On 06/05/14 19:05, Robert Haas wrote:
Which brings up another point: the behavior of non-shmem-connected workers is totally bizarre. An exit status other than 0 or 1 is not treated as a crash requiring a restart, but failure to disengage the deadman switch is still treated as a crash requiring a restart. Why? If the workers are not shmem-connected, then no crash requires a system-wide restart. Of course, there's the tiny problem that we aren't actually unmapping shared memory from supposedly non-shmem connected workers, which is a different bug, but ignoring that for the moment there's no reason for this logic to be like this.
Agreed.
What I'm inclined to do is change the logic so that: (1) After a crash-and-restart sequence, zero rw->rw_crashed_at, so that anything which is still registered gets restarted immediately.
Yes, that's quite obvious change which I missed completely :).
(2) If a shmem-connected backend fails to release the deadman switch or exits with an exit code other than 0 or 1, we crash-and-restart. A non-shmem-connected backend never causes a crash-and-restart.
+1
(3) When a background worker exits without triggering a crash-and-restart, an exit code of precisely 0 causes the worker to be unregistered; any other exit code has no special effect, so bgw_restart_time controls.
+1 -- Petr Jelinek http://www.2ndQuadrant.com/ PostgreSQL Development, 24x7 Support, Training & Services -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers