Do you recommend customers use v0.1.11 as the default version?

Yes it actually makes things simpler for users. Some of the improvements are:

• Single ResourceFlavor is now capable of supporting TAS and non-TAS workloads. Which means that only single quota is defined in the ClusterQueue.
• There's a new BestFit placement policy which increases performance for dense jobs.
• Workloads without any TAS annotation kueue.x-k8s.io/podset-preferred-topology or kueue.x-k8s.io/podset-required-topology will actually be scheduled with TAS but with a new unconstrained/auto mode and without rank ordering. This is needed to fight with fragmentation in clusters.

Do you recommend customers use v0.1.11 as the default version?

Yes it actually makes things simpler for users. Some of the improvements are:

• Single ResourceFlavor is now capable of supporting TAS and non-TAS workloads. Which means that only single quota is defined in the ClusterQueue.
• There's a new BestFit placement policy which increases performance for dense jobs.
• Workloads without any TAS annotation kueue.x-k8s.io/podset-preferred-topology or kueue.x-k8s.io/podset-required-topology will actually be scheduled with TAS but with a new unconstrained/auto mode and without rank ordering. This is needed to fight with fragmentation in clusters.

Thanks for confirming. Lets update the default version. We would also have to merge the Kueue v0.11.1 manifest in the manifest folder Will raise a PR and add you as a reviewer. Let's merge this PR after the manifest is merged.

Raised #3833 to support this PR. Please fix the errors highlighted by the test suite.

/gcbrun

@mwysokin can you fix the build errors happening due to documentation

Please update this doc: https://cloud.google.com/ai-hypercomputer/docs/workloads/schedule-gke-workloads-tas with the latest Kueue version and resolve the merge conflicts.

/gcbrun

/gcbrun