Re: ALTER TABLE ADD COLUMN fast default

Andrew Gierth <andrew@tao11.riddles.org.uk> writes:
> I just got through diagnosing a SEGV crash with someone on IRC, and the
> cause turned out to be exactly this - a table had (for some reason we
> could not determine within the available resources) lost its pg_attrdef
> record for the one column it had with a default (which was a serial
> column, so none of the fast-default code is actually implicated). Any
> attempt to alter the table resulted in a crash in equalTupleDesc on this
> line:
>             if (strcmp(defval1->adbin, defval2->adbin) != 0)
> due to trying to compare adbin values which were NULL pointers.

Ouch.

> Does equalTupleDesc need to be more defensive about this, or does the
> above check need to be reinstated?

Looking around at the other touches of AttrDefault.adbin in the backend
(of which there are not that many), some of them are prepared for it to be
NULL and some are not.  I don't immediately have a strong opinion whether
that should be an allowed state; but if it is not allowed then it's not
okay for AttrDefaultFetch to leave such a state behind.  Alternatively,
if it is allowed then equalTupleDesc is in the wrong.  We should choose
one definition and make all the relevant code match it.

            regards, tom lane