On Mon, Mar 7, 2011 at 11:53 PM, daveg <daveg@sonic.net> wrote:
>> Looking at the code, I don't see how that situation could arise, though.
>> The value calculated by GetOldestXmin() should never move backwards. And
>> GetOldestXmin() is called in lazy_vacuum_rel(), after it has acquired a
>> lock on the table, which should protect from a race condition where two
>> vacuums could run on the table one after another, in a way where the
>> later vacuum runs with an OldestXmin calculated before the first vacuum.
>>
>> Hmm, fiddling with vacuum_defer_cleanup_age on the fly could cause that,
>> though. You don't do that, do you?
>
> No.
>
> I've updated the patch to collect db and schema and added Merlins patch as
> well and run it for a while. The attached log is all the debug messages
> for pg_statistic page 333 from one database. I've also attached the two
> most recent page images for that particular page, the last digits in the
> filename are the hour and minute of when the page was saved.
Well from that log you definitely have OldestXmin going backwards. And
not by a little bit either. at 6:33 it set the all_visible flag and
then at 7:01 it was almost 1.3 million transactions earlier. In fact
to precisely the same value that was in use for a transaction at 1:38.
That seems like a bit of a coincidence though it's not repeated
earlier.
It also seems odd that it happens only with this one block of this one table.
What does SHOW ALL show for the current settings in effect? And what
was process 23896, are there any other log messages from it? When did
it start?
--
greg