On Wed, Oct 21, 2009 at 12:14 PM, Andrew Beekhof <[email protected]> wrote: > On Oct 21, 2009, at 12:11 PM, Colin wrote: >> On Wed, Oct 21, 2009 at 12:00 PM, Andrew Beekhof <[email protected]> wrote: >>> On Oct 21, 2009, at 11:50 AM, Colin wrote: >>>> On Wed, Oct 21, 2009 at 11:12 AM, Andrew Beekhof <[email protected]> >>>> wrote: >>>>> On Oct 21, 2009, at 11:05 AM, Colin wrote: >>>>> >>>>>> perhaps the section on adding cluster nodes should mention adjusting >>>>>> the expected-quorum-votes, otherwise you get something like: >>>>> >>>>> expected-quorum-votes is ignored for heartbeat based clusters. >>>>> i should remove it from the crm_mon output when not using openais >>>> >>>> Erm ... so how do I tell heartbeat how many nodes are needed? >>> >>> You don't. >>> It knows how many nodes it has ever seen and you get quorum when you have >>> at least half that. >> >> Ok -- is that a "greater equal" or a "greater than"? > > greater than. thats the definition of quorum > >> In a 8-node cluster we _don't_ want 4 nodes to have quorum, only 5 >> nodes or more, to strictly prevent a split brain... >> >> ...however a 2-node cluster wouldn't work with "greater than 50%", so >> presumably it's the other (which we don't want). > > heartbeat always pretends two-node clusters have quorum.
Ah, I see, that's why I was unsure about the "greater than" vs. "greater equal". Let's see whether I can summarise correctly: For Pacemaker + CoroSync/OpenAIS, the Pacemaker attribute "expected-quorum-votes" tells the cluster how many votes it needs for quorum. (What happens when the option is not set?) For Pacemaker + Heartbeat, the Pacemaker attribute "expected-quorum-votes" is ignored, and Heartbeat's idea of quorum is used, i.e. (a) for a two-node cluster (list of cluster nodes has two entries as per "cl_status listnodes") every node that is up and running believes to have quorum, whether alone or not, and (b) for an n-node cluster with n greater than two a partition with strictly greater than 50% of the nodes has quorum. (IIRC (a) can be changed by making the vote more complicated, but I'm not interested in two-node clusters and did not look into this.) Regards, Colin _______________________________________________ Pacemaker mailing list [email protected] http://oss.clusterlabs.org/mailman/listinfo/pacemaker
