test: add benchmark for select querybuilder #8955

draaglom · 2022-05-01T18:00:28Z

Description of change

Motivation: the query builder (and within it, replacePropertyNames and
associated functions) is pretty CPU intensive. For our workload, it's
one of the hottest functions in our entire stack.

While improved in #4760,
There are still outstanding issues relating to perf e.g. #3857

As we all know though, the first step in optimization is to measure
systematically ;)

https://wiki.c2.com/?ProfileBeforeOptimizing

On my machine, this benchmark runs in ~3550ms or about 0.35ms/query.
This tells us there's a way to go - on my stack, that's about 1/3 of a
typical query's latency!

Pull-Request Checklist

Code is up-to-date with the master branch
npm run format to apply prettier formatting
npm run test passes with this change
This pull request links relevant issues as Fixes #0000
There are new or updated unit tests validating the change
Documentation has been updated to reflect this change
The new commits follow conventions explained in CONTRIBUTING.md

Motivation: the query builder (and within it, replacePropertyNames and associated functions) is pretty CPU intensive. For our workload, it's one of the hottest functions in our entire stack. While improved in typeorm#4760, There are still outstanding issues relating to perf e.g. typeorm#3857 As we all know though, the first step in optimization is to measure systematically ;) https://wiki.c2.com/?ProfileBeforeOptimizing On my machine, this benchmark runs in ~3500ms or about 0.35ms/query. This tells us there's a way to go - on my stack, that's about 1/3 of a typical query's latency!

Context: the query builder is pretty CPU intensive, and can be slow - e.g. typeorm#3857 One of the things which makes this slow is `escapeRegExp` in the query builder: we freshly construct the same RegExp once per `replacePropertyName` invocation (many times per overall query!) and since the RegExp itself is constant -- we can lift it out and construct it once. Over-all this saves about 8% on our query build times as measured by typeorm#8955.

AlexMesser · 2022-05-20T11:51:52Z

thank you for contribution!

Context: the query builder is pretty CPU intensive, and can be slow - e.g. #3857 One of the things which makes this slow is `escapeRegExp` in the query builder: we freshly construct the same RegExp once per `replacePropertyName` invocation (many times per overall query!) and since the RegExp itself is constant -- we can lift it out and construct it once. Over-all this saves about 8% on our query build times as measured by #8955.

Digging further into typeorm#3857. See also typeorm#8955, typeorm#8956. As [previously discussed](typeorm#3857 (comment)), the query builder currently suffers from poor performance in two ways: quadratic numbers of operations with respect to total table/column counts, and poor constant factor performance (regexps can be expensive to build/run!) The constant-factor performance is the more tractable problem: no longer quadratically looping would be a chunky rewrite of the query builder, but we can locally refactor to be a bunch cheaper in terms of regexp operations. This change cuts the benchmark time here in ~half (yay!). We achieve this by simplifying the overall replacement regexp (we don't need our column names in there, since we already have a plain object where they're the keys to match against) so compilation of that is much cheaper, plus skipping the need to `escapeRegExp` every column as a result.

Digging further into #3857. See also #8955, #8956. As [previously discussed](#3857 (comment)), the query builder currently suffers from poor performance in two ways: quadratic numbers of operations with respect to total table/column counts, and poor constant factor performance (regexps can be expensive to build/run!) The constant-factor performance is the more tractable problem: no longer quadratically looping would be a chunky rewrite of the query builder, but we can locally refactor to be a bunch cheaper in terms of regexp operations. This change cuts the benchmark time here in ~half (yay!). We achieve this by simplifying the overall replacement regexp (we don't need our column names in there, since we already have a plain object where they're the keys to match against) so compilation of that is much cheaper, plus skipping the need to `escapeRegExp` every column as a result.