Tag: sql temp table

SQL Server 2017 CU 30: The Real Story With SelOnSeqPrj Fixes

Posted on by Erik Darling

I am a heading


Thanks for watching! Demo scripts below.

Demo Scripts

USE StackOverflow2013;
EXEC dbo.DropIndexes;
SET NOCOUNT ON;
DBCC FREEPROCCACHE;
GO 


CREATE INDEX 
   chunk 
ON dbo.Posts 
    (OwnerUserId, Score DESC) 
INCLUDE  
    (CreationDate, LastActivityDate)
WITH
    (MAXDOP = 8, SORT_IN_TEMPDB = ON, DATA_COMPRESSION = PAGE);
GO 

CREATE OR ALTER VIEW 
    dbo.PushyPaul
WITH SCHEMABINDING
AS
    SELECT 
        p.OwnerUserId,
        p.Score,
        p.CreationDate,
        p.LastActivityDate,
        PostRank = 
            DENSE_RANK() OVER
            ( 
               PARTITION BY 
                  p.OwnerUserId 
               ORDER BY     
                  p.Score DESC 
            )
    FROM dbo.Posts AS p;
GO 

SELECT 
    p.* 
FROM dbo.PushyPaul AS p
WHERE p.OwnerUserId = 22656;
GO 

CREATE OR ALTER PROCEDURE 
    dbo.StinkyPete 
(
    @UserId int
)
AS 
SET NOCOUNT, XACT_ABORT ON;
BEGIN
    SELECT 
        p.* 
    FROM dbo.PushyPaul AS p
    WHERE p.OwnerUserId = @UserId;
END;
GO 

EXEC dbo.StinkyPete 
    @UserId = 22656;



/*Start Here*/

ALTER DATABASE 
    StackOverflow2013 
SET PARAMETERIZATION SIMPLE;

DBCC TRACEOFF
(
    4199, 
    -1
);

ALTER DATABASE SCOPED CONFIGURATION 
    SET QUERY_OPTIMIZER_HOTFIXES = OFF;


SELECT 
    p.* 
FROM dbo.PushyPaul AS p
WHERE p.OwnerUserId = 22656
AND   1 = (SELECT 1); /*Avoid trivial plan/simple parameterization*/


/*Let's cause a problem!*/
ALTER DATABASE 
    StackOverflow2013 
SET PARAMETERIZATION FORCED;

SELECT 
    p.* 
FROM dbo.PushyPaul AS p
WHERE p.OwnerUserId = 22656
AND   1 = (SELECT 1); /*Avoid trivial plan/simple parameterization*/


/*Can we fix the problem?*/
DBCC TRACEON
(
    4199, 
    -1
);


SELECT 
    p.* 
FROM dbo.PushyPaul AS p
WHERE p.OwnerUserId = 22656
AND   1 = (SELECT 1); /*Avoid trivial plan/simple parameterization*/


/*That's kinda weird...*/
DBCC FREEPROCCACHE;


SELECT 
    p.* 
FROM dbo.PushyPaul AS p
WHERE p.OwnerUserId = 22656
AND   1 = (SELECT 1); /*Avoid trivial plan/simple parameterization*/


/*Turn Down Service*/
DBCC TRACEOFF
(
    4199, 
    -1
);

SELECT 
    p.* 
FROM dbo.PushyPaul AS p
WHERE p.OwnerUserId = 22656
AND   1 = (SELECT 1); /*Avoid trivial plan/simple parameterization*/


/*Okay then.*/


/*I'm different.*/
ALTER DATABASE SCOPED CONFIGURATION 
    SET QUERY_OPTIMIZER_HOTFIXES = ON;


SELECT 
    p.* 
FROM dbo.PushyPaul AS p
WHERE p.OwnerUserId = 22656
AND   1 = (SELECT 1); /*Avoid trivial plan/simple parameterization*/



/*Cleanup*/
ALTER DATABASE 
    StackOverflow2013 
SET PARAMETERIZATION SIMPLE;

ALTER DATABASE SCOPED CONFIGURATION 
    SET QUERY_OPTIMIZER_HOTFIXES = OFF;

DBCC TRACEOFF
(
    4199, 
    -1
);

Video Summary

In this video, I delve into a specific issue in Microsoft SQL Server 2017 CU30, where the documentation left out crucial details about how parameterized queries can affect query plans. I explain that running parameterized queries skips the cell on sequence project rule, preventing pushdown and causing full index scans instead of seeks. To demonstrate this, I walk through setting up an appropriate index and running both literal and parameterized queries to illustrate the difference in execution plans. The video also covers how trace flag 4199 affects query optimization but does not clear the plan cache, while the database scope configuration method does. This discrepancy highlights the importance of understanding these nuances for effective query tuning and optimization.

Full Transcript

Alright, I apologize if the lighting is a little bit weird. It’s a, there’s kind of a weird weather day out here, and the light is very bright and white, and then I turned on my ring light to try and compensate for that. I’m not sure how that’s gonna look, I’m not sure how that’s gonna go, but anyway. I, I, I need to follow up yesterday’s video about the, the, the Sell On Seek Project issue in Microsoft SQL Server 2017 CU30, because the, the, it turns out that the, the documentation in, in the, in this, in the cumulative update, shockingly, was, left, left some stuff to be desired, left, left some crucial elements out. Now.

This is still just saying the same thing that it said yesterday. In Microsoft SQL Server 2017, running parameterized query skips the cell on sequence project rule. Therefore, pushdown does not occur.

If you click on the little link there, nothing happens. It just takes you back to this, basically takes you to the bookmark of this issue. So that’s fun.

And that leaves out, like I said, a very crucial detail. Now, I’m going to walk back. Screw you, Mac Toolbar. Who does that?

Macs are the worst. If anyone ever tries to convince you to switch over to a Mac, burn them. Burn them like the witch they are.

Or warlock they are. Whatever it is. I don’t know. Anyway. Yesterday, we ran through this demo where we created an index that very well suits both the query that we’re going to run.

You know, owner user ID score, right? We got owner user ID and score and the windowing function. And creation date and last activity date in the select list. And later, we’re going to run some queries that filter on owner user ID with an equality predicate.

