Re: Big performance slowdown from 11.2 to 13.3

On Thu, Jul 22, 2021 at 10:35 AM Tom Lane <tgl(at)sss(dot)pgh(dot)pa(dot)us> wrote:
> Well, what we really ought to be using is size_t (a/k/a Size), at least
> for memory-space-related calculations. I don't have an opinion right
> now about what logtape.c ought to use. I do agree that avoiding "long"
> altogether would be a good ultimate goal.

I assume that we often use "long" in contexts where a signed integer
type is required. Maybe this is not true in the case of the work_mem
style calculations. But I know that it works that way in logtape.c,
where -1 is a sentinel value.

We already use int64 (not size_t) in tuplesort.c for roughly the same
reason: LACKMEM() needs to work with negative values, to handle
certain edge cases.

> In the short term though, the question is whether we want to regard this
> hashagg issue as something we need a fix for in v13/v14. The fact that
> it's Windows-only makes it slightly less pressing in my mind, but it's
> still a regression that some people are going to hit.

True. I worry about the potential for introducing new bugs on Windows
by backpatching a fix for this. Technically this restriction existed
in every other work_mem consumer on Windows. Of course this won't
matter much to users like Laurent.

--
Peter Geoghegan