, in the following query, the hint to use a Merge Join will be
ignored, as shown in the plan in Figure 7-3.
SELECT AddressID, City, StateProvinceID, ModifiedDate FROM Person.Address
WHERE City = 'Santa Fe'
OPTION (MERGE JOIN)
Listing 7-3.
Figure 7-3: Hint ignored in a bookmark lookup example.
As mentioned earlier, hints cannot force the Query Optimizer to generate invalid plans,
so the query in Listing 7-4 will not compile, as both Merge and Hash Joins require an
equality operator on the join predicate. Trying to execute this query will return the error
message shown in Listing 7-5.

227
Chapter 7: Hints
ON C.ContactID > I.ContactID
WHERE C.ContactID > 19974
OPTION (MERGE JOIN)
Listing 7-4.
Query processor could not produce a query plan because of the hints defined in this query.
Resubmit the query without specifying any hints and without using SET FORCEPLAN.
Listing 7-5.
However, as mentioned before, keep in mind that using the query-level hint will impact
the entire query. If you need explicit control over each join in a query, then you can use
ANSI-style join hints, the benefit of which is that a join type can be individually selected
for every join in the plan. However, be warned that using ANSI join hints will also add the
behavior of the FORCE ORDER hint, which asks to preserve the join order and aggregation
placement, as indicated by the query syntax. This behavior will be explained in the FORCE
ORDER section, later in this chapter.
In the meantime, let me show you an example. The following query without hints will
produce the execution plan shown in Figure 7-4:
Listing 7-6.

228
Chapter 7: Hints
Figure 7-4: Execution plan without hints.
The next query explicitly requests a Hash Join and a Nested Loops Join, and will produce
a different plan, as shown in Figure 7-5. Notice that the INNER keyword is required
this time.
INNER HASH JOIN Sales.Individual AS I
INNER LOOP JOIN Sales.Customer AS Cu
Listing 7-7.
In addition, the related warning (Listing 7-8) is shown in the Messages tab when the code
is executed using Management Studio.

229
Chapter 7: Hints
Figure 7-5: Execution plan with ANSI-style join hints.
Warning: The join order has been enforced because a local join hint is used.
Listing 7-8.
This warning indicates that not only was the join algorithm forced, but the join order
was forced as well, that is, the tables were joined using exactly the order specified in the
query text.
Aggregations
Just like join algorithms, aggregation algorithms can also be forced by using the GROUP
hints. Specifically, the ORDER GROUP hint requests that the Query Optimizer uses a
Stream Aggregate algorithm and the HASH GROUP hint requests a Hash Aggregate
algorithm. These hints can be specified only at the query level, so they will impact
all of the aggregation operations in the query. To see the effects of this, take a look
at the unhinted query in Listing 7-9, which produces the plan on Figure 7-6 using a
Stream Aggregate.

230
Chapter 7: Hints
SELECT SalesOrderID, COUNT(*)
Listing 7-9.
Figure 7-6: Execution plan using a Stream Aggregate.
Since the SalesOrderDetail table has a clustered index on the SalesOrderID
column, and so the data is already sorted on the GROUP BY column, using a Stream
Aggregate operator is the obvious choice. However, the following query will force a Hash
Aggregate operator, and will produce the plan shown in Figure 7-7, which will, of course,
make the query more expensive than necessary.
OPTION (HASH GROUP)
Listing 7-10.
Figure 7-7: Execution plan with a HASH GROUP hint.

231
Chapter 7: Hints
On the other hand, a scalar aggregation will always use a Stream Aggregate operator.
Trying to force a Hash Aggregate on a scalar aggregation, as in the following query, will
trigger the compilation error shown in Listing 7-5, complaining about the hints defined in
the query.
SELECT COUNT(*)
OPTION (HASH GROUP)
Listing 7-11.
FORCE ORDER
The FORCE ORDER hint can give the user full control over the join and aggregation place-
ment in an execution plan. Specifically, the FORCE ORDER hint asks the Query Optimizer
to preserve the join order and aggregation placement as indicated by the query syntax.
Notice, also, that the ANSI-style join hints explained before can also give you control
of the join order, in addition to control over the choice of the join algorithm. Both the
FORCE ORDER and ANSI-style join hints are very powerful, and because of that they need
to be used with caution. As explained earlier in this book, finding an optimum join order
is a critical part of the query optimization process, and also a challenging one, because
the sheer number of possible join orders can be huge even with queries involving only
a few tables. What this boils down to is that, by using the FORCE ORDER hint, you are
attempting to optimize the join order yourself.
You can use the FORCE ORDER hint to obtain any form of query, like left-deep trees, bushy
trees or right-deep trees, explained in Chapter 1, Introduction to Query Optimization. The
Query Optimizer will usually produce a left-deep tree plan, but you can force bushy trees
or right-deep trees by doing things like changing the location of the ON clause on the join
predicate, using subqueries, parenthesis, etc. Be aware that forcing join order does not
affect the simplification phase of query optimization, and some joins may still be removed
if needed, as explained in Chapter 5, The Optimization Process.

