Whenever I’m looking over query plans with clients, their eyes get drawn towards many things that I’ve learned to ignore over the years.
It’s not that they’re never important, it’s just that, you know… There’s usually more important stuff.
One of those things is compilation timeouts. Most people think that it’s time-based, and it means that their query timed out or took a long time to compile.
Not so! It’s purely a set number of steps the optimizer will take to figure out things like:
Join order
Join/Aggregate type
Index usage
Seeks vs Scans
Parallelism
And probably some other stuff that I just don’t have the Friday afternoon energy to think about any more.
But anyway, the point is that it’s not a sign that your query timed out, or even that plan compilation took a long time.
The initial number of steps allowed is based on the optimizer’s assessment of statement complexity, which includes the number of joins (of course), in case you were wondering.
From there each additional stage gets a set number of steps based on the number of steps that the previous stage took.
Plan Cache Script
You can use this script to look in your plan cache for plans that the optimizer has marked as having a timeout.
WITH
XMLNAMESPACES
(
DEFAULT 'http://schemas.microsoft.com/sqlserver/2004/07/showplan'
)
SELECT
query_text =
SUBSTRING
(
st.text,
qs.statement_start_offset / 2 + 1,
CASE qs.statement_start_offset
WHEN -1
THEN DATALENGTH(st.text)
ELSE qs.statement_end_offset
END - qs.statement_start_offset / 2 + 1
),
compile_time_ms =
qs.query_plan.value('(//StmtSimple/QueryPlan/@CompileTime)[1]', 'bigint'),
compile_cpu_ms =
qs.query_plan.value('(//StmtSimple/QueryPlan/@CompileCPU)[1]', 'bigint'),
compile_memory_mb =
qs.query_plan.value('(//StmtSimple/QueryPlan/@CompileMemory)[1]', 'bigint') / 1024.,
qs.query_plan,
qs.execution_count,
qs.total_worker_time,
qs.last_execution_time
FROM
(
SELECT TOP (10)
qs.plan_handle,
qs.sql_handle,
qs.statement_start_offset,
qs.statement_end_offset,
qs.last_execution_time,
qs.execution_count,
qs.total_worker_time,
qp.query_plan
FROM sys.dm_exec_query_stats AS qs
CROSS APPLY sys.dm_exec_query_plan(qs.plan_handle) AS qp
WHERE qp.query_plan.exist('//StmtSimple/@StatementOptmEarlyAbortReason[.="TimeOut"]') = 1
ORDER BY
total_worker_time / qs.execution_count DESC
) AS qs
CROSS APPLY sys.dm_exec_sql_text(qs.sql_handle) AS st;
There’s not a whole lot of sense to this query other than to prove a point. Here are some abridged results from a client system:
amanaplanacanalpanama
Despite all of these queries “timing out” during optimization phases, the longest compile time is 61 milliseconds.
Query Store Script
Like above, there’s not a lot of sense to this one. It is nice to be able to skip some of the additional XML shredding and go to some of the plan metadata stored in Query Store:
WITH
XMLNAMESPACES
(
DEFAULT 'http://schemas.microsoft.com/sqlserver/2004/07/showplan'
),
queries
AS
(
SELECT TOP (101)
parent_object_name =
ISNULL
(
OBJECT_NAME(qsq.object_id),
'No Parent Object'
),
qsqt.query_sql_text,
query_plan =
TRY_CAST(qsp.query_plan AS xml),
qsrs.last_execution_time,
qsrs.count_executions,
qsrs.avg_duration,
qsrs.avg_cpu_time,
avg_compile_duration_ms =
qsq.avg_compile_duration / 1000.,
avg_compile_memory_mb =
qsq.avg_compile_memory_kb / 1024.,
avg_optimize_cpu_time_ms =
qsq.avg_optimize_cpu_time / 1024.
FROM sys.query_store_runtime_stats AS qsrs
JOIN sys.query_store_plan AS qsp
ON qsp.plan_id = qsrs.plan_id
JOIN sys.query_store_query AS qsq
ON qsq.query_id = qsp.query_id
JOIN sys.query_store_query_text AS qsqt
ON qsqt.query_text_id = qsq.query_text_id
WHERE qsrs.last_execution_time >= DATEADD(DAY, -7, SYSDATETIME())
AND qsrs.avg_cpu_time >= (10 * 1000)
AND qsq.is_internal_query = 0
AND qsp.is_online_index_plan = 0
AND TRY_CAST(qsp.query_plan AS xml).exist('//StmtSimple/@StatementOptmEarlyAbortReason[.="TimeOut"]') = 1
ORDER BY
qsrs.avg_cpu_time DESC
)
SELECT
qs.query_sql_text,
qs.parent_object_name,
qs.query_plan,
qs.avg_compile_duration_ms,
qs.avg_optimize_cpu_time_ms,
qs.avg_compile_memory_mb,
qs.count_executions,
qs.avg_duration,
qs.avg_cpu_time,
qs.last_execution_time
FROM
queries AS qs
ORDER BY
qs.avg_cpu_time DESC
OPTION (RECOMPILE);
Also like above, the results bring back very short compile times.
So There
The point of this post was that you don’t need to worry about these timeouts from a plan compilation time perspective.
Of course, it may represent a plan quality issue, but that’s much harder to prove from first glances. You’d need to dig into that on your own Friday afternoon.
If you find user queries experiencing optimizer timeouts, it may solve the problem to simplify them as much as possible. Breaking long queries up into #temp tables is a popular solution for this.
Thanks for reading!
Going Further
If this is the kind of SQL Server stuff you love learning about, you’ll love my training. Blog readers get 25% off the Everything Bundle — over 100 hours of performance tuning content. Need hands-on help? I offer consulting engagements from targeted investigations to ongoing retainers. Want a quick sanity check before committing to a full engagement? Schedule a call — no commitment required.
SQL Server 2022 has a new feature in it to help with parameter sensitive query plans.
That is great. Parameter sensitivity, sometimes just called parameter sniffing, can be a real bear to track down, reproduce, and fix.
In a lot of the client work I do, I end up using dynamic SQL like this to get things to behave:
But with this new feature, you get some of the same fixes without having to interfere with the query at all.
How It Works
You can read the full documentation here. But you don’t read the documentation, and the docs are missing some details at the moment anyway.
It only works on equality predicates right now
It only works on one predicate per query
It only gives you three query plan choices, based on stats buckets
There’s also some additional notes in the docs that I’m going to reproduce here, because this is where you’re gonna get tripped up, if your scripts associate statements in the case with calling stored procedures, or using object identifiers from Query Store.
For each query variant mapping to a given dispatcher:
The query_plan_hash is unique. This column is available in sys.dm_exec_query_stats, and other Dynamic Management Views and catalog tables.
