Re: BUG #4958: Stats collector hung on WaitForMultipleObjectsEx while attempting to recv a datagram

Robert Haas <robertmhaas(at)gmail(dot)com> writes:
> Well, bad news. Dave built an installer off the 9.1 tip, we provided
> it to the customer, and they can still reproduce the problem. It's
> not clear whether we've narrowed the scope of the problem without
> eliminating it, or whether that fix just didn't help at all. But they
> still can't reproduce it on 9.2. This suggests that there's some
> other difference between 9.1 and 9.2 that is relevant here, but I'm
> not sure what. Any ideas?

I guess that's reassuring, in a way: it means that all the thrashing
about we did to get the latch-ified stats collector logic to work on
Windows actually brought us to a better place.

However, for the purposes of fixing the older branches, it's not so
nice. I think it means that we'd have to back-port some or all of the
latch-related changes in pgstat.c:

commit 966970ed636586b80739c0d21aec7561f0fafedd
commit f1ca51549e9e7045a2db95a61744334f1dbb4d64
commit d461d0502b6e20c6d4ec326ba2faec8212fd5dee
commit 9b63e9869ffaa4d6d3e8bf45086a765d8f310f1c

and maybe even some of these fixes in win32_latch.c:

commit e42a21b9e6c9b9e6346a34b62628d48ff2fc6ddf
commit f40022f1adaeff85b01d98fea38cf5aa16814aa7
commit 8ebc908c575b3378c9d4096412d16d91d107311c

At least, that's where I'd go looking if I were responsible for fixing
this, which is a responsibility I disclaim. Looking through this
history reminds me how much time I've already wasted on that miserable
excuse for an operating system.

regards, tom lane