On Wed, Feb 11, 2015 at 12:55 PM, Thom Brown <thom@linux.com> wrote:
> Today I witnessed a situation which appears to have gone down like this:
>
> - The primary server starting streaming WAL data from segment 00A8 to the
> standby
> - The standby server started receiving that data
> - Before 00A8 is finished, the wal sender process dies on the primary, but
> the archiver process continues, and 00A8 ends up being archived as usual
> - The primary continues to generate WAL and cleans up old WAL files from
> pg_xlog until 00A8 is gone.
> - The primary is restarted and the wal sender process is back up and running
> - The standby says "waiting for 00A8", which it can no longer get from the
> primary
> - 00A8 is in the standby's archive directory, but the standby is waiting for
> the rest of the segment from the primary via streaming replication, so
> doesn't check the archive
> - The standby is restarted
> - The standby goes back into recovery and eventually replays 00A8 and
> continues as normal.
>
> Should the standby be able to get feedback from the primary that the
> requested segment is no longer available, and therefore know to check its
> archive?
Last time I played around with this, if the standby requested a
segment from the master that was no longer present there, the standby
would immediately get an ERROR, which it seems like would get you out
of trouble. I wonder why that didn't happen in your case.
--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company
