Open Stack

>
> No argument here.  I agree that it's an optimization and you know what they
> say about premature optimization. :-)
>
> I do wonder whether the ongoing discussion about where container support
> should live in OpenStack is relevant to this as well.  Or would we not
> intend to manage the TripleO containers using OpenStack?  If we _do_ then
> that is going to be a dependency too - we won't want to commit to using
> containers until OpenStack as a whole has decided how we're going to support
> them.

This is the crux of it :) - and why I'm proposing we defer this until
folk have bandwidth to do a decent job of analysis.

But as a teaser - here's the stack, for a given arbitrary service -
overcloud Nova-compute:

Nova-compute
  deployed in container $UUID
    deployed by container-driver X
      on machine $UUID deployed by Ironic

So we have:
Ironic deploying the container substrate
The container substrate (e.g. nova + docker) deploying containers
The overcloud service itself

Thats three distinct hypervisors in play: are there three clouds, or
two clouds and three regions? Or two clouds and one multi-hypervisor
region?

How do we orchestrate graceful deployments of the container substrate
(consider - if nova-compute is running in a container, we'd have to
quiesce load on a machine at two separate levels...)?

Will VIF plugging, and iSCSI work properly in containers? There's some
reason to think that iSCSI won't, for one... we hit a wall with it
with nova-baremetal.

That may mean that we deploy a single overcloud via two different
hypervisors - simple service in containers and compute/storage to
baremetal).

And then there's the question of where the container substrate should
be owned: should it be part of the overcloud and deployed in the same
stack - e.g. 'we get full machines and for some of them we choose to
subdivide', or should it be an undercloud service - 'you can ask for
machines or containers'.

To my mind there are too many unknowns to meaningfully design at this
point. We'll need to do a bunch of experiments to learn the
constraints....

-Rob

-- 
Robert Collins <rbtcollins at hp.com>
Distinguished Technologist
HP Converged Cloud