One of the long-term goals of Nova has been to move resource claims from the compute service to the scheduler. This would help with build retries due to collisions in a pack-first placement configuration. This will also help with proper affinity/anti-affinity placement for server groups, which in turn can help us clean up a lot of technical debt and RPC/DB traffice between all of the compute services and the scheduler to sync the instance information for proper affinity placement.
Related to this, we plan on deprecating the CachingScheduler in the Pike release. Only the FilterScheduler is using the Placement service. However, OSIC testing has shown that we still don't get performance parity between the filter scheduler + placement that the caching scheduler gets at large scale, due to the lack of claims in the scheduler. So before we can deprecate/remove the caching scheduler, we need to support claims in the scheduler first.
Operators should be interested in this topic since it relates to scheduling, placement, and also cells v2 because of how multi-cell support is being handled in the scheduler, and what that means for retries.
The priority for claims in the scheduler came up after the Pike PTG so we didn't get a chance to discuss the details and plan at the PTG, so we need to take the opportunity for that at the Forum while operators are available to get their use cases, experience and perspective.
References:
* ML discussion: http://lists.openstack.org/pipermail/openstack-dev/2017-March/113839.html
* Spec: https://review.openstack.org/#/c/437424/