232
Chapter 7: Hints
If you do need to change the join order of a query for some reason, you can try starting
with the join order recommended by the Query Optimizer, and change only the part
which you think is suffering from a problem, such as cardinality estimation errors. You
can also follow the practices that the Query Optimizer itself would follow, as explained
in Chapter 2, The Execution Engine. For example, if you are forcing a Hash Join, select the
smallest table as the build input, or if you're forcing a Nested Loops Join, use small tables
in the outer input and the tables with indexes as the inner input. You could also start
by joining small tables first, or tables that can help to filter out the most possible number
of rows.
Let me show you an example. The query in Listing 7-12, without hints, will show you the
plan on Figure 7-8.
Listing 7-12.
Figure 7-8: Execution plan without hints.

233
Chapter 7: Hints
As you can see, the Query Optimizer does not follow the join order you have specified in
the query syntax; instead it found a better join order based on cost decisions. Now let's
see what happens if we change the query to use non-ANSI joins, by removing the ON
clauses and separating the table names with commas, and finally adding a FORCE ORDER
hint to the query. It will produce the plan in Figure 7-9.
FROM Person.Contact AS C, Sales.Individual AS I, Sales.Customer AS Cu
WHERE I.CustomerID = Cu.CustomerID
AND C.ContactID = I.ContactID
AND Cu.CustomerType = 'I'
Listing 7-13.
Figure 7-9: Execution plan with FORCE ORDER hint.
In this query using non-ANSI joins and the FORCE ORDER hint, the tables will be joined in
the order specified in the query, and by default will create a left-deep tree. On the other
hand, if you are using the FORCE ORDER hint in a query with ANSI joins, SQL Server will
consider the location of the ON clauses to define the location of the joins. As an example
of this phenomenon, the query in Listing 7-14 will create a similar plan to the one shown
in Figure 7-9 but, in this case, SQL Server is considering the location of the ON clauses

234
Chapter 7: Hints
and, as requested by the clause ON C.ContactID = I.ContactID, it's joining the
Contact and Individual tables first.
Listing 7-14.
In the next query, we are creating a right-deep tree (just to demonstrate that it's possible),
as we are requesting to join the Individual and Customer tables first, as requested
by the ON I.CustomerID = Cu.CustomerID clause. The resulting execution plan is
shown on Figure 7-10.
Listing 7-15.

235
Chapter 7: Hints
Figure 7-10: Plan forcing a right-deep tree.
In addition to taking control of join orders, as mentioned in the introduction of this
section, FORCE ORDER can also be used to force the order of aggregations. Consider this
unhinted example, which produces the plan seen in Figure 7-11:
Listing 7-16.
Figure 7-11: Plan with aggregation before the join.

236
Chapter 7: Hints
As you can see, in this case the Query Optimizer decided to perform the aggregation
before the join. (Remember that, as mentioned in Chapter 2, The Execution Engine, the
Query Optimizer can decide to perform aggregations before or after a join, depending on
the cost.) By adding a FORCE ORDER hint, as in the following query, the aggregation will
be moved to after the join, as shown in Figure 7-12.
Listing 7-17.
Figure 7-12: Plan with aggregation after the join.
Finally, a related statement, SET FORCEPLAN, can also be used to preserve the join order,
as indicated in the FROM clause of a query, but it will request Nested Loops Joins only.
A difference between that and the hints shown so far is that this statement needs to be
turned on, and will stay in effect until turned off. For more information regarding the SET
FORCEPLAN statement, please refer to Books Online.

