Re: pg_tablespace_location() error message

On Tue, Apr 10, 2012 at 07:57:30PM -0400, Tom Lane wrote:
> Bruce Momjian <bruce@momjian.us> writes:
> > On Tue, Apr 10, 2012 at 07:09:33PM -0400, Tom Lane wrote:
> >> Hm.  I have no objection to special-casing zero here, but what behavior
> >> do you want?  Should it return an empty string as we do for
> >> DEFAULTTABLESPACE_OID, or throw a different error?
> 
> > I have no idea.  The big problem is that we currently use '' for the
> > cluster default, while 0 means the database default.  I can't think of a
> > good return result --- I think it has to be an error of some kind.
> 
> If we expect this function to mainly be applied to pg_class.reltablespace,
> then it seems like it ought to understand that zero means "the database
> default" and substitute the database's default tablespace.  That might
> or might not be the same as the cluster default.

Well, do we really want to be reporting the _current_ data directory
location?  We do track tablespace symlink moves because we read the
symlinks now, so that isn't out of the question.  A bigger question is
whether returning '' for a database-default location is valid --- I am
thinking no.

> Alternatively, we could expect pg_upgrade to understand that and make
> the substitution itself, but if the same would be needed by most uses of
> the function, maybe we should just do it here.

I just applied a patch to pg_upgrade to do exactly that, and it is
needed in pre-9.2 as well because pg_upgrade was testing for an empty
spclocation, which could be '' or it could be NULL (from an outer join),
but pg_upgrade wasn't distinguising the two.  :-(

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