search

neo4j / neo4j-javascript-driver

Neo4j Bolt driver for JavaScript
https://neo4j.com/docs/javascript-manual/current/
Apache License 2.0
839 stars 148 forks source link

Optimize usage of Intl API to speed up response parsing with many datetime objects #1174

Closed vongruenigen closed 5 months ago

vongruenigen commented 5 months ago

Background

Lately, I've been upgrading a big business application that was using the latest 4.x version of the neo4j driver to use the latest 5.x version (we still use neo4j 4.x, but they are compatible according to docs). The upgrade itself was very smooth, but while testing everything afterwards, we noticed that (almost) all of our requests to the backend took considerably longer to finish (~2x).

After doing some investigation (mainly by using clinic flamegraphs) I noticed that there was a considerable increase in the time spent parsing the raw neo4j responses in the driver. Looking at it in more detail revealed that most of the increase stems from one particular codepath, namely from calls to getTimeInZoneId.

Looking at it almost immediately revealed the culprit, which is how the Intl API is used there. It seems that a new Intl.DateTimeFormat object is created for each date time returned in the response. The Intl API is notoriously slow afaik, hence we should reduce the usage of those APIs to an absolute minimum in hot code paths, such as response parsing. Also, since the application I was upgrading is basically doing nothing else than managing timestamps at its core, it made sense that we noticed the performance degradation in such a severe way.

Changes in this MR

I decided to try out to cache the DateTimeFormat to prevent intializing the formatter for a given time zone more than once, and it seems to have helped quite a lot (in our case the "big" requests got a speedup of 60-70%). I also checked for other usages of Intl in the code base, but luckily only found one other place, where it's used to check the validity of a given timezone string. I added caching there as well, though I'm not entirely sure if this is a case of premature optimization, since we personally didn't run into performance issues where this particular method was involved. I'll leave this up to you guys to decide if we should include those changes in this MR as well, or revert them.

Remarks/questions

Maybe there's a better way to achieve the same, and since I'm in no way an expert on the Intl API or this driver, I'm very open to suggestions and other ideas how to solve this (also because I'm a first time contributor, see below).

Looking forward to hearing from you guys, hopefully we can get this resolved as quickly as possible.

Thanks in advance! :)

CLA

I'm a first time contributor and the issue template said that I need to mention it here somewhere. I signed the CLA, let me know in case anything else is needed from my side.

vongruenigen commented 5 months ago

After some more digging I saw that the problem didn't occur the first time in a 5.x release, but rather in the release 4.4.7. We didn't notice until now because we had version 4.4.5 pinned in our npm package.

bigmontz commented 5 months ago

Hey @vongruenigen, good catch. I notice and register your CLA signature. The PR looks good, I will do some investigation here before get the approval and merge it.

Thanks for catching the regression and push the changes.

vongruenigen commented 5 months ago

Thanks a lot for the feedback and suggestions @bigmontz! I'll be on holiday for a week from today on but I'll try to find some spare time to incorporate your suggestions above. Will ping you again once I'm done.

vongruenigen commented 5 months ago

@bigmontz thanks a lot for the swift review and merging. Really looking forward to the new release! 👍