The plan_handle is unique. This column is available in sys.dm_exec_query_stats, sys.dm_exec_sql_text, sys.dm_exec_cached_plans, and in other Dynamic Management Views and Functions, and catalog tables.
The query_hash is common to other variants mapping to the same dispatcher, so it’s possible to determine aggregate resource usage for queries that differ only by input parameter values. This column is available in sys.dm_exec_query_stats, sys.query_store_query, and other Dynamic Management Views and catalog tables.
The sql_handle is unique due to special PSP optimization identifiers being added to the query text during compilation. This column is available in sys.dm_exec_query_stats, sys.dm_exec_sql_text, sys.dm_exec_cached_plans, and in other Dynamic Management Views and Functions, and catalog tables. The same handle information is available in the Query Store as the last_compile_batch_sql_handle column in the sys.query_store_query catalog table.
The query_id is unique in the Query Store. This column is available in sys.query_store_query, and other Query Store catalog tables.
The problem is that, sort of like dynamic SQL, this makes each different plan/statement impossible to tie back to the procedure.
What I’ve Tried
Here’s a procedure that is eligible for parameter sensitivity training:
CREATE OR ALTER PROCEDURE
dbo.SQL2022
(
@ParentId int
)
AS
BEGIN
SET NOCOUNT, XACT_ABORT ON;
SELECT TOP (10)
u.DisplayName,
p.*
FROM dbo.Posts AS p
JOIN dbo.Users AS u
ON p.OwnerUserId = u.Id
WHERE p.ParentId = @ParentId
ORDER BY u.Reputation DESC;
END;
GO
Here’s the cool part! If I run this stored procedure back to back like so, I’ll get two different query plans without recompiling or writing dynamic SQL, or anything else:
EXEC dbo.SQL2022
@ParentId = 184618;
GO
EXEC dbo.SQL2022
@ParentId = 0;
GO
amazing!
It happens because the queries look like this under the covers:
SELECT TOP (10)
u.DisplayName,
p.*
FROM dbo.Posts AS p
JOIN dbo.Users AS u
ON p.OwnerUserId = u.Id
WHERE p.ParentId = @ParentId
ORDER BY u.Reputation DESC
OPTION (PLAN PER VALUE(QueryVariantID = 1, predicate_range([StackOverflow2010].[dbo].[Posts].[ParentId] = @ParentId, 100.0, 1000000.0)))
SELECT TOP (10)
u.DisplayName,
p.*
FROM dbo.Posts AS p
JOIN dbo.Users AS u
ON p.OwnerUserId = u.Id
WHERE p.ParentId = @ParentId
ORDER BY u.Reputation DESC
OPTION (PLAN PER VALUE(QueryVariantID = 3, predicate_range([StackOverflow2010].[dbo].[Posts].[ParentId] = @ParentId, 100.0, 1000000.0)))
Where Things Break Down
Normally, sp_BlitzCache will go through whatever statements it picks up and associate them with the parent object:
But it doesn’t do that here, it just says that they’re regular ol’ statements:
do i know you?
The way that it attempts to identify queries belonging to objects is like so:
RAISERROR(N'Attempting to get stored procedure name for individual statements', 0, 1) WITH NOWAIT;
UPDATE p
SET QueryType = QueryType + ' (parent ' +
+ QUOTENAME(OBJECT_SCHEMA_NAME(s.object_id, s.database_id))
+ '.'
+ QUOTENAME(OBJECT_NAME(s.object_id, s.database_id)) + ')'
FROM ##BlitzCacheProcs p
JOIN sys.dm_exec_procedure_stats s ON p.SqlHandle = s.sql_handle
WHERE QueryType = 'Statement'
AND SPID = @@SPID
OPTION (RECOMPILE);
Since SQL handles no longer match, we’re screwed. I also looked into doing something like this, but there’s nothing here!
SELECT
p.plan_handle,
pa.attribute,
object_name =
OBJECT_NAME(CONVERT(int, pa.value)),
pa.value
FROM
(
SELECT 0x05000600B7F6C349E0824C498D02000001000000000000000000000000000000000000000000000000000000 --Proc plan handle
UNION ALL
SELECT 0x060006005859A71BB0304D498D02000001000000000000000000000000000000000000000000000000000000 --Query plan handle
UNION ALL
SELECT 0x06000600DCB1FC11A0224D498D02000001000000000000000000000000000000000000000000000000000000 --Query plan handle
) AS p (plan_handle)
CROSS APPLY sys.dm_exec_plan_attributes (p.plan_handle) AS pa
WHERE pa.attribute = 'objectid';
The object identifiers are 0 for these two queries.
One Giant Leap
This isn’t a complaint as much as it is a warning. If you’re a monitoring tool vendor, script writer, or script relier, this is gonna make things harder for you.
Perhaps it’s something that can or will be fixed in a future build, but I have no idea at all what’s going to happen with it.
Maybe we’ll have to figure out a different way to do the association, but stored procedures don’t get query hashes or query plan hashes, only the queries inside it do.
This is gonna be a tough one!
Thanks for reading!
Going Further
If this is the kind of SQL Server stuff you love learning about, you’ll love my training. Blog readers get 25% off the Everything Bundle — over 100 hours of performance tuning content. Need hands-on help? I offer consulting engagements from targeted investigations to ongoing retainers. Want a quick sanity check before committing to a full engagement? Schedule a call — no commitment required.
There are a million scripts out there that will give you all of the missing index requests for a database (or even a whole server).
Some will even try to prioritize based on the metrics logged along with each request.
Right now, most of you get:
Uses: How many times a query compiled that could have used the index
Average Query Cost: A unit-less cost used by the optimizer for choosing a plan
Impact: A metric relative to the unit-less cost of the operator the index will help
Breaking each of those down, the only one that has a concrete meaning is Uses, but that of course doesn’t mean that a query took a long time or is even terribly inefficient.
That leaves us with Average Query Cost, which is the sum of each operator’s estimated cost in the query plan, and Impact.
But where does Impact come from?
Impactful
Let’s look at a query plan with a missing index request to figure out what the Impact metric is tied to.
Here’s the relevant part of the plan:
sticky kid
And here’s the missing index request:
/*
The Query Processor estimates that implementing the following index could improve the query cost by 16.9141%.
*/
/*
USE [StackOverflow2013]
GO
CREATE NONCLUSTERED INDEX [<Name of Missing Index, sysname,>]
ON [dbo].[Comments] ([Score])
INCLUDE ([PostId])
GO
*/
Here’s the breakdown:
The optimizer estimates that hitting the Comments table will cost 762 query bucks, which is 17% of the total plan cost
The optimizer further estimates that hitting the Comments table with the suggested index will reduce the total plan cost by 16.9%
Here’s the relevant properties from the scan of the Comments table:
Indecisive
What I want you to take away from this is that, while hitting the Comments table may be 17% of the plan’s total estimated cost, the time spent scanning that index is not 17% of the plan’s total execution time, either in CPU or duration.
