CA DRA: handle device taints and tolerations (KEP-5055)

[x13n]

Which component are you using?:

/area cluster-autoscaler

Is your feature request designed to solve a problem? If so describe the problem this feature should solve.:

KEP-5055 adds support for Device taints to DRA. This means that the individual Devices exposed in ResourceSlices can be tainted similarly to how Nodes can be tainted today (there are tolerations, taint-based eviction, etc.). The feature is behind a separate feature gate and went to alpha in 1.33.

As part of the KEP, admins can now create patch objects (DeviceTaintRule) that automatically add taints to all Devices matching certain conditions. Cluster Autoscaler needs to apply the patches to ResourceSlices before exposing them to the DRA scheduler plugin via scheduler framework.

Describe the solution you'd like.:

We should update

autoscaler/cluster-autoscaler/simulator/dynamicresources/snapshot/snapshot_slice_lister.go

Line 25 in 9937f8f

func (s snapshotSliceLister) List() ([]*resourceapi.ResourceSlice, error) {

to apply the taint patches before returning ResourceSlices. Ideally we would reuse k8s.io/dynamic-resource-allocation/resourceslice/tracker if possible. We need to make sure that the patches are applied to the fake ResourceSlices created by CA as well, before they're used for anything else.

Furthermore, we should extend the existing Node taint handling to Device taints:

• Device taints should be filtered from template NodeInfos using the same logic as Node taints:
  • Well-known transient taints are filtered out.
  • Users can configure startup and status taints which are filtered out.
• Device taints applied by DeviceTaintRules shouldn't be filtered out from template NodeInfos if we can detect that they will apply to the new Node (e.g. if they apply to all Devices from a driver, regardless of the pool or other attributes).

Additional context.:

• device taints and tolerations (KEP 5055) kubernetes#130447
• DRA: device taints and tolerations enhancements#5055

[x13n]

/cc @towca

[MenD32]

I tried to start working on this, and this requires upgrading Kubernetes packages from v0.33.0.alpha to v0.33.0, which requires upgrading Go to v1.24.0. This can potentially affect a lot of other parts of CA. Should this be a different chore type issue/PR?

[x13n]

Yes, bumping deps is separate. I tried to do that in #7919, but many tests are failing and I didn't have time to get back to this yet.