FEAT(observability): enable HA with replica 2 and soft anti-affinity

- Add replicaCount: 2 to goldilocks, vpa, alertmanager
- Add replicas: 2 to loki singleBinary
- Add soft pod anti-affinity for node distribution
- Keep kube-state-metrics at replica 1 to prevent duplicate metrics

FIX(loki): revert to replica 1 for Single Binary mode

- Single Binary mode cannot run more than 1 replica without object storage
- Remove affinity configuration for single replica
- Keep filesystem storage backend

alertmanager/argocd.yaml

@@ -65,7 +65,7 @@ spec:
           repository: ghcr.io/prymitive/karma
           tag: v0.122
 
-        replicaCount: 1
+        replicaCount: 2
 
         resources:
           requests:
@@ -82,6 +82,16 @@ spec:
           - name: ALERTMANAGER_URI
             value: "http://alertmanager.alertmanager.svc.cluster.local:9093"
 
+        affinity:
+          podAntiAffinity:
+            preferredDuringSchedulingIgnoredDuringExecution:
+              - weight: 100
+                podAffinityTerm:
+                  labelSelector:
+                    matchLabels:
+                      app: karma
+                  topologyKey: kubernetes.io/hostname
+
         livenessProbe:
           httpGet:
             path: /health

goldilocks/helm-values.yaml

@@ -4,7 +4,7 @@
 # Dashboard configuration
 dashboard:
   enabled: true
-  replicaCount: 1
+  replicaCount: 2
 
   resources:
     requests:
@@ -52,3 +52,15 @@ controller:
 vpa:
   # Set to false since we're installing VPA separately
   enabled: false
+
+# Affinity - Soft Anti-Affinity to spread pods across nodes
+dashboard:
+  affinity:
+    podAntiAffinity:
+      preferredDuringSchedulingIgnoredDuringExecution:
+        - weight: 100
+          podAffinityTerm:
+            labelSelector:
+              matchLabels:
+                app.kubernetes.io/name: goldilocks
+            topologyKey: kubernetes.io/hostname

kube-state-metrics/helm-values.yaml

@@ -3,6 +3,10 @@
 
 fullnameOverride: kube-state-metrics
 
+# replicaCount: 1 (default)
+# Note: kube-state-metrics는 stateless이지만, 여러 replica는 동일한 메트릭을 중복 생성하므로
+# 단일 replica로 실행하는 것이 권장됩니다.
+
 resources:
   requests:
     cpu: 15m

loki/helm-values.yaml

@@ -37,6 +37,7 @@ deploymentMode: SingleBinary
 
 singleBinary:
   replicas: 1
+  # Note: Single Binary mode cannot run more than 1 replica without object storage backend
   persistence:
     enabled: true
     size: 5Gi  # Reduced from 10Gi to save storage
@@ -47,6 +48,7 @@ singleBinary:
       memory: 462Mi
     limits:
       memory: 462Mi
+  # Note: Affinity not needed for single replica
 
 # Disable components not needed in single binary mode
 backend:

vpa/helm-values.yaml

@@ -4,7 +4,7 @@
 # Recommender - recommends resource requests based on historical usage
 recommender:
   enabled: true
-  replicaCount: 1
+  replicaCount: 2
 
   resources:
     requests:
@@ -25,3 +25,15 @@ admissionController:
 
 # Optional: Configure default VPA behavior
 # metrics-server should be installed in the cluster
+
+# Affinity - Soft Anti-Affinity to spread pods across nodes
+recommender:
+  affinity:
+    podAntiAffinity:
+      preferredDuringSchedulingIgnoredDuringExecution:
+        - weight: 100
+          podAffinityTerm:
+            labelSelector:
+              matchLabels:
+                app: vpa-recommender
+            topologyKey: kubernetes.io/hostname