Autoscaling Controller · View 27 of 37 · 5 · Operations
Decisions
- Decision shards are spread across zones so losing one zone loses a third of the shards and no desired state.
- PostgreSQL runs zone-redundant with a synchronous standby: RPO 0, RTO ≤ 60 s.
- Already-declared capacity keeps converging throughout a zone loss, because the reconciler needs only the desired-state read.
Sizing
- Controller infrastructure budgeted under 1.5% of managed fleet compute spend.
- Signal collectors scale with sample rate, decision pods with target count, reconcilers with instance count — three independent axes.
Assumptions
- The controller's own AKS node pool is statically sized and excluded from its own management, to avoid a loop that scales itself.
- Private endpoints throughout; no public ingress to any control-plane component.
