TopolearnContainers lose their filesystem on restart; anything that must survive needs a volume.
A pod's volumes list makes storage available to its containers through volumeMounts. Some volume types are temporary: emptyDir lives as long as the pod (scratch space, sharing files between containers), configMap/secret project config as files.
For durable data, Kubernetes separates request from supply:
- A PersistentVolumeClaim (PVC, namespaced) asks for storage: size,
accessModes, optionalstorageClassName. - A PersistentVolume (PV, cluster-wide) is an actual piece of storage: a cloud disk, an NFS share.
- A StorageClass names a provisioner (the CSI driver for EBS, Persistent Disk, Azure Disk, Longhorn...) and its parameters. With dynamic provisioning, creating a PVC makes the driver create a disk and a PV, and binds them.
Access modes: ReadWriteOnce (one node at a time; typical block disks), ReadWriteOncePod (one pod), ReadOnlyMany, ReadWriteMany (shared filesystems like NFS or EFS).
reclaimPolicy decides what happens when the PVC is deleted: Delete (default for dynamic classes; the disk is gone) or Retain. volumeBindingMode: WaitForFirstConsumer delays creating the disk until a pod is scheduled, so it lands in the right zone.
PVCs can usually be expanded by increasing the size, if the class has allowVolumeExpansion: true; they can't shrink.
Example
apiVersion: v1
kind: PersistentVolumeClaim
metadata: {name: uploads}
spec:
accessModes: [ReadWriteOnce]
storageClassName: standard # kubectl get storageclass
resources:
requests: {storage: 10Gi}
---
apiVersion: v1
kind: Pod
metadata: {name: uploader}
spec:
containers:
- name: app
image: busybox:1.37
command: ["sh", "-c", "date >> /data/log; sleep 3600"]
volumeMounts: [{name: data, mountPath: /data}]
volumes:
- name: data
persistentVolumeClaim: {claimName: uploads}
# kubectl get pvc uploads # STATUS Bound, VOLUME pvc-3f1c...Remember
- PVC = request, PV = supply, StorageClass = how to provision
- ReadWriteOnce disks attach to one node at a time
- Default reclaimPolicy Delete removes the disk with the PVC
Watch out for
- Scaling a Deployment with one ReadWriteOnce PVC to several replicas on different nodes: the extra pods can't mount it.