On Fri, 2011-04-22 at 17:34 -0400, Bruce Momjian wrote:
> Tom Lane wrote:
> > Bruce Momjian <bruce@momjian.us> writes:
> > > I thought some more about this and I don't want autovacuum to run on the
> > > old server. This is because pg_dumpall --binary-upgrade --schema-only
> > > grabs the datfrozenxid for all the databases at the start, then connects
> > > to each database to gets the relfrozenxids. I don't want to risk any
> > > advancement of either of those during the pg_dumpall run.
> >
> > Why? It doesn't really matter --- if you grab a value that is older
> > than the latest, it's still valid. As Robert said, you're
> > over-engineering this, and thereby introducing potential failure modes,
> > for no gain.
>
> Uh, I am kind of paranoid about pg_upgrade because it is trying to do
> something Postgres was never designed to do. I am a little worried that
> we would be assuming that pg_dumpall always does the datfrozenxid first
> and if we ever did it last we would have relfrozenxids before the
> datfrozenxid. I am worried if we don't prevent autovacuum on the old
> server that pg_upgrade will be more fragile to changes in other parts of
> the system.
If we back-patch the "-b" to 8.3, then we can always use it on both the
old and new systems. Upgrading to the latest patch-level on both old and
new should be a prerequisite for pg_upgrade anyway.
That would turn the catalog check from a special case (use "-b"
sometimes, other times don't; which could cause fragility and bugs),
into just another sanity check with an easy workaround ("your postgres
doesn't support '-b', upgrade to the latest patch-level before
upgrading").
One of the things I like about the design of pg_upgrade is that it
doesn't seem to have a lot of special cases for different version
combinations.
What do you think?
Regards,Jeff Davis