On Wed, Mar 25, 2020 at 07:59:56PM -0700, Andres Freund wrote:
> FWIW, this kind of thing is why I think the added skipping logic is a
> bad idea. Silently skipping things like this (same with the "bogus"
> logic in datfrozenxid computation) is dangerous. I think we should
> seriously consider backing this change out.
That's actually what I would like to do at this stage as a first
step. It looks pretty clear that it does not help.
> And if not, then we should at least include enough detail in the message
> to be able to debug this.
Sure. In any attempt I have done until now I was easily able to skip
some jobs, but it should get easier with a higher number of concurrent
workers and a higher number of relations heavily updated. Thinking
about it, only catalog jobs were getting skipped in my own runs...
>> postgres=# SELECT datname, age(datfrozenxid), datfrozenxid FROM
>> pg_database ORDER BY age(datfrozenxid) DESC LIMIT 1;
>> datname | age | datfrozenxid
>> ----------+-----------+--------------
>> postgres | 202773709 | 4284570172
>
> And why should this lead to anti-wraparound vacuums not happening? This
> is older than the the cutoff age?
>
> xid 4284570172 having the age of 202 million xids suggests that
> ReadNewTransactionId() is approx 192376585. Which comports with the log
> saying: oldest xmin: 189591147.
Oops, sorry. My previous email was incorrect. It looked strange to
not see datfrozenxid being refreshed.
> Or are you saying that you conclude that the relcache entry is somehow
> out of date? It sure is interesting that all of the tables that hit the
> "skipping redundant vacuum" condition are shared tables.
Yeah, that's actually what I was thinking yesterday. In
heap_vacuum_rel(), xidFullScanLimit may be calculated right, but an
incorrect value of rd_rel->relminmxid or rd_rel->relfrozenxid could
lead to a job to become not aggressive. It should be actually easy
enough to check that.
--
Michael