From: "Joshua D. Drake" <jd@commandprompt.com>
> >Since pgpool has this capability, how about including a hook that allows
a
> >script to be run when pgpool detects a problem with the master? That
would
> >allow action to be taken to investigate further and, if required,
switchover
> >or failover and promote the slave to master.
> >
> >
> Because it is not the place of a connection management software
> to failover. Failover should happen only if:
>
> 1. You have a specific set of criteria that is matched via a
network/system
> management software.
>
> 2. You do it yourself.
I'm not suggesting that it's the place of pgpool to *force* a failover. I
am suggesting that one of the criteria that is likely to be useful is the
inability to connect to the master, and that's something that pgpool,
apparently, detects. It seems unnecessary to use completely different
failure-detection mechanisms for the purpose of failover to those used for
the connection management.
So all I'm looking for is a way for pgpool to shout if it detects a failure.
That could initiate the investigation of the other criteria required for
failover.
> The last thing in the world you need is to fail over to a slave because
> somebody accidently tripped over a network cord.
In our application, that's *exactly* what we need. We have a database that
receives data in a fairly continuous stream. If the datastream cannot be
written to the database, the database becomes worse than useless quite
rapidly. We need the ability to switchover or failover to another node as
master as soon as possible, to allow the datastream to be written to the
other node. We'll rebuild the "failed" master later, if necessary. But if
the failover doesn't happen promptly, we might as well rebuild the whole
cluster.
Julian Scarfe