You can see in the screenshot above that it takes around 450ms to perform the full scan of 24,534,730 rows.
Doubtful
In full, this query runs for around 23 seconds:
outta here
The estimated cost of hitting the Comments tables is not 17% of the execution time. That time lives elsewhere, which we’ll get to.
In the meantime, there are two more egregious problems to deal with:
The optimizer severely miscalculates the cost of scanning the Posts table at 70% (note the 860ms time here):
oh no no no
2. It buries other missing index requests in the properties of the root operator:
train tracks
Now, there are two other missing index requests listed here that are a) of higher “impact” b) not ordered by that impact number and c) even if both a and b were true, we know that adding those indexes would not substantially reduce the overall runtime of the stored procedure.
Assuming that we added every single missing index here, at best we would reduce the estimated cost of the plan by 87%, while only reducing the actual execution time of the plan by about 1.3 seconds out of 23 seconds.
Not a big win, here.
Hurtful
Examining where time is spent in this plan, this branch will stick out as the dominating factor:
baby don’t do it
Some interesting things to note here, while we’re talking about interesting things:
The scan of the Badges table takes 1.4 seconds, and has an estimated cost of 1%
The estimated cost of the eager index spool is 11%, but accounts for 20 seconds of elapsed time (less the 1.4 seconds for the scan of Badges)
There was no missing index request generated for the Badges table, despite the optimizer creating one on the fly
This is a bit of the danger in creating missing index requests without first validating which queries generated them, and where the benefit in having them would be.
In tomorrow’s post, we’ll look at how SQL Server 2019 makes figuring this stuff out easier.
Thanks for reading!
Going Further
If this is the kind of SQL Server stuff you love learning about, you’ll love my training. Blog readers get 25% off the Everything Bundle — over 100 hours of performance tuning content. Need hands-on help? I offer consulting engagements from targeted investigations to ongoing retainers. Want a quick sanity check before committing to a full engagement? Schedule a call — no commitment required.
You know those tables, right? The ones where developers went and got lazy or didn’t know any better and decided every string column was going to be gigantic.
They may have read, of course, that SQL Server’s super-smart variable length data types only consume necessary space.
It’s free real estate.
Except it isn’t, especially not when it comes to query memory grants.
The bigger a string column’s defined byte length is, the bigger the optimizer’s memory grant for it will be.
Memory Grant Primer
In case you need some background, the short story version is:
All queries ask for some memory for general execution needs
Sorts, Hashes, and Optimized Nested Loops ask for additional memory grants
Memory grants are decided based on things like number of rows, width of rows, and concurrently executing operators
Memory grants are divided by DOP, not multiplied by DOP
By default, any query can ask for up to 25% of max server memory for a memory grant
Approximately 75% of max server memory is available for memory grants at one
Needless to say, memory grants are very sensitive to misestimates by the optimizer. Going over can be especially painful, because that memory will most often get pulled from the buffer pool, and queries will end up going to disk more.
Underestimates often mean spills to disk, of course. Those are usually less painful, but can of course be a problem when they’re large enough. In particular, hash spills are worth paying extra attention to.
Memory grant feedback does supply some relief under modern query execution models. That’s a nice way of saying probably not what you have going on.
Query Noogies
Getting back to the point: It’s a real pain in the captain’s quarters to modify columns on big tables, even if it’s reducing the size.
SQL Server’s storage engine has to check page values to make sure you’re not gonna lose any data fidelity in the process. That’ a nice way of saying you’re not gonna truncate any strings.
But if you do something cute like run a MAX(LEN(StringCol) and see what you’re up against, you can use a view on top of your table to assuage SQL Server’s concerns about such things.
After all, functions are temporary. Data types are forever (usually).
An easy way to illustrate what I mean is to look at the details of these two queries:
SELECT TOP (1000)
p.Body
FROM dbo.Posts AS p
ORDER BY p.Score DESC
OPTION(RECOMPILE);
SELECT TOP (1000)
Body =
CONVERT
(
nvarchar(100),
p.Body
)
FROM dbo.Posts AS p
ORDER BY p.Score DESC
OPTION(RECOMPILE);
Some of this working is dependent on the query plan, so let’s look at those.
Pink Belly Plans
You can ignore the execution times here. The Body column is not a good representation of an oversized column.
It’s defined as nvarchar(max), but (if I’m remembering my Stack lore correctly) is internally limited to 30k characters. Many questions and answers are longer than 100 characters anyway, but on to the plans!
janitor
In the plan where the Body column isn’t converted to a smaller string length, the optimizer asks for a 16GB memory grant, and in the second plan the grant is reduced to ~3.5GB.
This is dependent on the compute scalar occurring prior to the Top N Sort operator, of course. This is where the convert function is applied to the Body column, and why the grant is reduced
If you were to build a view on top of the Posts table with this conversion, you could point queries to the view instead. That would get you the memory grant reduction without the pain of altering the column, or moving the data into a new table with the correct definition.
Thanks for reading!
Going Further
If this is the kind of SQL Server stuff you love learning about, you’ll love my training. Blog readers get 25% off the Everything Bundle — over 100 hours of performance tuning content. Need hands-on help? I offer consulting engagements from targeted investigations to ongoing retainers. Want a quick sanity check before committing to a full engagement? Schedule a call — no commitment required.
A client question that I get quite a bit is around why queries in production get a bad query plan that queries in dev, QA, or staging don’t get is typically answered by looking at statistics.
Primarily, it’s because of the cardinality estimates that queries get around ascending keys. It usually gets called the ascending key problem, but the gist is that:
You have a pretty big table
You’re using the legacy cardinality estimator
A whole bunch of rows get inserted, but not enough to trigger an auto stats update
You’re not using compatibility level >= 130 or trace flag 2371
Queries that look for values off an available histogram get a one row estimate using the legacy Cardinality Estimator or a 30% estimate using the default Cardinality Estimator
Which is a recipe for potentially bad query plans.
Reproductive Script
Here’s the full repro script. If you’re using a different Stack Overflow database, you’ll need to adjust the numbers.
