On Wed, Apr 27, 2016 at 11:51 AM, Andres Freund <andres@anarazel.de> wrote:
> On 2016-04-27 14:27:33 +0900, Michael Paquier wrote:
>> On Wed, Apr 27, 2016 at 12:25 AM, Robert Haas <robertmhaas@gmail.com> wrote:
>> > In other words, I think Masahiko Sawada's patch in the original post
>> > is pretty much right on target, except that we don't need to do that
>> > always, but rather only in the FPI case when the call to
>> > visibilitymap_pin() is being optimized away. If we solve the problem
>> > that way, I don't think we even need a new WAL record for this case,
>> > which is a non-trivial fringe benefit.
>>
>> The visibility map is not the only thing that need to be addressed,
>> no?
>
> If I understand Robert correctly his point is about fixing the smgr
> inval alone - without the invalidation fix that'd not be enough because
> the relcache info with outdated information (particularly relallvisible
> et al), would continue to be valid. Relcache invalidations imply an smgr
> one, but not the other way round.
>
> The reason the fix for nmsg > 0 && !markXidCommitted isn't entirely
> sufficient is because the smgr invalidation isn't transaction bound,
> i.e. sent out immediately. So, to have the same behaviour on master/HS,
> we need a way to send out the invalidiation properly in lockstep with
> replay.
What I'm confused about here is:
Masahiko Sawada posted a patch that fixes the problem for him, which
does not involve any new WAL record type. It also seems to be fixing
the problem in a way that is clean and consistent with what we've done
elsewhere.
The patch actually under discussion here manages to introduce a new
WAL record type without fixing that problem.
Therefore I include that the committed patch fixes some *other*
problem, not the one that this thread is ostensibly about.
--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company
