Scale-to-zero enforcement with e2e tests and HPA configuration

.github/workflows/ci-e2e-openshift.yaml

@@ -464,6 +464,8 @@ jobs:
           # HF_TOKEN is inherited from GITHUB_ENV (set in 'Get HF token from cluster secret' step)
           ENVIRONMENT: openshift
           INSTALL_GATEWAY_CTRLPLANE: "false"
+          # Disable benchmark mode - istioBench environment not available in llm-d helmfile
+          BENCHMARK_MODE: "false"
           E2E_TESTS_ENABLED: "true"
           NAMESPACE_SCOPED: "false"
           # Pass PR-specific namespaces to install script
@@ -473,6 +475,8 @@ jobs:
           CONTROLLER_INSTANCE: ${{ env.WVA_NAMESPACE }}
           # vLLM max-num-seqs for e2e testing (lower = easier to saturate)
           VLLM_MAX_NUM_SEQS: ${{ env.MAX_NUM_SEQS }}
+          # Disable benchmark mode - istioBench environment not available in llm-d helmfile
+          BENCHMARK_MODE: "false"
         run: |
           echo "Deploying WVA and llm-d infrastructure..."
           echo "  MODEL_ID: $MODEL_ID"
@@ -513,6 +517,8 @@ jobs:
           # HF_TOKEN is inherited from GITHUB_ENV
           ENVIRONMENT: openshift
           INSTALL_GATEWAY_CTRLPLANE: "false"
+          # Disable benchmark mode - istioBench environment not available in llm-d helmfile
+          BENCHMARK_MODE: "false"
           E2E_TESTS_ENABLED: "true"
           NAMESPACE_SCOPED: "false"
           # Override namespaces for Model B stack
@@ -526,6 +532,8 @@ jobs:
           DEPLOY_HPA: "false"
           # vLLM max-num-seqs for e2e testing (lower = easier to saturate)
           VLLM_MAX_NUM_SEQS: ${{ env.MAX_NUM_SEQS }}
+          # Disable benchmark mode - istioBench environment not available in llm-d helmfile
+          BENCHMARK_MODE: "false"
         run: |
           echo "Deploying Model B infrastructure in $LLMD_NAMESPACE_B..."
           echo "  MODEL_ID: $MODEL_ID"

.github/workflows/ci-pr-checks.yaml

@@ -44,6 +44,32 @@ jobs:
         run: |
           make test
 
+  # E2E tests run in parallel with different HPAScaleToZero feature gate settings
+  e2e-tests:
+    runs-on: ubuntu-latest
+    needs: lint-and-test
+    timeout-minutes: 60
+    strategy:
+      fail-fast: false
+      matrix:
+        scale_to_zero: [true, false]
+    name: e2e-tests (scale-to-zero=${{ matrix.scale_to_zero }})
+    steps:
+      - name: Checkout source
+        uses: actions/checkout@v4
+
+      - name: Extract Go version from go.mod
+        run: sed -En 's/^go (.*)$/GO_VERSION=\1/p' go.mod >> $GITHUB_ENV
+
+      - name: Set up Go with cache
+        uses: actions/setup-go@v5
+        with:
+          go-version: "${{ env.GO_VERSION }}"
+          cache-dependency-path: ./go.sum
+
+      - name: Install dependencies
+        run: go mod download
+
       - name: Install Kind
         run: |
           ARCH=$(uname -m)
@@ -57,7 +83,9 @@ jobs:
           sudo mv ./kind /usr/local/bin/kind
           kind version
 
-      - name: Run make test-e2e
+      - name: Run e2e tests
         shell: bash
+        env:
+          ENABLE_SCALE_TO_ZERO: ${{ matrix.scale_to_zero }}
         run: |
           make test-e2e

