Skip to content

KAI Scheduler Skill - kai-audit-queue-quota - Across the queue hierarchy, who's over-quota, borrowing, or starved? - Phase 1 · read #2057

Description

@Hagay-RunAI

What you would like to be added?

Labels: area/agent-skills · phase-1 · read · persona/admin

Question. "Across the queue hierarchy, who's over-quota, borrowing, or starved?" Persona. Admin · Category. Queue, Quota & Fair-Share

What it does. The admin fairness audit: names the starved-vs-borrowing queue pairs and the idle-quota reclaim candidates across the hierarchy, with quantified GPU counts — where fairness is being bent and where capacity can be recovered.

KAI signals (read). Queue CRD hierarchy (deserved / over-quota weight / limit); metrics queue_deserved_gpus, queue_allocated_gpus, queue_fair_share_gpu, queue_gpu_usage across all queues; reclaim metrics.

Acceptance criteria.

Lists starved↔borrowing queue pairs with GPU deltas.
Surfaces idle-quota reclaim candidates (low usage vs held quota).
Walks the full hierarchy, not a single queue.

When NOT to use. Single queue owner's self-service view → #2. Live queued/pending snapshot → #6.

Example. "Where can I reclaim GPUs this week?" → "Fairness violation: team-a starved (allocated 4 / deserved 16) while batch borrows 18 over-quota. Idle-quota candidate: archive holds 9 GPUs at ~3% usage."

Depends on 'kai-explain-queue-fairshare'.

Why is this needed?

Across the queue hierarchy, who's over-quota, borrowing, or starved?

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    enhancementNew feature or request

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions