Re: pg_upgrade if 'postgres' database is dropped

Tom Lane wrote:
> Bruce Momjian <bruce@momjian.us> writes:
> > Stephen Frost wrote:
> >> Yes, they would have removed it because they didn't want it.  As I
> >> recall, part of the agreement to create an extra database by default was
> >> that it could be removed if users didn't want it.  Turning around and
> >> then saying "but things won't work if it's not there" isn't exactly
> >> supporting users who decide to remove it.
> 
> > Well, you would have to remove it _after_ you did the pg_upgrade.
> 
> As far as the *target* cluster is concerned, I have no sympathy for
> someone who messes with its contents before running pg_upgrade.
> That's an RTFM matter: you're supposed to upgrade into a virgin
> just-initdb'd cluster.
> 
> However, it would be nice if pg_upgrade supported transferring from a
> *source* cluster that didn't have the postgres DB.

I have this C comment in pg_upgrade about this:
*  If someone removes the 'postgres' database from the old cluster and*  the new cluster has a 'postgres' database, the
numberof databases*  will not match.  We actually could upgrade such a setup, but it would*  violate the 1-to-1 mapping
ofdatabase counts, so we throw an error*  instead.  We would detect this as a database count mismatch during*  upgrade,
butwe want to detect it during the check phase and report*  the database name.
 

Is this worth fixing?  Another problem is that pg_dumpall doesn't remove
the postgres database in the new cluster if it was not in the old one,
so they are going to end up with a postgres database in the new cluster
anyway.  I could argue that pg_upgrade is better because reports it
cannot recreate the new cluster exactly, while pg_dumpall just keeps a
postgres database that was not in the old cluster.

> What about creating a new, single-purpose database in the source
> cluster and then removing it again after we're done?

That is not a problem --- I can easily use template1.

--  Bruce Momjian  <bruce@momjian.us>        http://momjian.us EnterpriseDB
http://enterprisedb.com
 + It's impossible for everything to be true. +