Makefile

@@ -153,9 +153,9 @@ test-e2e: manifests generate fmt vet ## Run the e2e tests. Expected an isolated
 		echo "Kind is not installed. Please install Kind manually."; \
 		exit 1; \
 	}
-	$(eval FOCUS_ARGS := $(if $(FOCUS),-ginkgo.focus="$(FOCUS)",-ginkgo.focus="Saturation Mode"))
+	$(eval FOCUS_ARGS := $(if $(FOCUS),-ginkgo.focus="$(FOCUS)",))
 	$(eval SKIP_ARGS := $(if $(SKIP),-ginkgo.skip="$(SKIP)",))
-	export COLLECTOR_V2=1 KUBECONFIG=$(KUBECONFIG) K8S_EXPECTED_VERSION=$(K8S_VERSION) && go test ./test/e2e-saturation-based/ -timeout 30m -v -ginkgo.v $(FOCUS_ARGS) $(SKIP_ARGS)
+	export COLLECTOR_V2=1 KUBECONFIG=$(KUBECONFIG) K8S_EXPECTED_VERSION=$(K8S_VERSION) && go test ./test/e2e-saturation-based/ -timeout 50m -v -ginkgo.v $(FOCUS_ARGS) $(SKIP_ARGS)
 
 # E2E tests on OpenShift cluster
 # Supports KUBECONFIG or in-cluster authentication (for self-hosted runners).
@@ -173,7 +173,7 @@ test-e2e-openshift: ## Run the e2e tests on OpenShift. Supports KUBECONFIG or in
 	DEPLOYMENT=$(DEPLOYMENT) \
 	REQUEST_RATE=$(REQUEST_RATE) \
 	NUM_PROMPTS=$(NUM_PROMPTS) \
-	go test ./test/e2e-openshift/ -timeout 30m -v -ginkgo.v $(FOCUS_ARGS) $(SKIP_ARGS)
+	go test ./test/e2e-openshift/ -timeout 50m -v -ginkgo.v $(FOCUS_ARGS) $(SKIP_ARGS)
 
 .PHONY: lint
 lint: golangci-lint ## Run golangci-lint linter

api/v1alpha1/variantautoscaling_types.go

@@ -53,7 +53,7 @@ type OptimizedAlloc struct {
 	Accelerator string `json:"accelerator"`
 
 	// NumReplicas is the number of replicas for the optimized allocation.
-	// +kubebuilder:validation:Minimum=1
+	// +kubebuilder:validation:Minimum=0
 	NumReplicas int `json:"numReplicas"`
 }
 

charts/workload-variant-autoscaler/crds/llmd.ai_variantautoscalings.yaml

@@ -187,7 +187,7 @@ spec:
                   numReplicas:
                     description: NumReplicas is the number of replicas for the optimized
                       allocation.
-                    minimum: 1
+                    minimum: 0
                     type: integer
                 required:
                 - accelerator

charts/workload-variant-autoscaler/templates/hpa.yaml

@@ -11,7 +11,7 @@ spec:
     apiVersion: apps/v1
     kind: Deployment
     name: {{ printf "%s-decode" .Values.llmd.modelName }}
-  # minReplicas: 0  # scale to zero - alpha feature
+  minReplicas: {{ .Values.hpa.minReplicas }}
   maxReplicas: {{ .Values.hpa.maxReplicas }}
   behavior:
     scaleUp:

charts/workload-variant-autoscaler/values-dev.yaml

@@ -66,6 +66,8 @@ va:
   sloTtft: 1000
 hpa:
   enabled: true
+  # minReplicas: 0 for scale-to-zero testing (requires HPAScaleToZero feature gate)
+  minReplicas: 0
   maxReplicas: 10
   targetAverageValue: "1"
 vllmService:

charts/workload-variant-autoscaler/values.yaml

@@ -84,6 +84,10 @@ va:
 
 hpa:
   enabled: true
+  # minReplicas: 0 enables scale-to-zero (requires HPAScaleToZero feature gate in k8s)
+  # minReplicas: 1 is the safe default that prevents scale-to-zero
+  # Set to 0 when wva.scaleToZero is enabled
+  minReplicas: 1
   maxReplicas: 10
   targetAverageValue: "1"
   # HPA scaling behavior configuration

