
The numbers are not flattering. Lucidity's State of Cloud Block Storage research puts average disk utilization at 30%, and its assessments across 600+ environments and 100+ PB of storage found the same average. In practice, many organizations pay for roughly three times the block storage they use.
This guide is for Kubernetes administrators, DevOps, ITOps, and FinOps teams running stateful workloads on cloud block storage, shared file storage, or multi-cloud clusters. It explains how StorageClasses shape cost and performance, then walks through evaluating, configuring, and governing them.
Key Takeaways
- A StorageClass is a provisioning policy, not a latency or IOPS guarantee.
- The cheapest class can cost more once slowdowns, overprovisioning, and retained volumes count.
- Match each workload to driver, volume type, access mode, topology, and reclaim policy.
- Name classes explicitly in PVCs, then monitor real utilization after launch.
What Is a Kubernetes StorageClass?
A StorageClass is a Kubernetes API resource that describes a category of dynamically provisioned storage. It names a provisioner, backend parameters, lifecycle policy, and binding behavior.
How the objects fit together
- StorageClass: defines how storage gets created.
- PersistentVolume (PV): the provisioned storage resource.
- PersistentVolumeClaim (PVC): a workload's request for storage.
- Pod: consumes the claim.
How dynamic provisioning works
- A PVC is created.
- The CSI driver named in the class provisions a volume.
- Kubernetes creates the PV and binds it to the PVC.
- The volume attaches to a node and mounts into the Pod.

The PVC's storageClassName picks the class. Omit it and the cluster's default class applies. Setting storageClassName: "" opts out of defaults. Leaning on an implicit default is how a team ends up with premium disks for a scratch workload.
Fields that matter
The Kubernetes documentation lists the full spec. These fields drive cost and risk:
| Field | Why it matters |
|---|---|
provisioner / parameters |
Driver-specific; names and values change by CSI driver and provider, so check current docs |
volumeBindingMode |
Immediate is the default; WaitForFirstConsumer waits for a Pod so node, zone, and affinity constraints are honored |
reclaimPolicy |
Defaults to Delete; Retain keeps the backing disk, and its bill |
allowVolumeExpansion |
Allows growing a PVC, never shrinking it |
mountOptions |
Not validated at creation; bad options cause mount failures |
Kubernetes validation rejects updates to parameters, provisioner, reclaimPolicy, and volumeBindingMode. Changing those means a new class and a migration plan. Make reclaimPolicy and volumeBindingMode explicit in every reviewed manifest rather than trusting defaults.
How StorageClass Choices Affect Cost and Performance
The performance dimensions
Latency, IOPS, throughput, burst behavior, durability, and availability matter differently by workload. A transactional database cares about tail latency. A log pipeline cares about throughput per dollar.
The class name guarantees none of this. Observed performance also depends on the CSI driver, volume type, provisioned settings, node placement, network path, filesystem, caching, and access pattern.
Published specs show the range. EBS gp3 includes 3,000 IOPS and 125 MiB/s, scaling to 80,000 IOPS and 2,000 MiB/s, according to AWS EBS documentation.
Azure Premium SSD v2 and Google Hyperdisk Balanced use a similar split between baseline and purchasable performance. These are provider limits, not benchmark results, and MB/s versus MiB/s differences are easy to miss.
The main cost drivers
Capacity, provisioned IOPS or throughput, snapshots, replication, and cross-zone transfer are separate line items. On Azure, Premium SSD v2 pricing in the East US example runs about $0.081/GiB, $0.0052 per provisioned IOPS above 3,000, and $0.041 per MB/s above 125, per Azure's pricing page.

Prices are estimates and vary by region, so check your own region on each provider's page.
Waste hides in three places:
- Oversized PVCs: you pay for provisioned space, not consumed space
- Idle volumes: unattached disks still bill, and
Retaindoes not stop the charge - Stale snapshots: incremental snapshots may keep billing while data remains referenced
Match tiers to workloads
| Workload | What to prioritize |
|---|---|
| Transactional databases | Low tail latency, provisioned IOPS, zone-aware binding |
| Latency-sensitive services | Consistent latency, premium SSD tiers |
| General application data | Balanced SSD, expansion enabled |
| Dev and test | Cheapest tier, Delete reclaim |
| Logs and backups | Throughput and capacity over IOPS |
| Shared files | RWX file storage such as EFS, Azure Files, or Filestore |
Block storage suits most single-writer stateful apps. Network file storage handles shared access. Local PVs give speed but pin Pods to a node through affinity. Object storage, such as the S3 Mountpoint CSI driver, is static-only and lacks full POSIX behavior, so it is not a drop-in replacement.
Balance price and risk
Cheap storage that causes timeouts, missed SLOs, or extra replicas is not cheap. Premium storage on every PVC is just as wasteful when most workloads barely touch the disk. Lucidity's research finds that 8-10% of disks are linked to downtime, the cost of getting this wrong in either direction.
Validate before committing
- Benchmark with tools like Kubestr, which CNCF describes as running fio against provisioned PVCs.
- Measure latency, IOPS, throughput, saturation, tail behavior, and failure recovery under your real access pattern.
- Review cost monthly using PVC metadata, billing data, utilization metrics, and owner feedback.
- Set thresholds yourself. Your SLOs and budget define "good enough."
How to Choose and Configure a StorageClass
A selection workflow
Gather these for each workload before picking anything:
- Capacity and growth expectations
- Access mode (RWO, RWX, RWOP)
- Latency sensitivity and IOPS/throughput profile
- Zone behavior, encryption, backup, and retention needs
- Budget

