No topology key found on hw nodes

[SNB-hz]

In clusters with hardware nodes, a new PVC and its workload can be stuck in Pending state if they are scheduled without nodeAffinity.

Steps to reproduce:

• run a cluster that includes a hardware worker, and label the hw node with instance.hetzner.cloud/is-root-server=true as mentioned in the README
• install CSI driver according to instructions
• apply the test-pvc and pod mentioned in the README, using the default storageClass with WaitForFirstConsumer volumeBindingMode

Expected Behaviour:

hcloud-csi-controller should provide the desired / required topology constaints to the k8s scheduler, which then schedules the pod on a node fulfilling the topology requirements.
As the hardware node does not run csi-driver and cannot mount hetzner cloud volumes, the workload should not be scheduled there.

Observed Behaviour:

• Both pvc and pod are stuck in Pending state.
• the container csi-provisioner of the CSI Controller deployment logs this Error:

'Warning' reason: 'ProvisioningFailed' failed to provision volume with StorageClass "hcloud-volumes": error generating accessibility requirements: no topology key found on CSINode hardwarenode.testcluster

More Info:

• the DaemonSet for hcloud-csi-node does not run on the hw node
• because of this, the csinode object for the node lists no driver:

kubectl get csinode
NAME                     DRIVERS       AGE
virtualnode.testcluster     1           1d
hardwarenode.testcluster    0           1d

• the csinode object of the virtual node looks ok:

kubectl get csinode virtualnode.testcluster -oyaml
apiVersion: storage.k8s.io/v1
kind: CSINode
...
spec:
  drivers:
  - allocatable:
      count: 16
    name: csi.hetzner.cloud
    nodeID: "12769030"
    topologyKeys:
    - csi.hetzner.cloud/location

• the csinode object of the hardware node does not have a driver and therefore no topology key, as the node intentionally runs no hcloud-csi-node pod due to the nodeAffinity:

kubectl get csinode hardwarenode.testcluster -oyaml
apiVersion: storage.k8s.io/v1
kind: CSINode
...
spec:
  drivers: null

Theory

It seems we are hitting this Issue in csi-provisioner.
As the hardware node has no csi-driver pod and therefore no driver or topology key listed, the csi-provisioner breaks. It is trying to build the preferred topology to give it to the scheduler, but as the hardware node has no topology key, the csi-provisioner fails. Pod and PVC cannot finish scheduling and remain in Pending state forever.

Workaround

This issue can be avoided by making sure the object that uses the PVC (StatefulSet, Pod etc.) cannot be scheduled on the hardware node in the first place. This can be done by specifying a nodeAffinity:

    spec:
      affinity:
        nodeAffinity:
          requiredDuringSchedulingIgnoredDuringExecution:
            nodeSelectorTerms:
            - matchExpressions:
              - key: instance.hetzner.cloud/is-root-server
                operator: NotIn
                values:
                - "true"

Proposed Solution

The external-provisioner Issue, lists a few possible solutions on the csi-driver side, such as running the csi-driver on all nodes, including hardware nodes.
CSI-controller would then need to be aware of which nodes are virtual or hardware when providing the topology preferences to the k8s scheduler.

[apricote]

Duplicate of #400