On Thu, Jul 21, 2016 at 5:19 PM, Kyotaro HORIGUCHI
<horiguchi.kyotaro@lab.ntt.co.jp> wrote:
> + * minRecoveryPoint can go behind the last checkpoint's redo location when
> + * the checkpoint writes out no buffer. This does no harm to performing a
> + * recovery but such inversion seems inconsistent from the view of the
> + * callers and prevents them from knowing WAL segments needed by sane
> + * calcuation. For the reason we return the last replayed point as the
> + * backup end location. Note that it can be greater than the exact backup
> + * end location or even the minimum recovery point of pg_control at the
> + * time. This is harmless for current uses.
It does not seem an improvement to me to mention a comment regarding
minRecoveryPoint in do_pg_stop_backup, especially knowing that the
patch that we have here uses the last replayed LSN and TLI and does
not care about that anymore.
> AFAICS pg_stop_backup returns a single value of LSN. I don't know
> where this comes from, but the attached patch fixes this and adds
> a mention on the "gap". The original description mentioned
> backup_label and tablespace_map but it seems not necessary. The
> following is the new content rewritten by the patch.
No, this second patch is wrong. This part of the docs refers to
non-exclusive backups, where 3 fields are returned. So the docs are
correct.
--
Michael