Re: Berserk Autovacuum (let's save next Mandrill)

On Tue, 2020-01-07 at 19:05 +0100, Tomas Vondra wrote:
> This patch is currently in "needs review" state, but that seems quite
> wrong - there's been a lot of discussions about how we might improve
> behavior for append-only-tables, but IMO there's no clear consensus nor
> a patch that we might review.
>
> So I think this should be either "waiting on author" or maybe "rejected
> with feedback". Is there any chance of getting a reviewable patch in the
> current commitfest? If not, I propose to mark it as RWF.
>
> I still hope we can improve this somehow in time for PG13. The policy is
> not to allow new non-trivial patches in the last CF, but hopefully this
> might be considered an old patch.

I think that no conclusion was reached because there are *many* things
that could be improved, and *many* interesting and ambitious ideas were
vented.

But I think it would be good to have *something* that addresses the immediate
problem (INSERT-only tables are autovacuumed too late), as long as
that does not have negative side-effects or blocks further improvements.

I don't feel totally well with the very simplistic approach of this
patch (use the same metric to trigger autoanalyze and autovacuum),
but what about this:

- a new table storage option autovacuum_vacuum_insert_threshold,
perhaps a GUC of the same name, by default deactivated.

- if tabentry->tuples_inserted exceeds this threshold, but not one
of the others, lauch autovacuum with index_cleanup=off.

How would you feel about that?

Yours,
Laurenz Albe