Ephemeral Volumes

This document describes ephemeral volumes in Kubernetes. Familiarity with volumes is suggested, in particular PersistentVolumeClaim and PersistentVolume.

Some application need additional storage but don't care whether that data is stored persistently across restarts. For example, caching services are often limited by memory size and can move infrequently used data into storage that is slower than memory with little impact on overall performance.

Other applications expect some read-only input data to be present in files, like configuration data or secret keys.

Ephemeral volumes are designed for these use cases. Because volumes follow the Pod's lifetime and get created and deleted along with the Pod, Pods can be stopped and restarted without being limited to where some persistent volume is available.

Ephemeral volumes are specified inline in the Pod spec, which simplifies application deployment and management.

Types of ephemeral volumes

Kubernetes supports several different kinds of ephemeral volumes for different purposes:

emptyDir: empty at Pod startup, with storage coming locally from the kubelet base directory (usually the root disk) or RAM
configMap, downwardAPI, secret: inject different kinds of Kubernetes data into a Pod
CSI ephemeral volumes: similar to the previous volume kinds, but provided by special CSI drivers which specifically support this feature
generic ephemeral volumes, which can be provided by all storage drivers that also support persistent volumes

emptyDir, configMap, downwardAPI, secret are provided as local ephemeral storage. They are managed by kubelet on each node.

CSI ephemeral volumes must be provided by third-party CSI storage drivers.

Generic ephemeral volumes can be provided by third-party CSI storage drivers, but also by any other storage driver that supports dynamic provisioning. Some CSI drivers are written specifically for CSI ephemeral volumes and do not support dynamic provisioning: those then cannot be used for generic ephemeral volumes.

The advantage of using third-party drivers is that they can offer functionality that Kubernetes itself does not support, for example storage with different performance characteristics than the disk that is managed by kubelet, or injecting different data.

CSI ephemeral volumes

FEATURE STATE: Kubernetes v1.25 [stable]

Note: CSI ephemeral volumes are only supported by a subset of CSI drivers. The Kubernetes CSI Drivers list shows which drivers support ephemeral volumes.

Conceptually, CSI ephemeral volumes are similar to configMap, downwardAPI and secret volume types: the storage is managed locally on each node and is created together with other local resources after a Pod has been scheduled onto a node. Kubernetes has no concept of rescheduling Pods anymore at this stage. Volume creation has to be unlikely to fail, otherwise Pod startup gets stuck. In particular, storage capacity aware Pod scheduling is not supported for these volumes. They are currently also not covered by the storage resource usage limits of a Pod, because that is something that kubelet can only enforce for storage that it manages itself.

Here's an example manifest for a Pod that uses CSI ephemeral storage:

kind: Pod
apiVersion: v1
metadata:
  name: my-csi-app
spec:
  containers:
    - name: my-frontend
      image: busybox:1.28
      volumeMounts:
      - mountPath: "/data"
        name: my-csi-inline-vol
      command: [ "sleep", "1000000" ]
  volumes:
    - name: my-csi-inline-vol
      csi:
        driver: inline.storage.kubernetes.io
        volumeAttributes:
          foo: bar

The volumeAttributes determine what volume is prepared by the driver. These attributes are specific to each driver and not standardized. See the documentation of each CSI driver for further instructions.

CSI driver restrictions

CSI ephemeral volumes allow users to provide volumeAttributes directly to the CSI driver as part of the Pod spec. A CSI driver allowing volumeAttributes that are typically restricted to administrators is NOT suitable for use in an inline ephemeral volume. For example, parameters that are normally defined in the StorageClass should not be exposed to users through the use of inline ephemeral volumes.

Cluster administrators who need to restrict the CSI drivers that are allowed to be used as inline volumes within a Pod spec may do so by:

Removing Ephemeral from volumeLifecycleModes in the CSIDriver spec, which prevents the driver from being used as an inline ephemeral volume.
Using an admission webhook to restrict how this driver is used.

