RFC #564: Is a Plan-time placement decision worth persisting before an actuation channel exists?

[znerol2]

Moved from #603, so the question can be discussed in the RFC category.

The question

The launcher passes no node selector, site or pool today. Is it worth recording a placement decision at Plan when nothing can yet apply it?

What the plan proposes

• Section 3.3 proposes persisting the constraint and not a resolution. A constraint can be site, colocation, anti-affinity or timing domain. A later compiler step would resolve it. A resolved node name should not be stored.
• Section 3.3 records the facts behind this. The chart accepts placement.nodeSelector and placement.tolerations. The launcher passes neither. The plan states this is not a general placement solution, and the compiler has nowhere to send its output yet.
• Section 3.3 warns that split-node mode adds its own per-role selectors. Any resolved selector has to coexist with them.
• Section 3.3 notes that a resolved node name cannot be verified today, because NetBox records neither facility nor node for an instance (NetBox doesn't record which facility or node a media function instance runs on — populate it and surface it #558).
• Section 1 defers capacity calculation at Plan. A plan-time figure is a projection that can go stale before Provision.

What would change the answer

• Evidence that persisted constraints become misleading while nothing enforces them.
• A first channel that carries a constraint into the chart values, which would make the decision testable.
• A decision on which constraints the compiler will support.

Links

• Discussion: https://github.com/orgs/dmfdeploy/discussions/564
• Plan: https://github.com/dmfdeploy/dmfdeploy/blob/main/docs/plans/DMF%20Media%20Workload%20Design%20Canvas%20Plan.md

This is a request for comment. Answering it does not start any work.