USE StackOverflow2013;
/*Figure out the 20% mark for stats updates using legacy compat levels*/
SELECT
c = COUNT_BIG(*),
c20 = CEILING(COUNT_BIG(*) * .20)
FROM dbo.Users AS u;
/*Stick that number of rows into a new table*/
SELECT TOP (493143)
u.*
INTO dbo.Users_Holder
FROM dbo.Users AS u
ORDER BY u.Id DESC;
/*Delete that number of rows from Users*/
WITH
del AS
(
SELECT TOP (493143)
u.*
FROM dbo.Users AS u
ORDER BY u.Id DESC
)
DELETE
FROM del;
/*I'm using this as a shortcut to turn off auto stats updates*/
UPDATE STATISTICS dbo.Users WITH NORECOMPUTE;
/*Put the rows back into the Users Table*/
SET IDENTITY_INSERT dbo.Users ON;
INSERT
dbo.Users
(
Id,
AboutMe,
Age,
CreationDate,
DisplayName,
DownVotes,
EmailHash,
LastAccessDate,
Location,
Reputation,
UpVotes,
Views,
WebsiteUrl,
AccountId
)
SELECT
uh.Id,
uh.AboutMe,
uh.Age,
uh.CreationDate,
uh.DisplayName,
uh.DownVotes,
uh.EmailHash,
uh.LastAccessDate,
uh.Location,
uh.Reputation,
uh.UpVotes,
uh.Views,
uh.WebsiteUrl,
uh.AccountId
FROM dbo.Users_Holder AS uh;
SET IDENTITY_INSERT dbo.Users OFF;
/*Figure out the minimum Id we put into the holder table*/
SELECT
m = MIN(uh.Id)
FROM dbo.Users_Holder AS uh;
/*Compare estimates*/
SELECT
c = COUNT_BIG(*)
FROM dbo.Users AS u
WHERE u.Id > 2623772
OPTION(USE HINT('FORCE_LEGACY_CARDINALITY_ESTIMATION'));
SELECT
c = COUNT_BIG(*)
FROM dbo.Users AS u
WHERE u.Id > 2623772
OPTION(USE HINT('FORCE_DEFAULT_CARDINALITY_ESTIMATION'));
/*Cleanup*/
UPDATE STATISTICS dbo.Users;
TRUNCATE TABLE dbo.Users_Holder;
Query Plans
Here are the plans for the stars of our show:
king push
In these query plans, you can see the legacy cardinality estimator gets a one row estimate, and the default cardinality estimator gets a 30% estimate.
There isn’t necessarily a one-is-better-than-the-other answer here, either. There are times when both can cause poor plan choices.
You can think of this scenario as being fairly similar to parameter sniffing, where one plan choice does not fit all executions well.
Checkout
There are a lot of ways that you can go about addressing this.
In some cases, you might be better off using trace flag 2371 to trigger more frequent auto stats updates on larger tables where the ~20% modification counter doesn’t get hit quickly enough. In others, you may want to force one estimator over the other depending on which gets you a better plan for most cases.
Another option is to add hints to the query in question to use the default cardinality estimator (FORCE_DEFAULT_CARDINALITY_ESTIMATION), or to generate quick stats for the index/statistics being used (ENABLE_HIST_AMENDMENT_FOR_ASC_KEYS). Documentation for both of those hints is available here. Along these lines, trace flags 2389, 2390, or 4139 may be useful as well.
Of course, you could also try to address any underlying query or index issues that may additionally contribute to poor plan choices, or just plan differences. A common problem in them is a seek + lookup plan for the one row estimate that doesn’t actually make sense when the actual number of rows and lookup executions are encountered at runtime.
Thanks for reading!
Going Further
If this is the kind of SQL Server stuff you love learning about, you’ll love my training. Blog readers get 25% off the Everything Bundle — over 100 hours of performance tuning content. Need hands-on help? I offer consulting engagements from targeted investigations to ongoing retainers. Want a quick sanity check before committing to a full engagement? Schedule a call — no commitment required.
There are metrics that I care and don’t care about when I’m looking for queries to tune.
Metrics I don’t care about:
Logical Reads
Costs
If a query does “a lot” of reads or has a high “cost”, I generally don’t care as long as they run quickly. Doing consistent physical reads is a slightly different story, but would probably fall more under server tuning or fixing memory grants.
Metrics I do care about:
CPU (taking parallelism into consideration)
Duration (compared to CPU)
Memory Grants (particularly when they’re not being fully utilized)
Writes (especially if it’s just a select)
Executions (mostly to track down scalar UDFs)
CPU and Duration
These two metrics get lumped together because they need to be compared in order to figure out what’s going on. First, you need to figure out what the minimum runtime of a query is that you want to tune.
In general, as query execution time gets faster, getting it to be much faster gets more difficult.
Bringing a query from 1 second to 100 milliseconds might be a small matter
Bringing that same query from 100 milliseconds to 1 millisecond might take more time than it’s worth
I say that because unless someone is querying SQL Server directly, smaller durations tend to be less detectable to end users. By the time they hit a button, send the request, receive the data, and have the application render it etc. they’re probably not aware of a 99 millisecond difference.
Of course, not everything is end-user centric. Other internal operations, especially any loop processing, might benefit greatly from reductions on the smaller side of things.
If duration and CPU are acceptable, leave it alone
If either is unacceptable, tune the darn thing
If CPU is much higher than duration, you have a parallel plan, and tuning is optional
If duration is much higher than CPU, you have blocking or another contention issue, and the query you’re looking at probably isn’t the problem
If duration and CPU are roughly equivalent, you either have a functional serial plan or a really crappy parallel plan
I give these the highest priority because reducing these is what makes queries faster, and reduces the surface area (execution time) of a query where something crappy might happen, like blocking, or deadlocks, or other resource contention.
Memory Grants
Using these as a tuning metric can have a lot of positive effects, depending on what kind of shape the system is in.
Consider a few scenarios:
PAGEIOLATCH_XX waits are high because large memory grants steal significant buffer pool space
RESOURCE_SEMAPHORE waits are high because queries suck up available memory space and prevent other queries from using it
Queries are getting too low of a memory grant and spilling significantly, which can slow them down and cause tempdb contention under high concurrency
Fixing memory grant issues can take many forms:
Getting better cardinality estimates for better overall grant estimates
Indexing to influence operator choices away from memory consumers
Using more appropriate string lengths to reduce memory grants
Fixing parallel skew issues that leaves some threads with inadequate memory
Rewriting the query to not ask for ordered data
Rewriting the query to ask for ordered data in smaller chunks
Rewriting the query to convert strings to better fitting byte lengths
That’s just some stuff I end up doing off the top of my head. There are probably more, but blog posts are only useful up to a certain length.
Like all other strings.
Writes and Selects
Modification queries are going to do writes. This seems intuitive and not at all shocking. If you have queries that are doing particularly large modifications, you could certainly look into tuning those, but it would be a standard exercise in query or index tuning.
Except that your index tuning adventure would most likely lead you to dropping unused and overlapping indexes to reduce the number of objects that you need to write to than to add an index.
But who knows. Boring anyway. I hear indexes tune themselves in the cloud.
When select queries do a large number of writes, then we’re talking about a much more interesting scenario.
Spills
Spools
Stats updates
Of course, stats updates are likely a pretty small write, but the read portion can certainly halt plan compilation for a good but on big tables.
Spills and Spools are going to be the real target here. If it’s a spill, you may find yourself tracking back to the memory grant section up above.
