Skip to content
This repository has been archived by the owner on Dec 22, 2018. It is now read-only.

Latest commit

 

History

History

demo

README.md

Demo

The beta dynamic provisioning feature allows administrators to define StorageClasses to enable Kubernetes to create PersistentVolumes on-demand. Kubernetes includes many provisioners to specify in StorageClasses definitions and now, with Kubernetes 1.5, also includes support for external or out-of-tree provisioners like nfs-provisioner.

nfs-provisioner creates NFS-backed PV's, leveraging the NFS volume plugin of Kubernetes, so given the ubiquity of NFS it will work almost anywhere. It's ideal for local clusters and dev work, any place a PV is wanted but not the manual work of creating one. We'll demonstrate how to get it quickly up and running, following a variation of the Kubernetes repo's NFS example.

The recommended way to run nfs-provisioner, which we'll demonstrate here, is as a single-instance stateful app, where we create a Deployment and back it with some persistent storage like a hostPath volume. We always create it in tandem with a matching service that has the necessary ports exposed. We'll see that when it's setup like so, the NFS server it runs to serve its PV's can maintain state and so survive pod restarts. The other ways to run are as a DaemonSet, standalone Docker container, or standalone binary, all documented here.

There are two main things one can customize here before creating the deployment: the provisioner name and the backing volume.

The provisioner name must follow the naming scheme <vendor name>/<provisioner name>, like for example kubernetes.io/gce-pd. It's specified here in the args field. This is the provisioner a StorageClass will specify later. We'll use the name example.com/nfs-tmp.

...
args:
  - "-provisioner=example.com/nfs"
...

The backing volume is the place mounted at /export where the nfs-provisioner instance stores its state and the data of every PV it provisions. So we can mount any volume there to specify that volume as the backing storage for provisioned PV's. We'll use a hostPath volume at /tmp/nfs-provisioner, so we need to make sure that the directory exists on the nodes our deployment's pod could be scheduled to and, if selinux is enforcing, that it is labelled appropriately.

...
  volumeMounts:
    - name: export-volume
      mountPath: /export
volumes:
  - name: export-volume
    hostPath:
      path: /tmp/nfs-provisioner
...
$ mkdir -p /tmp/nfs-provisioner
$ sudo chcon -Rt svirt_sandbox_file_t /tmp/nfs-provisioner

We create the deployment and its service.

$ kubectl create -f demo/deployment.yaml
service "nfs-provisioner" created
deployment "nfs-provisioner" created

Now, our instance of nfs-provisioner can be treated like any other provisioner: we specify its name in a StorageClass object and the provisioner will automatically create PersistentVolumes for PersistentVolumeClaims that ask for the StorageClass. We'll show all that.

We create a StorageClass that specifies our provisioner.

$ kubectl create -f demo/class.yaml
storageclass "example-nfs" created

We create a PersistentVolumeClaim asking for our StorageClass.

$ kubectl create -f demo/claim.yaml
persistentvolumeclaim "nfs" created

And a PersistentVolume is provisioned automatically and already bound to our claim. We didn't have to manually figure out the NFS server's IP, put that IP into a PV yaml, then create the yaml. We just had to deploy our nfs-provisioner and create a StorageClass for it, which are one-time steps.

$ kubectl get pv
NAME                                       CAPACITY   ACCESSMODES   RECLAIMPOLICY   STATUS      CLAIM         REASON    AGE
pvc-dce84888-7a9d-11e6-b1ee-5254001e0c1b   1Mi        RWX           Delete          Bound       default/nfs

If you don't see a PV bound to your PVC, check the deployment's provisioner pod's logs using kubectl logs and look for events in the PVC using kubectl describe.

Now we have an NFS-backed PVC & PV pair that is exactly like what is expected by the official Kubernetes NFS example, so we'll finish the example to show our storage works, can be shared, and persists. If you don't need that proof, you can skip ahead to the part where we discuss deleting and cleaning up the provisioner and its storage.

