12 Ways To Rewrite SQL Queries for Better Performance

Watch this week's video on YouTube

Over the past several week's I've been exploring ways to rewrite queries to improve execution performance.

I learned a lot of these techniques over time from trial an error, attending presentations, reading blog posts, speaking to other dbas and developers, etc... but never knew of a good resource that summarized these techniques in one place.

This post will be a quick round-up of everything I've covered so far, as well as 8 additional techniques that I use occasionally but don't necessarily require a full detailed post to explain them.

Why Rewrite Queries?

I often find myself working in environments where modifying indexes or changing server settings is out of the question when performance tuning. I usually run into these scenarios when dealing with:

  • Vendor databases
  • "Fragile" systems
  • Not enough disk space
  • Limited tooling/ad hoc analysis
  • Features limited by security software

While solving the root cause of a performance problem is always preferable, sometimes the only way I'm able to fix problems in these environments is by rewriting the queries.

I decided to write this summary post because it is a resource I would have loved to have when starting out. Sometimes it can be easy to get "writer's block" when trying to think of ways to rewrite a SQL query, so hopefully this list of techniques can provide ideas and get your creative juices flowing.

So, without further ado, here is a list of 12 techniques in no particular order that you can use to rewrite your queries to change their performance.

12 Ways to Refactor a Query to Change Performance

1. Window functions vs GROUP BY

Sometimes window functions rely a little too much on tempdb and blocking operators to accomplish what you ask of them. While using them is always my first choice because of their simple syntax, if they perform poorly you can usually rewrite them as an old-fashioned GROUP BY to achieve better performance.

2. Correlated subqueries vs derived tables

Many people like using correlated subqueries because the logic is often easy to understand, however switching to derived table queries often produces better performance due to their set-based nature.

3. IN vs UNION ALL

When filtering rows of data on multiple values in tables with skewed distributions and non-covering indexes, writing your logic into multiple statements joined with UNION ALLs can sometimes generate more efficient execution plans than just using IN or ORs.

4. Temporary Staging Tables

Sometimes the query optimizer struggles to generate an efficient execution plan for complex queries. Breaking a complex query into multiple steps that utilize temporary staging tables can provide SQL Server with more information about your data. They also cause you to write simpler queries which can cause the optimizer to generate more efficient execution plans as well as allow it to reuse result sets more easily.

5. Forcing Table Join Orders

Sometimes outdated statistics and other insufficient information can cause the SQL Server query optimizer to join tables in a less than ideal sequence. Adam Machanic has a fantastic presentation on forcing table join order with blocking operators without having to resort to join hints.

6. DISTINCT with few unique values

Using the DISTINCT operator is not always the fastest way to return the unique values in a dataset. In particular, Paul White uses recursive CTEs to return distinct values on large datasets with relatively few unique values. This is a great example of solving a problem using a very creative solution.

7. Eliminate UDFs

UDFs often cause poor query performance due to forcing serial plans and causing inaccurate estimates. One way to possibly improve the performance of queries that call UDFs is to try and inline the UDF logic directly into the main query. With SQL Server 2019 this will be something that happens automatically in a lot of cases, but as Brent Ozar points out you might occasionally have to manually inline a UDF's functionality to get the best performance.

8. Create UDFs

Sometimes a poorly configured server will parallelize queries too frequently and cause poorer performance than their serially equivalent plan. In those cases, putting the troublesome query logic into a scalar or multi-statement table-valued function might improve performance since they will force that part of the plan to run serially. Definitely not a best practice, but it is one way to force serial plans when you can't change the cost threshold for parallelism.

9. Data Compression

Not only does data compression save space, but on certain workloads it can actually improve performance. Since compressed data can be stored in fewer pages, read disk speeds are improved, but maybe more importantly the compressed data allows more to be stored in SQL Server's buffer pool, increasing the potential for SQL Server to reuse data already in memory.

10. Indexed Views

When you can't add new indexes to existing tables, you might be able to get away with creating a view on those tables and indexing the view instead. This works great for vendor databases where you can't touch any of the existing objects.

