Category: Query Tuning

Erik Darling see a marginal improvement between CTPs in SQL Server 2019’s ability to optimize scalar functions:

To test FROID, which is the codename for Microsoft’s initiative to inline those awful scalar valued function things that people have been griping about for like 20 years, I like to take functions I’ve seen used in real life and adapt them a bit to work in the Stack Overflow database.

Read on for a case where Erik saw major improvements.

Arthur Daniels helps us figure out which of SQL Server’s cardinality estimators your query used:

SQL Server 2008 is reaching end of support this year, so upgrading your SQL Server might be on your mind. One of the big changes when you upgrade your SQL Servers is upgrading the compatibility level, which by default will upgrade the cardinality estimator (CE).

This can change query performance, for better or for worse. This post won’t focus on whether it’s good or bad, but instead I want to show you how you can check to see what CE was used by your queries.

It’s not a 100% guarantee, but generally I’ve found the new estimator to be superior.

Erik Darling shows a couple examples of how the optimizer will sometimes pick a superior plan when dealing with complicated filters but not always:

Sometimes, the optimizer can take a query with a complex where clause, and turn it into two queries.

This only happens up to a certain point in complexity, and only if you have really specific indexes to allow these kinds of plan choices.

This is a case where I generally don’t trust the optimizer to get it right. Even when it does, I’d be concerned that this won’t be a stable solution and a minor change somewhere could result in regression to a bad plan.

Erik Darling explains that the nested loop operator is like a duck: there’s more going on beneath the surface than it lets on:

I’m going to talk about my favorite example, because it can cause a lot of confusion, and can hide a lot of the work it’s doing behind what appears to be a friendly little operator.

Something to keep in mind is that I’m looking at the actual plans. If you’re looking at estimated/cached plans, the information you get back may be inaccurate, or may only be accurate for the cached version of the plan. A query plan reused by with parameters that require a different amount of work may have very different numbers.

I like nested loop joins a lot, but there’s a big difference between a loop running a few dozen times and a loop running a couple hundred thousand times, even if the operator doesn’t show you that immediately.

Brent Ozar takes us through a realistic but nasty scenario:

When I do performance tuning for clients, I really pride myself on making as few changes as possible in order to make a tremendous difference. I consider it a failure when I have to tell someone to rewrite a bunch of queries from scratch.
However, there are some cases where I just can’t work around a perfect storm of anti-patterns. To understand why, let’s take an example. I’ve kept the general idea the same, but I’ve rewritten the entire example in the Stack Overflow database to protect the innocent.

I’ve seen cases similar to what Brent has. Developers understand encapsulation and minimizing code repetition, so they naturally want to do that with SQL, but the optimizer eventually gives up and picks a terrible plan. DRY is great for application code and normalization, but unfortunately, it’s not always great for T-SQL.

Erik Darling shows us a case where the same query can be nice and fast, but change one parameter and suddenly performance goes out the window:

In the original plan, the TOP asked for rows, and quickly got them.
In the second plan, the TOP kept asking for rows, getting them from the Votes table, and then losing them on the join to Posts.
There was no parameter sniffing, there were no out of date stats, no blocking, or any other oddities. It’s just plain bad luck because of the data’s relationship.

Read the whole thing.