Storage plugins follow a standard interface, and modern drivers add operational superpowers.

Like CRI and CNI, CSI lets any vendor plug in. The same PVC YAML works on Persistent Disk, EBS, Ceph or Longhorn.

Snapshot

A point-in-time copy of a PVC using the backend's native snapshot. The basis of backups and quick restores.

kind: VolumeSnapshot
spec:
  source: {persistentVolumeClaimName: data-pg-0}

Clone and restore

Create a new PVC pre-populated from a snapshot or another PVC. Spin up a copy of production data for debugging.

kind: PersistentVolumeClaim
spec:
  dataSource: {kind: VolumeSnapshot, name: snap1}

Expansion

Grow a PVC in place by editing its requested size. Needs allowVolumeExpansion: true on the class. You cannot shrink.

$ kubectl edit pvc data-pg-0
# spec.resources.requests.storage: 50Gi

StatefulSet ties it together

volumeClaimTemplates gives each replica its own PVC (data-pg-0, data-pg-1) that stays bound to that ordinal across restarts and rescheduling.

A snapshot is not an application backup

It captures disk blocks, which may be mid-write for a database. Use quiescing hooks or Velero with pre and post hooks (Deck 6), and keep copies off the cluster.

CSI adds snapshots, clones and online expansion

CSI is the third standard interface in the family. A CSI driver implements provisioning, attaching, mounting and optionally more. Because Kubernetes only speaks CSI, the same PVC manifest works on any backend whose driver is installed.

Three capabilities matter operationally. Snapshots take a point-in-time copy of a volume using the backend's native mechanism and are modelled by VolumeSnapshot objects. Cloning or restoring creates a new PVC whose dataSource is a snapshot or another PVC, which is excellent for reproducing production data in a test environment. Expansion lets you grow a claim by editing its requested size, provided the StorageClass allows it; depending on the driver and filesystem, the filesystem may grow online or after a Pod restart. Volumes cannot be shrunk.

A critical caution: a storage snapshot is crash-consistent at best. A database may be mid-transaction, so for application-consistent backups use database-aware hooks or tools such as Velero with pre and post hooks, and store backups outside the cluster and ideally outside the region. Deck 6 covers backup and disaster recovery.

StatefulSet volumeClaimTemplates is how all of this is applied to replicated stateful workloads: each ordinal has its own claim that stays attached to its identity.