Wrong CSI allocatable count with `kubernetes` metadata source

[horjulf]

/kind bug

What happened?

kubernetes metadata source doesn't lookup number of ENIs or EBS volumes, leading to a wrong CSI allocatable calculation in all instances since at least an EBS block device would be attached.

Values are statically set at:
https://github.com/kubernetes-sigs/aws-ebs-csi-driver/blob/master/pkg/cloud/metadata/k8s.go#L150

instanceInfo := Metadata{
		InstanceID:             instanceID,
		InstanceType:           instanceType,
		Region:                 region,
		AvailabilityZone:       availabilityZone,
		NumAttachedENIs:        1, // All nodes have at least 1 attached ENI, so we'll use that
		NumBlockDeviceMappings: 0,
	}

What you expected to happen?

Feature parity with IMDS metadata provider.

How to reproduce it (as minimally and precisely as possible)?

Helm chart values:

node:
  metadataSources: imds

Check allocatable count with:
kubectl get csinode <node> -o yaml

Anything else we need to know?:

Environment

• Kubernetes version (use kubectl version): 1.33
• Driver version: 1.45

[horjulf]

Related PR #2540

[torredil]

Hi @horjulf, thank you for reaching out about this. Today, the driver is unable to retrieve ENI information from the Kubernetes API, hence why 1 is assumed. As you've pointed out, this is not ideal as it can result in an inaccurate CSI Node allocatable value (which kube-scheduler uses for scheduling decisions).

The team is aware of this and is actively working on a solution to retrieve ENI count information for users that rely on Kubernetes metadata.

/priority backlog

[torredil]

/triage/accepted
/lifecycle/active