So this should be a totally seekable thing. So yesterday’s video, I showed you that if we use a literal value and we run that query, we get a nice seek. The literal value gets pushed down past the sequence project operator, seeks into the index.

But when we parameterize the query, that no longer happens. We scan the whole index, do the whole dense rank windowing function thing, and then filter out later. All right.

So we’re going to start here today. And we’re going to make sure that we are starting in the right place with none of this stuff going on. We want to make sure that none of these things are in effect when we run this. So I’m going to run this query, which is the same query that we ran yesterday, essentially.

But the reason I want to run it this way is with that one equals select one is to avoid SQL Server’s cost-based optimizer, trying to use a trivial plan or use simple parameterization on our query. And when we do that, we get this thing is a literal value.

And we can see that, you know, we have a sequence project, right? This is the SEQPRJ, part of that rule that gets skipped and all that. We got a couple of segments that I don’t really care about.

But then more importantly, we have the index seek into, again, our hero chunk. Anyway, let’s mess with that a little bit. Let’s cause a problem here.

So yesterday, I used a stored procedure to show you that a parameterized query would behave differently, even with the cumulative update installed, right? So let’s set parameterization to forced for this database.

And remember, under a simple parameterization, you pass in a literal value. It’s kind of up to the optimizer whether, you know, the trivial plan, simple parameterization kicks in and you actually get a simple parameterized query.

Under forced parameterization, under most circumstances, SQL Server will be like, oh, well, cool, we can throw this right at you, right? Turn that into a parameter magically for you.

All right. So now with parameterization force turned on, let’s run this thing. And this is where things sort of start to fall over, right? Because with forced parameterization turned on, we now have a query plan that looks like this.

I didn’t mean to have that tool tip pop up. Apologize there. But you’ll notice that this looks kind of funny, right?

Everything has these little spaces and stuff between and everything’s lowercase is God intended. So if anyone out there is watching and perhaps uses capitalized table aliases, perhaps this is, you know, a pretty good sign that that’s the wrong way to do things.

Just saying. But anyway, we have owner user ID equals at zero. And this is one of my favorite parts of simple parameterization is and at one equals select one.

So I’m not really sure where they came up with that. It’s just kind of cute for me. But anyway, the query plan looks a little bit different because we got this stuff up here to deal with that.

We actually have a startup expression predicate on the literal value one equaling the at one parameter. But, you know, that’s neither here nor there. The important part is down here where we now have that index scan that we saw yesterday.

Right? And that takes a couple seconds. And over here we have a filter operator. And that filter operator is where we figure out where that parameter value that we passed in gets applied.

Now, yesterday we had the stored procedure where it was called at user ID. Today the predicate is just going to be that at zero that we saw in the query text up here. Right?

That at zero. Okay. Okay. So, you know, when I was looking into it yesterday after I recorded the original video, something that threw me off and I thought was pretty funny was that, you know, a lot of these things are hidden behind trace flags. And now a very common one that a lot of these fixes get hidden behind is trace flag 4199.

4199 has been around, I don’t know, since like SQL Server. I think, I want to say 2008, but it might even be 2005. I refuse to try to find that literature at this point.

But 4199 hides a lot of the optimizer hot fixes that end up in SQL Server. So, this was like the first thing, like after I recorded yesterday’s video, I was like, okay, calm down. Send it yourself, Erik Darling.

Stop drinking. Well, that didn’t happen. But, so if you turn on this trace flag, something kind of funny happens at first. And that you turn on trace flag 4199 and you run the query again and you get the same query plan. All right.

And this might throw you off. All right. And why might this throw you off? Good question. I was just about to ask that. That was a great question. This is the next one that you answer in the video. So, the reason why you get the same query plan, this whole thing, is that turning on trace flag 4199, which enables optimizer hot fixes, doesn’t actually clear out the plan cache.

No, it does not. So, a trace flag that directly affects optimizer behavior does not clear out the plan cache. Why?

I don’t know. I’m going to pause for a moment. Hope I don’t make any mouth sounds with that. Do hate a mouth sound. But, let’s clear out the plan cache then.

Need a little pick me up there. Let’s clear out the plan cache and rerun this. My favorite characters ever is a rerun. But now, with trace flag 4199 enabled and a fresh plan generated for this query, we get the behavior that we would expect to see based on the documentation, which does not mention trace flag 4199. Out of the box with a little modification to the box there.

Tiny little difference. So, good, right? Sort of, I guess.

No one told you that. And that’s kind of depressing. But, let’s turn off trace flag 4199. Just to prove to you that that is the case, that 4199 does not do anything to the plan cache.

We turn that off, we’re actually still going to get the same query plan as last time, right? We get the seek plan again. So, that’s kind of annoying.

One thing that is different, and one thing that does clear out the plan cache and allow you to get the plan is to use the altered database scope configuration method of turning on optimizer hotfixes. Which is probably the preferred method, to be honest. Just because, you know, turning trace flags on and off is a little tricky.

You know, they don’t persevere restarts unless you, you know, set them at SQL Server startup. Or you have a startup store procedure run to flick those switches on. But, even with, like, stuff like trace flag 8048, you know, the startup procedure option isn’t quite as good because a bunch of other stuff gets initialized first.

So, anyway. Story for a different day. But, anyway.

So, you turn on optimizer hotfixes and all of it. And, you know, you will get the fresh plan and the plan cache and clear it out and get the seek plan and all that stuff. So, that’s sort of it for this one. If you want to see your parameters get pushed past the sequence project operator, you are going to need to enable trace flag 4199 and clear out the plan cache.

Or use the database scope configuration to set hotfixes on. So, moral of the story here. Well, I guess there’s maybe two or three of them.

We’ll see how many I think of as I start talking. One, Microsoft CU documentation is crap. Real bad.

Two, trace flag 4199 does not clear out the plan cache despite the fact that it directly affects the way the optimizer handles queries. Three, the database scope configuration for query optimizer hotfixes does clear out the plan cache. And, I guess, four, why the hell wouldn’t you make both of those things behave the same way?

