Tracking DRA Support for Beta

[sohankunkerkar]

Context

DRA support in Kueue is currently Alpha, supporting ResourceClaimTemplates with AllocationMode=ExactCount and DeviceClass mapping.

This issue is to discuss what's needed for Beta, aligned with upstream Kubernetes DRA development.

KEP: keps/2941-DRA

---

Gaps in Current Implementation

DRA Structured Parameters is GA in Kubernetes 1.34.

Affects Quota (Needs Implementation)

Capability	Description
AllocationMode=All	Request all devices on a node
Direct ResourceClaims	Reference pre-created claims
CEL Selectors	Filter devices by attributes (only with AllocationMode=All)

Does Not Affect Quota (Remove Rejection)

Capability	Description
Device Constraints	Co-location requirements - pass through to scheduler
Device Config	Driver configuration - pass through to scheduler

---

Upstream Features to Track

Feature	Kubernetes Status	Feature Gate
FirstAvailable	Beta	DRAPrioritizedList
Extended Resources	Alpha	DRAExtendedResource
Partitionable Devices	Alpha	DRAPartitionableDevices
Consumable Capacity	Alpha	DRAConsumableCapacity
Admin Access	Alpha	DRAAdminAccess
Device Taints	Alpha	DRADeviceTaints

---

Discussion

1. Which capabilities are required for Kueue DRA Beta?
2. CEL Selectors: conservative counting or CEL evaluation?
3. Should we include k8s Beta features (FirstAvailable)?
4. Any Alpha features we should wait for?

[sohankunkerkar]

/kind discussion

[k8s-ci-robot]

@sohankunkerkar: The label(s) kind/discussion cannot be applied, because the repository doesn't have them.

Details

In response to this:

/kind discussion

Instructions for interacting with me using PR comments are available here. If you have questions or suggestions related to my behavior, please file an issue against the kubernetes-sigs/prow repository.

[sohankunkerkar]

/kind feature

[sohankunkerkar]

cc @alaypatel07

[mimowo]

/priority important-soon

[kannon92]

I'd also like to see e2e tests for beta. I'm fine with using the fake dra driver but I think this feature should have e2e testing before beta.

[sohankunkerkar]

/assign

[sohankunkerkar]

Jotting down here for posterity. Things to consider for Beta:

• Support v1 API of DRA in core K8s
• Beta status for Extended Resources and Partitionable Devices in k8s
• Support integration with MultiKueue
• E2E tests using dra-example-driver

[kannon92]

I think Kueue DRA is using v1 APIs now.

9ac669d

[sohankunkerkar]

I think Kueue DRA is using v1 APIs now.

9ac669d

Yes, based on that I have raised #8421

[alaypatel07]

IMO, before going to beta, we should explore what the support for Extended Resources backed by DRA will look like in Kueue. There are overlapping concepts with DRA support in Kueue and Extended Resources backed by DRA in k8s, like the need for creating a mapping from a extended resources string -> to a device class. This work will prove out the gaps if any in the current solution that will help boost confidence for beta.