fix: Operator Size Increase

api/v2/weightsandbiases_types.go

@@ -541,7 +541,8 @@ type RedisConfig struct {
 }
 
 type RedisSentinelSpec struct {
-	Enabled bool                `json:"enabled"`
+	// +kubebuilder:default=true
+	Enabled *bool               `json:"enabled,omitempty"`
 	Config  RedisSentinelConfig `json:"config,omitempty"`
 }
 

config/crd/bases/apps.wandb.com_weightsandbiases.yaml

@@ -3864,9 +3864,8 @@ spec:
                                   type: object
                               type: object
                             enabled:
+                              default: true
                               type: boolean
-                          required:
-                          - enabled
                           type: object
                         storageSize:
                           type: string

config/rbac/role.yaml

@@ -396,3 +396,11 @@ rules:
   - securitycontextconstraints
   verbs:
   - use
+- apiGroups:
+  - storage.k8s.io
+  resources:
+  - storageclasses
+  verbs:
+  - get
+  - list
+  - watch

deploy/operator/templates/wandb-operator-wandb-role.yaml

@@ -4,6 +4,14 @@ kind: ClusterRole
 metadata:
   name: {{ .Release.Name }}-wandb
 rules:
+  - apiGroups:
+      - storage.k8s.io
+    resources:
+      - storageclasses
+    verbs:
+      - get
+      - list
+      - watch
   - apiGroups:
       - apps.wandb.com
     resources:
@@ -210,4 +218,4 @@ subjects:
   - kind: ServiceAccount
     name: {{ include "wandb-operator.fullname" . }}
     namespace: {{ .Release.Namespace }}
-{{- end }}
+{{- end }}

internal/controller/common/state.go

@@ -10,5 +10,5 @@ const (
 )
 
 var NotReadyStates = []string{
-	ErrorState, PendingState, UnavailableState,
+	ErrorState, PendingState, DegradedState, UnknownState, UnavailableState,
 }

internal/controller/infra/managed/clickhouse/altinity/keeper/read.go

@@ -48,7 +48,7 @@ func ReadState(
 		}}
 	}
 
-	return computeKeeperReadyCondition(ctx, podsRunning)
+	return computeKeeperReadyCondition(ctx, expectedKeeperPodCount(actual), podsRunning)
 }
 
 func keeperPodsRunningStatus(
@@ -70,7 +70,7 @@ func keeperPodsRunningStatus(
 	return result, nil
 }
 
-func computeKeeperReadyCondition(ctx context.Context, podsRunning map[string]bool) []metav1.Condition {
+func computeKeeperReadyCondition(ctx context.Context, expectedPodCount int, podsRunning map[string]bool) []metav1.Condition {
 	log := logx.GetSlog(ctx)
 
 	var runningCount, podCount int
@@ -80,19 +80,19 @@ func computeKeeperReadyCondition(ctx context.Context, podsRunning map[string]boo
 			runningCount++
 		}
 	}
-	log.Info("Keeper pods status", "running", runningCount, "total", podCount)
+	log.Info("Keeper pods status", "running", runningCount, "reported", podCount, "expected", expectedPodCount)
 
 	status := metav1.ConditionUnknown
 	reason := common.UnknownReason
 	message := ""
 	switch {
-	case podCount > 0 && podCount == runningCount:
+	case expectedPodCount > 0 && podCount == expectedPodCount && podCount == runningCount:
 		status = metav1.ConditionTrue
 		reason = common.ResourceExistsReason
-	case podCount > 0:
+	case expectedPodCount > 0 || podCount > 0:
 		status = metav1.ConditionFalse
 		reason = common.NoResourceReason
-		message = fmt.Sprintf("%d of %d keeper pods running", runningCount, podCount)
+		message = fmt.Sprintf("%d of %d expected keeper pods running (%d reported)", runningCount, expectedPodCount, podCount)
 	}
 
