# Bookstore — Part 12 ch.03 "Batch and gang scheduling": a Kueue ClusterQueue
# = the QUOTA + admission policy for the Bookstore ML batch workloads. Jobs/
# JobSets labelled onto a LocalQueue that points here are admitted (via
# spec.suspend) only when the WHOLE job fits this quota.
#
# !!! CRD-INTRINSIC DRY-RUN (identical precedent to raw-manifests/51-/70-/83-,
#     argocd/, operators/, chaos/) !!!
#   `ClusterQueue` is a Kueue CRD (kueue.x-k8s.io/v1beta2). WITHOUT Kueue
#   installed a client dry-run prints:
#     no matches for kind "ClusterQueue" in version "kueue.x-k8s.io/v1beta2"
#   EXPECTED and SCHEMA-CORRECT — install Kueue first (Part 12 ch.03 Hands-on:
#   pinned Helm, own namespace `kueue-system`). It also references the
#   `bookstore-ml-flavor` ResourceFlavor (apply that first). Schema verified
#   against Kueue kueue.x-k8s.io/v1beta2 ClusterQueue.
#
# QUOTA SHAPE:
#   cpu/memory  -> small, so the CPU-only 2-worker recommender "training"
#                  (recommender-jobset.yaml) actually runs on kind.
#   nvidia.com/gpu nominalQuota -> the MULTI-TENANT GPU FAIR-SHARE fence. On a
#                  GPU-less cluster this simply QUEUES any GPU-requesting job
#                  (no GPUs to admit against) — it does not error. On a GPU
#                  cluster it is the per-cohort fair-share budget; pair it with
#                  a HARD per-namespace ResourceQuota on requests.nvidia.com/gpu
#                  (Part 08 ch.04) as the inviolable ceiling (ch.03 Hands-on §4).
#
# Single-tenant here for the worked example. For real multi-tenant fair-share:
# give each team its own LocalQueue->ClusterQueue and group the ClusterQueues
# in a shared `cohort:` so idle quota is borrowable but reclaimable (Kueue
# preemption, built on the Part 04 ch.03 priority/preemption model).
apiVersion: kueue.x-k8s.io/v1beta2
kind: ClusterQueue
metadata:
  name: bookstore-ml-cq
  labels:
    app.kubernetes.io/part-of: bookstore-ml
spec:
  namespaceSelector: {}            # which namespaces may use it ({} = all;
                                   # the LocalQueue scopes it to bookstore-ml)
  resourceGroups:
    - coveredResources: ["cpu", "memory", "nvidia.com/gpu"]
      flavors:
        - name: bookstore-ml-flavor
          resources:
            - name: "cpu"
              nominalQuota: "4"        # small: CPU demo fits on kind
            - name: "memory"
              nominalQuota: 8Gi
            - name: "nvidia.com/gpu"
              nominalQuota: "4"        # GPU fair-share budget (queues if no GPUs)