Three, why wouldn’t a trace flag that changes optimizer behavior clear out the plan cache so that you can immediately see that optimizer behavior? That’s a little bit weird for me. I mean, I know, like, the database scope configuration thing, that cropped up around SQL Server 2016, I think.

So, we had, let’s see, like, probably three, four versions, major versions of SQL Server between, of trace flag 4199 not clearing out the plan cache. That’s, ain’t that cute as a boot. Anyway, I’m going to go finish this espresso, we’ll call it, and, I don’t know, wait five years for this video to render on my piece of crap Macintosh computer.

And, that’ll be, that’ll be my day. Just spend the day tending to the fire that, that occurs when, when I render a video. So, anyway, you all have a wonderful Saturday, or whatever day you end up watching this on.

I hope that, hope that you, hope that you are living your best lives. Thanks for watching.

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 2017 CU 30 Doesn’t Actually Fix The Problem With Views And Parameters

Posted on by Erik Darling

I am a heading


In the release notes for SQL Server 2017 CU30, there’s a note that it fixes a problem where parameters can’t be pushed passed Sequence Project operators:

“In Microsoft SQL Server 2017, running parameterized queries skips the SelOnSeqPrj rule. Therefore, pushdown does not occur.” But it doesn’t actually do that.

Paul White Original Post: The Problem with Window Functions and Views

The Problem In The Plan

Here are the good and bad plans, comparing using a literal value vs. a parameterized value:

SQL Server Query Plan
dunksville
  • In the plan with a literal value, the predicate is applied at the index seek, and the filtering is really quick.
  • In the plan with a parameterized value, the index is scanned, and applied at a filter way later in the query plan.

This is where the SelOnSeqPrj rule comes in: The parameter can’t be pushed past the Sequence Project operator like the literal value can.

Thanks for reading!

Video Summary

In this video, I delve into some of the known issues and updates in SQL Server 2017 CU30, focusing on one particular performance-related fix that caught my attention. Despite the title suggesting a discussion about 2000, we’re actually looking at modern SQL Server versions from 2022. I explore how running parameterized queries can sometimes skip certain seek optimizations, leading to suboptimal query plans. This issue has been around for quite some time and is something I’ve been highlighting in my work. The video demonstrates this with a practical example using SSMS, showing the difference between passing literal values versus parameters within stored procedures. It’s a reminder that while SQL Server continues to evolve, there are still areas where performance optimizations could be improved, especially when it comes to documentation and clear communication of these changes.

Full Transcript

Erik Darling here with Sir Erik Darling Data. And today I want to talk about SQL Server 2017 for some reason. Don’t ask me why. It’s 2000, midway through 2000, 2022. But we got CU30 for SQL Server 2017. Very exciting stuff in there. Just kidding. It’s not, mostly not very exciting. But there was one thing in there that caught my eye. Because it’s something that’s near and dear to my heart. Query performance stuff. I don’t know if you know that about me. I tend to, tend to traffic a bit in that area of the world. So, this is version, let’s, let’s use zoom it, proper human beings here. Will I wait for Mark Russinovich to release a new version that does screen recording? That’ll be nice. But let’s zoom in a little bit here. And let’s look at version 14.0.34. So, it’s, let’s go back to version 14.0.34.5.1.2. Wonderful. Get that sorted out. Well, if you, let’s go back. Thanks, Mac Toolbar for showing up and ruining my recording. Photo bombing piece of crap. Hate this thing. So, let’s go back over to SSMS real quick. And let’s just make sure that I am on SQL Server 2017 14.0.3451.2. So, we’re all sorted out there. That’s good for us. We got that all figured out. We’re doing, doing wonderful.

So, known issues in this update. What do we have going on here? What’s, what’s happening in this release? Well, uh, something about a latch timeout. Ooh, high availability. Don’t care. Ooh, trace flag. One, two, three, two, three. Great. We’re at 12,323 trace flag. Probably higher at this point. Uh, let’s see. Uh, match lock escalation, uh, change tracking. Who cares? Access violation occurred. When you try to truncate specific partitions using the partition function. Seems funny. Uh, dropping temp tables causes an unresolved deadlock and dump file. Ooh. Wow. Don’t drop those temp tables. Uh, let’s see. An assertion failure occurs when your query contains the merge statement. Big surprise.

Uh, let’s see. When you run dbcc checkdb with extended logical checks against a database by using the table valued function tbf that uses indexes. Here is the error message. Table percent ls does not exist. I’m going to pause here for a moment and ask you, why do we accept this? Why do we tolerate this? If we can’t get any sort of decent information about, uh, what fixes are out there for a piece of software, why can’t we get them in, in something that’s at least understandable?

Like, not everything has to be a book, but a complete thought would be nice. I don’t understand when this started happening or why this started happening, but the quality of the documentation for SQL Server is real, real broken. Uh, if you look at error, like, especially new error messages or new extended events, there is absolutely no oversight in the, in the, in the language used in there.

It’s full of typos and just like they saw one, uh, Aaron Bertrand brought one up to me yesterday where, uh, availability groups have a double dash between availability and groups. There’s, or always on or something like that. That has never been what they’ve been called or how they’ve been named or referred to.

And, uh, it, it, it really is just gone completely downhill. I don’t know whose idea that was. Maybe, maybe, maybe Postgres has just infiltrated Microsoft and they’re taking them down from the inside like termites.

I don’t know. Who knows? Tough to tell out there. It’s a, it’s a harsh world, isn’t it? But here’s the one that I want to talk about.

We’ll talk about this wonderful little thing right here. In Microsoft SQL Server 2017, running parameterized queries skips the sell on seek purge rule. Therefore, push down does not occur.

Well, thankfully, this is something that I’ve been demoing for years because it’s been a problem. Uh, I think the first time I ever read about it was in a Paul White blog post coming up on 10 years ago now. Crazy, right?

A 10-year-old performance bug in SQL Server. Well, I know they’re not busy fixing performance bugs and certainly not busy writing adequate documentation for anything. So here we are reading this.

