[rollandf]

When useCdi is enabled, the CDI annotation key was derived only from the CDI vendor and kind (cdi.k8s.io/nvidia.com_net-rdma), so every resource pool returned the same key from Allocate. Kubelet merges the annotations of all resources allocated to a container into a single map, so when a container requested several RDMA resources only one annotation survived and only one device was injected.

Include the pool (resource) name in the annotation key, e.g. cdi.k8s.io/nvidia.com_net-rdma_. If the pool name would produce an invalid key (over 63 characters, or not ending with an alphanumeric character), fall back to a short SHA-256 hash of the pool name.

Fixes: https://redmine.nvidia.com/issues/5369341

[greptile-apps]

Confidence Score: 5/5

[Medium impact] The PR appears safe to merge; no actionable issues were found.

Summary

Gives each resource pool a stable CDI annotation key so kubelet can keep annotations from multiple pools.

• CDI allocations keep a separate annotation for each pool.

No actionable issues were found. Tests were inspected but not run.

Diagram

%%{init: {'theme': 'neutral'}}%%
flowchart LR
  A[Allocate for each pool] --> B[Hash CDI prefix and pool name]
  B --> C[Create pool-specific annotation key]
  C --> D[Kubelet merges annotations]
  D --> E[Runtime reads unchanged CDI device names]

Loading

_{Reviews (2) · Last reviewed commit: "fix(cdi): use a unique CDI annotation ke..." · Reviewed by Greptile}