 	return []metav1.Condition{{
@@ -102,3 +102,21 @@ func computeKeeperReadyCondition(ctx context.Context, podsRunning map[string]boo
 		Message: message,
 	}}
 }
+
+func expectedKeeperPodCount(chk *chkv1.ClickHouseKeeperInstallation) int {
+	if chk == nil || chk.Spec.Configuration == nil {
+		return 0
+	}
+	var count int
+	for _, cluster := range chk.Spec.Configuration.Clusters {
+		if cluster == nil || cluster.Layout == nil {
+			continue
+		}
+		replicas := cluster.Layout.ReplicasCount
+		if replicas < 1 {
+			replicas = 1
+		}
+		count += replicas
+	}
+	return count
+}

internal/controller/infra/managed/clickhouse/altinity/keeper/read_test.go

@@ -10,20 +10,26 @@ import (
 
 var _ = Describe("Keeper readiness", func() {
 	It("is ready when all pods are running", func() {
-		conds := computeKeeperReadyCondition(context.Background(), map[string]bool{"a": true, "b": true, "c": true})
+		conds := computeKeeperReadyCondition(context.Background(), 3, map[string]bool{"a": true, "b": true, "c": true})
 		Expect(conds).To(HaveLen(1))
 		Expect(conds[0].Type).To(Equal(KeeperReportedReadyType))
 		Expect(conds[0].Status).To(Equal(metav1.ConditionTrue))
 	})
 
 	It("is not ready when some pods are not running", func() {
-		conds := computeKeeperReadyCondition(context.Background(), map[string]bool{"a": true, "b": false, "c": true})
+		conds := computeKeeperReadyCondition(context.Background(), 3, map[string]bool{"a": true, "b": false, "c": true})
 		Expect(conds[0].Status).To(Equal(metav1.ConditionFalse))
 		Expect(conds[0].Message).To(ContainSubstring("2 of 3"))
 	})
 
-	It("is unknown when no pods are reported yet", func() {
-		conds := computeKeeperReadyCondition(context.Background(), map[string]bool{})
-		Expect(conds[0].Status).To(Equal(metav1.ConditionUnknown))
+	It("is not ready when no desired pods are reported yet", func() {
+		conds := computeKeeperReadyCondition(context.Background(), 3, map[string]bool{})
+		Expect(conds[0].Status).To(Equal(metav1.ConditionFalse))
+	})
+
+	It("is not ready when fewer pods are reported than desired", func() {
+		conds := computeKeeperReadyCondition(context.Background(), 3, map[string]bool{"a": true})
+		Expect(conds[0].Status).To(Equal(metav1.ConditionFalse))
+		Expect(conds[0].Message).To(ContainSubstring("1 of 3 expected"))
 	})
 })

internal/controller/infra/managed/clickhouse/altinity/keeper/spec.go

@@ -42,6 +42,8 @@ func ToKeeperVendorSpec(
 	}
 
 	labels := common.BuildWandbLabels(wandb, KeeperModuleName)
+	settings := chiv1.NewSettings()
+	settings.Set("keeper_server/enable_reconfiguration", chiv1.NewSettingScalar("true"))
 
 	podSpec := corev1.PodSpec{
 		SecurityContext: keeperPodSecurityContext(),
@@ -70,6 +72,7 @@ func ToKeeperVendorSpec(
 		},
 		Spec: chkv1.ChkSpec{
 			Configuration: &chkv1.Configuration{
+				Settings: settings,
 				Clusters: []*chkv1.Cluster{
 					{
 						Name: ClusterName,

internal/controller/infra/managed/clickhouse/altinity/keeper/spec_test.go

@@ -41,6 +41,7 @@ var _ = Describe("Keeper vendor spec", func() {
 
 		Expect(chk.Spec.Configuration.Clusters).To(HaveLen(1))
 		Expect(chk.Spec.Configuration.Clusters[0].Layout.ReplicasCount).To(Equal(5))
+		Expect(chk.Spec.Configuration.Settings.Get("keeper_server/enable_reconfiguration").String()).To(Equal("true"))
 
 		Expect(chk.Spec.Templates.VolumeClaimTemplates).To(HaveLen(1))
 		storage := chk.Spec.Templates.VolumeClaimTemplates[0].Spec.Resources.Requests[corev1.ResourceStorage]