On Thu, Sep 24, 2015 at 11:47 AM, Teodor Sigaev <teodor@sigaev.ru> wrote:
> Hi!
>
> After reading thread
>
http://www.postgresql.org/message-id/flat/CAMkU=1xALfLhUUohFP5v33RzedLVb5aknNUjcEuM9KNBKrB6-Q@mail.gmail.com#CAMkU=1xALfLhUUohFP5v33RzedLVb5aknNUjcEuM9KNBKrB6-Q@mail.gmail.com
>
> I agree, that it's a bug, although, seems, with low probability. I'd like to
> suggest patch with introduces:
> 1 cleanup process could be only one at the same time
> 2 if cleanup called from regular insert sees waiting concurrent cleanup then
> it will stop process in hope that another cleanup is vacuum called.
>
> Insert-called cleanup process locks metapage with
> ConditionalLockPage(ExclusiveLock) and if it fails then just goes away,
> because other cleanup is already in progress. Also, insert-called cleanup
> process will sometimes do lock/conditional lock metapage. Vacuum-called
> cleanup process locks with a help of unconditional LockPage() and will not
> relock metapage during process, that gives a warranty to be a single cleanup
> process. Relock of insert-called cleanup process allows to leave a rest of
> work to vacuum if it runs right now. This reduces latency for insert and
> allows to vacuum to be sure that pending list is clear.
>
> Also, patch differentiates which limit of memory to use:
> autovacuum_work_mem, maintenance_work_mem or work_mem depending on call
> path.
Using a heavyweight lock on the metapage seems like a better idea to
me than the previous proposal of using the relation lock. But I think
it would be best to discuss this on the original thread instead of
starting a new one.
--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company
