What would you like to be added:

Improve accuracy of Topology-Aware Scheduling by respecting PodTemplate settings which are currently ignored by could block actual scheduling of pods at the kube-scheduler level. For example pod affinities and anti-affinities.

One option is tighter integration with kube-scheduler, similarly as ClusterAutoscaler does, however it may entail worse performance. Another option, is to gradually improve scheduling accuracy based on user feedback, but replicating the most commonly user properties (as currently we support tolerations).

Why is this needed:

TAS may currently schedule workloads which are unschedulable at the kube-scheduler level. Giving early feedback to users why such workloads cannot be scheduled would improve user experience.

Completion requirements:

This enhancement requires the following artifacts:

• Design doc
• API change
• Docs update

The artifacts should be linked in subsequent comments.