В Mon, 06 Oct 2014 10:27:49 -0400
Digimer <li...@alteeve.ca> пишет:
> On 06/10/14 02:11 AM, Andrei Borzenkov wrote:
> > On Mon, Oct 6, 2014 at 9:03 AM, Digimer <li...@alteeve.ca> wrote:
> >> If stonith was configured, after the time out, the first node would fence
> >> the second node ("unable to reach" != "off").
> >>
> >> Alternatively, you can set corosync to 'wait_for_all' and have the first
> >> node do nothing until it sees the peer.
> >>
> >
> > Am I right that wait_for_all is available only in corosync 2.x and not in
> > 1.x?
>
> You are correct, yes.
>
> >> To do otherwise would be to risk a split-brain. Each node needs to know the
> >> state of the peer in order to run services safely. By having both start at
> >> the same time, then they know what the other is doing. By disabling quorum,
> >> you allow one node to continue to operate when the other leaves, but it
> >> needs that initial connection to know for sure what it's doing.
> >>
> >
> > Does it apply to both corosync 1.x and 2.x or only to 2.x with
> > wait_for_all? Because I actually also was confused about precise
> > meaning of disabling quorum in pacemaker (setting no-quorum-policy:
> > ignore). So if I have two node cluster with pacemaker 1.x and corosync
> > 1.x with no-quorum-policy=ignore and no fencing - what happens when
> > one single node starts?
>
> Quorum tells the cluster that if a peer leaves (gracefully or was
> fenced), the remaining node is allowed to continue providing services.
>
> Stonith is needed to put a node that is in an unknown state into a known
> state; Be it because it couldn't reach the node when starting or because
> the node stopped responding.
>
> So quorum and stonith play rather different roles.
>
> Without stonith, regardless of quorum, you risk split-brains and/or data
> corruption. Operating a cluster without stonith is to operate a cluster
> in an undermined state and should never be done.
>
OK I try to rephrase. Is it possible to achieve the same effect as
wait_for_all in corosync 2.x with combination of pacemaker 1.1.x and
corosync 1.x? I.e. ensure that cluster does not come up *on the
first startup* until all nodes are present? So just make cluster nodes
wait for others to join instead of trying to stonith them?
_______________________________________________
Pacemaker mailing list: Pacemaker@oss.clusterlabs.org
http://oss.clusterlabs.org/mailman/listinfo/pacemaker
Project Home: http://www.clusterlabs.org
Getting started: http://www.clusterlabs.org/doc/Cluster_from_Scratch.pdf
Bugs: http://bugs.clusterlabs.org
