update helm chart to use preferred backends for gke

config/charts/llm-d-router-gateway/templates/gke.yaml

@@ -58,8 +58,30 @@ spec:
       grpcHealthCheck:
         portSpecification: "USE_FIXED_PORT"
         port: {{ $eppHealthPort }}
+{{- if .Values.router.epp.gkePreferredBackends.enabled }}
+---
+kind: HealthCheckPolicy
+apiVersion: networking.gke.io/v1
+metadata:
+  name: {{ printf "%s-backup" (include "llm-d-router.name" .) }}
+  namespace: {{ .Release.Namespace }}
+  labels:
+    {{- include "llm-d-router.labels" . | nindent 4 }}
+spec:
+  targetRef:
+    group: ""
+    kind: Service
+    name: {{ printf "%s-backup" (include "llm-d-router.name" .) }}
+  default:
+    config:
+      type: GRPC
+      grpcHealthCheck:
+        portSpecification: "USE_FIXED_PORT"
+        port: {{ $eppHealthPort }}
+{{- end }}
 {{- end }}
 ---
+{{- if not .Values.router.epp.gkePreferredBackends.enabled }}
 apiVersion: networking.gke.io/v1
 kind: GCPBackendPolicy
 metadata:
@@ -78,4 +100,52 @@ spec:
       enabled: true    # log all requests by default
 ---
 {{- end }}
+{{- if .Values.router.epp.gkePreferredBackends.enabled }}
+apiVersion: networking.gke.io/v1
+kind: GCPBackendPolicy
+metadata:
+  name: {{ .Release.Name }}
+  namespace: {{ .Release.Namespace }}
+spec:
+  targetRef:
+    group: ""
+    kind: Service
+    name: {{ include "llm-d-router.name" . }}
+  default:
+    timeoutSec: 300
+    backendPreference: PREFERRED
+    {{- if .Values.router.epp.gkePreferredBackends.balancingMode }}
+    balancingMode: {{ .Values.router.epp.gkePreferredBackends.balancingMode }}
+    {{- end }}
+    {{- if .Values.router.epp.gkePreferredBackends.maxRatePerEndpoint }}
+    maxRatePerEndpoint: {{ .Values.router.epp.gkePreferredBackends.maxRatePerEndpoint }}
+    {{- end }}
+    {{- if .Values.router.epp.gkePreferredBackends.capacityScalerPercent }}
+    capacityScalerPercent: {{ .Values.router.epp.gkePreferredBackends.capacityScalerPercent }}
+    {{- end }}
+---
+apiVersion: networking.gke.io/v1
+kind: GCPBackendPolicy
+metadata:
+  name: {{ printf "%s-backup" .Release.Name }}
+  namespace: {{ .Release.Namespace }}
+spec:
+  targetRef:
+    group: ""
+    kind: Service
+    name: {{ printf "%s-backup" (include "llm-d-router.name" .) }}
+  default:
+    timeoutSec: 300
+    backendPreference: DEFAULT
+    {{- if .Values.router.epp.gkePreferredBackends.balancingMode }}
+    balancingMode: {{ .Values.router.epp.gkePreferredBackends.balancingMode }}
+    {{- end }}
+    {{- if .Values.router.epp.gkePreferredBackends.maxRatePerEndpoint }}
+    maxRatePerEndpoint: {{ .Values.router.epp.gkePreferredBackends.maxRatePerEndpoint }}
+    {{- end }}
+    {{- if .Values.router.epp.gkePreferredBackends.capacityScalerPercent }}
+    capacityScalerPercent: {{ .Values.router.epp.gkePreferredBackends.capacityScalerPercent }}
+    {{- end }}
+{{- end }}
+{{- end }}
 {{- include "llm-d-router.gke" . -}}

