On Thu, Apr 2, 2015 at 2:55 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
> Robert Haas <robertmhaas@gmail.com> writes:
>> On Thu, Apr 2, 2015 at 2:40 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
>>> However, I'm having second thoughts about whether we've fully diagnosed
>>> this. Three out of the four failures we've seen in the buildfarm reported
>>> "cache lookup failed for access method 403", not "could not open relation
>>> with OID 2601" ... and I'm so far only able to replicate the latter
>>> symptom. It's really unclear how the former one could arise, because
>>> nothing that vacuum.sql does would change xmin of the rows in pg_am.
>
>> It probably changes the *relfilenode* of pg_am, because it runs VACUUM
>> FULL on that catalog. Perhaps some backend sees the old relfilenode
>> value and tries to a heap scan, interpreting the now-truncated file as
>> empty?
>
> Yeah, I came up with the same theory a few minutes later. Trying to
> reproduce on that basis.
>
> Actually, now that I think it through, the "could not open relation"
> error is pretty odd in itself. If we are trying to open pg_am using
> a stale catalog snapshot, it seems like we ought to reliably find its
> old pg_class tuple (the one with the obsolete relfilenode), rather than
> finding nothing. But the latter is the behavior I'm seeing.
What's to stop the old tuple from being HOT-pruned?
--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company