Uh, I’m not even going to bother with this one. Uh, I don’t know. Uh, index creation script fails. Cool. Great. Great write-up.

Whoever did that. Summer intern’s really working hard. Summer intern found a beer fridge, apparently. All right. Well, everyone’s working from home, so everything’s a beer fridge. Anyway, let’s go see if that actually is fixed.

So, uh, I’ve already created this index. I’m not going to sit there and make you watch me create an index over again. But just to make sure that we are on the same page here, uh, what did I do wrong? Oh, I didn’t, I didn’t highlight select.

There we go. My own quality is going downhill, I guess, too. So, uh, let’s take a look at the results here. So this thing was just restarted. Well, this, this time isn’t going to make any sense to you. It’s actually about 8.30 in the morning here.

But my VM is on West Coast time because I never bothered to change it because I don’t care. Uh, it’s a VM, right? There’s a cattle, not pets or whatever. Uh, but anyway, I am actually running the correct version of SQL Server to see this wonderful performance fix in action.

I guess I shouldn’t make fun of anyone else’s, uh, abilities and I can’t even say fix in action. Uh, but anyway, I, I’ve got an index on my, my, my post table called chunk. I forget why I called it that.

It was a long time ago, uh, but the index is on owner user ID and score descending. And it includes creation date and last act, last activity date. And that index matches up pretty well with the goals of this view, right? So we have a windowing function on owner user ID and score descending.

And my, my, uh, my formatting of this thing is a little, is disagreeable even to me. I don’t, I don’t like the way that turned out. I’m going to fix that right here in front of all of you.

All right. So now everything is on, got its own line. No one, no one has to share too much space. Everything’s maintaining proper distance. Uh, but then we’re selecting owner user ID, score, creation date, and last activity.

So that index works out pretty well for everything that we’re trying to do in there, right? We’ve got everything for our dense rank completely in order. And we’ve got, uh, our, our select list columns and the includes up there.

So joy to the world. An index has come. So what should that fix fix? Well, we’re going to turn on a query plan here and we’re going to run this select, right?

So we run this thing and we have an execution plan. Let’s zoom in on this execution plan and see what happened. Now, even though, uh, we’ve got a case of simple parameterization here, I don’t, I have a feeling this doesn’t stick.

Uh, I could, I could do some extra stuff to validate that, but, uh, I’ve already done that and it’s quite boring to watch. So we’re going to, we’re going to skip that part. But if we look down here in the query plan, because we’ve used a literal value and a simple parameterization didn’t, didn’t topple our query into the C.

Uh, we’ve got an index seek into our index called for some reason chunk. That takes 0.008 milliseconds. Wow.

What a great query tuner that Erik Darling is. We should hire him to tune all our queries. Well, maybe not so fast. Uh, so. That worked out pretty well.

Passing the literal value. Right. Everything got pushed down the query plan. Everything worked out great. Uh, happy, happy about that. But now let’s create a store procedure. All right.

Because if we go back to what that, that the cumulative update was talking about, this is when running a parameterized query. All right. Parameterized and literal value.

Well, even though it looked like it might have been simple parameterized was not actual parameterized. Right. There’s a literal value in there. So now let’s parameterize query.

Can’t, can’t get enough of the word parameterized. Makes me feel so very proper. So we’re going to run this procedure. Or we’re going to create this procedure here called stinky Pete.

I don’t know why Pete’s stinky. Same reason I don’t know why that index is chunky. Mysteries of the world. But here we have a parameter called user ID.

And we’re going to pass that parameter to our view down here. All right. Now, owner user ID equals user ID. Remember, we’ve got this wonderful index for some reason named chunk that leads with owner user ID. And so we should have, just like when we pass in a literal value, we should get a perfectly good seek to that owner user ID value.

But when I run this and a big reveal here, this does not finish in 0.008 milliseconds. In fact, this catastrophe drags on for seven seconds. And if we look at the difference in the plan, let’s zoom in real nice on that.

We have an index scan now on the post table. That takes 2.213 seconds. A bit of a far cry from the 0.008 milliseconds.

And that just tends to get worse as we move on in the plan to a 2.289 and then 4.719 and then 5.628 and then 6.054. So six seconds total for the query execution plus a little bit of time for SSMS to spit out and render our results. So, yeah, it’s pretty disappointing.

It said, hey, we fixed something and then the only demo I… Well, the best demo I have that shows the problem still has a problem. So thanks there.

Perhaps a little bit extra QA would have helped that one. Maybe that wasn’t even supposed to be in there. I don’t know. Maybe that will get pulled out of the release notes. I couldn’t tell you. No one from Microsoft talks to me anymore.

I don’t know why. I missed Joe’s sack. MungoDB got real lucky there. Well, anyway, it is 8.40 a.m. now on Friday.

And with that, I think it’s time to start drinking because there’s just no hope for the world. It’s going to be my new company tagline. There’s no hope for the world.

I’m waiting for Beer Gut Magazine to buy me out. Anyway, you have a nice day. I’m going to go pour something now. The first thing about the Shield.

For You and me now, let’s see, there are a few ways to can philosopher bless the world. It’s too late. This is a nicelegen to have an intro.

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.

Things SQL Server vNext Should Address: How Did I Do?

Posted on by Erik Darling

Mom I Did It

A while back, I wrote a bunch of posts about things I’d like to see vNext take care of. In this post, since it’s Friday and I don’t wanna do anything, will round those up and cover whether or they made it in or not.

Well, maybe I’ll need to update the list for future releases of SQL Server 2022.

Hmpf.

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 DOP Feedback: Related Extended Events

Posted on by Erik Darling

Short Round

Sort of on the heels of yesterday’s post, here are some Extended Events related to the DOP Feedback feature new to SQL Server 2022.

