Re: GIN fast-insert vs autovacuum scheduling

Alvaro Herrera <alvherre@commandprompt.com> writes:
> Tom Lane wrote:
>> On top of those issues, there are implementation problems in the
>> proposed relation_has_pending_indexes() check:

> I wonder if it's workable to have GIN send pgstats a message with number
> of fast-inserted tuples, and have autovacuum check that number as well
> as dead/live tuples.

> ISTM this shouldn't be considered part of either vacuum or analyze at
> all, and have autovacuum invoke it separately from both, with its own
> decision equations and such.  We could even have a scan over pg_class
> just for GIN indexes to implement this.

That's going in the wrong direction IMHO, because it's building
GIN-specific infrastructure into the core system.  There is no need for
any such infrastructure if we just drive it off a post-ANALYZE callback.
        regards, tom lane