11. Switch cardinality estimators

The newer cardinality estimator introduced in SQL Server 2014 improves the performance of many queries. However, in some specific cases it can make queries perform more slowly. In those cases, a simple query hint is all you need to force SQL Server to change back to the legacy cardinality estimator.

12. Copy the data

If you can't get better performance by rewriting a query, you can always copy the data you need to a new table in a location where you CAN create indexes and do whatever other helpful transformations you need to do ahead of time.

...And more

By no means is this list exhaustive. There are so many ways to rewrite queries, and not all of them will work all the time.

The key is to think about what the query optimizer knows about your data and why it's choosing the plan it is. Once you understand what it's doing, you can start getting creative with various query rewrites that address that issue.

Temporary Staging Tables

Watch this week's video on YouTube

SQL Server Spool operators are a mixed bag. On one hand, they can negatively impact performance when writing data to disk in tempdb. On the other hand, they allow filtered and transformed result sets to be temporarily staged, making it easier for that data to be reused again during that query execution.

The problem with the latter scenario is that SQL Server doesn't always decide to use a spool; often it's happy to re-read (and re-process) the same data repeatedly. When this happens, one option you have is to explicitly create your own temporary staging table that will help SQL Server cache data it needs to reuse.

This post is a continuation of my series to document ways of refactoring queries for improved performance. I'll be using the StackOverflow 2014 data dump for these examples if you want to play along at home.

No spools

Let's start by looking at the following query:

WITH January2010Badges AS ( 
    SELECT 
        UserId,
        Name,
        Date
    FROM 
        dbo.Badges 
    WHERE 
        Date >= '2010-01-01' 
        AND Date <= '2010-02-01' 
), Next10PopularQuestions AS ( 
    SELECT TOP 10 * FROM (SELECT UserId, Name, Date FROM January2010Badges WHERE Name = 'Popular Question' ORDER BY Date OFFSET 10 ROWS) t
), Next10NotableQuestions AS ( 
    SELECT TOP 10 * FROM (SELECT UserId, Name, Date FROM January2010Badges WHERE Name = 'Notable Question' ORDER BY Date OFFSET 10 ROWS) t
), Next10StellarQuestions AS ( 
    SELECT TOP 10 * FROM (SELECT UserId, Name, Date FROM January2010Badges WHERE Name = 'Stellar Question' ORDER BY Date OFFSET 10 ROWS) t
)
SELECT UserId, Name FROM Next10PopularQuestions 
UNION ALL 
SELECT UserId, Name FROM Next10NotableQuestions
UNION ALL 
SELECT UserId, Name FROM Next10StellarQuestions 

Note: This is not necessarily the most efficient way to write this query, but it makes for a good demo.

This query is returning offset results for different badges from one month of data in the dbo.Badges table. While the query is using a CTE to make the logic easy to understand (i.e. filter the data to just January 2010 results and then calculate our offsets based on those results), SQL Server isn't actually saving the results of our January2010Badges expression in tempdb to get reused. If we view the execution plan, we'll see it reading from our dbo.Badges clustered index three times:

2019-05-06-18-30-12

Table 'Badges'. Scan count 27, logical reads 151137, ...

That means every time SQL Server needs to run our offset logic in each "Next10..." expression, it needs to rescan the entire clustered index to first filter on the Date column and then the Name column. This results in about 150,000 logical reads.

Divide and Conquer

One potential solution would be to add a nonclustered index that would allow SQL Server to avoid scanning the entire clustered index three times. But since this series is about improving performance without adding permanent indexes (since sometimes you are stuck in scenarios where you can't easily add or modify an index), we'll look at mimicking a spool operation ourselves.

We'll use a temporary table to stage our filtered January 2010 results so SQL Server doesn't have to scan the clustered index each time it needs to perform logic on that subset of data. For years I've referred to this technique as "temporary staging tables" or "faking spools", but at a recent SQL Saturday Jeff Moden told me he refers to it as "Divide and Conquer". I think that's a great name, so I'll use it going forward. Thanks Jeff!

