Re: Updated backup APIs for non-exclusive backups

On 3/22/16 12:14 PM, Magnus Hagander wrote:
> On Tue, Mar 22, 2016 at 5:08 PM, David Steele <david(at)pgmasters(dot)net
> <mailto:david(at)pgmasters(dot)net>> wrote:
>
> On 3/19/16 8:15 AM, Magnus Hagander wrote:
>
> > I've attached an updated patch, which is rebased on current master and
> > includes the oid fix.
>
> Before doing a thorough review of this patch there are a few points I
> would like to consider:
>
> * I think it's really important to provide the stop time in some fashion
> when using this new technique. I would prefer a new column to be
> returned from pg_stop_backup() but I could live with STOP TIME being
> recorded in the label file. STOP TIME should probably be included in
> the label file anyway.
>
> Adding the stop time column should be a simple addition and I don't see
> a problem with that. I think I misunderstood your original request on
> that. Because you are just talking about returning a timestamptz with
> the "right now" value for when you called pg_stop_backup()? Or to be
> specific, just before pg_Stop_backup *finished*. Or do you mean when
> pg_stop_backup() started?

What would be ideal is the minimum time that could be used for PITR. In
an exclusive backup that's the time the end-of-backup record is written
to WAL. In a non-exlusive backup I'm not quite sure how that works.

> Doing it in the backup label file is obviously a different target, where
> we might need to consider backwards compatibility, Should we?

Physical backups can only be restored in the same version so I'm not
sure why it would be a problem? Do you mean for programs outside of
Postgres that are parsing this file?

> * It seems like STOP WAL LOCATION should now also be recorded in the
> label file. Preferably this would used by recovery to determine when
> the database has reach consistency but that could be a future patch.
> I'm not very happy with the current method of using pg_control to get
> this information as it assumes that pg_control is copied last (at least
> based on the code comments).
>
> That seems entirely out of scope for this patch, though. Doesn't mean it
> shouldn't be done, but that's a separate thing.

Fair enough.

--
-David
david(at)pgmasters(dot)net