Re: plan with result cache is very slow when work_mem is not enough

pá 7. 5. 2021 v 21:56 odesílatel David Rowley <dgrowleyml(at)gmail(dot)com> napsal:

> On Sat, 8 May 2021 at 07:18, Pavel Stehule <pavel(dot)stehule(at)gmail(dot)com>
> wrote:
> > yes, the slowdown is related to debug assertions
>
> With USE_ASSERT_CHECKING builds, I did add some code that verifies the
> memory tracking is set correctly when evicting from the cache. This
> code is pretty expensive as it loops over the entire cache to check
> the memory accounting every time we evict something from the cache.
> Originally, I had this code only run when some other constant was
> defined, but I ended up changing it to compile that code in for all
> assert enabled builds.
>
> I considered that it might be too expensive as you can see from the
> comment in [1]. I just wanted to get a few machines other than my own
> to verify that the memory accounting code was working as expected.
> There have been no complaints of any Assert failures yet, so maybe
> it's safe to consider either removing the code entirely or just having
> it run when some other more specific to purpose constant is defined.
> If we did the latter, then I'd have concerns that nothing would ever
> run the code to check the memory accounting, that's why I ended up
> changing it to run with USE_ASSERT_CHECKING builds.
>

I understand. I think this is too slow for generic assertions, because the
overhead is about 50x.

But I understand, so it may be necessary to have this code active some time.

Regards

Pavel

> David
>
> [1]
> https://github.com/postgres/postgres/blob/master/src/backend/executor/nodeResultCache.c#L305
>