237
Chapter 7: Hints
INDEX and FORCESEEK hints
The INDEX and FORCESEEK hints are table hints, and we'll consider each in turn.
The INDEX hint can be used to request the Query Optimizer to use a specific index
or indexes. Either the index id or the name of the index can be used as a target for the
Query Optimizer, but a name is the recommended way, as we do not have control of
the index id values for non-clustered indexes. However, if you still want to use index id
values, or you are interested in them for some other reason, they can be found on the
index_id column on sys.indexes; index id 0 is a heap, index id 1 is a clustered index,
and a value greater than 1 is a non-clustered index. On a query using a heap, using the
INDEX(0) hint results in a Table Scan operator being used, and INDEX(1) returns an
error message. However, a query with a clustered index can use both values: INDEX(0)
will force a Clustered Index Scan, and INDEX(1) can use either a Clustered Index Scan or
a Clustered Index Seek. On the other hand, the FORCESEEK hint can be used to force the
Query Optimizer to use an Index Seek operation, and can work on both clustered
or non-clustered indexes. It can also work in combination with the INDEX hint, as we'll
see later.
In addition to helping to improve the performance of your queries, in some cases you
may also want to consider using an index hint to minimize lock contention or deadlocks.
Notice that, when you use an INDEX hint, your query becomes dependent on the exist-
ence of the specified index, and it will not compile (or will stop working) if that index
is removed. Using FORCESEEK without an available index will also result in an error, as
shown later in this section.
You can also use the INDEX hint to avoid a bookmark lookup operation, as in the example
shown in Listing 7-18. Since the Query Optimizer estimates that only a few records will
be returned by the next query, it decides to use an Index Seek – Key Lookup combination,
as shown on Figure 7-13.

238
Chapter 7: Hints
Listing 7-18.
Figure 7-13: Plan without hints.
However, suppose that you want to avoid a bookmark lookup operation; you can use the
INDEX table hint to force a table scan instead (which could be the scan of either a heap
or a clustered index). The following query will force the use of a Clustered Index Scan
operator, as shown on the plan on Figure 7-14.
WITH (INDEX(0))
Listing 7-19.
Figure 7-14: Plan with an INDEX hint.
Using INDEX(1) in this example would give a similar result, as SQL Server cannot use the
clustered index to do an Index Seek operation; the clustered key is on SalesOrderID
and SalesOrderDetailID, so the only viable choice is to scan the clustered index.

239
Chapter 7: Hints
Of course, you can also force the opposite operation. In the following example, the Query
Optimizer estimates that a high number of records will be returned, and so it decides to
use a Clustered Index Scan, similar to the plan previously shown in Figure 7-14.
Listing 7-20.
Since we have an available index on ProductID (IX_SalesOrderDetail_ProductID),
we can force the plan to use that index.
WITH (INDEX(IX_SalesOrderDetail_ProductID))
Listing 7-21.
This query will produce a new plan, similar to the one in Figure 7-13 which, in fact, is
using an Index Seek on the IX_SalesOrderDetail_ProductID index, and a Key
Lookup to the base table, which in this case is the clustered index.
You can also achieve a similar result by forcing a seek using the FORCESEEK table hint,
which is new in SQL Server 2008. The following query will create a plan similar to the
one shown previously in Figure 7-13.
WITH (FORCESEEK)
Listing 7-22.

240
Chapter 7: Hints
You can even combine both hints to obtain the same results, as in the next query.
WITH (INDEX(IX_SalesOrderDetail_ProductID), FORCESEEK)
Listing 7-23.
Using FORCESEEK when SQL Server cannot do an Index Seek operation, as in the
following query, will not compile, and will instead return an error message.
WITH (FORCESEEK)
WHERE OrderQty = 1
Listing 7-24.
FAST N
FAST N is one of the so-called goal oriented hints, as it does not indicate what physical
operators to use, but instead just specifies what goal the plan is trying to achieve. This
hint is used to optimize a query to retrieve the first N rows of results as quickly as
possible. It can help in situations where only the first few rows returned by a query are
relevant, and perhaps you won't be using the remaining records of the query at all. The
price to pay for achieving this speed is that retrieving those remaining records may take
longer than if you had used a plan without this hint. In other words, since the query is
optimized to retrieve the first N records as soon as possible, retrieving all the records
returned by the query may be very expensive.

