Re: Problem with streaming replication, backups, and recovery (9.0.x)

On 29.03.2011 14:27, Fujii Masao wrote:
> On Tue, Mar 29, 2011 at 6:46 PM, hubert depesz lubaczewski
> <depesz(at)depesz(dot)com> wrote:
>>> Did you use recovery.conf to start standalone PostgreSQL? If not,
>>> recovery doesn't check whether it reaches the recovery ending position
>>> or not. So I guess no problem didn't happen.
>>
>> no, i don't use.
>>
>> hmm .. i am nearly 100% certain that previous pgs did in fact check if the end
>> of recovery is reached.
>
> Yes. In 8.4, that was checked only when starting recovery from the backup
> (i.e., which includes backup_label and backup history file) without
> recovery.conf.
> But in 9.0, the behavior was changed so that only archive recovery (i.e., with
> recovery.conf) checks that. IIRC, we don't have strong opinion about
> this change.
> We should revert, in order to make even crash recovery check whether it
> reaches the ending location?

Hmm, why did we change that? It seems like a mistake, the database is
not consistent until you reach the backup stop location, whether or not
you're doing archive recovery. +1 for reverting that, and backpatching
it as well.

"pg_basebackup -x", which includes all the WAL required to restore in
the pg_xlog directory of the base backup itself, is also affected.
Without the check that you reach the end-of-backup, an aborted base
backup will appear to restore fine, even though some WAL segments are
missing and the backup is incomplete.

--
Heikki Linnakangas
EnterpriseDB http://www.enterprisedb.com