This post is in a series covering the discussions at the Juno Nova mid-cycle meetup. This post will cover the current state of play of our scheduler refactoring efforts. The scheduler refactor has been running for a fair while now, dating back to at least the Hong Kong summit (so about 1.5 release cycles ago).
The original intent of the scheduler sub-team’s effort was to pull the scheduling code out of Nova so that it could be rapidly iterated on its own, with the eventual goal being to support a single scheduler across the various OpenStack services. For example, the scheduler that makes placement decisions about your instances could also be making decisions about the placement of your storage resources and could therefore ensure that they are co-located as much as possible.
During this process we realized that a big bang replacement is actually much harder than we thought, and the plan has morphed into being a multi-phase effort. The first step is to make the interface for the scheduler more clearly defined inside the Nova code base. For example, in previous releases, it was the scheduler that launched instances: the API would ask the scheduler to find available hypervisor nodes, and then the scheduler would instruct those nodes to boot the instances. We need to refactor this so that the scheduler picks a set of nodes, but then the API is the one which actually does the instance launch. That way, when the scheduler does move out it’s not trusted to perform actions that change hypervisor state, and the Nova code does that for it. This refactoring work is under way, along with work to isolate the SQL database accesses inside the scheduler.
I would like to set expectations that this work is what will land in Juno. It has little visible impact for users, but positions us to better solve these problems in Kilo.
We discussed the need to ensure that any new scheduler is at least as fast and accurate as the current one. Jay Pipes has volunteered to work with the scheduler sub-team to build a testing framework to validate this work. Jay also has some concerns about the resource tracker work that is being done at the moment that he is going to discuss with the scheduler sub-team. Since the mid-cycle meetup there has been a thread on the openstack-dev mailing list about similar resource tracker concerns (here), which might be of interest to people interested in scheduler work.
We also need to test our assumption at some point that other OpenStack services such as Neutron and Cinder would be even willing to share a scheduler service if a central one was implemented. We believe that Neutron is interested, but we shouldn’t be surprising our fellow OpenStack projects by just appearing with a complete solution. There is a plan to propose a cross-project session at the Paris summit to cover this work.
In the next post in this series we’ll discuss possibly the most controversial part of the mid-cycle meetup. The proposal for “slots” for landing blueprints during Kilo.