跳到主要内容
版本：最新版

Operate an Operator Deployment

This guide covers the day-two tasks for a SemanticRouter resource. For installation and a first deployment, see Deploy with the Kubernetes Operator.

Read reconciliation status​

kubectl get semanticrouter <name> -o wide
kubectl describe semanticrouter <name>
kubectl get semanticrouter <name> -o jsonpath='{.status.conditions}'

Use metadata.generation, status.observedGeneration, status conditions, and ready replicas together. A running controller does not mean the latest custom resource generation has been applied successfully.

Inspect the owned workload when reconciliation stalls:

kubectl get deployment,pod,service,configmap,pvc \
  -l app.kubernetes.io/instance=<name>
kubectl logs -n semantic-router-operator-system \
  deployment/semantic-router-operator-controller-manager

Update safely​

  1. Export the current custom resource and record the deployed image reference.
  2. Review CRD and release notes for schema changes.
  3. Apply the new custom resource or Operator version in a non-production environment.
  4. Wait for the observed generation and ready replicas to converge.
  5. Send real requests through every important entrypoint.
  6. Roll back the custom resource or image reference if readiness or routing regresses.

Pin image tags or digests. Avoid changing the Operator, Router image, routing policy, model pool, and storage backend in one rollout unless those changes are intentionally coupled and tested together.

Scale and availability​

Set fixed replicas through spec.replicas, or enable the HPA adapter:

spec:
  autoscaling:
    enabled: true
    minReplicas: 2
    maxReplicas: 10
    targetCPUUtilizationPercentage: 70

Use node affinity, tolerations, topology spread or anti-affinity, and a separately managed PodDisruptionBudget to match your availability target. Learning or other replica-local mutable state may require additional design; do not assume that adding replicas makes every stateful routing feature consistent.

Metrics and tracing​

The Router exposes Prometheus metrics on the configured metrics Service port (9190 by default):

kubectl port-forward service/<name> 9190:9190
curl -sS http://localhost:9190/metrics | head

Configure a ServiceMonitor or equivalent scraper with the labels used by your deployment. Enable OpenTelemetry through spec.config.observability and send traces to a collector reachable from the Router namespace. Keep trace sampling and captured attributes appropriate for the sensitivity of request data.

Common failures​

Backend discovery fails​

For service backends, verify the Service name, namespace, port, endpoints, and network policy. For KServe, verify that the InferenceService is ready and its predictor Service exists. For Llama Stack, inspect the labels on candidate Services.

kubectl get service,endpoints -n <backend-namespace>
kubectl describe inferenceservice <name> -n <backend-namespace>

Gateway mode has no route​

The Operator currently does not create an HTTPRoute. Verify the referenced Gateway, then apply a route that targets the Router Service on its API port.

kubectl get gateway -A
kubectl get httproute -A

Pod is in ImagePullBackOff​

Inspect pod events, confirm that the image exists, and provide an imagePullSecret when the registry requires authentication.

kubectl describe pod <pod-name>

See Restricted Network Environments for registry-mirror and cluster-egress guidance.

PVC remains pending​

Check that the requested StorageClass and access mode exist in the cluster and that the provisioner can satisfy the requested size:

kubectl get storageclass
kubectl describe pvc <pvc-name>

Changing storage settings can affect existing data. Back up durable state and follow the storage provider's migration procedure before replacing a claim.

Model artifact download fails​

Reference the download token from a Secret, check egress and certificate configuration, and inspect Router logs. Do not put the token directly in spec.env or a ConfigMap.

Deletion and data​

Before deleting a SemanticRouter, identify which state is ephemeral, which is stored in a PVC, and which is held by external Redis, Valkey, Milvus, Qdrant, or Postgres services. Deleting the custom resource removes managed Kubernetes objects according to their ownership and retention policy; it does not necessarily remove external data stores.

Back up durable data first, and inspect PVC reclaim policies before deleting claims or namespaces.

References​