Re: Lets (not) break all the things. Was: [pgsql-advocacy] 9.6 -> 10.0

On 21 June 2016 at 20:19, Robert Haas <robertmhaas(at)gmail(dot)com> wrote:

> On Mon, Jun 20, 2016 at 10:08 PM, Bruce Momjian <bruce(at)momjian(dot)us> wrote:
> > On Fri, May 20, 2016 at 07:40:53PM -0400, Robert Haas wrote:
> >> On Mon, May 16, 2016 at 3:36 AM, Bruce Momjian <bruce(at)momjian(dot)us>
> wrote:
> >> > On Sun, May 15, 2016 at 03:23:52PM -0500, Jim Nasby wrote:
> >> >> 2) There's no ability at all to revert, other than restore a backup.
> That
> >> >> means if you pull the trigger and discover some major performance
> problem,
> >> >> you have no choice but to deal with it (you can't switch back to the
> old
> >> >> version without losing data).
> >> >
> >> > In --link mode only
> >>
> >> No, not really. Once you let write transactions into the new cluster,
> >> there's no way to get back to the old server version no matter which
> >> option you used.
> >
> > Yes, there is, and it is documented:
> >
> > If you ran <command>pg_upgrade</command> <emphasis>without</>
> > <option>--link</> or did not start the new server, the
> > old cluster was not modified except that, if linking
> > started, a <literal>.old</> suffix was appended to
> > <filename>$PGDATA/global/pg_control</>. To reuse the old
> > cluster, possibly remove the <filename>.old</> suffix from
> > <filename>$PGDATA/global/pg_control</>; you can then restart the
> > old cluster.
> >
> > What is confusing you?
>
> I don't think I'm confused. Sure, you can do that, but the effects of
> any writes performed on the new cluster will not be there when you
> revert back to the old cluster. So you will have effectively lost
> data, unless you somehow have the ability to re-apply all of those
> write transactions somehow.
>

Also, if you run *with* --link, IIRC there's no guarantee that the old
version will be happy to see any new infomask bits etc introduced by the
new Pg. I think there will also be issues with oid to relfilenode mappings
in pg_class if the new cluster did any VACUUM FULLs or anything. It seems
likely to be a bit risky to fall back on the old cluster once you've
upgraded with --link . TBH it never even occurred to me that it'd be
possible at all until you mentioned.

I always thought of pg_upgrade as a one-way no-going-back process either
way, really. Either due to a fork in history (without --link) or due to
possibly incompatible datadir changes (with --link).

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