Here’s the complete text of a session to collect all of the related events that I’ve noticed so far:

CREATE EVENT SESSION 
    dop_feedback 
ON SERVER 
ADD EVENT 
    sqlserver.dop_feedback_eligible_query
    (
        ACTION(sqlserver.sql_text)
    ),
ADD EVENT 
    sqlserver.dop_feedback_provided
    (
        ACTION(sqlserver.sql_text)
    ),
ADD EVENT 
    sqlserver.dop_feedback_validation
    (
        ACTION(sqlserver.sql_text)
    ),
ADD EVENT 
    sqlserver.dop_feedback_reverted
    (
        ACTION(sqlserver.sql_text)
    )
ADD TARGET package0.event_file(SET filename=N'dop_feedback')
WITH 
(
    EVENT_RETENTION_MODE = ALLOW_SINGLE_EVENT_LOSS,
    MAX_DISPATCH_LATENCY = 5 SECONDS,
    STARTUP_STATE = ON
);

There is one additional event in the debug channel called maxdop_feedback_received, but the contents of it don’t appear immediately actionable.

Defining Moments

Here are the definitions for each of the events above:

  • dop_feedback_eligible_query: Reports when a query plan becomes eligible for dop feedback
  • dop_feedback_provided: Reports DOP feedback provided data for a query
  • dop_feedback_validation: Reports when the validation occurs for the query runtime stats against baseline or previous feedback stats
  • dop_feedback_reverted: This reports when a DOP feedback is reverted

Fairly straightforward, here. Also seems like a decent set of events that you’d wanna have in place.

Thanks, Microsoft.

MAPDOP

The map values for each of these events is also available:

+-----------------------+---------+--------------------------------------+
|         name          | map_key |              map_value               |
+-----------------------+---------+--------------------------------------+
| dop_calculation_stage |       0 | SetMaxDOP                            |
| dop_calculation_stage |       1 | SetTraceflag                         |
| dop_calculation_stage |       2 | CalculateBasedOnAvailableThreads     |
| dop_calculation_stage |       3 | PostCalculate                        |
| dop_feedback_state    |       0 | NotAnalyzed                          |
| dop_feedback_state    |       1 | NotEligible                          |
| dop_feedback_state    |       2 | InAnalysis                           |
| dop_feedback_state    |       3 | NoRecommendation                     |
| dop_feedback_state    |       4 | AnalysisStoppedDueToThrottling       |
| dop_feedback_state    |       5 | AnalysisStoppedDueToMaxResetsReached |
| dop_feedback_state    |       6 | AnalysisStoppedMinimumDOP            |
| dop_feedback_state    |       7 | PendingValidationTest                |
| dop_feedback_state    |       8 | VerificationRegressed                |
| dop_feedback_state    |       9 | RegressionDueToAbort                 |
| dop_feedback_state    |      10 | Stable                               |
| dop_statement_type    |       1 | Select                               |
| dop_statement_type    |       2 | Insert                               |
| dop_statement_type    |       3 | Update                               |
| dop_statement_type    |       4 | Delete                               |
| dop_statement_type    |       5 | Merge                                |
+-----------------------+---------+--------------------------------------+

Why two are zero-based and one is not is beyond what I can explain to you, here.

Perhaps that will be addressed in a future release.

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.

PASS Data Community Summit 2022: Come See My Session, The SQL Server Performance Tasting Menu

Posted on by Erik Darling

Workin’

On top of my precon, The Professional Performance Tuning Blueprint, I’ve also got a regular session selected!

The SQL Server Performance Tasting Menu:

You’re a DBA or Developer, and you’ve been using SQL Server for a few years.

You know there are different ways to make queries faster, but you’re not sure when to use them.

I’m Erik Darling, and I’ll be your sommelier for the evening.

Over several courses of delicious demos, I’ll show you the types of performance problems different tuning techniques pair well with, and which ones to avoid.

When we’re done, you’ll understand exactly what patterns to look for when you’re troubleshooting slow queries, and how to approach them.

You’ll have the secret recipe for gourmet queries.

Dates And Times

The PASS Data Community Summit is taking place in Seattle November 15-18, 2022 and online.

You can register here, to attend online or in-person. I’ll be there in all my fleshy goodness, and I hope to see you there too!

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.

Fixing Ordered Column Store Sorting In SQL Server 2022

Posted on by Erik Darling

Groove Is In The Heart

When Brent posted about the availability of, and disappointment with creating ordered column store indexes in SQL Server 2022, I got to work.

I can’t have my dear friend Brent being all distraught with all those fast cars around. That’s how accidents happen, and I fear he might leave the Blitz scripts to me in his will or something.

Anyway, I decided to dig in and see what was going on behind the scenes. Which of course, means query plans, and bothering people who are really good at debuggers.

Most of the problems that you’ll run into in SQL Server will come from sorting data.

Whenever I have to think about Sorts, I head to this post about all the different Sorts you might see in a query plan.

More on that later, though.

Cod Piece

In Paul’s post, he talks about using undocumented trace flag 8666 to get additional details about Sort operators.

Let’s do that. Paul is smart, though he is always completely wrong about which season it is.

DROP TABLE IF EXISTS
    dbo.Votes_CCI;

SELECT
    v.*
INTO dbo.Votes_CCI
FROM dbo.Votes AS v;

I’m using the Votes table because it’s nice and narrow and I don’t have to tinker with any string columns.

Strings in databases were a mistake, after all.

DBCC TRACEON(8666);
CREATE CLUSTERED COLUMNSTORE INDEX
    vcci
ON dbo.Votes_CCI
ORDER (Postid);
DBCC TRACEOFF(8666);

Here’s what we get back in the query plan:

SQL Server Query Plan
Tainted Sort

We’ve got a Soft Sort! What does our seasonally maladjusted friend say about those?

A “soft sort” uses only its primary memory grant and never spills. It doesn’t guarantee fully-sorted output. Each sort run using the available memory grant will be sorted. A “sort sort” represents a best effort given the resource available. This property can be used to infer that a Sort is implemented with CQScanPartitionSortNew without attaching a debugger. The meaning of the InMemory property flag shown above will be covered in part 2. It does not indicate whether a regular sort was performed in memory or not.

Well, with that attitude, it’s not surprising that there are so many overlapping buckets in the column store index. If it’s not good enough, what can you do?

Building the index with the Soft Sort here also leads to things being as bad as they were in Brent’s post.

Insert Debugging Here

Alas, there’s (almost) always a way. Microsoft keeps making these trace flag things.

