[sheepdog-users] Collie node kill and md

[Liu Yuan]

On 05/23/2013 06:21 PM, Valerio Pachera wrote:
> Ahah, sorry, without noticing I typed 'ex' during the check.
> After it, 'ex' program started.
> 
> The good news is that vdi check completed without any error and the
> guest works fine!
> Also a fsck on the guest hasn't report any problem.
> Again, no idea why I was getting problem before.
> 
> Note: before running 'vdi check'  I run 'vdi flush' on every vdi.
> 
> Is it relevant to flush cache before a vdi check?

The flush is not necessary if you star the guest in the same node as
before. Sheepdog will try to poke around object cache before backend
store. So the guest will always read/write the latest data.

> (in case of a clean shut down and after a kill -9)
> If yes, it may be a good idea to let 'vdi check' flush cache
> automatically before starting.

We can't inline 'flush' capability easily, suppose following scenario
  You have run guest A on node 1 and 2 and A was crashed on the two
nodes without manually flush/delete the object cache. So we have two
cache for A on node 1 and 2, collie can't tell which is the latest one.

Note, normal shutdown of guest will release(flush & delete) the cache
automatically.

Thanks,
Yuan