Spools, though! What interesting little creatures. I wrote a longer post about them here:
It has a bit of a link roundup of other posts on my site and others that talk about them, too.
But since we’re living in this now, let’s try to be present. Here’s the short story on spools that we might try to fix:
The Spools we typically care about are Table or Index
They can be eager or lazy
They’ll show up on the inner side of Nested Loops
SQL Server uses them as a temporary cache for data
They are a good indicator that something is amok with your query or indexes
For eager index spools, the story is pretty simple around creating a better index for SQL Server to use.
For lazy table spools, you have more options:
Give SQL Server unique data to work with
Get the optimizer to not choose nested loops
Use the NO_PERFORMANCE_SPOOL hint to test the query without spools
Of course, there are times where you’re better off with a spool than without. So don’t walk away feeling disheartened if that’s the case.
Executions
These are on the opposite end of the spectrum from most of the queries I go after. If a query runs enough, and fast enough, to truly rack up a high number of executions, there’s probably not a ton of tuning you could do.
Sure, sometimes there’s an index you could add or a better predicate you could write, but I’d consider it more beneficial to get the query to not run so much.
That might result in:
Rewriting functions as inline table valued functions
Handing the queries off to app developers for caching
To learn how I rewrite functions, check out this video
I know, you can’t rewrite every single function like this, but it’s a wonderful thing to do when you can.
Anything Other Than
Again, metrics I don’t ever look at are logical reads or costs.
Doing reads doesn’t necessarily mean that queries are slow, or that there’s anything you can fix
Costs are a meme metric that should be removed from query plans in favor of operator times
Well, okay, maybe not completely removed, but they shouldn’t be front and center anymore.
There are many other more reliable metrics to consider that are also far more interesting.
Thanks for reading!
Video Summary
In this video, I discuss the challenges and performance implications of using scalar UDFs in SQL Server queries, particularly focusing on a function called “no bueno.” I walk through how to rewrite this function as an inline table-valued function (TVF) to improve performance. By leveraging `GETDATE()` and Common Table Expressions (CTEs), we can avoid the limitations that prevent scalar UDFs from being inlined, leading to more efficient query execution plans. The video also delves into why functions should generally be avoided in WHERE clauses due to their tendency to hinder parallelism and increase execution time. With a bit of humor and personal anecdotes about moving offices to escape stalkers, I aim to make the topic both engaging and informative for viewers who are new or experienced with SQL Server.
Full Transcript
Erik Darling here with Erik Darling Data. And I believe this is take 297 of this video. So I’m sorry if these jokes sound a little bit rehearsed. I’m very tired. Kidding. I’m fine. Everything’s good. This is my first take. This is the first time I’m doing this. Don’t worry about me. I would never make that. I would never do anything bad. Now, I apologize if I look a little bit weird. I have this ring light up over here because my office is for some reason exceptionally dark. I don’t know. I don’t know. I don’t know. So sorry if I look weirder than usual, but I’m used to it. This is actually probably the last video that I record in this office because I have many, many stalkers. I have decided to move my office to a different location. And I can’t disclose the location of this office because, again, I have too many stalkers. I was actually voted by the nice folks at Beer Gut Magazine, the consultant most likely to get murdered by a stalker. So I’m trying to avoid that. I’m going to move my office. Well, it probably shows some internal shots of it in a blog post coming up. But for now, all you need to know is that I am disappearing from this room. I don’t know what the new room is going to, I don’t know what the setup is going to look like yet. But it’s going to lead to grand things like me actually having the will to do live streaming and live classes and stuff again because this office is so small that it drained my will to do those things. So there we have it. Anyway, let’s talk about what this function does. And this function is called no bueno because scalar UDFs are in general no bueno. Now it’s going to take a couple things, a user ID and a start date. And we’re going to sue and it’s going to return a thing, right? This returns an integer. Wowee. And you know what’s funny is like you have an integer, but like on the flip side, all you have is a big int. There’s no like big integer. I think that doesn’t work. So I don’t really understand. I don’t really understand that. But there we go. I don’t know. Blame the summer intern again. But inside the function, we’re going to supercharge it, right? We’re going to give it the schema binding and returns null on null input attributes, right? So make this function go as fast as we can with those. And then inside the function, we’re going to declare a few parameters. I guess that’s four, which is one more than a few. But what do I know? I’m a high school dropout. I don’t know what numbers are called anyway.
So the first thing this function is going to do after declaring those variables is we’re going to see if start date is null. And if it’s null, then we’re just going to subtract a year from get date. And of course, this get date function is what’s going to lead to us needing to rewrite the function because this get date being in here, well, guess what? That’s a limitation of Senor Freud. If we have an intrinsic function like get date in our scalar UDF, function can’t get in line.
Okay, then. I find that especially curious. Actually, there’s two things I find curious on this page. One is that if the get date thing, and two is the CTE thing. You want to know why I find these things curious is because both of these things break scalar UDF inline.
But when we rewrite this function, I mean, spoiler alert, major spoiler alert here. When we rewrite this function as an inline table valued function, we are going to use both get date and a CTE. So it doesn’t make a whole lot of sense why I can’t get done.
So you’ll sort of figure out if you write an inline table valued function, a scalar UDF is like, oh, I’m bad. Hands off. Hands off. Too much going on there. Oops, I clicked on the wrong window. Pretend that didn’t happen. Not re-recording this thing for the 297th time.
And then, okay, so check out that null thing. This is what screws us up. But then we’re going to get the creation date and last access date for whatever user we’re passed in. And then if we greater than or equal to whatever start date we pass in from the function, then we’re also going to get a count of their posts.
So good job us there. And then we are going to, you, set this average post per day return variable, return thing, as the total posts divided by the days between the creation date and last access date. And of course, we’re going to use the wonderful, fabulous, talented null if function to make sure that we don’t hit any divide by zero errors.
Very important defensive T-SQL there, right? T-SQL pro tip. Nerd. So anyway, let’s look at… I assume if you’re watching this channel, you’ve heard me make fun of functions many times before.
But in case we have any newcomers, let’s talk about why we need to get rid of functions a little bit. So first thing is that when you look at the query plan for a query that calls a scalar UDF, one thing that you’re going to see if you are not on…
Well, I guess one thing that you’re going to see if you are not on SQL Server 2019 and you are not getting your functions inlined is this non-parallel plan reason. This cannot generate valid parallel plan.
And of course, if we look at the plan cache, we’re going to see something else interesting. We are going to see this function, right? We see we have our top 10 query and we have this create function thing.
And if we scroll over a little bit, we’re going to see that that query executed once, but that function executed 10 times. So the more rows that have to get passed through that function, the more times that function is going to run, right?
So like if that query is… Like in this case, if we return the top 1,000, that function would have to execute 1,000 times. If we were filtering on that function, like if we were saying select top 10 from users who have more than 20 average posts per day, we would have to pass the entire users table.
