Re: WAL replication question

Kevin,

I understand the confusion. What I am being asked to do is the following:

1. Use pacemaker to determine who should be the Master node so it can assign a virtual IP to it.

2. Use the lsb:postgresql RA as a master resource to monitor the postgresql process on each node

3. Failover to the slave node when the master fails. Pacemaker moves over the virtual IP. I detect that in a process I run (script that checks the location of the virtual IP) and promote the slave to master using the trigger file.

4. When the failed server comes back up, I detect that it should no longer be master using the process and sync it to the new master, create the recovery.conf file and restart the postgres process.

The issue is when we restart the postgresql, Pacemaker takes that node "down" in alarm. The team lead does not want that. I am just wondering if it was possible to bring up the recovered server without restarting the postgresql process?

Thanks,
Keith

________________________________________
From: Kevin Grittner [kgrittn(at)mail(dot)com]
Sent: Monday, January 21, 2013 4:07 PM
To: Keith Ouellette; pgsql-novice(at)postgresql(dot)org
Subject: Re: [NOVICE] WAL replication question

Keith Ouellette wrote:

> I am using the trigger file capability to "promote" the Slave in
> the event that the Master fails.

> when the former Master comes back on line, is there a way to make
> it a slave (adding the recovery.conf file) without restarting the
> PostgreSQL process? I currently do this with a restart, which
> works well, but we are using Pacemaker and the PostgreSQL
> resource goes away and takes the node down until it comes up. We
> would rather the node not go down if possible.

That's confusing -- if the master node fails, how can it still be
up?

Once you fail over, you should take care that the old master node
is good and truly down until it can be started as a slave. You
will need to copy from the new master to the node which is to be
the new slave before you do that. And yes, add the recovery.conf.

-Kevin