Обсуждение: Contents of "backup_label" and "*.backup" in pg_wal location

On Fri, Nov 4, 2016 at 1:18 PM, Venkata B Nagothi <nag1010@gmail.com> wrote:
> I see the following contents in the file
> "000000010000000000000044.00000060.backup" which was generated in the pg_wal
> location during the online backup. When pg_stop_backup() is executed, the
> following content is written which includes the content copied from the
> backup_label file.
>
> [...]
>
> Can someone please help me know the importance of the above file?

It is not actually critical, and useful for debugging (you could say
the same about backup_label.old).

> How about having the same contents in the backup_label file as well?

> As of now, the backup_label file does not have any information related to
> when and at what position the backup actually completed.

Yes, and it is not actually possible to write the stop information
because when a backup finishes the backup_label is simply removed and
it has been included in the backup before it finishes. The role of
this file is to provide the LSN start location from which the backup
is able to replay things cleanly. The stop position, aka when
everything on disk is consistent, is determined at replay by the
XLOG_BACKUP_END record. This stop position is not something you can
know when the backup_label file is generated. And I am of course
talking about exclusive backups here.

For non-exclusive backups, it could be possible to add some
information related to the backup stop locations directly in the
backup_label file generated and returned to the caller, though the
information it would contain does not seem much pertinent as what
really matters is the LSN end location, and that's already a piece of
information available.
-- 
Michael

On Fri, Nov 4, 2016 at 7:04 PM, Venkata B Nagothi <nag1010@gmail.com> wrote:
> Sure. I will look at the possibility of using XLOG_BACKUP_END in my patch.
> I am looking at the possibility of keeping the backup_label at source until
> pg_stop_backup()
> is executed and then write the STOP information and then move it across to
> the backup location.

Non-exclusive backups already do that, except that as the backup is
already stopped at the moment the backup_label information is sent
back to the caller, and it is expected that it will be the caller that
will write its contents into the backed up PGDATA in a file named as
backup_label. Anyway, any design in this area for exclusive backups
would be really inconsistent. For example take the case of
pg_start_backup -> cp -> pg_stop_backup for an exclusive backup, when
are you going to update the backup_label file with the stop info?

> I see that when the START/STOP information is written to the WAL history
> file,
> the content from the backup_label file is being copied and I am thinking to
> do the same other way around.
>
> Am i making sense ? is that anyway not possible ?
>
> If this makes sense, then i would start working on an optimal design and
> look at the possibility of achieving this.

Before writing any code, I would be curious about the need behind
that, and you give no reason where this would help in this thread. Do
you actually want to time the timestamp when backup ends? This could
be added as a new field of pg_stop_backup for both the exclusive and
non-exclusive cases. Just an idea.
-- 
Michael