Alvaro Herrera <alvherre@commandprompt.com> writes:
> Tom Lane wrote:
>> Hmm, there is not any filter in ReindexDatabase() to exclude temp tables
>> of other backends, but it sure seems like there needs to be. CLUSTER
>> might have the same issue. I think we fixed this in VACUUM long ago,
>> but we need to check the other commands that grovel over all of a database.
> Was this ever fixed? I think it wasn't, because I don't see any check
> in ReindexDatabase. Here is a patch to add one.
No, that's still on the todo list. Your patch looks reasonable.
> I examined cluster.c and it does seem to be missing a check too. I'm
> not sure where to add one though; the best choice would be the place
> where the list of rels is built, but that scans only pg_index, so it
> doesn't have access to the namespace of each rel. So one idea would be
> to get the pg_class row for each candidate, but that seems slow.
> Another idea would be to just add all the candidates and silently skip
> the temp indexes in cluster_rel.
Yeah, an extra fetch of the pg_class row doesn't seem all that nice.
I think you'd want to check it in approximately the same two places
where pg_class_ownercheck() is applied (one for the 1-xact and one for
the multi-xact path).
Are there any other commands to be worried about? I can't see any
besides VACUUM/ANALYZE, and those seem covered.
regards, tom lane
