Re: [BUGS] Breakage with VACUUM ANALYSE + partitions

On Sat, Mar 26, 2016 at 8:39 AM, Andres Freund <andres(at)anarazel(dot)de> wrote:
> On 2016-03-25 12:02:05 -0400, Robert Haas wrote:
>> Gosh, that's surprising. I wonder if that just revealed an underlying
>> issue rather than creating it.
>
> I think that's the case; it's just somewhat unlikely to hit in other
> cases.
>
> If SMgrRelation->md_fd[n] is an empty relation, and mdread() or another
> routine is asking for a block in the second segment - which will error
> out. But even if the first segment is zero bytes, _mdfd_getseg() will
> dutifully try to open the second segment. Which will succeed in the case
> of a truncated relation, because we leave the truncated segment in
> place.
>
> ISTM that _mdfd_getseg better check the length of the last segment
> before opening the next one?

Well, I agree that it's pretty strange that _mdfd_getseg() makes no
such check, but I still don't think I understand what's going on here.
Backends shouldn't be requesting nonexistent blocks from a relation -
higher-level safeguards, like holding AccessExclusiveLock before
trying to complete a DROP or TRUNCATE - are supposed to prevent that.
If this patch is causing us to hold onto smgr references to a relation
on which we no longer hold locks, I think that's irretrievably broken
and should be reverted. I really doubt this will be the only thing
that goes wrong if you do that.

--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company