`KUBERNETES_ENABLE_POD_UID` does not add pod UID label in DRA-attributed metrics

[rajathagasthya]

Description

When GPU allocation happens through DRA (KUBERNETES_ENABLE_DRA=true), setting DCGM_EXPORTER_KUBERNETES_ENABLE_POD_UID=true does not add the UID label to metrics. The same configuration on a device-plugin node adds it as expected.

Expected behavior

With both KUBERNETES_ENABLE_DRA=true and
DCGM_EXPORTER_KUBERNETES_ENABLE_POD_UID=true, metrics attributed to a pod holding a DRA ResourceClaim should carry the UID label, the same as device-plugin allocated GPUs.

Actual behavior

DRA-attributed series carry pod, namespace, container, the dra_* labels, and (when enabled) pod labels, but no UID, regardless of the pod-UID setting.

Steps to reproduce

1. Node with DRA GPU allocation (e.g. GPU Operator with the DRA stack; NVIDIA DRA driver publishing ResourceSlices).
2. Run dcgm-exporter with KUBERNETES_ENABLE_DRA=true, DCGM_EXPORTER_KUBERNETES_ENABLE_POD_UID=true, a ServiceAccount token, and read access to resource.k8s.io/resourceslices and pods.
3. Start a pod that requests a GPU via a ResourceClaim and scrape /metrics.
4. Metrics for that pod have pod/namespace/container and dra_* labels but no UID label. Repeating with a device-plugin allocation on the same exporter produces the UID label.

Environment

Observed with dcgm-exporter 4.6.0-4.8.3 (GPU Operator DRA stack, Kubernetes with resource.k8s.io/v1).

[rvatkar]

@rajathagasthya Thank you for filling issue. DCGM exporter team in looking into it. For future internal Nvidia member's issues/PRs related to the DCGM exporter, we would appreciate it if you could also create an internal ticket/MR. This will help us triage, track, test and prioritize the request more effectively.

[rajathagasthya]

@nccurry I see PR #738 that addresses this issue is still open, but this issue is closed and mentioned in the release notes. Is this accurate?