241
Chapter 7: Hints
The Query Optimizer usually accomplishes this FAST N goal by avoiding any blocking
operators, like Sort, Hash Join or Hash Aggregation, so the client submitting the query
does not have to wait before the first records are produced. Let's see an example; run the
following query, which returns the plan shown in Figure 7-15.
ORDER BY ProductID
Listing 7-25.
Figure 7-15: Plan without a hint.
In this case, the Sort operator is the most effective way to get the records sorted by
ProductID if you want to see the entire query output. However, since Sort is a blocking
operator, SQL Server will not produce any record until the sort is completed. Now,
supposing that your application wants to see a page with 20 records at a time, you can use
the FAST hint to get these 20 records as quickly as possible, as seen in the next query.
ORDER BY ProductID
OPTION (FAST 20)
Listing 7-26.
This time, the new plan, seen in Figure 7-16, scans an available non-clustered index while
performing Key Lookups to the clustered table. Since this plan uses random I/Os, it would
be very expensive for the entire query, but it will achieve the goal of returning the first 20
records very quickly.

242
Chapter 7: Hints
Figure 7-16: Plan using a FAST N hint.
There is also a FASTFIRSTROW hint, but it is not as flexible as FAST N , as you can
specify any number for N. Essentially, FASTFIRSTROW would be the same as specifying
the FAST 1 hint.
NOEXPAND, EXPAND VIEWS hints
Before talking about the NOEXPAND and EXPAND VIEWS hints, let me explain the default
behavior of queries when using indexed views so that you can see how these hints can
change this behavior.
As explained in Chapter 5, The Optimization Process, SQL Server expands views in the
early steps of query optimization during binding, when a view reference is expanded
to include the view definition; for example, to directly include the tables used in
the view. This behavior is the same for every edition of SQL Server. Later on in the
optimization process, but only in the Enterprise edition, SQL Server may match the
query to an existing indexed view. So, basically, the view was expanded at the beginning
but was later matched to an existing indexed view. The EXPAND VIEWS hint removes the
matching step, thus making sure the views are expanded but not matched at the end of
the optimization process. So this hint only has an effect in SQL Server Enterprise edition.
On the other hand, the NOEXPAND hint asks SQL Server not to expand any views at all,
and to try to use any existing indexed view instead. This hint works in every SQL Server

243
Chapter 7: Hints
edition, and it is the only way (when using a SQL Server edition other than Enterprise) to
ask SQL Server to match an existing view.
Here's an example. Create an indexed view on AdventureWorks by running the
following code:
CREATE VIEW v_test WITH SCHEMABINDING AS
SELECT SalesOrderID, COUNT_BIG(*) as cnt
GO
CREATE UNIQUE CLUSTERED INDEX ix_test ON v_test(SalesOrderID);
Listing 7-27.
Next, run the following query:
Listing 7-28.
If you are using SQL Server Enterprise edition (or the Enterprise Evaluation or Developer
editions, which share the same database engine edition), you will get the following plan,
which actually matches the existing indexed view, as shown in the plan in Figure 7-17.
Figure 7-17: Plan using an existing indexed view.

244
Chapter 7: Hints
Alternatively, you can use the EXPAND VIEWS hint, as in the following query, to avoid
matching the index view. You will get the plan seen in Figure 7-18.
OPTION (EXPAND VIEWS)
Listing 7-29.
Figure 7-18: Plan using the EXPAND VIEWS hint.
Finally, drop the indexed view you just have created:
DROP VIEW v_test
Listing 7-30.
Plan Guides
There might be situations when you need to apply a hint to a query, but you are unable
or unwilling to change your query code or your application. As mentioned earlier, a
common situation where this occurs is if you are working with third-party code or
applications, which you cannot change.

245
Chapter 7: Hints
Plan guides, a new feature introduced with SQL Server 2005, can help you in these
instances. Plan guides essentially work by keeping a list of queries on the server, along
with the hints you want to apply to them. To use a plan guide, you need to provide
SQL Server with the query that you want to optimize, and either a query hint using the
OPTION clause, or an XML plan using the USE PLAN hint, which will be explained in the
next section. When the query is optimized, SQL Server will apply the hint requested
in the plan guide definition. You can also specify NULL as a hint in your plan guide to
remove an existing hint in your application.
As well as allowing you to apply hints to code which you can't or don't want to change,
plan guides make it easier to apply, update, and remove query hints. Plan guides can also
match queries in different contexts; for example, a stored procedure, a user-defined scalar
function, or a stand-alone statement which is not part of any database object.
You can use the sp_create_plan_guide stored procedure to create a plan guide, and
the sp_control_plan_guide to drop, enable or disable plan guides. For more details
on how to use these stored procedures, you should investigate Books Online, which has
much more detail than we could cover here. You can see which plan guides are defined in
your database by looking at the sys.plan_guides catalog view.
To make sure that the query in the plan guide definition matches the query being
executed, especially for stand-alone statements, you can use the Profiler's Plan Guide
Successful event class, which will show whether an execution plan was successfully
created using a plan guide. On the other hand, the Plan Guide Unsuccessful
event will show if SQL Server was unable to create an execution plan using a plan
guide, meaning that the query was instead optimized without it. You can see the Plan
Guide Unsuccessful event, for example, when trying to force a Merge or Hash Join
with a non-equality operator in the join condition, as shown in Listing 7-4, earlier in
this chapter.
Let's see an example of these events. Suppose we want to use plan guides to avoid a Merge
or Hash Join in our previous query, in order to avoid high memory usage. Before running
this code, open a SQL Server Profiler session, connect it to your instance of SQL Server,

