Re: Using pg_upgrade on log-shipping standby servers

On Thu, Jul 19, 2012 at 12:43:23PM -0400, Robert Haas wrote:
> On Thu, Jul 19, 2012 at 12:02 PM, Bruce Momjian <bruce(at)momjian(dot)us> wrote:
> > On Thu, Jul 19, 2012 at 09:41:29AM -0400, Robert Haas wrote:
> >> On Thu, Jul 19, 2012 at 2:38 AM, Bruce Momjian <bruce(at)momjian(dot)us> wrote:
> >> > No, the point is they run pg_upgrade on the stopped primary and stopped
> >> > standbys. Are those the same? I am not really sure.
> >>
> >> Of course not.
> >
> > OK, but why? When the clusters are stopped they are the same, you are
> > running the same initdb on both matchines, and running the same
> > pg_upgrade. What would cause the difference, other than the Database
> > System Identifier, which we can deal with? I don't think we can
> > guarantee they are the same, but what would guarantee they are
> > different?
>
> There isn't any guarantee that they are different. There's just no
> guarantee that they are the same, which is enough to make this idea a
> non-starter.
>
> In general, it's pretty easy to understand that if you perform the
> same series of inserts, updates, and deletes on two systems, you might
> not end up with the exact same binary contents. There are a lot of
> reasons for this: any concurrent activity whatsoever - even the exact
> timing of autovacuum - can cause the same tuples can end up in
> different places in the two systems. Now, admittedly, in the case of
> pg_upgrade, you're restoring the dump using a single process with
> absolutely no concurrent activity and even autovacuum disabled, so the
> chances of ending up with entirely identical binary contents are
> probably higher than average. But even there you could have
> checkpoints trigger at slightly different times while restoring the
> dumps, and of course checkpoints take buffer locks, and so now a HOT
> prune might happen on one machine but get skipped on the other one
> because the checkpointer has dropped the lock but not the pin, and now
> you're hosed.
>
> Even if you could control for that particular possibility, there are
> surely others now and there will be more in the future.

I think the checkpoint issue is the ideal killer --- thanks.

--
Bruce Momjian <bruce(at)momjian(dot)us> http://momjian.us
EnterpriseDB http://enterprisedb.com

+ It's impossible for everything to be true. +