PVC used by a job doesn't get resize after the pod of the job completed

[PhanLe1010]

Summary:
We have a setup in which the external-resizer is used with the storage provider that only supports offline expansion (e.g., only supports PluginCapability_VolumeExpansion_OFFLINE). We deployed a job that uses a PVC provisioned by the storage provider. While the job pod is running, we resize the PVC by modifying spec.resources.requests.storage. The PVC cannot be resized while the pod is running as expected. However, after the job pod is completed, the PVC still doesn't get resized. external-resizerdoesn't send resizing gRPC call to the storage provider. The PVC is stuck in this state forever until we manually delete the job pod.

Reproduce steps:

1. Deploy external-resizer together with a storage provider (we use Longhorn)

2. Don't set the --handle-volume-inuse-error flag for the external-resizer . It means that by default, external-resizer will handle handle volume in use error in resizer controller, link

3. Deploy a job that uses a PVC as below. The job creates a pod that will sleep for 2 minutes and complete.

   Click to open

   apiVersion: v1
   kind: PersistentVolumeClaim
   metadata:
     name: test-job-pvc
     namespace: default
   spec:
     accessModes:
       - ReadWriteOnce
     storageClassName: longhorn
     resources:
       requests:
         storage: 1Gi
   ---
   apiVersion: batch/v1
   kind: Job
   metadata:
     name: test-job
     namespace: default
   spec:
     backoffLimit: 1
     template:
       metadata:
         name: test-job
       spec:
         containers:
           - name: test-job
             image: ubuntu:latest
             imagePullPolicy: IfNotPresent
             securityContext:
               privileged: true
             command: ["/bin/sh"]
             args: ["-c", "echo 'sleep for 120s then exit'; sleep 120"]
             volumeMounts:
               - mountPath: /data
                 name: vol
         restartPolicy: OnFailure
         volumes:
           - name: vol
             persistentVolumeClaim:
               claimName: test-job-pvc
   

4. While the job pod become running, try to expand the PVC by editing the spec.resources.requests.storage

5. Observe that the resizing fail

6. Wait for the job pod to become completed.

7. Observer that that PVC stuck in the current state forever. It doesn't get resized because external-resizer doesn't attempt to make gRPC expanding call to the storage provider.

Expected Behavior:

Once the job pod is completed, the PVC is no longer consider to be in-used. Therefore external-resizer should attempt to make gRPC expanding call to the storage provider.

Propose:
We dig into the source code see that:

1. This checker prevent the external-resizer from retrying if the PVC has InUseErrors before AND it is in the ctrl.usedPVCs map
2. The problem is that the PVC is never removed from the ctrl.usedPVCs map when a pod move to completed phase. PVC is only removed when the pod is deleted, link
3. We think that the logic over here should be changed to handle the case when the pod become completed. I.e.,:

   func (ctrl *resizeController) updatePod(oldObj, newObj interface{}) {
       pod := parsePod(newObj)
       if pod == nil {
   	    return
       }
       
       if isPodTerminated(pod) {
   	    ctrl.usedPVCs.removePod(pod)
       } else {
   	    ctrl.usedPVCs.addPod(pod)
       }
   }
   

Evn:

• external-resizer v1.2.0
• Longhorn v1.2.2

[PhanLe1010]

/sig storage

[PhanLe1010]

We can open a PR if you guys agree with our proposal

[yasker]

ping @gnufied @mauriciopoppe

[mauriciopoppe]

I'm not familiar with the external-resizer, @msau42 do you know who can help?

[gnufied]

Hey sorry I was on PTO last week. The proposal looks good. Please open a PR.

[PhanLe1010]

@gnufied
PR is opened. Please allow the bot to test the PR

Do I also need to create back-ported PRs into release-1.2 and release-1.3?

[PhanLe1010]

@gnufied

I see that the fix is tagged for v1.4.0 release. However, v1.4.0 version requires CSI SPEC 1.5.0. Not all storage providers are using the new CSI SPEC 1.5.0. On the other hand, the bug also exists in external-resizer v1.2.0 and v1.3.0. Therefore, can we backport the fix to release-1.2, release-1.3 and release new versions v1.2.1, v1.3.1?

[gnufied]

The CSI spec is backward compatible and I don't recall we have made any core changes in features that affect volume expansion in CSI spec for awhile. So, a driver which does not implement 1.5.0 should still be able to run with external-resizer v1.4.0. I do not think there will be any issues or such.

We could backport the fix and open it in older releases too if that makes it easier.

[PhanLe1010]

We could backport the fix and open it in older releases too if that makes it easier

@gnufied It would be great if we can backport the fix to release-1.2, release-1.3 and release new versions v1.2.1, v1.3.1.

In our case, we are using external-resizer v1.2.0 with CSI Spec v1.2.0. Yeah, I understand that it could work with external-resizer v1.4.0 that uses CSI Spec 1.5.0, but it could also raise some unforeseen bugs that we are not fully aware of yet. I remembered we have some accidents with csi-provisioner before which makes us scared.

[gnufied]

Feel free to create backport request. I will make a release.

[PhanLe1010]

Thanks @gnufied
How do I create backport request?