Psssssp
Thanks for watching!
Video Summary
In this video, I delve into the exciting new parameter-sensitive plan optimization introduced in SQL Server 2022. After a bit of a hiatus from recording videos due to other commitments and personal projects, the release of SQL Server 2022 has reignited my passion for sharing knowledge with you all. The video explores how this feature works by running queries multiple times with different input values, demonstrating the creation of distinct plans based on parameter sensitivity. I also highlight some quirks in query store data handling and plan cache management that arise from this optimization, offering insights into what developers might encounter as they adapt to these changes.
Full Transcript
Hello, hello, hello. Erik Darling here with Erik Darling Data. And I know that it’s been a little while since I recorded anything. That’s mostly because I’ve been really busy. I’ve still been writing nice things for all you kind and pleasant folks out there, but recording, I don’t know, it takes a special kind of thing for me to want to record something. And a lot of the stuff that I’ve been writing about has just not been stuff that I’ve been like, yeah, let’s record it. Let’s do a video about it. So I’ve been slacking a little bit on there. And, but now we have SQL Server 2022, and it has rekindled my desire to record things. So you’re welcome. Thank you, Microsoft, for releasing SQL Server 2022. And, I’ve been digging a lot into the parameter-sensitive plan optimization in SQL Server 2022. Because, I mean, apart from it being a very cool thing that now we have this, like, new layer of intelligent query processing where, all of a sudden, like, we don’t have the same, well, we don’t have the same. We don’t always have the same issue with parameter. We have new issues to deal with, like, when it misses out on perhaps a potentially useful optimization.
But anyway, now we have at least the built-in heuristic ability to occasionally get some help with parameter sniffing when the optimizer deems a parameter, a quality predicate, sensitive enough or having a sufficient skewness to generate multiple plans. So, what I’m going to do is run this main query here in a loop 10 times with two different sets of input values. One is going to be a 1 and a 0 up here, and the other is going to be a 2 and a 184618 down here. So if I run this, first we’re going to clear out the plan cache. We’re going to set, we’re going to clear all the data out of query store because I want this to be very clear.
All right, I’m going to clear it to make it, shut up. And I’m also just going to make sure that everyone knows that my database is in compat level 160, which is the magical new SQL Server 2022 compat level. I mean, I guess it’s old news if you’re up in Azure, but I don’t know. I don’t think people in Azure care, obviously. If they did, they wouldn’t use Azure. So we’re going to run those in a loop, and it’s going to finish pretty quickly because it’s an expert query tuner.
Even when I have queries that are parameter sensitive, they still manage to finish very quickly when I get my hands on them. But the main stuff that we need to pay attention to in here is if you look up here, we’re going to have, and this is how you’re going to know that your plans have received the parameter sensitive plan optimization. You’re going to see this option plan per value, and you’re going to get some query variant ID attached to your query.
Right now, you only have three options, or you only get three plans as part of this, right? So you will certainly only create three parameters, different parameter sensitive plans based on like bucketized values. And you can see those bucketized values if you look a little bit further over in the text of that hint where you see this predicate range, right?
And so you have stuff at the low end of the range, and you’ll have stuff at the high end of the range, and then you’ll have stuff in the middle of the range, which would be your third plan. I’m at the very top and very bottom of this, so I only have the two variant IDs. But if we scroll down through this, we’re going to see two distinct plans.
Really, they’re very similar in a lot of ways. The only thing that’s really different are these estimates over here, because one of them will actually find 50,000 rows, and the other one will not find anywhere near 50,000 rows. They’ll only find 518 rows.
And you can also see, of course, which parameter predicate… Oh, I didn’t go over it further enough. I’m goofy. I’m all out of practice here. You can also see which parameter SQL Server decided was the sensitive one, the sad little lad who dropped his lolly and leaves. And decided that that’s the one that needed some extra help and attention.
Needed a Kleenex and some spit on the cheek. But if you look through this stuff, these two plans are basically going to go back and forth over and over again. It’s going to be all query variant 2 and all query variant 3 going all the way down to the bottom.
