CNCF Atlas

English

日本語

English

日本語

Appearance

Architecture ​

Big picture ​

KEDA runs as three binaries built by make build: the operator (manager), the metrics adapter, and the admission webhooks ([2], Makefile:211). The operator owns reconciliation, scaler polling, and the 0-to-1 / idle scaling the HPA cannot do. The adapter implements the Kubernetes External Metrics API but computes nothing itself; it forwards every metric request to the operator over gRPC. The HPA, created and managed by KEDA, still decides 1-to-N scaling.

Components ​

Operator (controller manager) ​

Entry point cmd/operator/main.go (func main at :69). It reconciles ScaledObject, ScaledJob, and TriggerAuthentication resources, polls scalers, and performs activation/deactivation. In the same process it starts the gRPC Metrics Service server: metricsservice.NewGrpcServer(&scaledHandler, ...) ([2], cmd/operator/main.go:355).

Metrics Adapter ​

Entry point cmd/adapter/main.go (func main at :226). It implements the External Metrics API on top of custom-metrics-apiserver and answers HPA queries. It holds no scaling state: it opens a gRPC client to the operator with metricsservice.NewGrpcClient(...) ([2], cmd/adapter/main.go:115) and wraps it in kedaprovider.NewProvider(...) ([2], cmd/adapter/main.go:126).

Admission Webhooks ​

Entry point cmd/webhooks/main.go. It validates ScaledObject and ScaledJob resources before they are admitted.

How a request flows ​

Tracing one metric query from HPA to value:

  1. The HPA calls the External Metrics API, hitting KedaProvider.GetExternalMetric ([2], pkg/provider/provider.go:75). The provider extracts the owning ScaledObject name from the label selector via selector.Get(kedav1alpha1.ScaledObjectOwnerAnnotation) ([2], pkg/provider/provider.go:99).
  2. The adapter does not compute the value. It forwards over gRPC: p.grpcClient.GetMetrics(ctx, scaledObjectName, namespace, info.Metric) ([2], pkg/provider/provider.go:107). If the connection is not yet ready it blocks on WaitForConnectionReady ([2], pkg/provider/provider.go:87).
  3. On the operator side the gRPC handler calls scaleHandler.GetScaledObjectMetrics ([2], pkg/scaling/scale_handler.go:585). It looks up the scalers cache with getScalersCacheForScaledObject ([2], pkg/scaling/scale_handler.go:590) and fans out across triggers, adding a goroutine per scaler (wg.Add(1) inside the loop at :635/:666).
  4. Each scaler implements the Scaler interface ([2], pkg/scalers/scaler.go:44). Concrete scalers come from the buildScaler switch keyed on the trigger type string ([2], pkg/scaling/scalers_builder.go:123).
  5. The result returns as an external_metrics.ExternalMetricValueList back through the adapter to the HPA, which decides the 1-to-N scale.

Separately, the operator drives the 0-to-1 and idle path itself, because the HPA floor is one replica. The reconcile-driven scale loop calls scaleExecutor.RequestScale ([2], pkg/scaling/executor/scale_scaledobjects.go:40), which branches on activity to scale up from zero/idle or down to zero/idle ([2], pkg/scaling/executor/scale_scaledobjects.go:73-117).

The ScaledObject reconcile body is reconcileScaledObject ([2], controllers/keda/scaledobject_controller.go:231), reached from Reconcile ([2], controllers/keda/scaledobject_controller.go:155). It checks pause state (:240), confirms the target is scalable via checkTargetResourceIsScalable (:280), validates triggers (:290), ensures the HPA via ensureHPAForScaledObjectExists (:301), and on a generation change kicks the scale loop with requestScaleLoop (:318).

Key design decisions ​

  • Operator and adapter split, with gRPC delegation. The adapter advertises the External Metrics API but owns no values; it delegates every read to the operator with GetMetrics ([2], pkg/provider/provider.go:107). Scaling state stays centralised in the operator and the adapter stays stateless.
  • 0-to-1 is the operator's job, not the HPA's. The HPA cannot scale below one replica, so KEDA performs activation and deactivation in the RequestScale branches ([2], pkg/scaling/executor/scale_scaledobjects.go:73-117) and leaves 1-to-N to the HPA. This division of labour is the core of KEDA.

Extension points ​

  • CRDs: ScaledObject, ScaledJob, TriggerAuthentication, and ClusterTriggerAuthentication, defined in apis/keda/v1alpha1/ ([2]).
  • Scaler interface: third-party and built-in scalers implement Scaler ([2], pkg/scalers/scaler.go:44); a push variant PushScaler adds Run(ctx, active chan<- bool) ([2], pkg/scalers/scaler.go:57).
  • Authentication providers: TriggerAuthenticationSpec resolves credentials from Pod Identity, Kubernetes secrets, HashiCorp Vault, Azure Key Vault, and others ([2], apis/keda/v1alpha1/triggerauthentication_types.go:75).