Merge bb0d217941 into 76d4f22aa3

Gating sidecar injection purely on cluster initialization state made
the operator's periodic EVALUATE drift-check disagree with the pod
spec stored at creation time, so it would roll out the primary right
after it finished bootstrapping, solely to strip the now-unneeded
sidecar.

Only stop adding the sidecar on TYPE_CREATE, once the cluster is
initialized; keep injecting it unconditionally on TYPE_EVALUATE so an
already-running pod never drifts against its own stored spec.

Signed-off-by: Marco Nenciarini <marco.nenciarini@enterprisedb.com>

internal/cnpgi/operator/lifecycle.go

@@ -330,18 +330,26 @@ func (impl LifecycleImplementation) collectAdditionalInstanceArgs(
 // two cases always inject it. A recovery-only cluster (only
 // RecoveryBarmanObjectName set, mirroring what pluginConfiguration.Validate()
 // accepts) only ever needs the sidecar for its one-time bootstrap restore, so
-// it's gated on cluster.IsInitialized() instead.
+// once the cluster has completed that initial bootstrap there's no reason to
+// keep carrying it on every future pod.
 //
-// Once the cluster finishes initializing, this makes the operator's own
+// That said, the two operation types this hook is invoked with can't be
-// drift-check (checkPodSpecIsOutdated) see the running pod's spec as outdated
+// treated the same way here. TYPE_CREATE fires only when a Pod is actually
-// and roll it out to drop the sidecar. That's deliberately accepted rather
+// about to be persisted (the bootstrap pod itself, a later replica, or any
-// than engineered around: it's one deterministic rollout using the same
+// pod recreated for an unrelated reason), so it's safe to gate it on
-// machinery the operator already uses for every other pod-spec change (a
+// cluster.IsInitialized(): a pod created after bootstrap simply won't carry
-// switchover if a replica is available, an in-place restart otherwise), not a
+// the sidecar. TYPE_EVALUATE, however, is also used by the operator's
-// new or fragile risk.
+// checkPodSpecIsOutdated to re-evaluate an already-running pod's spec for
+// drift on every reconcile; gating that the same way would make an
+// already-initialized cluster's freshly re-evaluated spec disagree with the
+// spec stored at the pod's creation, and the operator would roll out the
+// primary purely to strip the sidecar right after it finished bootstrapping.
+// So EVALUATE always keeps the sidecar for a recovery-only cluster, and only
+// CREATE actually stops adding it to pods created after initialization.
 func shouldInjectBarmanSidecar(
 	cluster *cnpgv1.Cluster,
 	pluginConfiguration *config.PluginConfiguration,
+	request *lifecycle.OperatorLifecycleRequest,
 ) bool {
 	if len(pluginConfiguration.BarmanObjectName) != 0 || len(pluginConfiguration.ReplicaSourceBarmanObjectName) != 0 {
 		return true
@@ -351,9 +359,13 @@ func shouldInjectBarmanSidecar(
 		return false
 	}
 
+	if request.GetOperationType().GetType() == lifecycle.OperatorOperationType_TYPE_CREATE {
 		return !cluster.IsInitialized()
 	}
 
+	return true
+}
+
 func reconcileInstancePod(
 	ctx context.Context,
 	cluster *cnpgv1.Cluster,
@@ -371,7 +383,7 @@ func reconcileInstancePod(
 
 	mutatedPod := pod.DeepCopy()
 
-	if shouldInjectBarmanSidecar(cluster, pluginConfiguration) {
+	if shouldInjectBarmanSidecar(cluster, pluginConfiguration, request) {
 		if err := reconcilePodSpec(
 			cluster,
 			&mutatedPod.Spec,

internal/cnpgi/operator/lifecycle_test.go

@@ -265,8 +265,8 @@ var _ = Describe("LifecycleImplementation", func() {
 			Expect(patch).To(ContainElement(HaveKeyWithValue("path", "/spec/initContainers")))
 		})
 
-		It("does not inject the sidecar for a recovery-only cluster that has "+
+		It("does not inject the sidecar into a new pod of a recovery-only cluster "+
-			"already completed its initial bootstrap", func(ctx SpecContext) {
+			"that has already completed its initial bootstrap", func(ctx SpecContext) {
 			recoveryOnlyConfig := &config.PluginConfiguration{
 				RecoveryBarmanObjectName: "minio-store-recovery",
 			}
@@ -279,6 +279,9 @@ var _ = Describe("LifecycleImplementation", func() {
 			podJSON, _ := json.Marshal(pod)
 			request := &lifecycle.OperatorLifecycleRequest{
 				ObjectDefinition: podJSON,
+				OperationType: &lifecycle.OperatorOperationType{
+					Type: lifecycle.OperatorOperationType_TYPE_CREATE,
+				},
 			}
 
 			response, err := reconcileInstancePod(ctx, cluster, request, recoveryOnlyConfig, sidecarConfiguration{})
@@ -287,6 +290,35 @@ var _ = Describe("LifecycleImplementation", func() {
 			Expect(response.JsonPatch).To(BeEmpty())
 		})
 
+		It("keeps injecting the sidecar on EVALUATE for an already-bootstrapped recovery-only "+
+			"cluster, so the operator never sees drift against the pod's stored spec and rolls "+
+			"it out just to strip the sidecar", func(ctx SpecContext) {
+			recoveryOnlyConfig := &config.PluginConfiguration{
+				RecoveryBarmanObjectName: "minio-store-recovery",
+			}
+			cluster.Status.LatestGeneratedNode = 1
+			pod := &corev1.Pod{
+				TypeMeta:   podTypeMeta,
+				ObjectMeta: metav1.ObjectMeta{Name: "test-pod"},
+				Spec:       corev1.PodSpec{Containers: []corev1.Container{{Name: "postgres"}}},
+			}
+			podJSON, _ := json.Marshal(pod)
+			request := &lifecycle.OperatorLifecycleRequest{
+				ObjectDefinition: podJSON,
+				OperationType: &lifecycle.OperatorOperationType{
+					Type: lifecycle.OperatorOperationType_TYPE_EVALUATE,
+				},
+			}
+
+			response, err := reconcileInstancePod(ctx, cluster, request, recoveryOnlyConfig, sidecarConfiguration{})
+			Expect(err).NotTo(HaveOccurred())
+			Expect(response).NotTo(BeNil())
+			Expect(response.JsonPatch).NotTo(BeEmpty())
+			var patch []map[string]interface{}
+			Expect(json.Unmarshal(response.JsonPatch, &patch)).To(Succeed())
+			Expect(patch).To(ContainElement(HaveKeyWithValue("path", "/spec/initContainers")))
+		})
+
 		It("does not mutate the pod when no object store is configured", func(ctx SpecContext) {
 			emptyConfig := &config.PluginConfiguration{}
 			pod := &corev1.Pod{