This is temporary workaround for Kubernetes local volume management until this feature will be implemented.
This provisioner requires custom kube-scheduler to be running in your cluster. You should edit your scheduler manifest and change container image there:
# ...
containers:
- image: image: nailgun/kube-scheduler-amd64:v1.6.2_hostpath.0
command:
- kube-scheduler
- # your args here
# ...- List of available image tags
- Sources for all image tags (based on upstream with one simple commit added)
After you have running custom scheduler, install the provisioner:
$ cd example
$ kubectl create -f sa.yaml -f cluster-roles.yaml -f cluster-role-bindings.yaml -f ds.yaml
Label all nodes that can provision hostPath volumes:
$ kubectl label node --all nailgun.name/hostpath=enabled
Now you can test it. You should annotate your nodes with information of available storage:
$ kubectl annotate node NODE_NAME hostpath.nailgun.name/ssd=/tmp/ssd
Then create test StorageClass, PVC and Pod:
$ kubectl create -f sc.yaml -f pvc.yaml -f test-pod.yaml
The test-pod will be always scheduled on the same node where the PVC has been provisioned.
Note StorageClass has hostPathName parameter equal to ssd which means that provisioner will search for a node having
hostpath.nailgun.name/ssd annotation and it will use hostPath specified in the annotation. If you have more than one node
with this annotation, node will be selected randomly.
You can also explicity specify node for PVC by adding nailgun.name/hostpath-node annotation to the PVC with the node name.
But the node should also match StorageClass.
This provisioner doesn't support any kind of storage isolation nor quotas. This is temporary solution to have working PVCs with local storage.
If a pod uses more then one PVC, they can be scheduled to different nodes and Pod will be failed to be scheduled.
Thanks to @TamalSaha for the idea and first implementation.