[ClusterLabs] Antw: Re: [EXT] Re: Two node cluster without fencing and no split brain?

[kgaillot at redhat.com]

On Wed, 2021-07-28 at 08:51 -0400, john tillman wrote:
> > Technically you could give one vote to one node and zero to the
> > other. 
> > If they lose contact only the server with one vote would make
> > quorum. 
> > The downside is that if the server with 1 vote goes down the entire
> > cluster comes to a halt.
> > 
> > 
> > That said, if both nodes can reach the same switch that they are
> > connected to each other through, why can't they reach each other?
> > 
> 
> "... why can't they reach each other?"  My question as well.
> 
> It feels like a very low probability thing to me.  Some
> blockage/filtering/delay of the cluster's "quorum packets" while ping
> packets were allowed through, perhaps caused by network
> congestion.  But
> I'm not a network engineer.  Any network engineers reading this care
> to
> comment?

It's not necessarily that they can't reach each other, but that one is
unresponsive. A kernel driver temporarily blocking activity, CPU or I/O
overload, or losing an essential disk drive (which won't affect
networking) can all cause a server to become unresponsive to cluster
traffic while still potentially having the ability to cause trouble if
resources are recovered elsewhere.

Having a separate network interface for fencing device access (ideally
on a separate physical card) is a good idea, so the interface is not a
single point of failure. Connecting that interface via a dedicated
switch, on a different UPS than the main switch, improves it even more.

A hardware watchdog is a good way to do fencing without all that
trouble.

> Thanks for echoing my thoughts and that interesting quorum-weight
> idea.
> 
> 
> > 
> > On 7/26/21 12:21 PM, john tillman wrote:
> > > They would continue running their resources and we would have
> > > split
> > > brain.
> > > 
> > > So there is no safe way to support a two node cluster 100% of the
> > > time.
> > > But when all you have are two nodes and a switch ... well, when
> > > life
> > > gives
> > > you lemons ...
> > 
> > _______________________________________________
> > Manage your subscription:
> > https://lists.clusterlabs.org/mailman/listinfo/users
> > 
> > ClusterLabs home: https://www.clusterlabs.org/
> > 
> 
> 
> _______________________________________________
> Manage your subscription:
> https://lists.clusterlabs.org/mailman/listinfo/users
> 
> ClusterLabs home: https://www.clusterlabs.org/
> 
-- 
Ken Gaillot <kgaillot at redhat.com>