specification: API Commons Rate Limits specificationVersion: '0.1' schema: https://raw.githubusercontent.com/api-evangelist/interface-research/main/schema/api-commons.yml#/$defs/RateLimits provider: KEDA providerId: keda created: '2026-05-04' # Provenance stamped 2026-08-11: this artifact was written by the API Evangelist # bulk sweep dated 2026-05-04, not harvested from the provider. See roadmap#35. method: generated modified: '2026-05-05' reconciled: true tags: - Autoscaling - CNCF - Event-Driven - Graduated - Kubernetes - Rate Limiting description: KEDA does not operate a hosted API; it is an in-cluster controller plus an external-metrics API server consumed by the Kubernetes HPA. There are no first-party, per-second / per-minute rate limits. Polling cadence and concurrent-event behavior are configured per ScaledObject via pollingInterval, cooldownPeriod, and per-scaler caps. sources: - https://keda.sh/docs/latest/concepts/scaling-deployments/ - https://keda.sh/docs/latest/operate/metrics-server/ responseCodes: throttled: 429 serviceUnavailable: 503 limits: - name: ScaledObject pollingInterval scope: scaledobject metric: poll_seconds limit: configured per ScaledObject (default 30s) notes: Effective request rate to the upstream event source is determined by pollingInterval and the number of ScaledObjects targeting it. - name: External Metrics API (in-cluster) scope: cluster metric: varies limit: bounded by Kubernetes apiserver and metrics-server resource limits, not by KEDA policies: - name: Polling Cadence description: Tune pollingInterval and cooldownPeriod per ScaledObject to balance scaling responsiveness against load on the event source. - name: Scaler-Side Limits description: Each scaler (Kafka, RabbitMQ, SQS, etc.) inherits the rate-limit semantics of the underlying source; KEDA itself does not throttle. - name: Backoff Strategy description: KEDA scalers apply exponential backoff on transient scaler errors before surfacing a ScalerError CloudEvent. maintainers: - FN: Kin Lane email: kin@apievangelist.com