Le 19/04/2013 16:01, Greg Woods a écrit : > I've got two-node clusters running Heartbeat (3.0.3-2.3.el5) and > Pacemaker (1.0.9.1-1.15.el5) from the clusterlabs repo on CentOS 5.9 > (yes, I know my Pacemaker is a little old, but I don't want to upgrade > unless there is some reason to believe it will help). These clusters run > Xen virtual machines that provide critical infrastructure services (such > as DNS). > > The problem is only that I don't know what is the best way to shut down > the entire cluster cleanly (e.g. as when we do electrical maintenance in > the data center and have a planned complete outage). Often one of the > nodes gets stuck at "Stopping HA Services" and eventually I just have to > hit the power button (physically or via IPMI). > > I can get a clean shutdown if I shut down one node first, then the > second, but that isn't really "clean" because it causes all the virtual > machines from the first node to stop and start on the second node, then > go down again. I'd rather have everything just go down and stay down. > > Is there a recommended method for taking a cluster out of service > cleanly? > > Thanks, > --Greg > >
The way I'd do it (maybe not the best): Put your cluster in maintenance-mode: $ crm configure property maintenance-mode=true That way, resources are not monitored and current state is remembered. Because resources are not monitored anymore, you won't be bothered with them moving from one node to another as you proceed with the shutdowns. As a side effect, you *must* properly shutdown from CLI all resources from CLI (pacemaker will not stop them in the shutdown process). It is very important in the case of shared storages and VMs to avoid data corruption. Then you just proceed with the shutdown of every server, node by node. Once you reboot all your nodes, the cluster is brought up but still in maintenance-mode (it is persisted in the CIB) so no resources will start. Verify that heartbeat connectivity is fine (who knows) and proceed to remove maintenance-mode: resource should start. I'm interested in a better way ;) -- Cheers, Florian Crouzat _______________________________________________ Linux-HA mailing list [email protected] http://lists.linux-ha.org/mailman/listinfo/linux-ha See also: http://linux-ha.org/ReportingProblems