There are a bunch of different ways to track them down, but figuring out the behavior of random trace flags that you may find just by enabling them isn’t easy.

One way to tie a trace flag to a behavior is to use WinDbg to step through different behaviors in action, and see if SQL Server checks to see if a trace flag is enabled when that behavior is performed.

If you catch that, you can be reasonably sure that the trace flag will have some impact on the behavior. Not all trace flags can be enabled at runtime. Some need to be enabled as startup options.

Sometimes it’s hours and hours of work to track this stuff down, and other times Paul White (b|t) already has notes on helpful ones.

The trace flag below, 2417, is present going back to SQL Server 2014, and can help with the Soft Sort issues we’re seeing when building ordered clustered column store indexes today.

Here’s another one:

DBCC TRACEON(8666, 2417);
CREATE CLUSTERED COLUMNSTORE INDEX
    vcci
ON dbo.Votes_CCI
ORDER (Postid)
WITH(MAXDOP = 1);
DBCC TRACEOFF(8666, 2417);

The MAXDOP 1 hint isn’t strictly necessary. With a parallel plan, you may see up to DOP overlapping row groups.

SQL Server Query Plan
community service

That’s why it was a popular maneuver to emulate this behavior by creating a clustered row store index, and then create a clustered column store index over it with drop existing and a MAXDOP 1 hint.

At DOP 1, you don’t see that overlap. It takes a lot longer of course — 3 minutes instead of 30 or so seconds — which is a real bummer. But without it, you could see DOP over lapping rowgroups.

If you want All The Pretty Little Rowgroups, this is what you have to do.

Anyway, the result using sp_BlitzIndex looks a lot better now:

EXEC sp_BlitzIndex
    @TableName = 'Votes_CCI';
SQL Server Query Results
capture the flag

How nice.

You can also use undocumented and unsupported trace flag 11621, which is

[A] feature flag for the ‘partition sort on column store order’ so the end result is similar, but via a different mechanism to 2417.
A partition sort is useful in general to prevent unnecessary switching between partitions. If you sort the stream by partition, you process all the rows for one before moving on to the next. A soft sort is ok there because it’s just a performance optimization. Worst case, you end up switching between partitions quite often because the sort ran out of memory, but correct results will still occur.

Chain Gang

A “reasonable” alternative to trace flags maybe to adjust the index create memory configuration option. If we set it down to the minimum value, we get a “helpful” error message:

EXEC sys.sp_configure 
    'index create memory', 
    704;

RECONFIGURE;

As promised:

Msg 8606, Level 17, State 1, Line 31

This index operation requires 123208 KB of memory per DOP.

The total requirement of 985800 KB for DOP of 8 is greater than the sp_configure value of 704 KB set for the advanced server configuration option “index create memory (KB)”.

Increase this setting or reduce DOP and rerun the query.

If you get the actual execution plan for the clustered column store index create or rebuild with the Soft Sort disabled and look at the memory grant, you get a reasonable estimate for what to set index create memory to.

Changing it does two things:

  • Avoids the very low memory grant that Soft Sorts receive, and causes the uneven row groups
  • The Soft Sort keeps the index create from going above that index create memory number

Setting index create memory for this particular index creation/rebuild to 5,561,824 gets you the nice, even row groups (at MAXDOP 1) that we saw when disabling the Soft Sort entirely.

Bottom line, here is that uneven row groups happen with column store indexes when there’s a:

  • Parallel create/rebuild
  • Low memory grant create/rebuild

If this sort of thing is particularly important to you, you could adjust index create memory to a value that allows the Soft Sort adequate memory.

But that’s a hell of a lot of work, and I hope Microsoft just fixes this in a later build.

Reality Bites

The cute thing here is that, while this syntactical functionality has been available in Azure Cloud Nonsense© for some time, no one uses that, so no one cares.

The bits for this were technically available in SQL Server 2019 as well, but I’m not telling you how to do that. It’s not supported, and bad things might happen if you use it.

I mean, bad things happen in SQL Server 2022 where it’s supported unless you use an undocumented trace flag, but… Uh. I dunno.

This trace flag seems to set things back to how things worked in the Before Times, though, which is probably how they should have stayed.

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.

Considerations For Paging Queries In SQL Server With Batch Mode (Don’t Use OFFSET/FETCH)

Posted on by Erik Darling

First Things First

The first SQL Server blog posts that I ever read while trying to solve a specific problem here these two:

They sort of changed my life a little, despite the author’s aversion to the letter Z. So that’s cool. Can’t have everything.

To this day, though, I see people screw up paging queries in numerous ways.

  • Selecting all the columns in one go
  • Adding in joins when exists will do
  • Sticking a DISTINCT on there just because
  • Thinking a view will solve some problem
  • Piles and piles of UDFs
  • Local variables for TOP or OFFSET/FETCH
  • Not paying attention to indexing

It’s sort of like every other query I see, except with additional complications.

Especially cute for a query slathered in NOLOCK hints is the oft-accompanying concern that “data might change and people might see something weird when they query for the next page”.

Okay, pal. Now you’re concerned.

Modern Love

A while back I recorded a video about using nonclustered column store indexes to improve the performance of paging queries:

While a lot of the details in there are still true, I want to talk about something slightly different today. While nonclustered column store indexes make great data sources for queries with unpredictable search predicates, they’re not strictly necessary to get batch mode anymore.

With SQL Server 2019, you can get batch mode on row store indexes, as long as you’re on Enterprise Edition, and in compatibility level 150.

Deal with it.

The thing is, how you structure your paging queries can definitely hurt your chances of getting that optimization.

Saddened Face

The bummer here is that the paging technique that I learned from Paul’s articles (linked above) doesn’t seem to qualify for batch mode on row store without a column store index in place, so they don’t make the demo cut here.

The good news is that if you’re going to approach this with any degree of hope for performance, you’re gonna be using a column store index anyway.

The two methods we’re going to look at are OFFSET/FETCH and a more traditional ROW_NUMBER query.

