Deployment / ReplicaSet / StatefulSet / DaemonSet

Deployment

Replicas mismatch (spec vs available)

kube_deployment_spec_replicas != kube_deployment_status_replicas_available

Generation mismatch (rollout non completato/osservato)

kube_deployment_status_observed_generation != kube_deployment_metadata_generation

ReplicaSet

Replicas mismatch

kube_replicaset_spec_replicas != kube_replicaset_status_ready_replicas

StatefulSet

Down (ready < desired)

kube_statefulset_replicas != kube_statefulset_status_replicas_ready > 0

Replicas mismatch

kube_statefulset_status_replicas_ready != kube_statefulset_status_replicas

Generation mismatch

kube_statefulset_status_observed_generation != kube_statefulset_metadata_generation

Update non rollato su tutte le repliche

max without (revision) (
  kube_statefulset_status_current_revision
  unless kube_statefulset_status_update_revision
)
* (kube_statefulset_replicas != kube_statefulset_status_replicas_updated)

DaemonSet

Rollout bloccato

kube_daemonset_status_number_ready
/ kube_daemonset_status_desired_number_scheduled * 100 < 100
or
kube_daemonset_status_desired_number_scheduled - kube_daemonset_status_current_number_scheduled > 0

Misscheduled (pod schedulati dove non dovrebbero)

kube_daemonset_status_number_misscheduled > 0

CPU/Memoria aggregati per workload (via label del pod)

Per aggregare al livello Deployment serve joinare kube_pod_info/kube_replicaset_owner con le metriche cAdvisor. Esempio via owner_kind/owner_name di kube_pod_owner:

sum by (namespace, owner_name) (
  rate(container_cpu_usage_seconds_total{container!="", container!="POD"}[5m])
)
* on(namespace, pod) group_left(owner_name)
  kube_pod_owner{owner_kind="ReplicaSet"}

In pratica: più veloce filtrare per pod=~"<deploy-name>-.*" in ambienti WR quando i nomi sono prevedibili (naming convention del chart/deployment).