Re: Startup process on a hot standby crashes with an error "invalid memory alloc request size 1073741824" while replaying "Standby/LOCK" records

Kyotaro Horiguchi <horikyota(dot)ntt(at)gmail(dot)com> writes:
> At Tue, 4 Oct 2022 17:15:31 -0700, Nathan Bossart <nathandbossart(at)gmail(dot)com> wrote in
>> On Tue, Oct 04, 2022 at 07:53:11PM -0400, Tom Lane wrote:
>>> After further thought, maybe it'd be better to do it as attached,
>>> with one long-lived hash table for all the locks.

> First one is straight forward outcome from the current implement but I
> like the new one. I agree that it is natural and that the expected
> overhead per (typical) transaction is lower than both the first one
> and doing the same operation on a list. I don't think that space
> inefficiency in that extent doesn't matter since it is the startup
> process.

To get some hard numbers about this, I made a quick hack to collect
getrusage() numbers for the startup process (patch attached for
documentation purposes). I then ran the recovery/t/027_stream_regress.pl
test a few times and collected the stats (also attached). This seems
like a reasonably decent baseline test, since the core regression tests
certainly take lots of AccessExclusiveLocks what with all the DDL
involved, though they shouldn't ever take large numbers at once. Also
they don't run long enough for any lock list bloat to occur, so these
numbers don't reflect a case where the patches would provide benefit.

If you look hard, there's maybe about a 1% user-CPU penalty for patch 2,
although that's well below the run-to-run variation so it's hard to be
sure that it's real. The same comments apply to the max resident size
stats. So I'm comforted that there's not a significant penalty here.

I'll go ahead with patch 2 if there's not objection.

One other point to discuss: should we consider back-patching? I've
got mixed feelings about that myself. I don't think that cases where
this helps significantly are at all mainstream, so I'm kind of leaning
to "patch HEAD only".

regards, tom lane