Re: Autovacuum fails to keep visibility map up-to-date in mostly-insert-only-tables

On 2014-10-09 18:16:46 -0300, Alvaro Herrera wrote:
> Andres Freund wrote:
> > On 2014-10-09 18:03:00 -0300, Alvaro Herrera wrote:
> > > Bruce Momjian wrote:
> > >
> > > > I agree this is a serious problem. We have discussed various options,
> > > > but have not decided on anything. The TODO list has:
> > > >
> > > > https://wiki.postgresql.org/wiki/Todo
> > > >
> > > > Improve setting of visibility map bits for read-only and insert-only
> > > > workloads
> > > >
> > > > http://www.postgresql.org/message-id/20130906001437.GA29264@momjian.us
> > >
> > > I hate to repeat myself, but I think autovacuum could be modified to run
> > > actions other than vacuum and analyze. In this specific case we could
> > > be running a table scan that checks only pages that don't have the
> > > all-visible bit set, and see if it can be set.
> >
> > Isn't that *precisely* what a plain vacuum run does?
>
> Well, it also scans for dead tuples, removes them, and needs to go
> through indexes to remove their references.

IIRC it doesn't do most of that if that there's no need. And if it's a
insert only table without rollbacks. I *do* think there's some
optimizations we could make in general.

> I'm thinking in something
> very lightweight. Otherwise, why don't we just reduce the
> vacuum_scale_factor default to something very small, so that vacuum is
> triggered more often?

The problem here is that that doesn't trigger for inserts. Just for
updates/deletes or rollbacks.

Greetings,

Andres Freund

--
Andres Freund http://www.2ndQuadrant.com/
PostgreSQL Development, 24x7 Support, Training & Services