
Introduction
Stateful workloads on AKS need fast rollback points before upgrades, schema changes, and risky deploys. VolumeSnapshots deliver that. But every snapshot, restored PVC, and abandoned disk keeps billing until someone deletes it.
Teams often find the second problem months later: dozens of forgotten snapshots, restored disks nobody owns, and retained backend copies that no kubectl get command shows.
This guide covers CSI-backed Azure Disk volumes on AKS. You'll create a VolumeSnapshot, restore it to a new PVC, validate the data, cut over safely, and then apply lifecycle controls to contain cost and clutter.
Key Takeaways
- A VolumeSnapshot is a declarative request that the Azure Disk CSI driver fulfills in Azure.
- Restoring always creates a new PVC; it never overwrites the original.
- A snapshot is a recovery point, not a full backup or application-consistent database copy.
- Sprawl cleanup must cover Kubernetes objects, Azure snapshots, restored disks, and orphaned volumes.
Kubernetes Volume Snapshots on AKS: Architecture, Scope, and Prerequisites
What AKS adds to Kubernetes
Kubernetes is the open-source orchestration platform. AKS is Microsoft's managed service: Azure creates, configures, and maintains the control plane.
AKS documentation covers managed CSI drivers and the snapshot-controller enable/disable flags.
Check the current docs before installing snapshot CRDs or controllers yourself. Microsoft says to uninstall an existing open-source CSI driver before enabling the managed Azure one, and a manual install can conflict with managed components.
Snapshot resource model
Each object has a counterpart in the volume model:
| Snapshot object | Volume equivalent | Role |
|---|---|---|
VolumeSnapshot |
PVC | The user's request |
VolumeSnapshotContent |
PV | The provisioned snapshot |
VolumeSnapshotClass |
StorageClass | Driver and parameters |
The API group is snapshot.storage.k8s.io/v1, and the driver is disk.csi.azure.com. A Pod cannot mount a VolumeSnapshot. You restore it into a new PVC whose dataSource points at the snapshot. The original PVC stays separate, so keep it until validation finishes.
Confirm storage and driver compatibility
Identify the source PVC, its bound PV, the StorageClass, and the driver before you snapshot:
kubectl get pvc -n <namespace>
kubectl get pv
kubectl get sc
kubectl describe pvc <pvc-name> -n <namespace>
Preflight checklist:
- The source PVC shows
Bound. - The StorageClass provisioner is
disk.csi.azure.com. - The snapshot class uses the same driver.
- Your AKS version is supported (
az aks get-versions --location <region> --output table). - The access mode is
ReadWriteOnce, as required by the Azure Disk CSI driver.
Keep these concerns separate. Upstream driver compatibility, AKS release availability, snapshot placement, and PVC restoration each need their own answer. Cross-region snapshot creation exists in newer driver versions, but a full direct-restore path across regions and subscriptions is not always documented. Test your restore path before production use.
Define consistency and recovery expectations
A plain storage snapshot is crash-consistent. It captures what is on disk at that moment, much like a power loss. Microsoft's documentation recommends ensuring the application is not writing to the source disk first.
For databases, use a coordinated mechanism:
- AKS Backup documents custom hooks that freeze a MySQL Pod, snapshot the volume, then unfreeze it.
- SQL Server's T-SQL snapshot backup suspends writes and records backup metadata around the storage snapshot.
- Native dumps, log shipping, or an operator are also valid options.
Prepare access, naming, and retention controls
The AKS control-plane identity manages Azure Disk CSI and holds Contributor on the node resource group. The driver needs actions such as Microsoft.Compute/snapshots/read, write, and delete, plus matching disk actions. With customer-managed keys, the cluster identity also needs read access to the disk encryption set.
Before taking production snapshots, decide on:
- A naming convention, such as
<app>-<env>-<purpose>-<yyyymmdd> - Owner, environment, and expiry labels
- Recovery objectives
- A starting deletion policy
Create and Restore an AKS Volume Snapshot Safely
Define the VolumeSnapshotClass
This example comes from the Azure Disk CSI repository:
apiVersion: snapshot.storage.k8s.io/v1
kind: VolumeSnapshotClass
metadata:
name: csi-azuredisk-vsc
driver: disk.csi.azure.com
deletionPolicy: Delete
parameters:
incremental: "true"
incremental defaults to true. Note that subscriptionID requires resourceGroup and incremental: "false". Verify parameters against current docs before reusing this class everywhere.
The deletion policy decides what survives:
Delete: removing the VolumeSnapshot also removes the backend snapshot and content object.Retain: the backend snapshot andVolumeSnapshotContentremain. Safer for critical points, but a prime source of sprawl.
Create and verify the snapshot
apiVersion: snapshot.storage.k8s.io/v1
kind: VolumeSnapshot
metadata:
name: azuredisk-volume-snapshot
spec:
volumeSnapshotClassName: csi-azuredisk-vsc
source:
persistentVolumeClaimName: pvc-azuredisk
kubectl apply -f snapshot.yaml
kubectl get volumesnapshot
kubectl describe volumesnapshot azuredisk-volume-snapshot
A healthy result shows Ready To Use: true, a bound content object, creation metadata, and no warning events. Kubernetes also postpones deleting the source PVC until snapshot creation completes or aborts.
Restore the snapshot to a new PVC
apiVersion: v1
kind: PersistentVolumeClaim
metadata:
name: pvc-azuredisk-snapshot-restored
spec:
accessModes:
- ReadWriteOnce
storageClassName: managed-csi
resources:
requests:
storage: 10Gi
dataSource:
name: azuredisk-volume-snapshot
kind: VolumeSnapshot
apiGroup: snapshot.storage.k8s.io
Create it in the same namespace as the snapshot. Choose a compatible StorageClass, and request at least the snapshot's restore size. The 10Gi value is only the example, not a verified universal minimum.
Validate the restored data before cutover
A PVC reaching Bound proves nothing about your application. Mount it in an isolated validation Pod, then check:
- Expected files and directory structure
- Checksums on known datasets
- Database startup and schema validation
- Transaction-log replay
- A write-then-read test
Cut over deliberately and preserve rollback options
- Scale down or pause the original workload.
- Point the Deployment at the restored PVC.
- Confirm health and run application checks.
- Keep the original PVC and snapshot until the rollback window closes.
StatefulSet warning: each Pod gets a PVC per
volumeClaimTemplatesentry, and a rescheduled Pod reuses its associated claim. A separately named restored PVC is not a documented in-place replacement. Plan a deliberate scale-down, claim migration, or workload recreation.

