Verify and troubleshoot

Start with the workload and fleet status. All commands name the kubeconfig context explicitly.

kubectl --context YOUR_CONTEXT -n celld-system rollout status deployment/celld-celld-operator --timeout=180s
kubectl --context YOUR_CONTEXT -n fleets get statefulset/my-fleet,service/my-fleet
kubectl --context YOUR_CONTEXT -n fleets describe celldfleet my-fleet
kubectl --context YOUR_CONTEXT -n fleets get pods -o wide

A healthy first run has the operator Deployment available, three ready runtime Pods, a my-fleet Service on port 8080, and Ready=True on the fleet. Ready follows the current workload’s applied replicas and generation; compare it with status.desiredReplicas and status.appliedReplicas when a change is in progress. Conditions and Events are operational observations, not proof that a previous writer was safely stopped or an acknowledged write survived a failure.

Symptom	Check next
Blocked=True, NamespaceAccessDenied	Check fleets RoleBinding and its subject. For Helm release celld, it must name celld-system/celld-celld-operator.
Blocked=True, NetworkPolicy not attested	Verify CNI enforcement, then set chart networkPolicyEnforced=true.
Pods Pending	kubectl --context YOUR_CONTEXT -n fleets describe pod POD_NAME shows zone, capacity or pull errors. Strict placement will not silently relax.
Pods start but are not ready	Inspect Pod logs and bucket access. For PersistentFleet see runtime recovery.
Ready=True but application request fails	Confirm celld deploy used the fleet bucket, inspect runtime logs, and use an authorized client label for in-cluster traffic.
Blocked=True, storage reservation conflict	Use a new, dedicated bucket. A reservation is bound to a fleet UID and is not released by routine deletion.

For a focused status view:

kubectl --context YOUR_CONTEXT -n fleets get celldfleet my-fleet -o json | jq '.status.conditions'
kubectl --context YOUR_CONTEXT -n fleets logs statefulset/my-fleet --tail=100
kubectl --context YOUR_CONTEXT -n celld-system logs deployment/celld-celld-operator --tail=100

Events appear when blocker state changes, so a quiet fleet may have no new Events. The conditions reference defines the current types and reasons. Optional Prometheus metrics are configured through chart values; Metrics Server is needed for capacity recommendations, not for this first fleet.

The repository’s disposable make integration harness runs with kind and MinIO. It is useful to test the implementation locally; its Kind storage driver differs from EBS. Do not use --local-test against EKS. If a lifecycle change keeps waiting, preserve the storage reservation and affected disks and follow lifecycle troubleshooting.