Re: intermittent failures in Cygwin from select_parallel tests

Robert Haas <robertmhaas(at)gmail(dot)com> writes:
> On Wed, Jun 14, 2017 at 3:33 PM, Tom Lane <tgl(at)sss(dot)pgh(dot)pa(dot)us> wrote:
>> So the first problem here is the lack of supporting information for the
>> 'could not map' failure.

> Hmm. I think I believed at the time I wrote dsm_attach() that
> somebody might want to try to soldier on after failing to map a DSM,
> but that doesn't seem very likely any more.

Well, if they do, they shouldn't be passing elevel == ERROR.

>> AFAICS, this must mean either that dsm_attach()
>> returned without ever calling dsm_impl_op() at all, or that
>> dsm_impl_op()'s Windows codepath encountered ERROR_ALREADY_EXISTS or
>> ERROR_ACCESS_DENIED. It's far from clear why those cases should be
>> treated as a silent fail.

> There's a good reason for that, though. See
> 419113dfdc4c729f6c763cc30a9b02ee68a7da94.

But surely the silent treatment should only apply to DSM_OP_CREATE?
We're not going to retry anything else.

>> It's even less clear why dsm_attach's early
>> exit cases don't produce any messages. But since we're left not knowing
>> what happened, the messaging design here is clearly inadequate.

> I don't know what you mean by this. The function only has one
> early-exit case, the comment for which I quoted above.

OK, s/cases/case/, but the problem remains. We don't know what happened.
We cannot have more than one case in this code where nothing gets logged.

>> It's not very clear how that happened, but my
>> bet is that the postmaster incremented parallel_terminate_count more than
>> once while cleaning up after the crashed worker. It looks to me like
>> there's nothing stopping BackgroundWorkerStateChange from incrementing it
>> and then the eventual ForgetBackgroundWorker call from incrementing it
>> again. I haven't traced through things to identify why this might only
>> occur in a worker-failure scenario, but surely we want to make sure that
>> the counter increment happens once and only once per worker.

> Yeah -- if that can happen, it's definitely a bug.

My first thought about fixing it is that we should remove that code from
BackgroundWorkerStateChange altogether. The parallel_terminate_count
increment should happen in, and only in, ForgetBackgroundWorker. There
seems little reason to risk bugs by trying to do it a bit earlier.

regards, tom lane