# Wave 6 — the model servers behind api.riotpiao.com (namespace `llm-serving`).
#
# Syncs before wave 7 (api-gw), so the predictor Services exist before the
# gateway that routes to them. KServe itself is part of the substrate; this Application
# owns only the InferenceServices.
#
# Adopted from live state on 2026-08-15. These five had been `kubectl apply`-ed
# by hand — no ArgoCD ownership, present in no repo — so every change to them
# was drift by definition. Each manifest was exported from the cluster and
# verified with `kubectl diff` returning empty before this file existed; the
# first sync therefore adopted them without restarting anything.
apiVersion: argoproj.io/v1alpha1
kind: Application
metadata:
  name: llm-serving
  namespace: argocd
  annotations:
    argocd.argoproj.io/sync-wave: "6"
spec:
  project: homelab
  revisionHistoryLimit: 3
  source:
    repoURL: https://forgejo.riotpiao.com/rock/homelab.git
    targetRevision: main
    path: k8s/apps/llm-serving
  destination:
    server: https://kubernetes.default.svc
    namespace: llm-serving
  syncPolicy:
    automated:
      # `prune: false` here, unlike every other Application in this repo, and it
      # is not an oversight.
      #
      # ArgoCD tracks ownership with the `argocd.argoproj.io/instance` label
      # (argocd-cm `application.instanceLabelKey`). KServe copies an
      # InferenceService's labels onto the Deployment and Service it generates —
      # visible today as `app.kubernetes.io/name` and `part-of` on
      # `ornith-predictor`. So once ArgoCD labels an InferenceService, KServe
      # propagates that tracking label to children that are not in git, ArgoCD
      # reads them as extraneous, prunes them, and KServe recreates them. That
      # loop churns GPU pods.
      #
      # Deleting an InferenceService therefore means deleting the file AND
      # removing the object, rather than relying on prune.
      prune: false
      selfHeal: true
    syncOptions:
      - CreateNamespace=true
      # KServe CRDs are large; server-side apply avoids the
      # "metadata.annotations: Too long" failure client-side apply hits, and is
      # the correct mode for adopting objects an operator also writes to.
      - ServerSideApply=true
    retry:
      limit: 3
      backoff:
        duration: 10s
        factor: 2
        maxDuration: 3m