246
Chapter 7: Hints
select the blank template to start a new trace definition, and select both Plan Guide
Successful and Plan Guide Unsuccessful on the Performance section of the Events
tab, and then start the trace.
Next, create the following stored procedure:
CREATE PROCEDURE test
AS
FROM Person.Contact AS C JOIN Sales.Individual AS I
Listing 7-31.
Before creating a plan guide, execute the stored procedure and display its execution plan
to verify that it is using a Hash Join operator.
EXEC test
Listing 7-32.
Once you've confirmed that, create a plan guide to force the query to use a Nested
Loops Join.
EXEC sp_create_plan_guide
@name = N'plan_guide_test',
@stmt = N'SELECT FirstName, LastName
ON C.ContactID = I.ContactID',
@type = N'OBJECT',
@module_or_batch = N'test',
@params = NULL,
@hints = N'OPTION (LOOP JOIN)';
Listing 7-33.

247
Chapter 7: Hints
Now, if you execute the stored procedure again, you can verify that it is now using a
Nested Loops Join operator, as shown in the plan in Figure 7-19.
Figure 7-19: Plan using a plan guide.
In addition, during this execution SQL Server Profiler should capture a Plan Guide
Successful event, showing that SQL Server was able to use the defined plan guide.
The TextData column will show the name of the plan guide, which in this case is
plan_guide_test, as shown in Figure 7-20.
Figure 7-20: Capturing a Plan Guide Successful event.

248
Chapter 7: Hints
Once you've created your plan guide, you can enable or disable it at any time. For
example, the following statement will disable the previous plan guide, and the stored
procedure will again use a Hash Join when executed.
EXEC sp_control_plan_guide N'DISABLE', N'plan_guide_test';
Listing 7-34.
To enable the plan guide again, use:
EXEC sp_control_plan_guide N'ENABLE', N'plan_guide_test';
Listing 7-35.
Finally, to clean up, drop both the plan guide and the stored procedure. Note that
you need to drop the plan guide first, as you cannot drop a stored procedure that it is
currently referenced by a plan guide.
EXEC sp_control_plan_guide N'DROP', N'plan_guide_test';
DROP PROCEDURE test
Listing 7-36.
USE PLAN
Finally, let's take a look at the USE PLAN query hint, which was introduced with SQL
Server 2005. This takes the use of hints to the extreme by allowing the user to specify an
entire execution plan as a target to be used to optimize a query. The USE PLAN hint is
useful when you know that a better plan than the Query Optimizer's suggestion exists.

249
Chapter 7: Hints
This can be the case, for example, when a better performing plan was produced in the
past, or in a different system, or even in a previous version of SQL Server.
The plan should be specified in XML format, and you will usually use SQL Server itself
to generate the XML text for the desired plan, as it can be extremely difficult to write an
XML plan manually.
The USE PLAN hint can force most of the specified plan properties, including the tree
structure, join order, join algorithms, aggregations, sorting and unions, and index
operations like scans, seeks and intersections, so that only the transformation rules
that can be useful in finding the desired plan are executed. In addition, USE PLAN now
supports UPDATE statements (INSERT, UPDATE, DELETE and MERGE), which was not
the case when the hint was first introduced in SQL Server 2005. Some statements still
not supported include full-text or distributed queries, and queries with dynamic, keyset-
driven and forward-only cursors.
Suppose we have the same query we saw in the plan guides section, which produces a
Hash Join ...
Listing 7-37.
… and suppose that you want SQL Server to use a different execution plan, which we can
generate using a hint:
OPTION (LOOP JOIN)
Listing 7-38.