Then map requirements to a current CSI driver and volume type. Name the class for its purpose, such as db-gp3-retain, rather than "standard" or "fast," and document why it fits.
Example StorageClass
This is based on the official EKS Auto Mode example. Swap provisioner and parameters for your platform.
apiVersion: storage.k8s.io/v1
kind: StorageClass
metadata:
name: app-gp3-encrypted
provisioner: ebs.csi.eks.amazonaws.com # use ebs.csi.aws.com on standard EKS
volumeBindingMode: WaitForFirstConsumer
reclaimPolicy: Delete
allowVolumeExpansion: true
parameters:
type: gp3
encrypted: "true"
metadata.name is what PVCs reference. WaitForFirstConsumer delays provisioning until the scheduler picks a node. allowVolumeExpansion permits growth later.
Warning: don't copy deprecated in-tree provisioners such as awsElasticBlockStore, azureDisk, or gcePersistentDisk into production, and verify every parameter against the current Kubernetes and provider docs.
Create and consume a PVC
apiVersion: v1
kind: PersistentVolumeClaim
metadata:
name: orders-data
spec:
storageClassName: app-gp3-encrypted
accessModes: ["ReadWriteOnce"]
resources:
requests:
storage: 100Gi
A Deployment, StatefulSet, or Pod references it under spec.volumes[].persistentVolumeClaim.claimName, in the same namespace. The Pod never configures the PV directly. Storage choices live in the claim and class.
Set and verify the default class
kubectl get storageclass
kubectl patch storageclass standard -p '{"metadata":{"annotations":{"storageclass.kubernetes.io/is-default-class":"false"}}}'
kubectl patch storageclass app-gp3-encrypted -p '{"metadata":{"annotations":{"storageclass.kubernetes.io/is-default-class":"true"}}}'
kubectl get storageclass
Changing the default affects future PVCs that omit storageClassName. With multiple defaults, Kubernetes picks the most recently created. Use explicit names for anything with strict cost, performance, or retention needs.
Validate binding, topology, and cleanup
kubectl describe storageclassandkubectl get pvc,pvshow binding state.kubectl describe podsurfaces scheduling conflicts, attach failures, and mount errors.- Pending claims usually mean a missing driver, a zone mismatch, or an unsupported parameter.
In a non-production cluster, test expansion, reclaim behavior, snapshot and restore, and deletion before opening the class to everyone. A Released PV does not mean the underlying disk is gone.
StorageClass Optimization and Operational Best Practices
Monitor utilization, performance, and cost
Kubelet exposes kubelet_volume_stats_used_bytes, kubelet_volume_stats_capacity_bytes, and kubelet_volume_stats_available_bytes, labeled by namespace and PVC. Compare requested to consumed capacity, and track:

- Growth trends
- Latency and IOPS
- Filesystem fullness
- Snapshots and retained volumes
For FinOps, assign ownership through namespaces, labels, and environments. OpenCost can attach persistent-storage cost to workloads at the Pod level. Use those cost signals to ask each owner whether the workload still needs its tier.
Reduce waste without weakening reliability
- Create separate classes only for materially different requirements
- Size PVCs to realistic need and expand under control
- Document each reclaim policy
- Set snapshot lifecycles, since VolumeSnapshotClass
deletionPolicyis separate from PV reclaim - Audit idle and orphaned volumes regularly
Reserve Retain for data that needs deliberate recovery. For temporary or reproducible workloads, Delete is fine once you've tested it.
Govern multi-cloud storage
Across AWS, Azure, and Google Cloud, standardize what you can: naming, tagging, ownership, encryption, backup expectations, and test methods. Let provider-specific parameters differ. The FinOps Foundation's Allocation capability covers using tags, labels, and naming metadata across environments.

Lucidity can sit alongside StorageClasses as an optional intelligence layer for cloud block storage. Lumen scores disk tiers against IOPS, throughput, latency, and cost, and flags idle disks (unattached, reserved, unmounted, or zero-I/O).
AutoScaler expands and shrinks disks with zero downtime. It does not replace Kubernetes, your CSI driver, or application-level storage testing.
Conclusion
The right StorageClass is a workload decision. Match I/O, durability, and cost to each app instead of defaulting to a generic "standard" or "premium" label.
From there:
- Define requirements for the workload
- Pick a current CSI-backed class and reference it explicitly in PVCs
- Test with real workloads before broad rollout
- Watch utilization and spend, and review lifecycle policies on a schedule
Frequently Asked Questions
How do I set a StorageClass as the default?
Set storageclass.kubernetes.io/is-default-class: "true" on the class and patch the existing default to "false". Confirm with kubectl get storageclass, where the default shows (default).
What are storage classes?
They describe storage types and provisioning behavior for dynamically created PersistentVolumes, including provisioner, parameters, reclaim policy, and binding mode.
What are PV and PVC in Kubernetes?
A PersistentVolume is the provisioned storage resource. A PersistentVolumeClaim is an application's request for storage, bound one-to-one to a PV.
What is the difference between a storage class and a persistent volume?
The StorageClass defines how storage gets provisioned. The PV is the resulting storage resource bound to a claim.
Can Kubernetes use NFS storage?
Yes, through a CSI driver such as nfs.csi.k8s.io or an external provisioner. Both need an existing NFS server, so plan for access modes, permissions, network availability, and performance.
Does a Kubernetes volume have to be configured in a pod?
A Pod references a PVC in its volume configuration. The storage itself and its provisioning are defined separately through the PV, PVC, StorageClass, and CSI driver.