Kubernetes graph

A StatefulSet database: headless Service, per-Pod volumes, snapshots and restore

The objects around a replicated database on Kubernetes: a StatefulSet with a headless Service and one PersistentVolumeClaim per Pod, a PodDisruptionBudget, and a VolumeSnapshot restored into a new claim.

A StatefulSet gives each replica a fixed name and its own volume. db-0 is always db-0, it always gets the claim data-db-0, and it has a DNS name of its own. That is what a database needs and what a Deployment can't give.

The second half of the graph is the backup path: a CSI snapshot of one replica's volume, and a new claim created from it.

Click a resource to see what it does, or follow the setup step by step.

You write itCreated by a controllerComes with an add-onreferencesselects by labelcreates

How it works

All 10 resources and what they do

Service dbv1: A headless Service (clusterIP: None). It has no virtual IP: DNS returns the Pod addresses, and each Pod gets a name such as db-0.db.data.svc.cluster.local. Replicas use those names to find each other.
StatefulSet dbapps/v1: Creates db-0, db-1 and db-2 in order, each from the same template, and one claim per Pod from volumeClaimTemplates. serviceName ties the Pods to the headless Service for their DNS names.
Pod db-0, db-1, db-2v1: Stable names with an ordinal. db-1 starts only when db-0 is Ready, and a rolling update goes from the highest ordinal down. A Pod that is replaced comes back with the same name and the same claim.
PersistentVolumeClaim data-db-0, data-db-1, data-db-2v1: One per Pod, named <template>-<pod>. The StatefulSet creates them and, with the retention policy on Retain, never deletes them, not on scale-down and not when the StatefulSet is deleted.
StorageClass gp3storage.k8s.io/v1: Provisions each volume when its Pod is scheduled (WaitForFirstConsumer), in that Pod's zone. A volume lives in one zone, so its Pod is tied to that zone for good.
PodDisruptionBudget dbpolicy/v1: Lets a node drain or an autoscaler evict one database Pod at a time. A second eviction is refused until the first Pod is Ready again, which protects the quorum.
VolumeSnapshotClass ebs-snapshotssnapshot.storage.k8s.io/v1: Which CSI driver takes snapshots and what happens to the cloud snapshot when the object is deleted. Retain keeps it.
VolumeSnapshot db-0-2026-11-13snapshot.storage.k8s.io/v1: A request for a point-in-time copy of one claim. status.readyToUse turns true when the driver has finished. It is crash-consistent: what Postgres would find after a power cut.
VolumeSnapshotContent snapcontent-8c1f…snapshot.storage.k8s.io/v1: The cluster-scoped record of the real snapshot, with its ID at the cloud provider. It is to a VolumeSnapshot what a PersistentVolume is to a claim.
PersistentVolumeClaim data-db-restore-0v1: A new claim whose dataSource is the snapshot. The driver creates a volume from it, at least as large as the original. Named data-<statefulset>-0, it is adopted by a StatefulSet called db-restore.

Common mistakes

kubectl delete statefulset db leaves the claims and the data. Deleting a claim with a Delete reclaim policy destroys the disk. Know which one you are typing.
volumeClaimTemplates can't be edited. To grow the volumes, expand each claim (the class needs allowVolumeExpansion), then recreate the StatefulSet with --cascade=orphan and the new size.
A volume is zonal. If its zone has no node with room, the Pod stays Pending with volume node affinity conflict, however empty the other zones are.
A snapshot of a running database is crash-consistent only. For a clean one, use the database's own backup tool or take the snapshot from a replica you paused.
VolumeSnapshot needs the snapshot CRDs and the snapshot controller, which many clusters don't install with the CSI driver. Without them the object is accepted by nothing and kubectl get volumesnapshot fails.

The manifests you write

Everything above that is marked "You write it", in one file. The rest comes from add-ons and controllers.

apiVersion: v1
kind: Service
metadata:
  name: db
  namespace: data
spec:
  clusterIP: None
  selector:
    app: db
  ports:
    - name: postgres
      port: 5432
---
apiVersion: apps/v1
kind: StatefulSet
metadata:
  name: db
  namespace: data
spec:
  serviceName: db
  replicas: 3
  selector:
    matchLabels:
      app: db
  persistentVolumeClaimRetentionPolicy:
    whenDeleted: Retain
    whenScaled: Retain
  template:
    metadata:
      labels:
        app: db
    spec:
      terminationGracePeriodSeconds: 60
      containers:
        - name: postgres
          image: postgres:18
          ports:
            - containerPort: 5432
          envFrom:
            - secretRef:
                name: db-credentials
          volumeMounts:
            - name: data
              mountPath: /var/lib/postgresql/data
  volumeClaimTemplates:
    - metadata:
        name: data
      spec:
        accessModes: ["ReadWriteOnce"]
        storageClassName: gp3
        resources:
          requests:
            storage: 100Gi
---
apiVersion: policy/v1
kind: PodDisruptionBudget
metadata:
  name: db
  namespace: data
spec:
  maxUnavailable: 1
  selector:
    matchLabels:
      app: db
---
apiVersion: snapshot.storage.k8s.io/v1
kind: VolumeSnapshotClass
metadata:
  name: ebs-snapshots
driver: ebs.csi.aws.com
deletionPolicy: Retain
---
apiVersion: snapshot.storage.k8s.io/v1
kind: VolumeSnapshot
metadata:
  name: db-0-2026-11-13
  namespace: data
spec:
  volumeSnapshotClassName: ebs-snapshots
  source:
    persistentVolumeClaimName: data-db-0
---
apiVersion: v1
kind: PersistentVolumeClaim
metadata:
  name: data-db-restore-0
  namespace: data
spec:
  accessModes: ["ReadWriteOnce"]
  storageClassName: gp3
  resources:
    requests:
      storage: 100Gi
  dataSource:
    apiGroup: snapshot.storage.k8s.io
    kind: VolumeSnapshot
    name: db-0-2026-11-13

Keep going

Kubernetes quizTest yourself on Pods, Deployments, Services, scheduling, storage, networking, security and kubectl.Try it in the playgroundRecover a StatefulSet Pod from a dead node Try it in the playgroundRecover data from a Released PersistentVolume Try it in the playgroundDrain a node without breaking a PodDisruptionBudget ArticleHow to debug OOMKilled containers in Kubernetes ArticleCVE-2026-123158 (*sshole): the OpenSSH password bypass on Alpine ArticleHow to drain Kubernetes nodes safely with PodDisruptionBudgets