Note: "permalinks" may not be as permanent as we would like,
direct links of old sources may well be a few messages off.
Il 23/12/2016 09:48, Roland Kammerer ha scritto: ... >>> ...There I would need more information. The version >>> we ship and the one from proxmox are very different. In between these >>> versions the architecture of DM was changed. Just as an example: If you >>> have two nodes in your cluster and you started DM only on one node, it >>> will never start up successfully. Or if only a minority of your overall >>> nodes is operational, drbdmanage will refuse to start, as it waits for a >>> majority. >> For the cluster I will do the test on: it is a three node PVE 4.4-2 >> cluster, vm resides on two main nodes, while the third is dedicated to >> provide quorum and DRBD9 replication only. I will start the test on this >> node. >> >> First question. Is it possible to mix and match versions of drbdmanage >> on the cluster? >> I really would like to update the third node only at first, but now I >> don't think it is possible, given the differences between proxmox >> provided drbdmanage and linbit's one. > The obvious: You could update *all*, there are entries in the control > volume (the .drbdctrl resource acting as cluster DB) that are not used > any more, but they are just ignored on the new version. Ok, nice to know > A mix-and-match isn't a good idea in that scenario. Let's assume you > update the 3rd node only, then this would happen: > > Assuming it is a node that has access to the .drbdctrl, it would try to > become the leader (i.e., switching .drbdctrl to primary). ok > Here it > depends if the resource > is already primary on another node or not. Let's > assume the "good" case, it loses the leader election and decides to > become a satellite. The old nodes would not talk to the new node via > TCP, but the new node expects the leader to talk to it, it would never > receive it's initial configuration. It would just idle around. In order > to remove any potential leftover (autogenerated .res files), the first > thing a node does is to remove all the old autogenerated .res files. So > by starting the new version, you lost all your DM .res files and can not > even manually "drbdadm up" them. For completeness the "bad" case: The > two old nodes idle around and have .drbdctrl as Secondary. The new one > comes and decides to become the leader and switches .drbdctrl to > Primary. The old nodes don't expect that, and would try to switch > .drbdctrl to Primary to store some information (how things worked in the > old version), but a leader never releases the .drbdctrl (new > architecture). Ok, this gives me a good hint about the differences between old and new behaviour. > > So, this is essentially a stop-the-world and then upgrade scenario. I agree. Will try directly that. Bye, rob