Thread: what to do after a failover

On Wed, Jan 08, 2020 at 11:06:28PM -0500, Rita wrote:
> I run a master and standby setup with Postgresql 11. The systems are
> identical from a hardware and software setup.  If the master goes down I
> can do a pg_ctl promote on the standby and point my applications to use the
> standby (new master).
>
> Once the original master is online, when is an appropriate time to fail
> back over? And are there any other things besides promote after the
> failover is done?

Make sure that you still have an HA configuration able to handle
multiple degrees of failures with always standbys available after a
promotion.

The options available to rebuild your HA configuration after a
failover depend on the version of PostgreSQL you are using.  After a
failover the most simple solution would be to always recreate a new
standby from a base backup taken from the freshly-promoted primary,
though it can be costly depending on your instance.  You could also
use pg_rewind (available in core since 9.5) to recycle the previous
primary and reuse it as a standby of the new promoted custer.  Note
that there are community-based solutions for such things, like
pg_auto_failover or pacemaker-based stuff just to name two.  These
rely on more complex architectures, where a third node is present to
monitor the others (any sane HA infra ought to do at least that to be
honest).
--
Michael

On Thu, 9 Jan 2020 06:55:18 -0500
Rita <rmorgan466@gmail.com> wrote:

> Thanks for the response.
> I am using Postgresql 11.
> I want something simple and I have a strong preference toward using stock
> tools. After the promotion and the original master comes online, I was
> thinking of doing a pg_basebackup to sync. Any thoughts about that? 

If you can afford that, this is the cleanest and easiest procedure you could
find.

Note that pg_basebackup need an empty PGDATA, so it will have to transfert the
whole instance from new promoted primary to the original one.

Regards,

On Thu, Jan 09, 2020 at 03:14:59PM +0100, Jehan-Guillaume de Rorthais wrote:
> If you can afford that, this is the cleanest and easiest procedure you could
> find.
>
> Note that pg_basebackup need an empty PGDATA, so it will have to transfert the
> whole instance from new promoted primary to the original one.

Simple is easier to understand.  Now the larger your instance, the
longer it takes to copy a base backup and the longer your reduce the
availability of your cluster.  So be careful with what you choose.
--
Michael