One limit to remember: these snapshots are Azure-backend artifacts and not portable backups. For cross-cluster, cross-region, or cross-subscription needs, evaluate AKS Backup, Velero, replication, or database-native backups.
Cut Snapshot and Volume Sprawl in AKS
Map every resource that can become orphaned
Sprawl lives in several places at once:
VolumeSnapshotandVolumeSnapshotContentobjects- Azure backend snapshots
- Restored managed disks
- PVCs and PVs
- Unattached disks
Deleting one Kubernetes object doesn't guarantee every backend resource disappears, especially under Retain. Reconcile kubectl inventory against Azure.
az snapshot list inventories snapshots, and Microsoft's FinOps storage guidance shows Resource Graph queries for disks with an empty managedBy and snapshots older than 30 days. That 30 days is an example threshold, not an Azure default.
Create an ownership and retention policy
Label or tag every snapshot with:
- Application and environment
- Owner and purpose
- Creation date and expiry date
- Change ticket and recovery tier
The driver supports snapshot tags, and Azure Policy can require or inherit tags, though it won't expire anything on its own.
Separate two classes:
- Short-lived pre-upgrade points, with expiry in days
- Longer-lived compliance or DR copies, with named approvers for extensions and legal holds
Automate cleanup with safety checks
Build a scheduled job or GitOps workflow that:
- Lists snapshots past their expiry label.
- Skips the newest known-good recovery point and any active restore source.
- Runs in dry-run mode first.
- Deletes, then records the result.
Plan for failure too. Objects can sit in Terminating until finalizers clear. Hunt for orphaned VolumeSnapshotContent objects, and check Azure for backend snapshots that outlived their Kubernetes resources.
Measure storage growth and cost attribution
Per Microsoft's billing documentation, the first incremental snapshot holds all used data, and later ones hold only changes. Charges follow used data, not provisioned size. Incremental snapshots bill on Standard HDD storage regardless of source disk type.
A restored disk is separate. It bills by provisioned capacity, so a tiny snapshot delta doesn't mean a tiny restored-disk bill.
Track these signals:
- Snapshot age and growth
- Restore duration
- Failed deletions
- Unattached disks
- Spend by application or environment, using Cost Management tag grouping

Azure Advisor's preview unattached-disk recommendation helps too, but it warns that deletion is irreversible.
Use storage visibility to find adjacent waste
Snapshot policies don't catch idle disks around your cluster. Lucidity is a cloud block storage management layer with visibility across Azure and other clouds. Its Lumen product surfaces unattached, reserved, unmounted, and zero-I/O disks, with evidence-backed cleanup recommendations.
Keep the boundary clear. Lucidity isn't a backup service, so snapshot retention and application-aware backup policy stay governed by your own rules.
Production Checklist: Monitoring, Testing, and Failure Boundaries
Monitor:
readyToUsestatus, with an alert when a snapshot stays unready past your threshold- Snapshot creation latency and warning events
- CSI controller logs (the
azurediskcontainer inkube-system) - Azure Activity Log, which keeps 90 days by default
- PVC provisioning failures, restore duration, and cleanup failures
Test: run scheduled restore drills in an isolated namespace or test cluster. Record RPO and RTO, run application-level validation, confirm encryption-key access, and prove you can clean up afterward. AKS storage best practices call for regular restore testing.
Know the boundary: use VolumeSnapshots for fast rollback and cloning. For cluster deletion, account compromise, region failure, or compliance retention, add application-native backups, off-cluster copies, replication, or AKS Backup. Only its Vault tier supports cross-region restore, with documented RPO up to 24 hours in the primary region.

Frequently Asked Questions
What is the purpose of a snapshot?
A snapshot captures point-in-time storage state for rollback, recovery, or cloning. It may not be application-consistent, and it isn't a standalone backup.
What is AKS vs Kubernetes?
Kubernetes is the open-source orchestration platform. AKS is Microsoft's managed service where Azure runs the control plane, so you manage workloads, storage classes, and snapshot policy.
What is the difference between an Azure restore point and a snapshot?
A VM restore point captures VM configuration plus disk restore points. A managed disk snapshot is an Azure disk artifact, and a VolumeSnapshot is a Kubernetes request bound to one. They aren't interchangeable.
How do I restore a PVC from a VolumeSnapshot on AKS?
Create a new PVC with dataSource.kind: VolumeSnapshot and a compatible StorageClass. Wait for it to bind, validate the data in an isolated Pod, then cut over deliberately.
Are Kubernetes volume snapshots a complete backup?
No. They sit in the same Azure backend and failure domain, and plain snapshots aren't application-consistent. Add independent backups or replication for broader disaster recovery.
How do I prevent unused snapshots and disks from increasing AKS costs?
Apply ownership labels and expiry dates, choose Delete or Retain deliberately, and reconcile Azure inventory regularly. Monitor spend in Cost Management and alert on unattached or idle disks.