config/charts/routerlib/templates/_deployment.yaml

@@ -1,31 +1,4 @@
-{{- define "llm-d-epp.deployment" -}}
-apiVersion: apps/v1
-kind: Deployment
-metadata:
-  name: {{ include "llm-d-router.name" . }}
-  namespace: {{ .Release.Namespace }}
-  labels:
-    {{- include "llm-d-router.labels" . | nindent 4 }}
-    {{- include "llm-d-router.modeLabels" . | nindent 4 }}
-spec:
-  replicas: {{ .Values.router.epp.replicas | default 1 }}
-  strategy:
-    # The current recommended EPP deployment pattern is to have a single active replica. This ensures
-    # optimal performance of the stateful operations such prefix cache aware scorer.
-    # The Recreate strategy the old replica is killed immediately, and allow the new replica(s) to
-    # quickly take over. This is particularly important in the high availability set up with leader
-    # election, as the rolling update strategy would prevent the old leader being killed because
-    # otherwise the maxUnavailable would be 100%.
-    type: Recreate
-  selector:
-    matchLabels:
-      {{- include "llm-d-router.selectorLabels" . | nindent 6 }}
-  template:
-    metadata:
-      labels:
-        {{- include "llm-d-router.selectorLabels" . | nindent 8 }}
-        {{- include "llm-d-router.modeLabels" . | nindent 8 }}
-    spec:
+{{- define "llm-d-epp.deployment-pod-spec" -}}
       {{- $proxy := include "llm-d-router.proxy" . | fromYaml | default dict }}
       {{- $proxyType := include "llm-d-router.proxyType" . | trim }}
       {{- $proxyMode := include "llm-d-router.proxyMode" . | trim }}
@@ -129,7 +102,7 @@ spec:
               - "json"
               - --config-file
               - "/config/{{ .Values.router.epp.pluginsConfigFile }}"
-          {{- if gt (.Values.router.epp.replicas | int) 1 }}
+          {{- if and (gt (.Values.router.epp.replicas | int) 1) (not .Values.router.epp.gkePreferredBackends.enabled) }}
               - --ha-enable-leader-election
           {{- end }}
           {{- $grpcHealthPort := .Values.router.epp.grpcHealthPort | default 9003 }}
@@ -305,5 +278,67 @@ spec:
       tolerations:
         {{- toYaml .Values.router.epp.tolerations | nindent 8 }}
       {{- end }}
+{{- end }}
+
+{{- define "llm-d-epp.deployment" -}}
+apiVersion: apps/v1
+kind: Deployment
+metadata:
+  name: {{ include "llm-d-router.name" . }}
+  namespace: {{ .Release.Namespace }}
+  labels:
+    {{- include "llm-d-router.labels" . | nindent 4 }}
+    {{- include "llm-d-router.modeLabels" . | nindent 4 }}
+spec:
+  {{- $replicas := .Values.router.epp.replicas | default 1 }}
+  {{- if .Values.router.epp.gkePreferredBackends.enabled }}
+    {{- $replicas = .Values.router.epp.gkePreferredBackends.preferredReplicas | default 1 }}
+  {{- end }}
+  replicas: {{ $replicas }}
+  strategy:
+    type: Recreate
+  selector:
+    matchLabels:
+      {{- if .Values.router.epp.gkePreferredBackends.enabled }}
+      llm-d-router-gateway: {{ include "llm-d-router.name" . }}
+      {{- else }}
+      {{- include "llm-d-router.selectorLabels" . | nindent 6 }}
+      {{- end }}
+  template:
+    metadata:
+      labels:
+        {{- if .Values.router.epp.gkePreferredBackends.enabled }}
+        llm-d-router-gateway: {{ include "llm-d-router.name" . }}
+        {{- else }}
+        {{- include "llm-d-router.selectorLabels" . | nindent 8 }}
+        {{- end }}
+        {{- include "llm-d-router.modeLabels" . | nindent 8 }}
+    spec:
+{{ include "llm-d-epp.deployment-pod-spec" . }}
+{{- if .Values.router.epp.gkePreferredBackends.enabled }}
+---
+apiVersion: apps/v1
+kind: Deployment
+metadata:
+  name: {{ printf "%s-backup" (include "llm-d-router.name" .) }}
+  namespace: {{ .Release.Namespace }}
+  labels:
+    {{- include "llm-d-router.labels" . | nindent 4 }}
+    {{- include "llm-d-router.modeLabels" . | nindent 4 }}
+spec:
+  replicas: {{ .Values.router.epp.gkePreferredBackends.defaultReplicas | default 1 }}
+  strategy:
+    type: Recreate
+  selector:
+    matchLabels:
+      llm-d-router-gateway: {{ printf "%s-backup" (include "llm-d-router.name" .) }}
+  template:
+    metadata:
+      labels:
+        llm-d-router-gateway: {{ printf "%s-backup" (include "llm-d-router.name" .) }}
+        {{- include "llm-d-router.modeLabels" . | nindent 8 }}
+    spec:
+{{ include "llm-d-epp.deployment-pod-spec" . }}
+{{- end }}
 ---
 {{- end }}