Generic ephemeral volumes

FEATURE STATE: Kubernetes v1.23 [stable]

Generic ephemeral volumes are similar to emptyDir volumes in the sense that they provide a per-pod directory for scratch data that is usually empty after provisioning. But they may also have additional features:

Storage can be local or network-attached.
Volumes can have a fixed size that Pods are not able to exceed.
Volumes may have some initial data, depending on the driver and parameters.
Typical operations on volumes are supported assuming that the driver supports them, including snapshotting, cloning, resizing, and storage capacity tracking.

Example:

kind: Pod
apiVersion: v1
metadata:
  name: my-app
spec:
  containers:
    - name: my-frontend
      image: busybox:1.28
      volumeMounts:
      - mountPath: "/scratch"
        name: scratch-volume
      command: [ "sleep", "1000000" ]
  volumes:
    - name: scratch-volume
      ephemeral:
        volumeClaimTemplate:
          metadata:
            labels:
              type: my-frontend-volume
          spec:
            accessModes: [ "ReadWriteOnce" ]
            storageClassName: "scratch-storage-class"
            resources:
              requests:
                storage: 1Gi

Lifecycle and PersistentVolumeClaim

The key design idea is that the parameters for a volume claim are allowed inside a volume source of the Pod. Labels, annotations and the whole set of fields for a PersistentVolumeClaim are supported. When such a Pod gets created, the ephemeral volume controller then creates an actual PersistentVolumeClaim object in the same namespace as the Pod and ensures that the PersistentVolumeClaim gets deleted when the Pod gets deleted.

That triggers volume binding and/or provisioning, either immediately if the StorageClass uses immediate volume binding or when the Pod is tentatively scheduled onto a node (WaitForFirstConsumer volume binding mode). The latter is recommended for generic ephemeral volumes because then the scheduler is free to choose a suitable node for the Pod. With immediate binding, the scheduler is forced to select a node that has access to the volume once it is available.

In terms of resource ownership, a Pod that has generic ephemeral storage is the owner of the PersistentVolumeClaim(s) that provide that ephemeral storage. When the Pod is deleted, the Kubernetes garbage collector deletes the PVC, which then usually triggers deletion of the volume because the default reclaim policy of storage classes is to delete volumes. You can create quasi-ephemeral local storage using a StorageClass with a reclaim policy of retain: the storage outlives the Pod, and in this case you need to ensure that volume clean up happens separately.

While these PVCs exist, they can be used like any other PVC. In particular, they can be referenced as data source in volume cloning or snapshotting. The PVC object also holds the current status of the volume.

PersistentVolumeClaim naming

Naming of the automatically created PVCs is deterministic: the name is a combination of the Pod name and volume name, with a hyphen (-) in the middle. In the example above, the PVC name will be my-app-scratch-volume. This deterministic naming makes it easier to interact with the PVC because one does not have to search for it once the Pod name and volume name are known.

The deterministic naming also introduces a potential conflict between different Pods (a Pod "pod-a" with volume "scratch" and another Pod with name "pod" and volume "a-scratch" both end up with the same PVC name "pod-a-scratch") and between Pods and manually created PVCs.

Such conflicts are detected: a PVC is only used for an ephemeral volume if it was created for the Pod. This check is based on the ownership relationship. An existing PVC is not overwritten or modified. But this does not resolve the conflict because without the right PVC, the Pod cannot start.

Caution: Take care when naming Pods and volumes inside the same namespace, so that these conflicts can't occur.

Security

Using generic ephemeral volumes allows users to create PVCs indirectly if they can create Pods, even if they do not have permission to create PVCs directly. Cluster administrators must be aware of this. If this does not fit their security model, they should use an admission webhook that rejects objects like Pods that have a generic ephemeral volume.

The normal namespace quota for PVCs still applies, so even if users are allowed to use this new mechanism, they cannot use it to circumvent other policies.

What's next