Well, that’s 2.4 million rows. We have to pass 2.4 million rows through that function in order to generate a resultant filter on it. That would just be ugly. That would be horrible.
And I don’t want to make you sit through a video where I do that, so I’m not going to do that. Bottom line, don’t put functions in your where clause. Actually, yeah, just don’t put functions in your where clause. That’s it. But of course, the real performance hit from this function doesn’t really come from what the function itself does.
The function itself is pretty fast when it runs on its own. There’s a bike gang outside. This might be more of my stalkers coming to get me. But thanks, bike gang. Bike sounds like farts.
Doesn’t sound like a tough bike at all. But anyway, where functions really start to hurt queries is when they prevent a larger query that could and should go parallel from going parallel using multiple chords.
They force that query to run single-threaded, so that query just runs for a longer time, just having one thread have to deal with a whole bunch of rows. You can see now this thing, even though it still returns 10 rows, and we found a bunch of really low-impact people, zero rows per day.
This thing all runs for 11 and a half seconds. And of course, our scalar UDF forces the query to run serially. And if we went and looked at the plan cache again, we’d probably see 20 executions of that function now because we did another top 10.
But that’s about all the times that I want to go look at the plan cache right now. So we’re not going to do that again. I don’t need to prove that to you twice. You can figure it out on your own. You’re a smart person.
You’re capable. Everyone loves you. So let’s rewrite this function using two constructs that force inlining to not work. Let’s use the getDate function, and let’s use CTE.
So now, you don’t necessarily need to do this. And I totally cop to the fact that, you know, you could probably rewrite this using fewer CTE or, like, you know, move things around a little bit.
But when I’m rewriting a function, I like to rewrite things in exactly the order that the scalar UDF does them in so that when I’m reviewing the logic, if I, like, run something and I see, oh, that’s different results, then I can compare apples to apples, like, the steps that I’m taking, and I can figure out where the problem is, right?
So the first thing, the first CTE is going to do what the last, the first procedural bit of logic and the other function did, which is set start date to a year ago if start date is null, right?
So that’s the first thing we’re going to do. Then down here in the user dates CTE, we’re going to do exactly what we did before. We’re going to get the creation date and the last access date for the user ID, right? And we’re going to have to cross join that start date CTE, which is fine because it’s only ever going to return one row, and then filter on that S.startDate column.
The next thing we’re going to do is go get the total posts, right? So we’re going to go select count from posts in here, and we’re going to go get all the posts for that user.
And then we’re just going, the final thing that we’re going to do, we’re all done with the CTE, is we’re going to get the average posts per day, which is the total posts from the table above. And we’re going to null if, to have some divide by zero protection.
We’re going to take the creation date and the last access date from above, and then we’re going to select from users. And again, we’re going to cross join here, but again, the cross join is okay because total posts is only ever going to be one row, and the result is only ever going to be one row.
So we’re all right there, right? So let’s create this function. Create or alter, I suppose. And now let’s double check our work. So one thing that’s very important to do whenever you’re rewriting code like this is make sure that it’s logically equivalent.
So let’s take the top 100 people from the post table, and let’s just spot check to make sure that we’re returning the same rows or the same average posts per day for all of these.
And just a quick spot check there looks pretty good. So I assume that I nailed the logic on that. It wasn’t an overly complex query, so we’re good there. And now let’s look at the results and make sure that we return the correct results from both of these, right?
So we’re going to run these. And one thing that you might notice is that the inline table-valued function version, right? Now, one thing that’s very important is that how we call the inline table-valued function is a little bit different than how we call the scalar UDF.
Since it’s returning a table, we have to put it into sort of a subquery-looking thing like this, where we see the average post per day is equal to open paren, select from the inline table-valued function, passing in those same two columns, close paren, and all that good stuff.
And then one thing that we’re going to notice here is that the inline table-valued function plan sure runs a lot faster, right? That finishes in 3.1 seconds. And, of course, the scalar UDF plan still, well, actually, that was 12.5 seconds that time.
I don’t know where we got that extra second from after the last execution, but, wow, that’s scary stuff, isn’t it? But, you know, I think the main difference if we were going to look at these two queries, I mean, aside from the fact that the plan shapes might be a tiny bit different, or maybe a lot bit different, is that this query up here was allowed to go parallel.
We can see all sorts of parallel operators in this plan where they didn’t exist in this plan. And that is, of course, one of the things that sets this query that has to sort of process a lot of rows to get down to that top 10.
It sets it free so that it can go and be fast and all that other stuff. So, anyway, that’s it for me. I’m going to go get some food now because I am starving.
It’s 2 in the afternoon. I haven’t eaten all day. I know it’s hard to believe. I think I just wake up and snort caviar, and I wish someday. Someday I’m going to wake up and snort caviar.
I learned somewhat recently that the best way to eat caviar is off the little space between your thumb and finger. The skin contact is supposed to do something. So, I don’t know.
I don’t know that I have refined enough a palette to taste the difference, but I’m willing to do what the cool kids tell me to do so that I fit in, which doesn’t explain why I’m doing SQL Server stuff at all.
I got some thinking to do. Boy. Anyway, thanks for watching. I hope you learned something.
I hope you enjoyed yourself. I’m going to be doing some more stuff like this and be talking about getting around some of those limitations that we looked at in Freud, different ways you can rewrite functions to sort of get around those.
And, I don’t know. Goodbye. Thanks. Again. Bye.
Going Further
If this is the kind of SQL Server stuff you love learning about, you’ll love my training. Blog readers get 25% off the Everything Bundle — over 100 hours of performance tuning content. Need hands-on help? I offer consulting engagements from targeted investigations to ongoing retainers. Want a quick sanity check before committing to a full engagement? Schedule a call — no commitment required.
I’m a really big fan of using operator properties for a lot of things, at least visually. Where things sort of fall down for that is copying and pasting things out.
For some stuff, you still need to head down to the XML.
Let’s say you have a stored procedure that accepts a bunch of parameters. The rest of this one isn’t important, but here you go:
CREATE OR ALTER PROCEDURE
dbo.AwesomeSearchProcedure
(
@OwnerUserId int = NULL,
@CreationDate datetime = NULL,
@LastActivityDate datetime = NULL,
@PostTypeId int = NULL,
@Score int = NULL,
@Title nvarchar(250) = NULL,
@Body nvarchar(MAX) = NULL
)
A Plan Appears
Let’s say we grab a query plan for this thing from the plan cache or query store. We can get the properties of the select operator and see compile time values:
Again — nice visually — but it doesn’t do much for us if we want to recreate executing the stored procedure to get an actual execution plan.
It’s also not terrible helpful if we want to simulate a parameter sniffing situation, because we only have the compile time values, not the run time values.
Bummer. But whatever.