First let's divide our query so that we insert our January 2010 data into its own temporary table:

DROP TABLE IF EXISTS #January2010Badges;
CREATE TABLE #January2010Badges
(
    UserId int,
    Name nvarchar(40),
    Date datetime
    CONSTRAINT PK_NameDateUserId PRIMARY KEY CLUSTERED (Name,Date,UserId)
);

INSERT INTO #January2010Badges
SELECT
    UserId,
    Name,
    Date
FROM 
    dbo.Badges
WHERE 
    Date >= '2010-01-01' 
    AND Date <= '2010-02-01'; 

You'll notice I added a clustered primary key which will index the data in an order that will make filtering easier.

Next, we conquer by changing the rest of our query to read from our newly created temp table:

WITH Next10PopularQuestions AS ( 
    SELECT TOP 10 * FROM (SELECT UserId, Name, Date FROM #January2010Badges WHERE Name = 'Popular Question' ORDER BY Date OFFSET 10 ROWS) t
), Next10NotableQuestions AS ( 
    SELECT TOP 10 * FROM (SELECT UserId, Name, Date FROM #January2010Badges WHERE Name = 'Notable Question' ORDER BY Date OFFSET 10 ROWS) t
), Next10StellarQuestions AS ( 
    SELECT TOP 10 * FROM (SELECT UserId, Name, Date FROM #January2010Badges WHERE Name = 'Stellar Question' ORDER BY Date OFFSET 10 ROWS) t
)
SELECT UserId, Name FROM Next10PopularQuestions 
UNION ALL 
SELECT UserId, Name FROM Next10NotableQuestions 
UNION ALL 
SELECT UserId, Name FROM Next10StellarQuestions 

Running this all together, we get the following plans and logical read counts:

2019-05-06-18-35-56

Table 'Badges'. Scan count 9, logical reads 50379, ...

(42317 rows affected)

(20 rows affected)
Table '#January2010Badges______________________________00000000003B'. Scan count 3, logical reads 12, ...

In this version of the query, SQL Server scans the clustered index a single time and saves that data to a temporary table. In the subsequent SELECTs, it seeks from this much smaller temporary table instead of going back to the clustered index, reducing the total amount of reads to 50379 + 12 = 50392: about a third of what the original query was doing.

Temporary Staged Data

At the end of day, you can hope that SQL Server creates a spool to temporarily stage or data, or you can be explicit about it and do it yourself. Either option is going to increase usage on your tempdb database, but at least by defining the temporary table yourself you can customize and index it to achieve maximum reuse and performance for your queries.

It's important to note that this is not a technique you want to abuse: writing and reading too much data from tempdb can cause contention problems that can make you worse off than having allowed SQL Server to scan your clustered index three times. However, when implemented sparingly and for good reasons, this technique can greatly improve the performance of certain queries.

IN vs UNION ALL

Watch this week's video on YouTube

When you need to filter query results on multiple values, you probably use an IN() statement or multiple predicates separated by ORs:

WHERE Col1 IN ('A','B','C')

or

WHERE Col1 = 'A' OR Col1 = 'B' OR Col1 = 'C'

While SQL Server will generate the same query plan for either syntax, there is another technique you can try that can sometimes can improve performance under certain conditions: UNION ALL.

This post is a continuation of my series to document ways of refactoring queries for improved performance. I'll be using the StackOverflow 2014 data dump for these examples if you want to play along at home.

Lookups and Scans

Let's say we have the following index on our dbo.Badges table:

CREATE NONCLUSTERED INDEX [IX_Badges] ON [dbo].[Badges] ([Name]) INCLUDE ([UserId]);

Next let's run these two separate queries:

/* Query 1 */
SELECT 
    Name, UserId, Date 
FROM 
    dbo.Badges 
WHERE 
    Name = 'Benefactor'
OPTION(MAXDOP 1)

/* Query 2 */
SELECT 
    Name, UserId, Date 
FROM 
    dbo.Badges 
WHERE 
    Name = 'Research Assistant'
OPTION(MAXDOP 1)

Note I'm enforcing MAXDOP 1 here to remove any performance differences due to parallelism in these demos.

