On May 12, 2016 16:57, "Bruce Momjian" <bruce@momjian.us> wrote:
>
> On Thu, May 12, 2016 at 10:37:28AM -0400, Robert Haas wrote:
> > 3. I think we need to replace pg_upgrade with a real in-place upgrade
> > scheme so that you just fire up the new version of the server on your
> > old data directory, and it rejiggers things in place without needing
> > to create a new cluster and migrate stuff over to it. I think that
> > actually making this work is a huge engineering effort, and I have no
> > plans to undertake it in the near term, but I think it has to be done.
> > pg_upgrade isn't reliable enough, and using pglogical means you need a
> > second machine. Maybe everybody should run with a standby, but not
> > everyone does.
>
> I don't see why you can't have the pg_logical slave be on the same
> server as the master for an upgrade. It will double the write volume
> while it is active, but assuming it is setup only to perform a major
> version upgrade, it should be fine.
>
I think that's a pretty bad assumption. A lot, if not most, of the people who actually need zero downtime upgrades don't have 50% extra space and in particular not 50% extra performance on their servers to throw at that.
Can it be made to work? Sure. But I definitely agree with Robert that we need "real" in place upgrades at some point.
/Magnus