> pg_upgrade *does* overwrite the destination pg_log, and what that
> means is that incoming tuples in user relations should be fine.
> What's at risk is recently-committed tuples in the system relations,
> notably the metadata that pg_upgrade has just inserted for those
> user relations.
>
> The point of the VACUUM is to try to ensure that everything
> in the system relations is marked as certainly committed (or
> certainly dead) before we discard the pg_log information.
> I don't recall ever hearing from Vadim about whether that
> is a trustworthy way of doing it, however.
>
> One thing that occurs to me just now is that we probably need
> to vacuum *each* database in the new installation. The patch
> I added to pg_dump doesn't do the job because it only vacuums
> whichever database was dumped last by pg_dumpall...
I see what you are saying now. pg_upgrade basically replaces the system
tables, but keeps the user data and pg_log. So, if you do initdb, and
create your user table, then recover the user data tables and pg_log,
and if pg_log has a transaction marked as aborted that has the same
number as one of the user create table statements, it would not see the
table. I see why the vacuum is needed.
I wrote pg_upgrade as an attempt to do upgrades without dumping. I
heard so little about it when it was introduced, I thought it was not
really being used. When I disabled it for 6.5, I found out how many
people were using it without incident.
-- Bruce Momjian | http://www.op.net/~candle maillist@candle.pha.pa.us | (610)
853-3000+ If your life is a hard drive, | 830 Blythe Avenue + Christ can be your backup. | Drexel Hill,
Pennsylvania19026