David G Johnston <david.g.johnston@gmail.com> writes:
> Scott Marlowe-2 wrote
>> Are you sure these aren't to prevent wrap-around?
> This is my other theory, and probably the correct one. Is this something
> that can be discovered from the logs (since nothing in the query provides
> this detail)?
In all recent versions of Postgres, the pg_stat_activity line should
include "(to prevent wraparound)" if that's what autovac thinks it's
doing.
However ... it looks to me like if the autovac launcher gets launched
forcibly because of a possible wraparound issue, it will keep running
normally after that even if autovacuum is off :-(. You need to SIGHUP
it to get it to quit again. (A second bug is that it will then honor the
SIGHUP and quit, even if the wraparound emergency is still present.)
So I'm betting that what happened, particularly given the OP's apparent
insistence on managing vacuums manually, is that a wraparound vacuum
was forced on some relatively small table and then the launcher kept
doing its thing after that. Since log_autovacuum_min_duration was
set to 0, there should be some evidence for this theory in the logs
(although it looks like the log entries *don't* say whether the
vacuum was for wraparound :-().
regards, tom lane