Storage
Keep files beyond a container restart with emptyDir, and beyond a Pod with PersistentVolumeClaims.
Why volumes exist
Files a container writes go to a temporary layer tied to that container. If the Pod crashes, is deleted, or is updated, that filesystem is wiped, and the replacement starts fresh.
Attach an external Volume and have the container write there. Storage outside the container survives container replacement.
| Volume type | Lifetime | Best for | If the Pod is deleted |
|---|---|---|---|
emptyDir |
As long as the Pod | Scratch space, caches, and sharing files between containers in the same Pod | Data is erased |
| PersistentVolumeClaim (PVC) | Independent of the Pod (cloud disk, NFS, SAN, and similar storage) | Databases and uploaded files | Data is kept and reattaches to the new Pod |
emptyDir
An emptyDir is a temporary scratch folder created on the node when the Pod starts. It starts empty.
- Use it as a temporary workspace or cache, or to share files between containers in the same Pod. An app and a logging sidecar can mount the same
emptyDir. - A container restart keeps the data. Deleting or evicting the Pod erases it.
apiVersion: v1
kind: Pod
metadata:
name: scratch-pod
spec:
containers:
- name: app
image: busybox
command: ["sleep", "3600"]
volumeMounts:
- name: cache # mount the volume at /data
mountPath: /data
volumes:
- name: cache # create an emptyDir as the source
emptyDir: {} # {} means default settings
PersistentVolume and PersistentVolumeClaim
Storage is split into two objects, like a locker and a ticket. The Pod mounts a PersistentVolumeClaim. The claim binds to a PersistentVolume, which is the actual storage. A StorageClass can create that volume on demand.
| Object | Represents | Analogy | Lifetime |
|---|---|---|---|
| PV | The actual disk, share, or folder (AWS EBS, NFS, or a local disk) | The real locker | Outlives Pods |
| PVC | A request such as “I need 100 Mi of storage” | A storage ticket | Managed by the developer or Pod |
| StorageClass | An automated PV generator | A storage clerk | Cluster-wide infrastructure |
Your Pod references the PVC, not the PV. That decouples the app from the underlying hardware.
Dynamic provisioning: you create a PVC, for example 100Mi. A StorageClass notices, the cluster creates a matching PV, and the two are linked as Bound.
Beginner trap: a PVC stuck in
Pendingis common. On many clusters, including kind, the default StorageClass usesWaitForFirstConsumer. The PV is not created until a Pod actually uses the PVC, so Kubernetes knows which node to place it on. The PVC staysPendinguntil a Pod attaches, then flips toBound.
apiVersion: v1
kind: Pod
metadata:
name: pvc-pod
spec:
containers:
- name: app
image: busybox
command: ["sleep", "3600"]
volumeMounts:
- name: store
mountPath: /data
volumes:
- name: store
persistentVolumeClaim:
claimName: data-pvc
The data-pvc claim itself is a small manifest with accessModes and resources.requests.storage, the same shape as the volumeClaimTemplates entry in a StatefulSet.