[Linux-cluster] Tiebreaker IP Address

[chris barry]

On Wed, 2008-01-23 at 08:00 -0600, Brian Kroth wrote:
> Barry Brimer <lists at brimer.org>:
> > As mentioned before, fencing is "pulling the plug" .. if fencing is set up 
> > correctly, the node will reboot and rejoin the cluster.
> 
> Unless it's been simply IO fenced (unplugged from the shared storage).  In
> that case you'll need to reboot and verify its "live" status in the
> cluster before you unfence it.  In my experience though the node tends to
> freak out if it's only been IO fenced and an unclean reboot is necessary
> anyways.  

This has been my experience as well. In a (still running) circa 2002
Sistina GFS FC fenced cluster we're running (but moving off of),
invariably after one node hangs, it's IO fenced, and I have to
power-cycle it, and then typically one or two days later the whole
cluster becomes unstable, and I have to reboot both nodes to clean
things up. After that, typical uptime has been 100 - 150 days.

> This is probably what you meant by "if fencing is setup
> correctly" - both IO and Power fencing methods should be used.