Armand du Plessis wrote: > We're running into a scenario where despite doing a manual vacuum as a superuser the relfrozenxid for > one relation now dangerously close to wraparound is not getting reset. > > It's a Postgres 9.2.3 cluster. We shutdown other access to the machine while running the VACUUM to > ensure it could complete quick enough with an aggressive vacuum (vacuum_cost_limit 10000 and no > delay). The previous autovacuum was running for days without completing. > > There's also no old transactions in either pg_prepared_xacts or pg_stat_activity.
> The tail-end of the vacuum log:
[...]
> DETAIL: 701995 dead row versions cannot be removed yet.
That's your problem. Somebody must have been using these rows, unless there is a bug. That's an awfully high number too.
