On 2013-07-22 19:09:13 -0400, Tom Lane wrote:
> Andres Freund <andres@2ndquadrant.com> writes:
> > On 2013-07-17 10:11:28 -0400, Tom Lane wrote:
> >> The buildfarm members that use -DCLOBBER_CACHE_ALWAYS say this patch
> >> is broken.
>
> > Jagarundi still tells that story.
>
> Uh, no. Jagarundi was perfectly happy for several build cycles after
> I committed 405a468. The buildfarm history shows that it started
> complaining again after this change:
>
> f01d1ae Add infrastructure for mapping relfilenodes to relation OIDs
Huh? That's rather strange. No code from that patch is even exececuted
until the alter_table regression tests way later. So I can't see how
that patch could cause the matview error. Which is pretty clearly using
an freed relcache entry.
I think this may be a timing issue.
> > At least something like the following patch seems to be required.
>
> This does look like a necessary change, but I suspect there is also
> something rotten in f01d1ae.
The alter table failure lateron seems to be "ERROR: relation "tmp"
already exists", which just seems to be a consequence of incomplete
cleanup due to the earlier crashes.
Greetings,
Andres Freund
-- Andres Freund http://www.2ndQuadrant.com/PostgreSQL Development, 24x7 Support, Training &
Services