We setup the fake backend that updates index.html on the NFS server every 10 seconds. And check that our mounts are working.

$ kubectl create -f demo/nfs-busybox-rc.yaml
$ kubectl get pod -l name=nfs-busybox
NAME                READY     STATUS    RESTARTS   AGE
nfs-busybox-h782l   1/1       Running   0          13m
nfs-busybox-nul47   1/1       Running   0          13m
$ kubectl exec nfs-busybox-h782l -- cat /mnt/index.html
Mon Dec 19 18:10:09 UTC 2016
nfs-busybox-h782l

We setup the web server that reads from the NFS share and runs a simple web server on it. And check that nginx is serving the data, the index.html from above, appropriately.

$ kubectl create -f demo/nfs-web-rc.yaml
$ kubectl create -f demo/nfs-web-service.yaml
$ kubectl get pod -l name=nfs-busybox
NAME                READY     STATUS    RESTARTS   AGE
nfs-busybox-h782l   1/1       Running   0          13m
nfs-busybox-nul47   1/1       Running   0          13m
$ kubectl get services nfs-web
NAME      CLUSTER-IP   EXTERNAL-IP   PORT(S)   AGE
nfs-web   10.0.0.187   <none>        80/TCP    7s
$ kubectl exec nfs-busybox-h782l -- wget -qO- http://10.0.0.187
Mon Dec 19 18:11:51 UTC 2016
nfs-busybox-nul47

We see that the PV created by our nfs-provisioner works, let's now show that it will continue to work even after our nfs-provisioner pod restarts. Because of how NFS works, anything that has shares mounted will hang as it tries to access or unmount them while the NFS server is down. Recall that all our nfs-provisioner instance's state and data persists in the volume we mounted at /export, so it should recover and its shares become accessible again when it, and the NFS server it runs, restarts. We'll simulate this situation.

We scale the deployment down to 0 replicas.

$ kubectl scale --replicas=0 deployment/nfs-provisioner
deployment "nfs-provisioner" scaled

We try the same check from before that nginx is serving the data, and we see it hangs indefinitely as it tries to read the share.

$ kubectl exec nfs-busybox-h782l -- wget -qO- http://10.0.0.187
...
^C

We scale the deployment back up to 1 replica.

$ kubectl scale --replicas=1 deployment/nfs-provisioner
deployment "nfs-provisioner" scaled

And after a brief delay all should be working again.

$ kubectl exec nfs-busybox-h782l -- wget -qO- http://10.0.0.187
Mon Dec 19 18:21:49 UTC 2016
nfs-busybox-nul47

Now we'll show how to delete the storage provisioned by our nfs-provisioner once we're done with it. Let's first delete the fake backend and web server that are using the PVC.

$ kubectl delete rc nfs-busybox nfs-web
replicationcontroller "nfs-busybox" deleted
replicationcontroller "nfs-web" deleted
$ kubectl delete service nfs-web
service "nfs-web" deleted

Once all those pods have disappeared and so we are confident they have unmounted the NFS share, we can safely delete the PVC. The provisioned PV the PVC is bound to has the ReclaimPolicy Delete, so when we delete the PVC, the PV and its data will be automatically deleted by our nfs-provisioner.

$ kubectl delete pvc nfs
persistentvolumeclaim "nfs" deleted
$ kubectl get pv

Note that deleting an nfs-provisioner instance won't delete the PV's it created, so before we do so we need to make sure none still exist as they would be useless for as long as the provisioner is gone.

$ kubectl delete deployment nfs-provisioner
deployment "nfs-provisioner" deleted
$ kubectl delete service nfs-provisioner
service "nfs-provisioner" deleted

Thanks for following along. If at any point things didn't work correctly, check the provisioner pod's logs using kubectl logs and look for events in the PV's and PVC's using kubectl describe. If you are interested in Kubernetes storage-related things like this, head to the Storage SIG. If you are interested in writing your own external provisioner, all the code is available for you to read or fork, and better documentation on how to do it is in the works.