Re: Pg_upgrade and toast tables bug discovered

On Tue, Aug  5, 2014 at 07:31:21PM -0400, Bruce Momjian wrote:
> On Thu, Jul 10, 2014 at 06:38:26PM -0400, Bruce Momjian wrote:
> > On Thu, Jul 10, 2014 at 06:17:14PM -0400, Bruce Momjian wrote:
> > > Well, we are going to need to call internal C functions, often bypassing
> > > their typical call sites and the assumption about locking, etc.  Perhaps
> > > this could be done from a plpgsql function.  We could add and drop a
> > > dummy column to force TOAST table creation --- we would then only need a
> > > way to detect if a function _needs_ a TOAST table, which was skipped in
> > > binary upgrade mode previously.
> > > 
> > > That might be a minimalistic approach.
> > 
> > I have thought some more on this.  I thought I would need to open
> > pg_class in C and do complex backend stuff, but I now realize I can do
> > it from libpq, and just call ALTER TABLE and I think that always
> > auto-checks if a TOAST table is needed.  All I have to do is query
> > pg_class from libpq, then construct ALTER TABLE commands for each item,
> > and it will optionally create the TOAST table if needed.  I just have to
> > use a no-op ALTER TABLE command, like SET STATISTICS.
> 
> Attached is a completed patch which handles oid conflicts in pg_class
> and pg_type for TOAST tables that were not needed in the old cluster but
> are needed in the new one.  I was able to recreate a failure case and
> this fixed it.
> 
> The patch need to be backpatched because I am getting more-frequent bug
> reports from users using pg_upgrade to leave now-end-of-life'ed 8.4. 
> There is not a good work-around for pg_upgrade failures without this
> fix, but at least pg_upgrade throws an error.

Patch applied through 9.3, with an additional Assert check. 9.2 code was
different enough that there was too high a risk for backpatching.

--  Bruce Momjian  <bruce@momjian.us>        http://momjian.us EnterpriseDB
http://enterprisedb.com
 + Everyone has their own god. +