Re: Legacy GiST invalid tuples

Hi,

On 2018-07-04 16:43:19 -0400, Alvaro Herrera wrote:
> On 2018-Jul-04, Tom Lane wrote:
>
> > Alvaro Herrera <alvherre(at)2ndquadrant(dot)com> writes:
> > > Requiring a scan of all indexes during pg_upgrade might increase the
> > > upgrade time prohibitively for some sites, so I don't think that's a
> > > good solution.
> >
> > Perhaps VACUUM could be taught to clean up invalid entries? That
> > wouldn't help Andrey's unstated goal of being able to re-use the bits
> > for some other purpose in v12, but it might be practical to re-use
> > them sometime sooner than v17.
>
> We tried to clean up HEAP_MOVED_IN / HEAP_MOVED_OFF a long time ago, but
> gave up :-) I recall Andres posted a write-up about adding columns to
> pg_class to indicate "what version did last vacuum this whole table", as
> a marker for features that are no longer needed.

Right. I plan to pick that up as soon as I've some free cycles. I'd not
object to somebody else working on it first though.

> Maybe that idea can be reused for this purpose.

Yes, I think that's explicitly in scope.

> I'm thinking pg_upgrade can use its --check phase to look for indexes
> marked as older than X (possibly containing invalid tuples), so it
> instructs the user to run vacuum on it prior to the upgrade.

I kinda wondered about making the feature back-patchable (by storing the
version in reloptions or such), but I think that'd just make it less
likely to get in.

> I think the soonest this can work is to add the column in pg12 so that
> it can be used to upgrade to pg13.

I don't think we can easily require that everyone pg_upgrading to v13+
upgrades to v12 first?

Greetings,

Andres Freund