Open Stack

On Thu, 2012-11-22 at 11:34 +0100, Thierry Carrez wrote:
> Monty Taylor wrote:
> > On 11/21/2012 09:13 AM, Doug Hellmann wrote:
> > 
> >>     My understanding of option 5 is that the library would be use by
> >>     ceilometer to access Nova, not the other way around ?
> >>
> >> I thought that option represented moving the virtualization code out of
> >> nova into a library that was used by both components for talking to the
> >> hypervisor. That introduces a new dependency, but neither component is
> >> communicating with the other so there is still no dependency between
> >> nova and ceilometer (in fact, removing our current dependency on nova is
> >> part of the motivation for these changes).
> > 
> > I'm a fan of this. It uses existing software engineering concepts
> > (libraries) that are well understood with well understood machinery. Of
> > course, the proof is in the pudding (what does the patch actually look
> > like) - but I think this has the best chance of being something that
> > doesn't turn crazy.
> 
> This just looks a lot like an Oslo (common code) library. Is there any
> reason why we couldn't just use Oslo for this ? The code is common
> between Nova and Ceilometer, and should become a separate library
> (oslo-virt ?) used by both. That way we reuse a well-known model and
> there is no reason to create a special case, and I'm completely fine
> with it.
> 
> The versioning scheme discussion then becomes the same as the oslo
> libraries versioning scheme discussion, and I suggest we continue there:
> 
> http://lists.openstack.org/pipermail/openstack-dev/2012-November/002709.html
> 
> Additional questions on the oslo route: is it OK to move to oslo, today,
> code that is currently only used in one core project (Nova), to care for
> the need of an incubated project (Ceilometer) ? Could we be using
> oslo-incubator copy/update mechanism in the first stage of this ? I
> guess we need markmc's view on this.

I went back over the thread and replied:

  http://lists.openstack.org/pipermail/openstack-dev/2012-November/003222.html

AFAICT Ceilometer is looking to re-use approximately 3% of Nova's
libvirt driver. There's no way the Nova project should incur the
complexity (separate team, releases, cross-project changes, stable APIs,
etc.) of managing *all* of the virt driver code as a separate project
just for the sake of the re-use of that 3%.

What might make sense is a little "talk to the hypervisor and get
CPU/disk/NIC usage data" library which both Ceilometer and the nova virt
drivers would use. That could be an Oslo library but I really doubt that
would be worth it.

Finally, I think it's worth investigating in more depth whether Nova can
publish notifications with the data required for Ceilometer to meter
this stuff.

Cheers,
Mark.