Hi,
On 2019-03-29 20:51:38 +0900, Michael Paquier wrote:
> So, coming back to this thread, and studying the problem again, it
> looks that the diagnostic that a non-aggressive, anti-wraparound
> vacuum could be triggered because the worker sees trouble in the
> force because of some activity happening in parallel. Hence, if we
> face this case, it looks right to skip the vacuum for this relation.
>
> Attached is an updated patch with a better error message, more
> comments, and the removal of the anti-wraparound non-aggressive log
> which was added in 28a8fa9. The error message could be better I
> guess. Suggestions are welcome.
> diff --git a/src/backend/access/heap/vacuumlazy.c
> b/src/backend/access/heap/vacuumlazy.c
> index 5c554f9465..82be8c81f3 100644
> --- a/src/backend/access/heap/vacuumlazy.c
> +++ b/src/backend/access/heap/vacuumlazy.c
> @@ -248,6 +248,25 @@ heap_vacuum_rel(Relation onerel, VacuumParams *params,
> if (params->options & VACOPT_DISABLE_PAGE_SKIPPING)
> aggressive = true;
>
> + /*
> + * When running an anti-wraparound vacuum, we expect relfrozenxid to be
> + * old enough so as aggressive is always set. If this is not the case,
> + * it could be possible that another concurrent vacuum process has done
> + * the work for this relation so that relfrozenxid in relcache has
> + * already been moved forward enough, causing this vacuum run to be
> + * non-aggressive. If that happens, note that this relation no longer
> + * needs to be vacuumed, so just skip it.
> + */
> + if (params->is_wraparound && !aggressive)
> + {
> + ereport(LOG,
> + (errmsg_internal("found vacuum to prevent
> wraparound of table \"%s.%s.%s\" to be not aggressive, so skipping",
> +
> get_database_name(MyDatabaseId),
> +
> get_namespace_name(RelationGetNamespace(onerel)),
> +
> RelationGetRelationName(onerel))));
> + return;
> + }
> +
Which valid scenario can lead to this? Neither the comment, nor commit
message explain it. Unless you're thinking of scenarios where autovacuum
and manual vacuum are mixed, I don't really see valid reasons? Normally
autovacuum's locking + the table_recheck_autovac() check should prevent
problematic scenarios.
I do see a few scenarios that can trigger this - but they all more or
less are bugs.
It doesn't strike me as a good idea to work around such bugs by silently
neutering heap_vacuum_rel(). The likelihood of that temporarily covering
up more severe problems seems significant - they're likely to then later
bite you with a cluster shutdown.
Greetings,
Andres Freund
