Skip to content

Capacity and Performance

Capacity work asks whether a resource can meet measured demand without wasting cost or exhausting a dependency. FDAI combines current utilization, forecast evidence, minimum floors, dependency checks, and promotion state before it can propose a scaling action.

  • Current utilization and saturation by resource and window.
  • Demand trend, forecast horizon, uncertainty, and expected breach time.
  • Minimum and maximum capacity plus warm-capacity floors.
  • Quota, regional availability, and dependent-resource constraints.
  • Workload SLO and error-budget impact.
  • Cost estimate and rollback or scale-back path.

Missing or stale telemetry produces unavailable or held for review evidence. It does not produce a zero-demand assumption.

Freyr evaluates capacity while Njord evaluates cost. Their advice can conflict, such as scale up for reliability versus scale down for efficiency. Specialists remain advisory; Forseti and the safety check apply the configured precedence and autonomy ceiling.

Forseti emits a cross-vertical arbitration request when advice conflicts. Odin applies the versioned priority policy from the rule catalog and returns one reproducible arbitration result before Forseti makes the decision. The default policy prefers SLO protection over cost and architecture advice, but a deployment can supply a reviewed policy without changing agent code. The arbitration result is evidence; it cannot relax a safety check ceiling.

Example: low utilization suggests scale-down while an SLO forecast shows an imminent capacity breach -> the configured priority policy preserves the SLO floor -> what-if still checks quota and dependencies -> the safety check decides observation mode, approval, or promoted execution.

  1. A detector or scheduled evaluation emits a capacity detected issue.
  2. The detected issue correlates with workload SLO, current changes, and incidents.
  3. What-if verifies quota, dependencies, floors, and expected effect.
  4. A typed scale proposal carries scope, batch, rate, stop condition, and rollback.
  5. Observation mode evidence and promotion state determine whether the proposal can reach approval or promoted auto behavior.

Never scale below a declared safety floor, strand a dependency, exceed quota, or treat a forecast as execution authority. Per-resource locks and bounded batch changes prevent competing scale actions from racing.

Runtime checkIf it passesIf it fails or is unknown
Demand and SLO evidence is freshContinue to what-ifHold with unavailable evidence
Quota and dependency checks passBuild a typed proposalNo proposal
Floor, batch, and rate limits holdContinue to the safety checkDeny or reduce scope
Lock and idempotency claim succeedApply at most onceRetry safely or no-op
Stop condition remains healthyContinue the bounded batchStop and follow rollback policy
To learn aboutRead
How forecasts are formedObservability, detection, and forecasting
How workload impact is measuredSLOs and error budgets
How cost and capacity interactCost Governance
How actions are promotedObserve, then enable changes