Re: Simplify backend terminate and wait logic in postgres_fdw test

Bharath Rupireddy <bharath(dot)rupireddyforpostgres(at)gmail(dot)com> writes:
> On Tue, May 4, 2021 at 4:12 AM Tom Lane <tgl(at)sss(dot)pgh(dot)pa(dot)us> wrote:
>> The buildfarm is showing that one of these test queries is not stable
>> under CLOBBER_CACHE_ALWAYS:

> I can reproduce the issue with the failing case. Issue is that the
> backend pid will be null in the pg_stat_activity because of the cache
> invalidation that happens at the beginning of the query and hence
> pg_terminate_backend returns null on null input.

No, that's nonsense: if it were happening that way, the query would
return one row with a NULL result, but actually it's returning no
rows. What's actually happening, it seems, is that because
pgfdw_inval_callback is constantly getting called due to cache
flushes, we invariably drop remote connections immediately during
transaction termination (cf pgfdw_xact_callback). Thus, by the time
we inspect pg_stat_activity, there is no remote session to terminate.

I don't like your patch because what it effectively does is mask
whether termination happened or not; if there were a bug there
causing that not to happen, the test would still appear to pass.

I think the most expedient fix, if we want to keep this test, is
just to transiently disable debug_invalidate_system_caches_always.
(That option wasn't available before v14, but fortunately we
don't need a fix for the back branches.)

I believe the attached will do the trick, but I'm running the test
with debug_invalidate_system_caches_always turned on to verify
that. Should be done in an hour or so...

regards, tom lane