All; While most discussion of MONs, and their failure modes revolves around the failure of the MONs themselves, the recommendation for od numbers of MONs has nothing to do with the loss of one or more MONs. It's actually in response to the split brain problem.
Imagine you have the following (where 's" is a switch): s1--mon1 | s2--mon2 | s3--mon3 | s4--mon4 Now imagine what happens when the link between s2 and s3 breaks (imagine accidently pulling the wrong cable, port failure on switch, WAN fiber cut, etc.) s1--mon1 | s2--mon2 X s3--mon3 | s4--mon4 All 4 MONs are alive, but which has the official state of the cluster? Which MON(s) can make decisions on behalf of the cluster? Now imagine a similar situation for 3 MONs: s1--mon1 X s2--mon2 | s3--mon3 or: s1--mon1 | s2--mon2 X s3--mon3 The cluster can continue. Similarly imagine 5 MONs: s1--mon1 X s2--mon2 | s3--mon3 | s4--mon4 | s5--mon5 or: s1--mon1 | s2--mon2 X s3--mon3 | s4--mon4 | s5--mon5 or: s1--mon1 | s2--mon2 | s3--mon3 X s4--mon4 | s5--mon5 or: s1--mon1 | s2--mon2 | s3--mon3 | s4--mon4 X s5--mon5 In each case, one side retains a quorum; enough MONs to definitively make decisions on behalf of the cluster. Note that it is just as important, in solving the split-brain problem, to recognize when you are NOT in the quorum (and thus should not make decisions), as to recognize when you are. Within a single datacenter it is relatively easy to ensure that this kind of failure shouldn't occur (ring-style switch stacking for instance), but imagine that you cluster covers a good portion of the Eastern U.S., with MON(s) in Philadelphia, New York, and Baltimore. Can you achieve redundant interconnects without going through the same fiber bundler? Thank you, Dominic L. Hilsbos, MBA Director - Information Technology Perform Air International Inc. [email protected] www.PerformAir.com -----Original Message----- From: ceph-users [mailto:[email protected]] On Behalf Of Alfredo Daniel Rezinovsky Sent: Monday, August 05, 2019 3:28 AM To: ceph-users Subject: [ceph-users] even number of monitors With 3 monitors, paxos needs at least 2 to reach consensus about the cluster status With 4 monitors, more than half is 3. The only problem I can see here is that I will have only 1 spare monitor. There's any other problem with and even number of monitors? -- Alfrenovsky _______________________________________________ ceph-users mailing list [email protected] http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com _______________________________________________ ceph-users mailing list [email protected] http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com
