Re: kevent latch paths don't handle postmaster death well

On 14/10/2020 21:58, Tom Lane wrote:
> I noticed that chipmunk failed [1] with a rather interesting log:
>
> 2020-10-14 08:57:01.661 EEST [27048:6] pg_regress/prepared_xacts LOG: statement: UPDATE pxtest1 SET foobar = 'bbb' WHERE foobar = 'aaa';
> 2020-10-14 08:57:01.721 EEST [27048:7] pg_regress/prepared_xacts LOG: statement: SELECT * FROM pxtest1;
> 2020-10-14 08:57:01.823 EEST [27048:8] pg_regress/prepared_xacts FATAL: postmaster exited during a parallel transaction
> TRAP: FailedAssertion("entry->trans == NULL", File: "pgstat.c", Line: 909, PID: 27048)
> 2020-10-14 08:57:01.861 EEST [27051:1] ERROR: could not attach to dynamic shared area
> 2020-10-14 08:57:01.861 EEST [27051:2] STATEMENT: SELECT * FROM pxtest1;
>
> I do not know what happened to the postmaster, but seeing that chipmunk
> is a very small machine running a pretty old Linux kernel, it's plausible
> to guess that the OOM killer decided to pick on the postmaster. (I wonder
> whether Heikki has taken any steps to prevent that on that machine.)

For the record, it was not the OOM killer. It was the buildfarm cron job
that did it:

Oct 14 08:57:01 raspberrypi /USR/SBIN/CRON[27050]: (pgbfarm) CMD
(killall -q -9 postgres; cd /home/pgbfarm/build-farm-client/ &&
./run_branches.pl --run-all)

Apparently building and testing all the branches is now taking slightly
more than 24 h on that system, so the next day's cron job kills the
previous tests. I'm going to change the cron schedule so that it runs
only every other day.

- Heikki