Re: BUG #11638: Transaction safety fails when constraints are dropped and analyze is done

On 2014-10-15 15:02:43 +0900, Michael Paquier wrote:
> On Wed, Oct 15, 2014 at 2:12 PM, Andres Freund <andres@2ndquadrant.com>
> wrote:
>
> > I'm not sure which danger you're seeing here. Imo we need to choose
> > between heap_inplace/heap_update for VACUUM/ANALYZE because one is
> > allowed to run in a transaction, and the other is not. It simply *can't*
> > be safe for ANALYZE to set things like relhastriggers = false using
> > heap_inplace().
> > There's problems with both it rolling back and thus undoing the action
> > that allowed relhastriggers = false to be set and scenarios where it's
> > not ok that other backends can see that value before the transaction
> > committed.
>
> Hm, I was wondering about the potential effects of VACUUM FULL or VACUUM
> ANALYZE, but as they cannot run in a tx block...

Also they all take ShareUpdateExclusive locks preventing them from
running concurrently.

> Btw, I have just put my hands on this code and made the attached to
> make vac_update_relstats able to do a transactional update. It looks
> to work fine with only a check on the flags of vacuum statement.

Have you tested that the problem's now fixed?

Imo this is complex enough to deserve a regression test. Can you add
one?

Greetings,

Andres Freund

--
 Andres Freund                       http://www.2ndQuadrant.com/
 PostgreSQL Development, 24x7 Support, Training & Services