Local persistent volumes allows users to access local storage through the standard PVC interface in a simple and portable way. The PV contains node affinity information that the system uses to schedule pods to the correct nodes.

An external static provisioner and a related bootstrapper are available to help simplify local storage management once the local volumes are configured.

Current status: 1.7 - Alpha

What works:

• Create a PV specifying a directory with node affinity.
• Pod using the PVC that is bound to this PV will always get scheduled to that node.
• External static provisioner daemonset that discovers local directories, creates, cleans up and deletes PVs.

What doesn't work and workarounds:

• Multiple local PVCs in a single pod.
  • Goal for 1.8.
  • No known workarounds.
• PVC binding does not consider pod scheduling requirements and may make suboptimal or incorrect decisions.
  • Goal for 1.8.
  • Workarounds:
    • Run your pods that require local storage first.
    • Give your pods high priority.
    • Run a workaround controller that unbinds PVCs for pods that are stuck pending. TODO: add link
• External provisioner cannot correctly detect capacity of mounts added after it has been started.
  • This requires mount propagation to work, which is targeted for 1.8.
  • Workaround: Before adding any new mount points, stop the daemonset, add the new mount points, start the daemonset.
• Fsgroup conflict if multiple pods using the same PVC specify different fsgroup
  • Workaround: Don't do this!

Future features:

• Local block devices as a volume source, with partitioning and fs formatting
• Pod accessing local raw block device
• Local PV health monitoring, taints and tolerations
• Inline PV (use dedicated local disk as ephemeral storage)
• Dynamic provisioning for shared local persistent storage

KUBE_FEATURE_GATES="PersistentLocalVolumes=true" NODE_LOCAL_SSDS=<n> cluster/kube-up.sh

gcloud alpha container cluster create ... --local-ssd-count=<n>
gcloud alpha container node-pools create ... --local-ssd-count=<n>

1. Partition and format the disks on each node according to your application's requirements.
2. Mount all the filesystems under one directory per StorageClass. The directories are specified in a configmap, see below. By default, the discovery directory is /mnt/disks and storage class is local-storage.
3. Configure the Kubernetes API Server with the PersistentLocalVolumes feature gate.

1. Create /mnt/disks directory and mount several volumes into its subdirectories. The example below uses three ram disks to simulate real local volumes:

$ mkdir /mnt/disks
$ for vol in vol1 vol2 vol3; do
    mkdir /mnt/disks/$vol
    mount -t tmpfs $vol /mnt/disks/$vol
done

2. Run the local cluster.

$ ALLOW_PRIVILEGED=true LOG_LEVEL=5 FEATURE_GATES=PersistentLocalVolumes=true hack/local-up-cluster.sh

3. Continue with Creating local persistent volumes below.

This is optional, only for automated creation and cleanup of local volumes. See bootstrapper/ and provisioner/ for details and sample configuration files.

1. Create an admin account with cluster admin priviledge:

$ kubectl create -f bootstrapper/deployment/kubernetes/admin-account.yaml

2. Create a ConfigMap with your local storage configuration details:

$ kubectl create -f bootstrapper/deployment/kubernetes/example-config.yaml

3. Launch the bootstrapper, which in turn creates static provisioner daemonset:

$ kubectl create -f bootstrapper/deployment/kubernetes/bootstrapper.yaml

If you don't use the external provisioner, then you have to create the local PVs manually. Example PV:

apiVersion: v1
kind: PersistentVolume
metadata:
  name: example-local-pv
  annotations:
    "volume.alpha.kubernetes.io/node-affinity": '{
      "requiredDuringSchedulingIgnoredDuringExecution": {
        "nodeSelectorTerms": [
          { "matchExpressions": [
            { "key": "kubernetes.io/hostname",
              "operator": "In",
              "values": ["my-node"]
            }
          ]}
         ]}
        }'
spec:
  capacity:
    storage: 5Gi
  accessModes:
  - ReadWriteOnce
  persistentVolumeReclaimPolicy: Delete
  storageClassName: local-storage
  local:
    path: /mnt/disks/ssd1

In the PVC, specify the StorageClass of your local PVs.

kind: PersistentVolumeClaim
apiVersion: v1
metadata:
  name: example-local-claim
spec:
  accessModes:
  - ReadWriteOnce
  resources:
    requests:
      storage: 5Gi
  storageClassName: local-storage

go run hack/e2e.go -- -v --test --test_args="--ginkgo.focus=\[Feature:LocalPersistentVolumes\]"

• For IO isolation, a whole disk per volume is recommended
• For capacity isolation, separate partitions per volume is recommended
• Avoid recreating nodes with the same node name while there are still old PVs with that node's affinity specified. Otherwise, the system could think that the new node contains the old PVs.

When you want to decommission the local volume, here is a possible workflow.

1. Stop the pods that are using the volume
2. Remove the local volume from the node (ie unmounting, pulling out the disk, etc)
3. Delete the PVC
4. The provisioner will try to cleanup the volume, but will fail since the volume no longer exists
5. Manually delete the PV object