250
Chapter 7: Hints
You can force this new plan to use a Nested Loops Join instead of a Hash Join. In order
to accomplish that, display the new XML plan (by right-clicking on the graphical plan
and selecting Show Execution Plan XML …), copy it to an editor, replace all of the single
quotes with double quotes, and then copy the plan to the query, as shown below.
OPTION (USE PLAN N'<?xml version="1.0" encoding="utf-16"?> …
</ShowPlanXML>')
Listing 7-39.
Of course, the XML plan is too long to display here, so I've just displayed the start and
end. Make sure the query ends with ‘) after the XML plan. Running the SELECT state-
ment above will request SQL Server to try to use the indicated plan, and the query will be
executed with a Nested Loops Join, as requested in the provided XML execution plan.
You can combine both plan guides and the USE PLAN query hint to force a specific execu-
tion plan in a situation where you don't want to change the text of the original query. The
following (and final) query will use the same example included in Listing 7-33 in the plan
guides section, together with the XML plan generated a moment ago. Note the use of two
single quotes before the XML plan specification, meaning that, this time, the query text
needs to end with ’’)’.
EXEC sp_create_plan_guide
@name = N'plan_guide_test',
@stmt = N'SELECT FirstName, LastName
ON C.ContactID = I.ContactID',
@type = N'OBJECT',
@module_or_batch = N'test',
@params = NULL,
@hints = N'OPTION (USE PLAN N''<?xml version="1.0" encoding="utf-16"?> …
Listing 7-40.

251
Chapter 7: Hints
Finally, bear in mind that, when the USE PLAN hint is used directly in a query, an invalid
plan will make the query fail. However, when the USE PLAN hint is used in a plan guide,
an invalid plan will simply compile the query without the requested hint, as mentioned in
the previous section.
Summary
The Query Optimizer typically selects a good execution plan for your queries, but there
may still be cases when you are not getting good performance from a selected plan, even
after extensive troubleshooting. Although hints can be used to improve the performance
of a query in these cases by directly taking control of the execution plan selection, they
should always be used with caution, and only as a last resort. You should also be aware
that code using hints will require additional maintenance, and is significantly less flexible
to changes in your database, application or software upgrades.
This chapter explained how to use hints to force join algorithms, join order, aggregations,
indexes for both scan or seek operations, and the use of indexed views, among other
behaviors. We also examined the use of plan guides to implement hints without changing
the code of your (or third-party) applications, and the ability of the USE PLAN hint to
specify an entire XML plan as the target of the optimization.
Finally, my hope is that the chapters of this book have provided you with the knowl-
edge needed to write better queries, and to give the Query Optimizer the information
it needs to produce efficient execution plans. At the same time, I hope you've seen more
about how to get the information you need to diagnose and troubleshoot the cases when
(despite your best efforts) you are not getting a good plan. In addition, having seen how
the Query Optimizer works, and some of the limitations this complex piece of software
still faces today, you can be better prepared to decide when and how hints can be used to
improve the performance of your queries.

About Red Gate
You know those annoying jobs that spoil your day
whenever they come up?
Writing out scripts to update your production database,
or trawling through code to see why it’s running so slow.
Red Gate makes tools to fix those problems for you.
Many of our tools are now industry standards. In fact,
at the last count, we had over 650,000 users.
But we try to go beyond that. We want to support you
and the rest of the SQL Server and .NET communities in
any way we can.
First, we publish a library of free books on .NET and
SQL Server. You’re reading one of them now. You can
get dozens more from www.red-gate.com/books
Second, we commission and edit rigorously accurate articles from experts on the
front line of application and database development. We publish them in our online
journal Simple Talk, which is read by millions of technology professionals each year.
On SQL Server Central, we host the largest SQL Server community
in the world. As well as lively forums, it puts out a daily dose of
distilled SQL Server know-how through its newsletter, which now
has nearly a million subscribers (and counting).
Third, we organize and sponsor events (about 50,000 of you
came to them last year), including SQL in the City, a free event
for SQL Server users in the US and Europe.
So, if you want more free books and articles, or to get
sponsorship, or to try some tools that make your life easier,
then head over to www.red-gate.com

Inside the sql server query optimizer

More Related Content

Similar to Inside the sql server query optimizer (20)

More from Kesavan Munuswamy (18)

Recently uploaded (20)

Inside the sql server query optimizer