XML Time!
If we right click and select “show execution plan XML”, we can scroll way down to the bottom to find the XML fragment that holds what the properties display:
This still isn’t awesome, because we have to do some surgery on the XML itself to get values out.
It’s even worse if we have a parameterized application query, because not only do we need to make a DECLARE to assign values to these variables but we need to turn the query itself into dynamic SQL.
For most things, I absolutely adore using operator properties. For some things, you still need the XML.
It’d be nice if there were some fancy copy and paste magic that would do that for you, but so far it doesn’t exist.
Thanks for reading!
Going Further
If this is the kind of SQL Server stuff you love learning about, you’ll love my training. Blog readers get 25% off the Everything Bundle — over 100 hours of performance tuning content. Need hands-on help? I offer consulting engagements from targeted investigations to ongoing retainers. Want a quick sanity check before committing to a full engagement? Schedule a call — no commitment required.
There’s a lot of stuff flying around in a query plan. Data and what not.
Sure, you can hover over operators and arrows and see some stuff, but if you really wanna see stuff — I mean REALLY wanna see stuff — you gotta get into the properties.
You can access those in two ways:
Hit F4 (not F5 again; you already suffered through that)
Right click on any operator in the query plan and hit Properties
And that, my dear friend, will unlock many mysteries in your query plans.
Start With Select
I think one of the most interesting places to start is with the root operator (select, insert, update, delete), because there’s so much in there.
Here’s an abridged list of things you can see from the properties of the root node of an Actual Execution Plan:
Compile metrics: CPU, duration, memory
Degree Of Parallelism
Detailed Memory Grant information
Stats Usage
Query Time Stats in CPU and duration (including UDF times)
Parameter compile and runtime values
Nonparallel Plan reasons
Set Options
Warnings
CPU thread usage
Wait Stats
There’s more in there too, but holy cow! All the stuff you can learn here is fantastic. You might not be able to solve all your problems looking here, but it’s as good a place to start as any.
Plus, this is where you can get a sense of just how long your query ran for, and start tracking down the most troublesome operators.
Follow The Time
I’ve said before that operator costs are basically useless, and you should be following the operator times to figure out where things get wonky.
For some operators, just looking at the tool tip is enough. For example, if you have an operator that piles up a bunch of execution time because of a spill, the spill details are right in front of you.
contagious
But other times, operator properties expose things that aren’t surfaced at the tool tip.
Skew Manchu
Take skewed parallelism, for example. There are no visual indicators that it happened (maybe there should be, but given the warnings we get now, I’m not sure I trust that summer intern).
year of the spider
But you know, it might be nice to know about stuff like this. Each thread is supposed to get an equal portion of the query memory grant, and if work is distributed unevenly, you can end up with weird, random performance issues.
This is something I almost always spot check in parallel plans. In a perfect world, duration would be CPU➗DOP. Life rarely ends up perfect, which is why it’s worth a look.
I To The O
You can also see I/O stats at the operator level, logical and physical. This is why I kinda laugh at folks who still use SET STATISTICS TIME, IO ON; — you can get that all in one place — your query plan.
ding!
You can interpret things in the same way, it’s just a little easier to chew.
Thanks for reading!
Going Further
If this is the kind of SQL Server stuff you love learning about, you’ll love my training. Blog readers get 25% off the Everything Bundle — over 100 hours of performance tuning content. Need hands-on help? I offer consulting engagements from targeted investigations to ongoing retainers. Want a quick sanity check before committing to a full engagement? Schedule a call — no commitment required.
The good news is that SQL Server’s query plans will attempt to warn you about problems. The bad news is that most of the warnings only show up in Actual Execution Plans. The worse news is that a lot of the warnings that try to be helpful in Estimated Execution plans can be pretty misleading.
Certain of these are considered runtime issues, and are only available in Actual Execution Plans, like:
Spills to tempdb
Memory Grants
I’ve never seen the “Spatial Guess” warning in the wild, which probably speaks to the fact that you can measure Spatial data/index adoption in numbers that are very close to zero. I’ve also never seen the Full Update For Online Index Build warning.
Then there are some others like Columns With No Statistics, Plan Affecting Converts, No Join Predicate, and Unmatched Indexes.
Let’s talk about those a little.
Columns With No Statistics
I almost never look at these, unless they’re from queries hitting indexed views.
The only time SQL Server will generate statistics on columns in an indexed view is when you use the NOEXPAND hint in your query. That might be very helpful to know about, especially if you don’t have useful secondary indexes on your indexed view.
If you see this in plans that aren’t hitting an indexed view, it’s likely that SQL Server is complaining that multi-column statistics are missing. If your query has a small number of predicates, it might be possible to figure out which combination and order will satisfy the optimizer, but it’s often not worth the time involved.
Like I said, I rarely look at these. Though one time it did clue me in to the fact that a database had auto create stats disabled.
So I guess it’s nice once every 15 years or so.
Plan Affecting Converts
There are two of these:
Ones that might affect cardinality estimates
Ones that might affect your ability to seek into an index
Cardinality Affecting
The problem I have with the cardinality estimation warning is that it shows up when it’s totally useless.
SELECT TOP (1)
Id = CONVERT(varchar(1), u.Id)
FROM dbo.Users AS u;
fine2me
Like I said, misleading.
Seek Affecting
These can be misleading, but I often pay a bit more attention to them. They can be a good indicator of data type issues in comparison operations.
Where they’re misleading is when they tell you they mighta-coulda done a seek, when you don’t have an index that would support a seek.
SELECT TOP (1)
u.Id
FROM dbo.Users AS u
WHERE u.Reputation = CONVERT(sql_variant, N'138');
knot4you
Of course, without an index on Reputation, what am I going to seek to?
Nothing. Nothing at all.
No Join Predicate
This one is almost a joke, I think.
Back when people wrote “old style joins”, they could have missed a predicate, or something. Like so:
SELECT
c = COUNT_BIG(*)
FROM dbo.Users AS u,
dbo.Badges AS b,
dbo.Comments AS c
WHERE u.Id = b.UserId;
/*Oops no join on comments!*/
Except there’s no warning in this query plan for a missing join predicate.
well okay
But if we change the query to this, it’ll show up:
SELECT
u.Id
FROM dbo.Users AS u,
dbo.Badges AS b,
dbo.Comments AS c
WHERE u.Id = b.UserId;
/*Oops no join on comments!*/
greatly
But let’s take a query that has a join predicate:
SELECT TOP (1)
b.*
FROM dbo.Comments AS c
JOIN dbo.Badges AS b
ON c.UserId = b.UserId
WHERE b.UserId = 22656;
We still get that warning:
tough chickens
We still get a missing join predicate, even though we have a join predicate. The predicate is implied here, because of the where clause.
But apparently the check for this is only at the Nested Loops Join. No attempt is made to track pushed predicates any further.
run for your life
If there were, the warning would not appear.
