Discarded
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:
Understand Your Plan: Operators That Write Data (Spools, Spools, Spools)
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.