Re: Using pg_upgrade on log-shipping standby servers

On Tue, Jul 17, 2012 at 11:55 AM, Jeff Davis <pgsql(at)j-davis(dot)com> wrote:
> On Tue, 2012-07-17 at 01:02 -0700, Daniel Farina wrote:
>> Could pg_upgrade emit WAL segment(s) to provide continuity of a
>> timeline? So something like:
>
> By "segments" did you mean "records"?

Yes. It would be nicer not to have to tie it to the WAL segment file size.

>> * Take down the writable primary for pg_upgrade
>> * Some WAL is emitted and possibly archived
>> * The old version, when reaching the special pg_upgrade WAL, could
>> exit or report its situation having paused replay (as clearly, it
>> cannot proceed). Unsure.
>
> I don't really understand this step.

"Some WAL is emitted and possibly archived" needs a subject in that fragment:

"pg_upgrade somehow (directly, or indirectly) emits and/or archives
WAL used to complete binary-upgrade". That means that it should
appear in the WAL stream and be subject to archive_command, like any
other WAL.

The sticky part is what the standby should do when it encounters the
special wal-upgrade records. It should probably pause replay to allow
some other program to stop the old postgres version and start the new
version with the same cluster.

>> * Start up a new version of postgres on the same cluster at that
>> point, which plays the upgrade-WAL.
>>
>> I see this being pretty mechanically intensive, but right now my hands
>> are completely tied as to achieving total continuity of my archives,
>> costing a base-backup's worth of risk window upon upgrade.
>
> Does "continuity of archives" mean "avoid downtime" or "maintain a
> single WAL sequence".

The latter.

--
fdr