Using Openstack scheduler hints with juju
william.reade at canonical.com
Mon Feb 4 14:40:31 UTC 2013
On Mon, 2013-02-04 at 14:00 +0000, James Troup wrote:
> William Reade <william.reade at canonical.com> writes:
> > On Fri, 2013-02-01 at 15:23 +0100, William Reade wrote:
> > For the purposes of argument, how well would something like the following work to fulfil the goals that led to the os-scheduler-hints work?
> >> * Units of service X should be deployed to instances running on
> >> different physical machines ("I want my mongos to be separated").
> FWIW, this is the use case that we care about - a lot.
> > juju set-constraints -s mongo avoid-services=mongo
> While this would work and be better than nothing, it would be more
> strict than necessary.
I hadn't imagined that level of strictness. I'd expect a provider to
make a best effort to separate units of a service as best it can, but to
start doubling/tripling/etc up as necessary when it was inevitable.
Maybe the names "prefer-seek-services" and "prefer-avoid-services" would
be clearer -- would that be good enough for you?
> One way of describing what I'm looking for might be:
> juju set-constraints -s haproxy min-physical-redundancy=2
> (Or if we want to get real fancy, it could accept arguments of the form
> 'n+1' - how we define 'n' in that case is left as an exercise to the
> reader ;-)
Well, we do -- but probably not just yet ;p. The trouble is that I
wouldn't be keen on "min-physical-redundancy" *unless* it were expressed
as a function of unit count -- and that feels a bit like overkill at
this stage ;).
More information about the Juju