Stephen Frost <sfrost@snowman.net> writes:
> I've not quite tracked it down, but I would caution against blaming this
> commit- when doing some parallel regression test runs, I was seeing
> failures also, but they've not been consistent and I was trying to get
> something else done so didn't focus on them, so they may have been
> failures due to my environment, but might also be some kind of race
> condition in an earlier commit (my guess at the moment is actually the
> INOUT arguments for procedures commit...).
crake has now failed twice at the same spot, so it's looking reproducible
to me. I've enabled TestUpgradeXversion on longfin and will try to
reproduce the problem there, if Andrew doesn't manage to scrape the
relevant log out of crake first. (It'll probably take a few hours
for me to get results.)
The broader picture is that we've been having irreproducible failures on
the buildfarm since around December. Some of them are in postgres_fdw,
and I'd originally suspected a problem there, but we've also seen more
than one select_parallel failure like this one:
https://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=longfin&dt=2018-03-11%2023%3A25%3A46
My gut instinct right now is that those are related, since they both
look like "plan change with no visible triggering cause". I have no
ideas about a plausible explanation.
regards, tom lane