The nonclustered index doesn't cover these queries - while SQL Server can seek the index for the Name predicate in the WHERE clause, it can't retrieve all the columns in the SELECT from the index alone. This leaves SQL Server with a tough choice to make:

  1. Does it scan the whole clustered index to return all the required columns for the rows requested?
  2. Does it seek to the matching records in the nonclustered index and then perform a key lookup to retrieve the remaining data?

So, what does SQL Server decide to do?

2019-04-26-15-46-29

For Query 1, SQL Server thinks that reading the entire clustered index and returning only the rows where Name = 'Benefactor' is the best option.

SQL Server takes a different approach for Query 2 however, using the non-covering nonclustered indexes to find the records with Name = 'Research Assistant' and then going to look up the Date values in the clustered index via a Key Lookup

The reason SQL server chooses these two different plans is because it thinks it will be faster to return smaller number of records with a Seek + Key Lookup approach ("Research Assistant", 127 rows), but faster to return a larger number of records with a Scan ("Benefactor", 17935 rows).

Kimberly Tripp has an excellent post that defines where this "tipping point" from a key lookup to a clustered index scan typically occurs, but the important thing to keep in mind for this post is that we can sometimes use SQL Server's ability to switch between these two approaches to our advantage.

Combining Queries with IN

So, what plan does SQL Server generate when we combine our two queries into one?

SELECT 
    Name, UserId, Date 
FROM 
    dbo.Badges 
WHERE 
    Name IN ('Benefactor','Research Assistant')
OPTION(MAXDOP 1)

2019-04-25-21-39-29

Interestingly enough SQL Server decides to retrieve the requested rows from the nonclustered index and then go lookup the remaining Date column in the clustered index.

If we look at the page reads (SET STATISTICS IO ON;) we'll see SQL Server had to read 85500 pages to return the data requested:

(18062 rows affected)
Table 'Badges'. Scan count 2, logical reads 85500, physical reads 20, read-ahead reads 33103, ...

Without correcting our index to include the Date column, is there some way we can achieve the same results with better performance?

UNION ALL

In this case it's possible to rewrite our query logic to use UNION ALL instead of IN/ORs:

SELECT 
    Name,UserId,Date 
FROM 
    dbo.Badges 
WHERE 
    Name = 'Benefactor' 
UNION ALL
SELECT 
    Name,UserId,Date 
FROM 
    dbo.Badges 
WHERE 
    Name = 'Research Assistant'
OPTION(MAXDOP 1)

2019-04-25-21-40-09

We get the same exact results through a hybrid execution plan.

In this case, our plan mirrors what SQL Server did when running our original two queries separately:

  • The rows where Name = 'Benefactor' are returned by scanning the clustered index.
  • The nonclustered index is seeked with clustered index lookups for the Name = 'Research Assistant' records.

Looking at the IO statistics for this UNION ALL query:

(18062 rows affected)
Table 'Badges'. Scan count 2, logical reads 50120, physical reads 6, read-ahead reads 49649, ...

Even though this query reads the whole clustered index to get the Benefactor rows, the total number of logical reads is still smaller than the seek/key lookup pattern seen in the combined query with IN(). This UNION ALL version gives SQL Server the ability to build a hybrid execution plan, combining two different techniques to generate a plan with fewer overall reads.

IN or UNION ALL?

There's no way to know for sure without trying each variation.

But if you have a slow performing query that is filtering on multiple values within a column, it might be worth trying to get SQL Server to use a different plan by rewriting the query.

Correlated Subqueries vs Derived Tables

Watch this week's video on YouTube

Correlated subqueries provide an intuitive syntax for writing queries that return related data. However, they often perform poorly due to needing to execute once for every value they join on.

The good news is that many correlated subqueries can be rewritten to use a derived table for improved performance.

This post is a continuation of my series to document ways of refactoring queries for improved performance. I'll be using the StackOverflow 2014 data dump for these examples if you want to play along at home.

When was each user's first badge awarded?

StackOverflow awards users badges for things like asking good questions, hitting certain vote thresholds, and more.

I want to write a query that figures out on what date did each user receive their first badge.

