[Linux-cluster] Quorum question
Bowie Bailey
Bowie_Bailey at BUC.com
Wed Dec 21 13:53:49 UTC 2005
Patrick Caulfield <mailto:pcaulfie at redhat.com> wrote:
> Bowie Bailey wrote:
> > Graham Wood <mailto:gwood at dragonhold.org> wrote:
> >
> > > > Can I just drop the "two_node" definition for a 3-node cluster
> > > > to force it to keep running with only one node?
> > >
> > > If you're looking at GFS, then this arrangement is almost
> > > definitely going to fry the data in the partition - which will
> > > take the system down for you permanently.
> > >
> > > Imagine that the 3 nodes lose communication (but all three are
> > > still running) - they're all going to reply the logs from the
> > > other two, and then start writing to the shared filesystem as if
> > > they were the only ones in the cluster.
> > >
> > > Which will corrupt the GFS very quickly.
> >
> > Isn't that what fencing is supposed to take care of? Maybe I'm not
> > understanding how this all works together.
>
> No, because fencing has to be done by one of the cluster nodes. And
> the cluster must be quorate to fence another node - otherwise it
> could be an isolated node fencing the valid part.
Ok, that makes sense. How does this work with a two-node cluster?
> > What I will have is three nodes. Two that actively use the data in
> > the shared storage and one node that handles backups.
> >
> > The backup node is not critical and could be down at any time for a
> > number of reasons. I want to make sure that if the backup node is
> > down and one of the other nodes crashes, that the one remaining
> > node will continue to be able to access the data in the GFS.
Is there a way to make my setup work the way I want?
--
Bowie
More information about the Linux-cluster
mailing list