Placement constraints

What I want to achieve:

Ensure replicas are spread across availability zones and never run in the same.

current approach:

I am currently doing this through placement preferences.

The problem:

However, they are that. Preferences. If one zone becomes unavailable, the other replica will be scheduled in the other zone which already contains a replica.

This is particularly troublesome, as the service is backed by persistent storage (ebs), which is not replicated across availability zones; meaning that the replica will start up with an empty volume in that zone.

Is it possible to set a placement preference similar to how constraints work? So that I can guarantee a (one) replica will always run in a distinct zone, if the other zone is not available then do not start the replica? If not, is there an alternative way of achieving this?

The issue names no files, tests, or entry points; begin by locating the scheduler's placement-preference and constraint handling. Determine whether hard availability-zone separation is feasible for replicas backed by EBS, then define the expected behavior or an alternative approach.