On 2015-06-02 11:49:56 -0400, Robert Haas wrote:
> On Tue, Jun 2, 2015 at 11:44 AM, Andres Freund <andres@anarazel.de> wrote:
> > On 2015-06-02 11:37:02 -0400, Robert Haas wrote:
> >> The exact circumstances under which we're willing to replace a
> >> relminmxid with a newly-computed one that differs are not altogether
> >> clear to me, but there's an "if" statement protecting that logic, so
> >> there are some circumstances in which we'll leave the existing value
> >> intact.
> >
> > I guess we'd have to change that then.
>
> Yeah, but first we'd need to assess why it's like that. Tom was the
> one who installed the current logic, but I haven't been able to fully
> understand it.
We're talking about:
/* Similarly for relminmxid */
if (MultiXactIdIsValid(minmulti) &&
pgcform->relminmxid != minmulti &&
(MultiXactIdPrecedes(pgcform->relminmxid, minmulti) ||
MultiXactIdPrecedes(ReadNextMultiXactId(), pgcform->relminmxid)))
{
pgcform->relminmxid = minmulti;
dirty = true;
}
right? Before that change (78db307bb/87f830e0ce) we only updated
relminmxid if the new value was newer than the old one. That's to avoid
values from going backwards, e.g. when a relation is first VACUUM
FREEZEd and then a normal VACUUM is performed (these values are
unfortunately based on the cutoff values instead of the observed
minima). The new thing is the || MultiXactIdPrecedes(ReadNextMultiXactId(), pgcform->relminmxid)
which is there to recognize values from the future. E.g. the 1
errorneously left in place by pg_upgrade.
Makes sense?
> >> It would similarly do so when the oldest MXID reference in the
> >> relation is in fact 1, but that value can't be vacuumed away yet.
> >
> > I'd thought of just forcing consumption of one multixact in that
> > case. Not pretty, but imo acceptable.
>
> What if multixact 1 still has living members?
I think we should just trigger the logic if 1 is below the multi
freezing horizon - in which case a) the 1 will automatically be
replaced, because the horizon is newer b) it can't have a living member
anyway.
- Andres