As you may have picked up from the title, one will turn out better, and it’s not the OFFSET/FETCH variety. Especially as you get larger, or go deeper into results, it becomes a real boat anchor.

Anyway, let’s examine, as they say in France.

Barfset Wretch

This is the best way of writing this query that I can come up with.

DECLARE 
    @page_number int = 1,
    @page_size int = 1000;

WITH 
    paging AS
(
    SELECT 
        p.Id
    FROM dbo.Posts AS p
    ORDER BY 
        p.LastActivityDate, 
        p.Id 
    OFFSET ((@page_number - 1) * @page_size) 
    ROW FETCH NEXT (@page_size) ROWS ONLY
)
SELECT 
    p.*
FROM paging AS pg
JOIN dbo.Posts AS p
    ON pg.id = p.Id
ORDER BY 
    p.LastActivityDate,
    p.Id
OPTION (RECOMPILE);

Note that the local variables don’t come into play so much here because of the recompile hint.

Still, just to grab 1000 rows, this query takes just about 4 seconds.

SQL Server Query Plan
what took you so long?

This is not so good.

Examine!

Hero Number

The better-performing query here with the batch mode on row store enhancement(!) is using a single filtered ROW_NUMBER to grab the rows we care about.

DECLARE 
    @page_number int = 1,
    @page_size int = 1000;

WITH 
    fetching AS
(
    SELECT 
        p.Id, 
        n = 
            ROW_NUMBER() OVER 
            ( 
                ORDER BY
                    p.LastActivityDate, 
                    p.Id 
            )
    FROM dbo.Posts AS p
)
SELECT 
    p.*
FROM fetching AS f
JOIN dbo.Posts AS p
    ON f.Id = p.Id
WHERE f.n > ((@page_number - 1) * @page_size)
AND   f.n < ((@page_number * @page_size) + 1)
ORDER BY 
    p.LastActivityDate,
    p.Id
OPTION (RECOMPILE);

Again, this is about the best I can write the query. Maybe you have a better way. Maybe you don’t.

Mine takes a shade under 2 seconds. Twice as fast. Examine!

SQL Server Query Plan
cell tv

I’ll take twice as fast any day of the week.

Compare/Contrast

The OFFSET/FETCH query plan is all in row mode, while the ROW_NUMBER query has batch mode elements.

You can see this by eyeballing the plan: it has a window aggregate operator, and an adaptive join. There are other batch mode operators here, but none have visual cues in the graphical elements of the plan.

This is part of what makes things faster, of course. The differences can be even more profound when you add in the “real life” stuff that paging queries usually require. Filtering, joining, other sorting elements, etc.

Anyway, the point here is that how you write your paging queries from the start can make a big difference in how they end up, performance-wise.

Newer versions of SQL Server where certain behaviors are locked behind heuristics (absent column store indexes being present in some manner) can be especially fickle.

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 Parameter Sensitive Plan Optimization: Does Not Care To Fix Your Local Variable Problems

Posted on by Erik Darling

–To Fix Parameter Sniffing

There are some code comments you see that really set the stage for how tuning a query is going to go.

Usually one misgiving about how SQL Server works gives way to a whole levee-breaking bevy of other ones and three days later you can’t feel your legs but dammit it’s done.

Okay, maybe it was three hours, but it felt like three days. Something about the gravitation pull of these black hole queries.

One fix I’ve been wishing for, or wish I’ve been fixing for, is a cure for local variables. I’d even be cool if Forced Parameterization was that cure, but you know…

Time will tell.

Husk

Let’s say we’ve got this stored procedure, which does something similar to the “I’m gonna fix parameter sniffing with a local variable hey why is everything around me turning to brimstone before my very eyes?” idea, but with… less of an end-of-times vibe.

CREATE OR ALTER PROCEDURE 
    dbo.IndexTuningMaster
( 
    @OwnerUserId int,
    @ParentId int, 
    @PostTypeId int 
)
AS
BEGIN
SET NOCOUNT, XACT_ABORT ON;

    /*Someone passed in bad data and we got a bad query plan,
      and we have to make sure that doesn't happen again*/
    
    DECLARE 
        @ParentIdFix int = 
            CASE 
                WHEN @ParentId < 0 
                THEN 0 
                ELSE @ParentId 
            END;
    
    SELECT TOP (1) 
        p.*
    FROM dbo.Posts AS p
    WHERE p.ParentId = @ParentIdFix
    AND   p.PostTypeId = @PostTypeId
    AND   p.OwnerUserId = @OwnerUserId
    ORDER BY 
        p.Score DESC, 
        p.Id DESC;

END;

How bad could a top 1 query be, anyway?

Fortune Teller

When we run this query like so and so:

EXEC dbo.IndexTuningMaster 
    @OwnerUserId = 22656, 
    @ParentId = 0, 
    @PostTypeId = 1;

EXEC dbo.IndexTuningMaster 
    @OwnerUserId = 22656, 
    @ParentId = 184618, 
    @PostTypeId = 2;

We come up with zip zero zilch none nada:

SQL Server Query Plan
still playing

We get a super low guess for both. obviously that guess hurts a large set of matched data far worse than a small one, but the important thing here is that both queries receive the same bad guess.

This is a direct side effect of the local variable’s poor estimate, which PSP isn’t quite yet ready to go up against.

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 Parameter Sensitive Plan Optimization: How PSP Can Help Some Queries With IF Branches

Posted on by Erik Darling

Time Served

I’ve spent a bit of time talking about how IF branches can break query performance really badly in SQL Server.

While the Parameter Sensitive Plan (PSP) optimization won’t fix every problem with this lazy coding habit, it can fix some of them in very specific circumstances, assuming:

  • The parameter is eligible for PSP
  • The parameter is present across IF branches

We’re going to use a simple one parameter example to illustrate the potential utility here.

After all, if I make these things too complicated, someone might leave a comment question.

The horror

IFTTT

Here’s the procedure we’re using. The point is to execute one branch if @Reputation parameter is equal to one, and another branch if it equals something else.

In the bad old days, both queries would get a plan optimized at compile time, and neither one would get the performance boost that you hoped for.

