[sheepdog-users] Upgrade procedure?

[Liu Yuan]

On 06/25/2012 04:11 PM, icez network wrote:
> I'm currently using sheepdog v0.3.0 (2 copies) with qemu 1.0 (fedora 17
> backported to centos 6.2) on my pre-production cluster running on 4
> machines. The current status seems to be fine with some rarely qemu
> segfault mentioned earlier by 'David Douard'. The performance is
> impressive and the cluster recovery process seems working properly when
> i got storage problem last week on one node.
> 

This problem is really a headache. Kazutaka has some un-submiteed
patches on his QEMU tree as mentioned in the thread David Douard started.

The good news is that with those patches, the QEMU run much stable than
before. But bad news is that, the problem still exists (QEMU segfault)

> Now I'm testing the recent sheepdog on git with my test cluster (running
> in 3 VMs with 2 copies) and found that the storage backend has been
> changed from 'simple' to 'farm'. I still haven't test the upgrade
> procedure from v0.3.0 yet since my test cluster have some task to do so
> may I ask if there's any 'ONLINE' upgrade procedure to upgrade sheepdog
> from v0.3.0 to current master version on git? The process I planned to
> use is by following:
> 
> - kill 1 node and remove from cluster
> - upgrade to the recent sheepdog version.
> - clear all the target backend folder used by old sheepdog. (rm -rf
> /var/sheepdog ?)
> - start sheepdog again and wait until cluster recovery process done.
> - then do the above process with each other machines until complete.
> 
> Is there anything I should take care of?

Online upgrade is not allowed for current implementation. Well, not
technically, it is actually very easy to allow different Sheep run with
different backends from code perspective. But for now we don't have
internal protocol/version compatibility check, so it is at least not
safe to run different version sheep.

The good news is that Christoph will go to implement version/protocol
compatibility check feature, I think we can see it in the near future.

Thanks,
Yuan