On Thu, May 9, 2013 at 06:19:31PM -0400, Bruce Momjian wrote:
> > pg_upgrade already deals with the new code deciding not to create a
> > toast table (by forcing it to do so anyway in binary upgrade mode).
>
> Yes, a good point I had forgotten. postgres --binary-upgrade mode can
> force the toast table to be created to match the old cluster; see
> toasting.c::create_toast_table():
>
> /*
> * Check to see whether the table actually needs a TOAST table.
> *
> * If an update-in-place toast relfilenode is specified, force toast file
> * creation even if it seems not to need one.
> */
> if (!needs_toast_table(rel) &&
> (!IsBinaryUpgrade ||
> !OidIsValid(binary_upgrade_next_toast_pg_class_oid)))
> return false;
>
> > It's only the other case that's problematic -- but then AFAICS fixing
> > that is just a SMOP.
>
> Yes, it is this opposite case where the _new_ cluster wants a TOAST
> table that the old cluster doesn't have, which is what Evan is
> reporting.
So, if we eventually agree we need to be able to _suppress_ creation of
the TOAST table on the new cluster, I propose we do it in a similar way
to how we force TOAST creation, by having pg_dump set a backend variable
that is then tested in the backend to suppress TOAST table creation.
I don't think we know enough about the cause of this pg_upgrade failure
to know if this is necessary.
-- Bruce Momjian <bruce@momjian.us> http://momjian.us EnterpriseDB
http://enterprisedb.com
+ It's impossible for everything to be true. +
