How placement works

#Strategies

#Constraints

Requests carry CPU, memory, disk, preferred or required node, region, storage locality, and constraints. Constraint types are affinity, anti-affinity, region, node, and label, each with in, not-in, exists, or not-exists operators. Required constraints filter hard; soft ones add bounded bonuses. A preferred node adds +0.30, storage matches add +0.15, mismatches cost −0.50, and replicas of the same app repel at −0.1 each. Missing data fails closed: unknown totals, unknown operators, and unresolvable affinity all reject rather than guess.

#Explainability

Every evaluation can report which nodes were filtered and why, per-node score breakdowns, and replica-by-replica explanations. The dashboard surfaces this so a surprising placement is debuggable instead of mysterious. Reservations must be made durable before the next placement, or two schedulers can promise the same capacity twice.