Using a correlated subquery, I might write my query as follows:

SET STATISTICS IO, TIME ON;

SELECT DISTINCT
    UserId,
    FirstBadgeDate = (SELECT MIN(Date) FROM dbo.Badges i WHERE o.UserId = i.UserId)
FROM
    dbo.Badges o

The syntax of the correlated subquery here makes it clear that for each UserId we want to return the MIN(Date) associated with that UserId from the badges table.

Looking at the execution plan and time and IO statistics (abbreviated for clarity) we see:

2019-04-18-07-11-57

(1318413 rows affected)
Table 'Worktable'. Scan count 0, logical reads 0, ...
Table 'Workfile'. Scan count 0, logical reads 0, ...
Table 'Badges'. Scan count 2, logical reads 43862, ...

(1 row affected)

 SQL Server Execution Times:
   CPU time = 3625 ms,  elapsed time = 8347 ms.

So, what's going on here? We read ~8 million rows of data from our index on the dbo.Badges table and then calculate the MIN(Date) for each UserId. This is the "correlated" part of our query, which then gets joined back to the dbo.Badges table using a Hash Match join operator.

Our join doesn't eliminate any rows so the ~8 million rows continue flowing through until near the very end where we have another Hash Match operator, this time being used to dedupe the rows for the DISTINCT part of query, reducing the final result to ~1 million rows.

Eliminating the Correlated Subquery

What would things look like if we rewrote this correlated subquery as a derived table in the FROM clause?

SELECT DISTINCT
    o.UserId,
    FirstBadgeDate
FROM
    dbo.Badges o
    INNER JOIN 
        (SELECT 
            UserId, 
            MIN(Date) as FirstBadgeDate 
        FROM 
            dbo.Badges GROUP BY UserId
        ) i
    ON o.UserId = i.UserId

2019-04-18-07-26-36

(1318413 rows affected)
Table 'Workfile'. Scan count 0, logical reads 0, ...
Table 'Worktable'. Scan count 0, logical reads 0, ...
Table 'Badges'. Scan count 2, logical reads 43862, ...

(1 row affected)

 SQL Server Execution Times:
   CPU time = 2516 ms,  elapsed time = 5350 ms.

If we look at the IO statistics, it's interesting to note that there is no difference in reads between these two queries.

Looking at the CPU time statistics however, this derived table query consistently comes in about 33% faster than the correlated subquery example. Why is that?

Looking at the execution plan reveals some details: in this plan, you can see we read in from the dbo.Badges index and go straight into a Hash Match operator. The top stream is deduping our data on UserId, taking it from ~8 million rows to ~1 million rows. The bottom stream does the same deduping while also calculating the MIN(DATE) for each UserId grouping.

When both of those streams join together, the final hash match operator is only joining ~1 million rows with ~1 million rows (as opposed to the first query that was joining ~8 million rows with ~1 million rows).

This last join is the reason for the performance improvement: because this execution plan can reduce the number of rows sooner the final join ends up having to do less work. Additionally, the records were already distinct going into the join, saving us from an extra deduping step.

Further Reducing Redundancy

You may have noticed that both of these queries are a little redundant: they both call on the dbo.Badges table unnecessarily. The best option to improve query performance would be to rewrite it as:

SELECT 
    UserId, 
    MIN(Date) as FirstBadgeDate 
FROM 
    dbo.Badges 
GROUP BY 
    UserId

2019-04-18-07-48-58-1

While this is the most efficient query of the three, most real-world queries and scenarios aren't this easy to simplify.

When your queries have more joins, WHERE clauses, and more, knowing how to refactor from a correlated subquery to a derived table query is critical to potentially improving performance.

Window Functions vs GROUP BYs

Watch this week's video on YouTube

There are many options available for improving the performance of a query: indexes, statistics, configuration settings, etc...

However, not all environments allow you to use those features (eg. vendor databases), leaving query rewriting as the only option.

This is the first post in a series to document common ways to refactor queries without otherwise altering the database. The goal of these posts will be to provide examples of performance pitfalls in queries and how to rewrite those queries to generate different query plans that (hopefully) improve performance.

I'll be using the StackOverflow 2014 data dump for these examples if you want to play along at home.

Who was first to earn each badge?

StackOverflow awards users badges for things like asking good questions, hitting certain vote thresholds, and more.