config/charts/routerlib/templates/_helpers.tpl

@@ -397,11 +397,25 @@ EPP resource validations
 {{/*
 EPP generic validations
 */}}
+{{- define "llm-d-router.validations.epp.preferredBackends" -}}
+{{- if .Values.router.epp.gkePreferredBackends.enabled }}
+  {{- $preferredReplicas := .Values.router.epp.gkePreferredBackends.preferredReplicas | default 1 | int }}
+  {{- $defaultReplicas := .Values.router.epp.gkePreferredBackends.defaultReplicas | default 1 | int }}
+  {{- if lt $preferredReplicas 1 }}
+    {{- fail ".Values.router.epp.gkePreferredBackends.preferredReplicas must be at least 1 when gkePreferredBackends.enabled is true" }}
+  {{- end }}
+  {{- if lt $defaultReplicas 1 }}
+    {{- fail ".Values.router.epp.gkePreferredBackends.defaultReplicas must be at least 1 when gkePreferredBackends.enabled is true" }}
+  {{- end }}
+{{- end }}
+{{- end -}}
+
 {{- define "llm-d-router.validations.epp" -}}
 {{- include "llm-d-router.validations.deprecations" . }}
 {{- include "llm-d-router.validations.epp.resources" . }}
 {{- include "llm-d-router.validations.epp.inferenceObjectives" . }}
 {{- include "llm-d-router.validations.epp.tokenizer" . }}
+{{- include "llm-d-router.validations.epp.preferredBackends" . }}
 {{- end -}}
 
 {{/*

config/charts/routerlib/templates/_inferencepool.yaml

@@ -26,7 +26,11 @@ spec:
       {{- end }}
       {{- end }}
   endpointPickerRef:
+    {{- if .Values.router.epp.gkePreferredBackends.enabled }}
+    name: {{ printf "%s-backup" (include "llm-d-router.name" .) }}
+    {{- else }}
     name: {{ include "llm-d-router.name" . }}
+    {{- end }}
     port:
       number: {{ .Values.router.epp.extProcPort | default 9002 }}
     failureMode: {{ .Values.router.inferencePool.failureMode | default "FailOpen" }}

config/charts/routerlib/templates/_service.yaml

@@ -6,13 +6,22 @@ metadata:
   namespace: {{ .Release.Namespace }}
   labels:
     {{- include "llm-d-router.labels" . | nindent 4 }}
+  {{- if .Values.router.epp.gkePreferredBackends.enabled }}
+  annotations:
+    cloud.google.com/neg: '{"exposed_ports":{"{{ .Values.router.epp.extProcPort | default 9002 }}":{}}}'
+  {{- end }}
 spec:
   selector:
+    {{- if .Values.router.epp.gkePreferredBackends.enabled }}
+    llm-d-router-gateway: {{ include "llm-d-router.name" . }}
+    {{- else }}
     {{- include "llm-d-router.selectorLabels" . | nindent 4 }}
+    {{- end }}
   ports:
     - name: grpc-ext-proc
       protocol: TCP
       port: {{ .Values.router.epp.extProcPort | default 9002 }}
+      appProtocol: kubernetes.io/h2c
     - name: http-metrics
       protocol: TCP
       port: {{ .Values.router.metricsPort | default 9090 }}
@@ -21,4 +30,24 @@ spec:
     {{- end }}
   type: ClusterIP
 ---
+{{- if .Values.router.epp.gkePreferredBackends.enabled }}
+apiVersion: v1
+kind: Service
+metadata:
+  name: {{ printf "%s-backup" (include "llm-d-router.name" .) }}
+  namespace: {{ .Release.Namespace }}
+  labels:
+    {{- include "llm-d-router.labels" . | nindent 4 }}
+  annotations:
+    cloud.google.com/neg: '{"exposed_ports":{"{{ .Values.router.epp.extProcPort | default 9002 }}":{}}}'
+spec:
+  selector:
+    llm-d-router-gateway: {{ printf "%s-backup" (include "llm-d-router.name" .) }}
+  ports:
+    - name: grpc-ext-proc
+      protocol: TCP
+      port: {{ .Values.router.epp.extProcPort | default 9002 }}
+      appProtocol: kubernetes.io/h2c
+  type: ClusterIP
+{{- end }}
 {{- end }}

config/charts/routerlib/values.yaml

@@ -5,6 +5,19 @@ clusterDomain: cluster.local
 
 epp:
   replicas: 1
+  gkePreferredBackends:
+    enabled: false
+    # Number of primary active pod instances pinned to the PREFERRED load balancing tier. 
+    # Setting >1 scales concurrent active routing.
+    preferredReplicas: 1
+    # Number of standby pod instances pinned to the DEFAULT load balancing tier. 
+    # Setting >1 scales warm standby failover capacity.
+    defaultReplicas: 1
+    # Benchmark capacity and threshold settings for PREFERRED load balancing tier.
+    # When incoming traffic exceeds capacityScalerPercent of maxRatePerEndpoint, excess traffic spills over to DEFAULT standby pods.
+    balancingMode: RATE
+    maxRatePerEndpoint: 100
+    capacityScalerPercent: 100
   image:
     registry: ghcr.io/llm-d
     repository: llm-d-router-endpoint-picker-dev