fix: a default chart install could not create a cell, or say why

.github/workflows/ci.yaml

@@ -58,6 +58,42 @@ jobs:
         run: |
           helm lint charts/gpucellpool
           helm template gpucellpool charts/gpucellpool --namespace gpucellpool-system > /dev/null
+      # The release workflow pushes the git tag verbatim (vX.Y.Z) while appVersion
+      # carries no leading v, so a default install once asked for a tag that had
+      # never been published and failed on its own image. Assert the two agree.
+      - name: Default image tag matches what the release workflow publishes
+        run: |
+          want="v$(grep '^appVersion:' charts/gpucellpool/Chart.yaml | awk '{print $2}' | tr -d '\"')"
+          got=$(helm template gpucellpool charts/gpucellpool \
+                  | grep -oE 'gpucellpool/manager:[^"]+' | head -1 | cut -d: -f2)
+          echo "appVersion implies $want; chart renders $got"
+          test "$want" = "$got"
+
+      # Both of these shipped broken in v0.1.0 and neither could fail a test that
+      # ran with admin credentials: the operator RENDERS a per-cell bootstrap
+      # Secret, and its recorder writes through events.k8s.io. A read-only secrets
+      # grant meant no chart-installed pool could ever create a cell; a
+      # core-group-only events grant meant every event was silently rejected.
+      - name: Chart RBAC permits what the operator actually does
+        run: |
+          helm template gpucellpool charts/gpucellpool --show-only templates/rbac.yaml > /tmp/rbac.yaml
+          python3 -c '
+          import sys, yaml
+          need = {"secrets": {"create", "update"}, "events": {"create"}}
+          have = {k: set() for k in need}
+          for doc in yaml.safe_load_all(open("/tmp/rbac.yaml")):
+              for rule in (doc or {}).get("rules", []):
+                  if "events" in rule.get("resources", []) and "events.k8s.io" not in rule.get("apiGroups", []):
+                      continue
+                  for res in rule.get("resources", []):
+                      if res in have:
+                          have[res] |= set(rule.get("verbs", []))
+          for res, verbs in need.items():
+              missing = verbs - have[res]
+              assert not missing, "%s missing %s (have %s)" % (res, sorted(missing), sorted(have[res]))
+          print("rbac ok:", {k: sorted(v) for k, v in have.items()})
+          '
+
       # The chart ships a copy of the CRD; helm upgrade never updates crds/, so a
       # stale copy would install an old schema on a fresh cluster and silently
       # drop fields.

charts/gpucellpool/rules.yaml

@@ -1,18 +1,21 @@
 - apiGroups:
   - ""
   resources:
-  - events
+  - secrets
   verbs:
   - create
-  - patch
+  - get
+  - list
+  - update
+  - watch
 - apiGroups:
   - ""
+  - events.k8s.io
   resources:
-  - secrets
+  - events
   verbs:
-  - get
-  - list
-  - watch
+  - create
+  - patch
 - apiGroups:
   - bootstrap.cluster.x-k8s.io
   resources:

charts/gpucellpool/templates/deployment.yaml

@@ -24,7 +24,13 @@ spec:
           type: RuntimeDefault
       containers:
         - name: manager
-          image: "{{ .Values.image.repository }}:{{ .Values.image.tag | default .Chart.AppVersion }}"
+          {{- /*
+          The release workflow pushes the git tag verbatim (v0.1.0), while appVersion
+          carries no leading v (0.1.0, and the workflow asserts that). Defaulting to
+          appVersion alone therefore asked for a tag that does not exist: a default
+          install of the released chart failed with ImagePullBackOff on its own image.
+          */}}
+          image: "{{ .Values.image.repository }}:{{ .Values.image.tag | default (printf "v%s" .Chart.AppVersion) }}"
           imagePullPolicy: {{ .Values.image.pullPolicy }}
           args:
             - --health-probe-bind-address=:8081

charts/gpucellpool/templates/monitoring/servicemonitor.yaml

@@ -31,6 +31,14 @@ spec:
     - port: metrics
       path: /metrics
       interval: {{ $sm.interval | default "30s" }}
+      # Every gpucell_* series carries pool and namespace labels describing the POOL.
+      # Prometheus overwrites exposed labels with target labels by default, so our
+      # namespace was being replaced by the operator's own namespace and ours pushed
+      # to exported_namespace — measured against a live Prometheus. Every alert then
+      # named the wrong namespace. honorLabels keeps the describing labels, which is
+      # exactly what kube-prometheus-stack does for kube-state-metrics, for exactly
+      # this reason.
+      honorLabels: true
       {{- if .Values.metrics.secure }}
       # The manager serves HTTPS with a self-signed certificate, so verification
       # is skipped. It does not authenticate scrapers — see issue #11 and the

config/rbac/role.yaml

@@ -7,18 +7,21 @@ rules:
 - apiGroups:
   - ""
   resources:
-  - events
+  - secrets
   verbs:
   - create
-  - patch
+  - get
+  - list
+  - update
+  - watch
 - apiGroups:
   - ""
+  - events.k8s.io
   resources:
-  - secrets
+  - events
   verbs:
-  - get
-  - list
-  - watch
+  - create
+  - patch
 - apiGroups:
   - bootstrap.cluster.x-k8s.io
   resources:

internal/controller/controller.go

@@ -62,8 +62,20 @@ type GPUCellPoolReconciler struct {
 // +kubebuilder:rbac:groups=cluster.x-k8s.io,resources=clusters,verbs=get;list;watch
 // +kubebuilder:rbac:groups=infrastructure.cluster.x-k8s.io,resources=kubeswiftmachines,verbs=get;list;watch;create;update;patch;delete
 // +kubebuilder:rbac:groups=bootstrap.cluster.x-k8s.io,resources=*,verbs=get;list;watch;create;update;patch;delete
-// +kubebuilder:rbac:groups="",resources=secrets,verbs=get;list;watch
+// Secrets need create+update, not only read: the operator RENDERS a per-cell
+// bootstrap Secret from the user's join template (reconcileBootstrapSecret). A
+// read-only grant let the chart install, pass every test that used admin
+// credentials, and then fail on the first cell with "secrets is forbidden".
+// Delete is deliberately absent — the per-cell Secret carries the pool's
+// ownerRef and is garbage-collected.
+// +kubebuilder:rbac:groups="",resources=secrets,verbs=get;list;watch;create;update
+// Events need BOTH groups. The core-group grant alone looks right and is
+// useless: the recorder writes through the events.k8s.io API, so every event
+// this operator emits was rejected with "events.events.k8s.io is forbidden" —
+// visible only in the manager log, never to the operator reading `kubectl
+// describe gpucellpool`. The core grant stays for clients that still read there.
 // +kubebuilder:rbac:groups="",resources=events,verbs=create;patch
+// +kubebuilder:rbac:groups=events.k8s.io,resources=events,verbs=create;patch
 
 // Reconcile implements the loop in docs/design/gpucellpool-reconciliation.md §4.
 func (r *GPUCellPoolReconciler) Reconcile(ctx context.Context, req ctrl.Request) (ctrl.Result, error) {