StatefulSet Storage Patterns

Identity-Linked Storage

StatefulSet members need disks that survive rescheduling: mysql-0 must find its data wherever it lands next. volumeClaimTemplates deliver this by generating one PVC per ordinal:

volumeClaimTemplates:
- metadata: { name: data }
  spec:
    accessModes: [ReadWriteOnce]
    storageClassName: fast-ssd
    resources: { requests: { storage: 50Gi } }

The generated claims follow data-<pod-name> — data-mysql-0, data-mysql-1 — and are never deleted with the set by default. On reschedule, kubelet reuses the ordinal's existing claim: identity binds storage, so the disk follows the member rather than the node.

Scaling out creates the next claim alongside the next pod; scaling in keeps data for a possible scale-back.

Quorum-Safe Patterns

Clustered stores need coordination beyond disks:

• Ordered updates plus a PDB — restart one member at a time; a PodDisruptionBudget of minAvailable: 2 on a three-member set stops drains from taking quorum.
• Honest readiness — a member that joined but cannot vote must fail readiness, or the controller rolls through the fleet too fast.
• DNS-based formation — first pod bootstraps, later pods join via headless names like mysql-0.mysql.

# companion PodDisruptionBudget
minAvailable: 2

Write-ahead logs or index data can live in a second claim template (wal-<pod>), giving two volumes under one identity. Use podManagementPolicy: Parallel for faster initial creation while keeping ordered updates.

Pitfalls and Interview Angles

Pitfall 1 — careless delete and recreate. PVCs outlive the StatefulSet by design; recreating with the same name reattaches old volumes with old data, schema migrations included. Know your cleanup procedure before deploying.

Pitfall 2 — stuck volume attachments after node loss. An RWO volume attached to a dead node must detach before the pod can start elsewhere; multi-attach errors in describe mean the CSI driver's detach has not completed.

Interview-ready: explain how volumeClaimTemplates link identity to storage, walk a safe scale-out and scale-in, and describe quorum protection as the trio of PodDisruptionBudget, truthful readiness gates, and ordered updates.