perf: Backport of #3889 to v12

[mkleczek]

See #3889

@steve-chavez @taimoorzaeem
I've prepared a backport of asynchronous JWT cache purge to 12.
Looks like there is consistent (small but noticable) performance gain.

Done this as it is not clear when next major is going to be released (and if any JWT caching improvements are going to be included).

[wolfgangwalther]

Looks like there is consistent (small but noticable) performance gain.

How did you test that?

[mkleczek]

Looks like there is consistent (small but noticable) performance gain.

How did you test that?

https://github.com/PostgREST/postgrest/actions/runs/13158254898

[wolfgangwalther]

I see. The loadtest results have no meaning for this kind of PR:

• First of all the main column obviously must be ignored.
• The difference between v12.2.7 and HEAD is not different from the noise we usually have when running this on the same commit. Example: https://github.com/PostgREST/postgrest/actions/runs/13140690883 - the latest commit on main, which is a simple refactor. 403 -> 404 there, 465 -> 467 here. Not significant.
• The loadtest runs all requests with the exact same JWT.
• Since those JWT are hardcoded - they are surely set up so that they never expire.

Thus, the loadtest is testing none of what would be relevant for this PR.

[mkleczek]

It is difficult for me to judge the effectiveness of existing load tests.
The results are consistent with tests performed earlier on main:
https://github.com/PostgREST/postgrest/actions/runs/13151068303

So it seems that even in case of small JWT cache with fresh tokens taking purgeExpired from the hot path provides response time gains.

[steve-chavez]

I don't think we should backport this. The MVar/locking logic looks risky, it could cause breakage.

Additionally a perf commit is not critical to backpatch. The gains don't look clear as well.

[mkleczek]

@steve-chavez
not questioning your decision but could you elaborate a little on the risks of MVar locking logic? I find it pretty straightforward.

[steve-chavez]

So it seems that even in case of small JWT cache with fresh tokens taking purgeExpired from the hot path provides response time gains.

Yes, here #3889 (comment) it's revealed there's indeed perf gains.

but could you elaborate a little on the risks of MVar locking logic? I find it pretty straightforward.

Mostly that it doesn't seem good for maintenance to have it like that, I was thinking it should be in a function. It's looking like this logic could be inside the debounce added in #3802 or perhaps another debounce.

---

Let's continue on #3889. This should go into v13.