# Bookstore — Part 12 ch.07 "ML pipelines and workflows": the nightly
# retraining `CronWorkflow`. References the `recommender-pipeline`
# WorkflowTemplate in this dir; CronWorkflow = "CronJob for Argo Workflows".
#
# !!! CRD-INTRINSIC DRY-RUN (identical precedent to recommender-workflow.yaml
#     in this dir, raw-manifests/51-/70-/83-, argocd/, operators/, chaos/,
#     ml/batch/, ml/serve/recommender-inferenceservice.yaml) !!!
#   `CronWorkflow` is an Argo Workflows CRD (argoproj.io/v1alpha1). WITHOUT
#   Argo Workflows installed a client dry-run prints:
#     no matches for kind "CronWorkflow" in version "argoproj.io/v1alpha1"
#   EXPECTED and SCHEMA-CORRECT — install Argo Workflows first (Part 12 ch.07
#   Hands-on: pinned Helm `argo/argo-workflows --version 0.42.0` -> ns `argo`).
#   Schema verified against argoproj.io/v1alpha1 (schedule / timezone /
#   concurrencyPolicy / workflowSpec.workflowTemplateRef).
#
# WHAT IT DOES
#   Runs the `recommender-pipeline` WorkflowTemplate nightly at 02:00 UTC.
#   This is the canonical "scheduled retrain" — the Periodic Job pattern
#   (Ibryam) applied to ML pipelines.
#
# WHY a CronWorkflow (and not a Kubernetes CronJob spawning `argo submit`)
#   Both work; CronWorkflow is the Argo-native shape: same Workflow spec,
#   same RBAC, same UI/CLI/observability as ad-hoc runs. The Kubernetes
#   CronJob approach reintroduces "pipeline" as something a separate
#   controller needs to know about — exactly the integration cost
#   Workflows + CronWorkflow remove.
#
# Apply (after Argo Workflows is installed AND recommender-workflow.yaml is
# applied; see pipeline/README.md):
#   kubectl apply -f examples/bookstore/ml/pipeline/recommender-cronworkflow.yaml
#   kubectl get cronworkflow -n bookstore-ml
#   # Trigger one off-schedule for testing:
#   argo cron list  -n bookstore-ml
#   argo submit --from cronworkflow/recommender-nightly -n bookstore-ml
apiVersion: argoproj.io/v1alpha1
kind: CronWorkflow
metadata:
  name: recommender-nightly
  namespace: bookstore-ml
  labels:
    app.kubernetes.io/part-of: bookstore-ml
    app.kubernetes.io/component: ml-pipeline
    ml.bookstore/path: argo-cronworkflow
spec:
  # 02:00 UTC nightly. Quoted explicitly so YAML doesn't reinterpret.
  schedule: "0 2 * * *"
  timezone: "UTC"
  # Replace: never run two retrains at once (the next nightly waits for the
  # previous to finish). For the Bookstore the runtime is < 1 min; in prod
  # set this to `Forbid` if a too-long run should NOT be killed.
  concurrencyPolicy: "Replace"
  # Bound the time the controller waits to submit a missed schedule
  # (e.g. after a controller restart). 10 minutes is sane for nightly.
  startingDeadlineSeconds: 600
  # Keep a few finished histories — Argo UI navigates these.
  successfulJobsHistoryLimit: 3
  failedJobsHistoryLimit: 1
  workflowSpec:
    # Reference the WorkflowTemplate by name; no spec duplication.
    workflowTemplateRef:
      name: recommender-pipeline
    # All other workflow-level overrides (parameters, ttl, …) inherit from
    # the WorkflowTemplate. Override `seed` per-run if you want
    # non-deterministic retraining across nights — left at default here.