[Sheepdog] partition recovery algorithm

[Chris Webb]

Dietmar Maurer <dietmar at proxmox.com> writes:

> [Jean-Marie BEGUIN <boubouzzzz at orange.fr> wrote:]
> > dog is nice coded in java.
> > why recoding it ? 
> > performance ?
> 
> yes

I'm less concerned about performance, since dog isn't on the fast path, but
more about

> jvm is very heavy compared to C

and also

> redhat cluster and pacemaker runs on top of corosync. And I am sure you
> don't want to configure the cluster and fencing twice.

Some background: in our (ElasticHosts') clusters we have very small,
lightweight host images which can run from an initramfs. Adding a JVM would
more than double the size of this image, as well as the obvious memory
footprint issues: essentially be a complete show-stopper in our application.

At present, we store our disk images locally on LVM volumes distributed
across the cluster, accessing via iSCSI for the non-local case. However,
this means that single machine failure => inaccessibility of the images
stored there, so a move to a properly distributed block store has always
been on the cards.

We had begun writing a distributed store ourselves, but I was very
interested to see sheepdog announced on the qemu list. This project is
further along than we were, and we're quite excited about the idea of
getting involved and contributing to sheepdog instead of continuing to
develop our own in-house solution.

Cheers,

Chris.