Unmatched Indexes
If you create filtered indexes, you should know a couple things:
It’s always a good idea to have the column(s) you’re filter(ing) on somewhere in the index definition (key or include, whatever)
If query predicate(s) are parameterized on the column(s) you’re filter(ing) on, the optimizer probably won’t choose your filtered index
I say probably because recompile hints and unsafe dynamic SQL may prompt it to use your filtered index. But the bottom line here is parameters and filtered indexes are not friends in some circumstances.
Here’s a filtered index:
CREATE INDEX
cigarettes
ON dbo.Users
(Reputation)
WHERE
(Reputation >= 1000000)
WITH
(SORT_IN_TEMPDB = ON, DATA_COMPRESSION = PAGE);
And here’s a query that should use it:
SELECT
u.Reputation
FROM dbo.Users AS u
WHERE u.Reputation >= 1000000;
BUUUUUUUUUUUUUUUT!
combine
SQL Server warns us we didn’t. This is an artifact of Simple Parameterization, which happens early on in the Trivial Plan optimization phase.
It’s very misleading, that.
Warnings And Other Drugs
In this post we covered common scenarios when plan warnings just don’t add up to much of a such. Does that mean you should always ignore them? No, but also don’t be surprised if your investigation turns up zilch.
If you’re interested in learning more about spills, check out the Spills category of my blog. I’ve got a ton of posts about them.
At this point, you’re probably wondering why people bother with execution plans. I’m sort of with you; everything up to the actual version feels futile and useless, and seems to lie to you.
Hopefully Microsoft invests more in making the types of feedback mechanisms behind gathering plans and runtime metrics easier for casual users in future versions of SQL Server.
Thanks for reading!
Going Further
If this is the kind of SQL Server stuff you love learning about, you’ll love my training. Blog readers get 25% off the Everything Bundle — over 100 hours of performance tuning content. Need hands-on help? I offer consulting engagements from targeted investigations to ongoing retainers. Want a quick sanity check before committing to a full engagement? Schedule a call — no commitment required.
The Users table looks like this. It’s important for me to show you this up front, because column ordinal position in the table is important for understanding missing index requests.
what-whatcha need?
Keep this in mind — the columns aren’t in alphabetical order, or how selective they are, or by data type, etc.
They’re in the order that they are when the table was created, and then if any of them were added later on.
That’s all.
Long Time
Let’s take this query:
SELECT TOP (10000)
u.Id,
u.AccountId,
u.DisplayName,
u.Reputation,
u.Views,
u.CreationDate
FROM dbo.Users AS u
WHERE u.Views = 0
ORDER BY u.CreationDate DESC;
Is it very useful? No. But it’ll help us paint the right picture. The query plan doesn’t matter, because it’s just a clustered index scan, and it’ll be a clustered index scan for every other permutation, too.
The missing index for this query is like so:
CREATE NONCLUSTERED INDEX [<Name of Missing Index, sysname,>]
ON [dbo].[Users] ([Views])
INCLUDE ([CreationDate],[DisplayName],[Reputation],[AccountId])
Only the Views column is in the key of the index, even though CreationDate is an order by column.
In this case, it would be beneficial to have it as the second key column, because it would be sorted for free for us after an equality predicate.
You may also notice that Id is not part of the definition too. That’s because it’s the clustered index, so it will be inherited by any nonclustered indexes we create.
Normal
Okay, now let’s look at this query, with a slightly different where clause:
SELECT TOP (10000)
u.Id,
u.AccountId,
u.DisplayName,
u.Reputation,
u.Views,
u.CreationDate
FROM dbo.Users AS u
WHERE u.Views = 0
AND u.Reputation = 1
ORDER BY u.CreationDate DESC;
We’re adding another predicate on Reputation = 1 here. The missing index request now looks like this:
CREATE NONCLUSTERED INDEX [<Name of Missing Index, sysname,>]
ON [dbo].[Users] ([Reputation],[Views])
INCLUDE ([CreationDate],[DisplayName],[AccountId])
Neither one of these predicates is particularly selective (7,954,119 and 6,197,417, respectively) but Reputation ends up first in the key column list because its ordinal position in the table is first.
Frequency
How about if we add a really selective predicate to our query?
SELECT TOP (10000)
u.Id,
u.AccountId,
u.DisplayName,
u.Reputation,
u.Views,
u.CreationDate
FROM dbo.Users AS u
WHERE u.AccountId = 12462842
AND u.Views = 0
ORDER BY u.CreationDate DESC;
Now our missing index request looks like this:
CREATE NONCLUSTERED INDEX [<Name of Missing Index, sysname,>]
ON [dbo].[Users] ([Views],[AccountId])
Two things happened:
We don’t have any included columns anymore
AccountId is second in the key columns
This is amusing because the missing index request machine seems to understand that this will only ever one row via the equality predicate on AccountId but it still gets enumerated as the second index key column.
In other words, it doesn’t put the most selective column first. It gives you an index designed, like other examples, based on the column’s ordinal position in the table.
Nothing else, at least not so far.
Inequality
Where missing index requests will change column order is when it comes to inequality predicates. That doesn’t just mean not equal to, oh no no no.
That covers any “range” predicate: <, <=, >, >=, <> or !=, and IS NOT NULL.
Take this query for example:
SELECT TOP (10000)
u.Id,
u.AccountId,
u.DisplayName,
u.Reputation,
u.Views,
u.CreationDate
FROM dbo.Users AS u
WHERE u.AccountId = 0
AND u.Reputation < 0
AND u.Views < 0
ORDER BY u.CreationDate DESC;
The missing index request looks like this:
CREATE NONCLUSTERED INDEX [<Name of Missing Index, sysname,>]
ON [dbo].[Users] ([AccountId],[Reputation],[Views])
Now AccountId is the leading key column, but Reputation and Views are still in ordinal position order as inequality predicates.
Wink Wink
Now, look, I started off by saying that missing index requests aren’t perfect, and that’s okay. They’re not meant to replace a well-trained person. They’re meant to help the hapless when it comes to fixing slow queries.
As you get more comfortable with indexes and how to create them to make queries go faster, you’ll start to see deficiencies in missing index requests.
But you don’t want the optimizer spending a long time in the index matching/missing index request phases. That’s a bad use of its time.
As you progress, you’ll start treating missing index requests like a crying baby: something might need attention, but it’s up to you as an adult DBA or developer to figure out what that is.
Thanks for reading!
Going Further
If this is the kind of SQL Server stuff you love learning about, you’ll love my training. Blog readers get 25% off the Everything Bundle — over 100 hours of performance tuning content. Need hands-on help? I offer consulting engagements from targeted investigations to ongoing retainers. Want a quick sanity check before committing to a full engagement? Schedule a call — no commitment required.
