Autoscaling Controller · View 12 of 37 · 3 · Data
Decisions
- The desired-state record is authoritative. A controller restart re-derives nothing and simply resumes reconciling towards the stored value.
- Signals are inputs, not state. Losing 15 days of signal history costs a counterfactual replay and nothing else.
- The journal is append-only and immutable, with the queryable copy in Data Explorer and the evidential copy in WORM Blob.
Why four stores and not one
- Their durability requirements differ by three orders of magnitude, and one store sized for the strictest of them would cost more than the fleet it manages.
- 200,000 samples/second and 150 journal records/second have nothing in common as write profiles.
Risks
- Two authoritative stores in one PostgreSQL instance makes that instance the control plane's single point of failure. Zone-redundant HA with a synchronous standby is the mitigation, and RTO ≤ 60 s the accepted exposure.
- During that RTO the fleet holds capacity and cannot change size. That is the deliberate trade: no change beats a wrong change.