Open Stack

On Sat, 2015-02-07 at 00:44 +0000, Adrian Otto wrote:
> Magnum Team,
> 
> In our initial spec, we addressed the subject of resource scheduling. Our plan was to begin with a naive scheduler that places resources on a specified Node and can sequentially fill Nodes if one is not specified.
> 
> Magnum supports multiple conductor backends[1], one of which is our Kubernetes backend. We also have a native Docker backend that we would like to enhance so that when pods or containers are created, the target nodes can be selected according to user-supplied filters. Some examples of this are:
> 
> Constraint, Affinity, Anti-Affinity, Health
> 
> We have multiple options for solving this challenge. Here are a few:
> 
> 1) Cherry pick scheduler code from Nova, which already has a working a
> filter scheduler design. 
> 2) Integrate swarmd to leverage its scheduler[2]. 
> 3) Wait for the Gantt, when Nova Scheduler to be moved out of Nova.
> This is expected to happen about a year from now, possibly sooner.
> 4) Write our own filter scheduler, inspired by Nova.
> 
> I suggest that we deserve to have a scheduling implementation for our
> native docker backend before Gantt is ready. It’s unrealistic that the
> Magnum team will be able to accelerate Gantt’s progress, as
> significant changes must be made in Nova for this to happen. The Nova
> team is best equipped to do this. It requires active participation
> from Nova’s core review team, which has limited bandwidth, and other
> priorities to focus on. I think we unanimously agree that we would
> prefer to use Gantt, if it were available sooner.
> 
> I suggest we also rule out option 4, because it amounts to
> re-inventing the wheel.
> 
> That leaves us with options 1 and 2 in the short term. The
> disadvantage of either of these approaches is that we will likely need
> to remove them and replace them with Gantt (or a derivative work) once
> it matures. The advantage of option 1 is that python code already
> exists for this, and we know it works well in Nova. We could cherry
> pick that over, and drop it directly into Magnum. The advantage of
> option 2 is that we leverage the talents of the developers working on
> Swarm, which is better than option 4. New features are likely to
> surface in parallel with our efforts, so we would enjoy the benefits
> of those without expending work in our own project.
> 
> So, how do you feel about options 1 and 2? Which do you feel is more
> suitable for Magnum? What other options should we consider that might
> be better than either of these choices?
> 
> I have a slight preference for option 2 - integrating with Swarm, but
> I could be persuaded to pick option 1, or something even more
> brilliant. Please discuss.

Got to say that Option 1 looks far preferable.  As you say, we have to
switch to Gantt eventually, so it might end up being an expensive and
difficult retro fit with Option 2.  With Option 1, we look mostly like
the Nova scheduler, so can let them take the initial hit of doing the
shift to Gantt and slipstream in their wake once the major pain points
are ironed out.

James