And that’s the end of it, right? So these both executed 10 times a piece, just sort of in that loop. Now, where things get interesting…
If you read my blog, you would have seen a post sort of recently about how the parameter sensitive plan optimization, at least right now, as of this build, this first CTP build, kind of messes up plan cache stuff. Because when a store procedure calls a query and the query executes, it used to be, or at least it currently is, that you can figure out… Like there’s a parent object ID that gets tied back from the query to the store procedure that call this.
You can be like, oh, this query came from the store procedure. Great. That messes that up. They all get weird different SQL handles. Everything’s kind of a weird shamble over there.
There’s sort of a similar situation in query store. It’s not exactly the same. I mean, it is the same with you can’t tie statements back to the store procedure that called them. But there’s an additional sort of bit of weirdness that I want to talk you through here.
So what I’m going to do is execute these queries. And there’s nothing all that special or interesting about these queries. I’m not going to explain them because querying query store data is the hell on earth. I feel bad for anyone who does it.
I wrote spquickie store so that you wouldn’t have to do it because I care about you deeply as people. And I want you to have long and happy lives so you don’t have to think about these things. So looking through query store data, we have three distinct executions of this query, which is weird because we only had two plans, but I’m going to show you where that third thing comes from.
So if we blow this column up a little bit, and I wish there was a better way to do this without running some dumb substring parsing, we’ll see in this section that top plan, that ends the score descending clause of the query. The two ones below it have the plan per value, predicate range, blah, blah, blah stuff at the end of it.
All well and good. Where things get kind of interesting though, is if we, come on, move it, some scroll bars, not anyone’s friend, is when we look through other parts of this, right?
So we have query ID 123, plan ID 123, query text ID 123, okay, fine. Plan ID 1 doesn’t show up in runtime stats, right? So just kind of like a silly run, like query on, to get data out of the runtime stats view.
Query ID 1 doesn’t show up in there, right? We have, or sorry, plan ID 1 doesn’t show up there. We have plan ID 2 and 3 in there, right? So these two are in there.
This top one isn’t. Down at the bottom, and I will explain this query because this one is relatively simple to explain. We’re hitting a new view in SQL Server 2022 called sys.queryStoreVariant. And that is brand new.
And this is the one that tells us when a plan had a variant or a deviant in it somewhere. What’s kind of interesting here is that we have query variant query ID 2 and 3, right? So these query IDs would match up to these query IDs here.
Query ID 1 isn’t in there, right? Because it doesn’t have that option plan per value stuff in it. We have parent query ID, which does match up to that. And then we have dispatcher plan ID, which only shows you one.
Okay. Here’s where things get a little tricky. If we were to look for a plan, if we were to purely look for plans that have a variant in them, right? That were, you know, dispatched some other query plans.
This is all we get back. And as far as I know, I’m the first person to talk about this. So, you know, go me. Gold star, A plus, A for effort, all that stuff. We get this single operator back.
Multiple plan. Right? That’s all we get back here. Not a whole lot of detail. And if we go look at the XML, there’s a little bit more to see, but it’s kind of just weird stuff. You know, I mean, it’s all, you know, XML, which sucks to look at.
But we have this parameter sensitive predicate thing in here. We have a little bit of information about the, you know, the query. Like we can see what the query text is.
Just doesn’t have the parameter sensitive stuff tacked onto it. We can see the index that got used. We can see the boundary values for the parameter sensitivity stuff. And we can see the predicate that was considered parameter sensitive. Right?
So we get, there’s some, there’s some stuff in there, but it doesn’t show us the whole execution plan. If we click on any other, any of those other query plans, we’ll get the plan back. Like I said, there was the one that has a missing index attached to it. And then this one down here is the one that doesn’t have the missing index attached to it.
So those are the two variants that we saw when I executed the other scripts. So you have to be a little bit careful how you identify stuff in query store. Obviously you’d want to use the query variant query ID.
If you wanted to find the actual statements that had the varying plans in them. But right now getting, getting this dispatch or plan ID is just kind of useless. I wish there was another column that maybe had query variant plan ID in it so that we could say, Oh, this query was attached to this, you know, variant on the plan.