In the good news days that you’ll probably get to experience around 2025, things are different!

CREATE OR ALTER PROCEDURE 
    dbo.IFTTT 
(
    @Reputation int
)
AS 
BEGIN
SET NOCOUNT, XACT_ABORT ON;

SET STATISTICS XML ON;  

    IF @Reputation = 1
    BEGIN
        SELECT 
            u.Id, 
            u.DisplayName, 
            u.Reputation, 
            u.CreationDate
        FROM dbo.Users AS u
        WHERE u.Reputation = @Reputation;
    END;

    IF @Reputation > 1
    BEGIN
        SELECT 
            u.Id, 
            u.DisplayName, 
            u.Reputation, 
            u.CreationDate
        FROM dbo.Users AS u
        WHERE u.Reputation = @Reputation;
    END;

SET STATISTICS XML OFF; 

END;
GO

Johnson & Johnson

If we execute these queries back to back, each one gets a new plan:

EXEC dbo.IFTTT 
    @Reputation = 1;
GO 

EXEC dbo.IFTTT 
    @Reputation = 2;
GO
SQL Server Query Plan
psychic driving

Optimize For You

The reason why is in the resulting queries, as usual. The Reputation column has enough skew present to trigger the PSP optimization, so executions with differently-bucketed parameter values end up with different plans.

option (PLAN PER VALUE(QueryVariantID = 3, predicate_range([StackOverflow2013].[dbo].[Users].[Reputation] = @Reputation, 100.0, 1000000.0)))

option (PLAN PER VALUE(QueryVariantID = 2, predicate_range([StackOverflow2013].[dbo].[Users].[Reputation] = @Reputation, 100.0, 1000000.0)))

And of course, each plan has different compile and runtime values:

SQL Server Query Plan
care

If I were to run this demo in a compatibility level under 160, this would all look totally different.

This is one change I’m sort of interested to see the play-out on.

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 Parameter Sensitive Plan Optimization: Sometimes There’s Nothing To Fix

Posted on by Erik Darling

Best Intentions

After seeing places where the Parameter Sensitive Plan (PSP) optimization quite stubbornly refuses to kick in, it’s somewhat amusing to see it kick in where it can’t possibly have any positive impact.

Even though some parameters are responsible for filtering on columns with highly skewed data, certain other factors may be present that don’t allow for the type of plan quality issues you might run into under normal parameter sensitivity scenarios:

  • Adequate indexing
  • Row goals
  • Other filtering elements

This isn’t to say that they can always prevent problems, but they certainly tend to reduce risks much of the time.

If only everything were always ever perfect, you know?

Setup

Let’s start by examining some data in the Posts table.

First, PostTypeIds:

SQL Server Query Results
resultant

Questions and answers are the main types of Posts. The data is clearly skewed, here, and in my testing this does qualify for PSP on its own.

The thing is, there are several attributes that Questions can have that Answers can’t. One of those is a ParentId. Looking through how the top 15 or so of those counts breaks down:

SQL Server Query Results
hitherto

Okay, so! Wikis don’t have ParentIds, neither do Moderator Nominations. More importantly, Questions don’t.

The Question with the Most answers is Id 184618, with 518. A far cry from the next-nearest Post Types, and light years from the number of Questions with a ParentId of zero.

More important than loving your data is knowing your data.

To Query A Butterfly

Let’s say we have this query:

SELECT TOP (5000)
    p.Id,
    p.OwnerUserId,
    p.Score
FROM dbo.Posts AS p
WHERE p.PostTypeId = @po
AND   p.ParentId = @pa
ORDER BY 
    p.Score DESC;

The three things we care about getting done are:

  • Filtering to PostTypeId
  • Filtering to ParentId
  • Ordering by Score

Either of these indexes would be suitable for that:

CREATE INDEX 
    popa
ON dbo.Posts
(
    PostTypeId,
    ParentId,
    Score DESC
)
WITH
(
    SORT_IN_TEMPDB = ON,
    DATA_COMPRESSION = PAGE
);

CREATE INDEX 
    papo
ON dbo.Posts
(
    ParentId,
    PostTypeId,
    Score DESC
)
WITH
(
    SORT_IN_TEMPDB = ON,
    DATA_COMPRESSION = PAGE
);

With No PSP At All

Under compatibility level 150, we can run the query in a variety of ways and get nearly identical performance results:

SQL Server Query Plan
PostTypeId = 1, ParentId = 0
SQL Server Query Plan
PostTypeId = 2, ParentId = 184618

There’s a 27 millisecond difference between the two to find the first 5000 rows that match both predicates. You would have to run these in a very long loop to accumulate a meaningful overall difference.

In this case, both queries use and reuse the same execution plan. You can see that in the estimates.

With All The PSP

Switching to compat level 160, the queries are injected with the PLAN PER VALUE hint.

SELECT TOP (5000)
    p.Id,
    p.OwnerUserId,
    p.Score
FROM dbo.Posts AS p
WHERE p.PostTypeId = @po
AND   p.ParentId = @pa
ORDER BY 
    p.Score DESC 
OPTION 
(
    PLAN PER VALUE
    (
        QueryVariantID = 2, 
        predicate_range
        (
            [StackOverflow2013].[dbo].[Posts].[PostTypeId] = @po, 
            100.0, 
            10000000.0
        )
    )
)

SELECT TOP (5000)
    p.Id,
    p.OwnerUserId,
    p.Score
FROM dbo.Posts AS p
WHERE p.PostTypeId = @po
AND   p.ParentId = @pa
ORDER BY 
    p.Score DESC 
OPTION 
(
    PLAN PER VALUE
    (
        QueryVariantID = 3, 
        predicate_range
        (
            [StackOverflow2013].[dbo].[Posts].[PostTypeId] = @po, 
            100.0, 
            10000000.0
        )
    )
)

The thing is, both queries end up with identical execution times to when there was no PSP involved at all.

In other words, there is no parameter sensitivity in this scenario, despite there being skew in the column data.

Even searching for the “big” result — Questions with a ParentId of zero, finishes in <30 milliseconds.

Ah well. Gotta train the models somehow.

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.