Ceph » rbd

The simple distribution policy should just attempt to assign <number of images> / <number of up instances> to each rbd-mirror instance. The mapping produced by the policy should be stable (i.e. adding or removing an instance won't cause a complete re-shuffle) and persistently recorded. The mapping should be separate from the policy since it should be able to be re-used between policies. The mapping needs to record the current and future states of the mapping (i.e. currently unassigned, requested attach to instance XYZ).

The goal will to use an interface that will support cleanly swapping out the "simple" policy with one that can handle hotspots and other logic under Phase 3. Design for this configuration-level swap-out now so that work can continue on Phase 3 even after freeze.