Re: [HACKERS] Replication slots and isolation levels

On 2015-10-30 13:42:19 +0100, Michael Paquier wrote:
> On Fri, Oct 30, 2015 at 12:40 PM, Vladimir Borodin wrote:
> > On Thu, Oct 29, 2015 at 3:29 PM, Oleksii Kliukin wrote:
> >> Could it be a consequence of how REPEATABLE READ transactions handle
> >> snapshots? With REPEATABLE READ the snapshot is acquired only once at the
> >> beginning of a transaction; a READ COMMITTED transaction re-evaluates its
> >> snapshot with each new command.
> >
> > I still don’t fully understand why is it so (the problem occurs while
> > running only one SELECT-statement in READ COMMITED so only one snapshot is
> > taken), but if is expected behavior shouldn’t the documentation mention that
> > using READ COMMITTED (which is the default) you may still get conflicts with
> > recovery while using replication slots?
>
> Replication slots and hot_standby_feedback are two different unrelated
> concepts, slots being aimed at retaining WAL.

Uh. Slots also retain the xmin horizon if hot_standby_feedback is
enabled on the standby?

> I guess that's the origin of your confusion:
> http://www.postgresql.org/message-id/20150616192141.GD2626@alap3.anarazel.de

That just says what I said above, I don't see how this makes replication
slots and hs feedback unrelated?

Greetings,

Andres Freund