I want to write a query that figures out who is the first person awarded each badge. In cases where there is a tie for the first person to receive that badge, I want to return the user with the lowest UserId.

Window functions make this type of question easy to write a query for:

SELECT DISTINCT
    Name,
    FIRST_VALUE(UserId) OVER (PARTITION BY Name ORDER BY Date,UserId) AS UserId
FROM
    dbo.Badges b
ORDER BY
    Name,UserId

If you've used FIRST_VALUE before, this query should be easy to interpret: for each badge Name, return the first UserId sorted by Date (earliest date to receive the badge) and UserId (pick the lowest UserId when there are ties on Date).

This query was easy to write and is simple to understand. However, the performance is not great: it takes 46 seconds to finish returning results on my machine.

2019-04-11-20-45-45

Note: I assumed this table started off with the following index:

CREATE NONCLUSTERED INDEX IX_Badges__Name_Date_UserId ON [dbo].[Badges] (Name,Date,UserId);

Why so slow?

If we SET STATISTICS IO ON we'll notice that SQL Server reads 46767 pages from a nonclustered index. Since we aren't filtering our data, there's not much we can do to make that faster.

Reading right to left, next up we see two Segment operators. These don't add much overhead since our data is sorted on our segments/groups, so making SQL Server identify when our sorted rows change values is trivial.

Next up is the Window Spool operator which "Expands each row into the set of rows that represent the window associated with it." While it looks innocent by having a low relative cost, this operator is writing 8 million rows/reading 16 million rows (because of how Window Spool works) from tempdb. Ouch.

After that the Stream Aggregate operator and Compute Scalar operators check to see if the first value in each window being returned from the Window Spool is null and then return the first non-null value. These operations are also relatively painless since the data flowing through is already sorted.

The Hash Match operator then dedupes the data for our DISTINCT and then we sort the remaining ~2k rows for our output.

So while our query looks simple, the fact that our whole table of data is getting written to and read from tempdb before being deduped and sorted is a real performance killer.

Removing tempdb usage the old-fashioned way

When I say "the old fashioned way", I mean rewriting our window function to use more traditional aggregate functions and a GROUP BY:

SELECT
    b.Name,
    MIN(b.UserId) AS UserId
FROM
    dbo.Badges b
    INNER JOIN
    (
    SELECT
        Name,
        MIN(Date) AS Date
    FROM
        dbo.Badges
    GROUP BY
        Name
    ) m
        ON b.Name = m.Name
        AND b.Date = m.Date
GROUP BY
    b.Name
ORDER BY
    Name,UserId

I think by most people's standards, this query is not as easy to read. While not overly complex, it does take up a lot more screen space and is complicated by multiple GROUP BYs and a derived table.

And while the query may look ugly on the outside, it's what lies below the surface that really matters:

2019-04-11-20-49-58

What a beautifully simple execution plan. And it finishes executing almost instantly.

Let's break down what's going on. First, we start with similar Index Scan and Segment operators as the previous query so no real difference there.

At this point you may have noticed that while the written query uses two GROUP BYs and two MIN functions that are then joined together, there are not two Index Scans, two sets of aggregations, and no join happening in the execution plan.

SQL Server can use an optimization with the Top operator that allows it to take the sorted data and return only the Name and UserId rows for the top Name and Date values within a group (essentially matching the MIN logic). This is a great example of how the optimizer can take a declarative SQL query and decide how to efficiently return the data needs.

At this point, the Top operator filters our 8 million rows down to around 30k rows. 30k rows get deduped a lot faster with our Stream Aggregate operator, and since the data is already sorted we don't need an extra Sort operator.

Overall, this second query runs so much better than the original because SQL Server doesn't have to go to tempdb for any operations - all the data is pre-sorted in the index and can flow through.

So I shouldn't use Window Functions?

Not necessarily - it comes down to a trade offs.

I almost always start with a window function because of how easy they are to write and read. Plus I think they are fun to write as well.

However, if the window function is having to read/write a lot of data to tempdb and it's affecting the overall performance of your query, a rewrite may be necessary.

In that case, I much rather take more verbose syntax to get a 2000x performance boost.