That might be kind of nice. The other thing that’s kind of interesting to look at in here is some of the hashes, right? So somehow these all have the same query hash, which is interesting.
I mean, because of the two of them have the option plan stuff at the end of them. So it almost doesn’t make sense for them to have that, but maybe, maybe that’s not considered as part of the hash value in there. Maybe that gets, maybe that gets sent in afterwards.
I don’t know. And then we have the query plan hash, which, you know, I guess that makes sense because, you know, we have this one where, you know, the, obviously that just that single multiple plan operator is not going to hash out to the same stuff. But then, you know, I mean, the other two are the same because it’s fundamentally the same plan.
It’s a blah, blah, key lookup. I mean, there was no difference in plan shape or anything. So in situations where the, the plan vary, the plan, the different plans generated had different shapes, different operators, parallel, not parallel, stuff like that. These would likely, these would likely vary.
What’s interesting to me though, is we have this query hash over here where they all end up being the same, but then quite obviously for last compile batch SQL handle and statement SQL handle, these are all very different values, right? This one’s 02, blah, blah, 51, 87, 72.
So these are all very different, right? These are all like just completely different SQL handles for things, right? These are all, these are all way different values. And even the statement SQL handles over, over here, obviously quite different values of the most will cooperate. Thank you very much, Mr. Mouse.
You are so useful today. Almost as useful as my liver. These all have very different values, right? So like, there’s a lot of stuff that’s just sort of spread out in here where, you know, before, there would at least be like some semblance of, of like, stuff you could figure out, right?
You could like, say, oh, the SQL handle for this query hasn’t changed. So this is where stuff gets a little bit trickier to identify and kind of get, get to the bottom of in here. So I guess the main thing here is be on the lookout for this, these kinds of changes.
Be on the lookout for maybe your query store scripts not working or getting you the value data that you want back as you start getting these new views into things. Obviously this dispatcher plan ID that just shows you the multiple plan operators, not very useful, right?
Like that’s not a very good query plan for you to use. The parent query ID likewise, that’s just going to get you this thing back, right? Query ID one, right?
That’s just going to get you that, you know, like that, the, the initial query without any of the parameter sensitive stuff attached to it. But you really want to be using this as query variant query ID, because this is what’s going to get you to these column to rather to these rows over here that have the more interesting plans. So, you know, be aware of this stuff, you know, as much as I love the idea of fixing parameter sniffing, I wish that there was a little bit less sort of nonsense, nonsense and shenanigans going on behind the scenes with all of the, the, the different values that, the, the queries generate and stuff.
I understand there has to be some separation. It wouldn’t make sense for everything to be just kind of congealed in together, but I do wish that there were a little bit more tooling or visibility into some of this stuff. For example, I mean, like, like I said earlier, it would be really helpful to have a query variant plan ID in here to track us to the plan IDs that we care about where there are variations.
And, you know, even in query store, you know, I talked about the plan cache earlier, even in query store, anything that gets called in the store procedure has a zero for a parent object ID when it gets a parameter sensitive plan optimization. So it’s going to be interesting seeing if that’s kind of stuff gets fixed or worked out as, as the CTPs roll out and we get closer to closer to RTM time. But anyway, sort of an initial observation, maybe at some point in the future, this video will be completely wrong.
Uh, that would be nice, but who knows? Uh, anyway, uh, I’m going to go now cause it’s, um, yeah, again, it’s seven, like seven 30 on a Friday. I’m going to, I’m going to go, uh, have, pour me a nice glass of something into my, my beer gut magazine, uh, mug.
And, uh, you know, maybe, maybe, maybe I’ll remember recording this video tomorrow. Maybe not. We’ll see. It’s all, it’s all a coin toss. It’s all up in the air.
Anyway, thanks for watching. Uh, I will see you hopefully in another video soon. Now the SQL Server 2022 stuff does have me, uh, back on to, uh, to actually record videos again. So I don’t know. I’ll have to work on that.
You have to brush off the rust and anything. Anyway, goodbye.
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.