cmd/main.go

@@ -43,7 +43,6 @@ import (
 	"sigs.k8s.io/controller-runtime/pkg/webhook"
 
 	llmdVariantAutoscalingV1alpha1 "github.com/llm-d-incubation/workload-variant-autoscaler/api/v1alpha1"
-	"github.com/llm-d-incubation/workload-variant-autoscaler/internal/collector/registration"
 	"github.com/llm-d-incubation/workload-variant-autoscaler/internal/collector/source"
 	"github.com/llm-d-incubation/workload-variant-autoscaler/internal/collector/source/prometheus"
 	"github.com/llm-d-incubation/workload-variant-autoscaler/internal/config"
@@ -379,9 +378,6 @@ func main() {
 			os.Exit(1)
 		}
 
-		// Register scale-to-zero queries with the prometheus source
-		registration.RegisterScaleToZeroQueries(sourceRegistry)
-
 		engine := saturation.NewEngine(
 			mgr.GetClient(),
 			mgr.GetScheme(),

config/crd/bases/llmd.ai_variantautoscalings.yaml

@@ -187,7 +187,7 @@ spec:
                   numReplicas:
                     description: NumReplicas is the number of replicas for the optimized
                       allocation.
-                    minimum: 1
+                    minimum: 0
                     type: integer
                 required:
                 - accelerator

deploy/install.sh

@@ -57,7 +57,7 @@ TTFT_AVERAGE_LATENCY_MS=${TTFT_AVERAGE_LATENCY_MS:-200}
 
 # Gateway Configuration
 GATEWAY_PROVIDER=${GATEWAY_PROVIDER:-"istio"} # Options: kgateway, istio
-BENCHMARK_MODE=${BENCHMARK_MODE:-"true"} # if true, updates to Istio config for benchmark
+BENCHMARK_MODE=${BENCHMARK_MODE:-"false"} # if true, updates to Istio config for benchmark (istioBench env required in helmfile)
 # Save original value to detect if explicitly set via environment variable
 INSTALL_GATEWAY_CTRLPLANE_ORIGINAL="${INSTALL_GATEWAY_CTRLPLANE:-}"
 INSTALL_GATEWAY_CTRLPLANE="${INSTALL_GATEWAY_CTRLPLANE:-false}"
@@ -81,6 +81,9 @@ DEPLOY_PROMETHEUS_ADAPTER=${DEPLOY_PROMETHEUS_ADAPTER:-true}
 DEPLOY_VA=${DEPLOY_VA:-true}
 DEPLOY_HPA=${DEPLOY_HPA:-true}
 HPA_STABILIZATION_SECONDS=${HPA_STABILIZATION_SECONDS:-240}
+# HPA minReplicas: 0 enables scale-to-zero (requires HPAScaleToZero feature gate)
+# Default to 1 for safety; set to 0 for scale-to-zero testing
+HPA_MIN_REPLICAS=${HPA_MIN_REPLICAS:-1}
 SKIP_CHECKS=${SKIP_CHECKS:-false}
 E2E_TESTS_ENABLED=${E2E_TESTS_ENABLED:-false}
 # vLLM max-num-seqs (max concurrent sequences per replica, lower = easier to saturate for testing)
@@ -145,6 +148,7 @@ Environment Variables:
   DEPLOY_VA                    Deploy VariantAutoscaling (default: true)
   DEPLOY_HPA                   Deploy HPA (default: true)
   HPA_STABILIZATION_SECONDS    HPA stabilization window in seconds (default: 240)
+  HPA_MIN_REPLICAS             HPA minReplicas (default: 1, set to 0 for scale-to-zero)
   UNDEPLOY                     Undeploy mode (default: false)
   DELETE_NAMESPACES            Delete namespaces after undeploy (default: false)
   CONTROLLER_INSTANCE          Controller instance label for multi-controller isolation (optional)
@@ -427,6 +431,7 @@ deploy_wva_controller() {
         --set va.sloTpot=$SLO_TPOT \
         --set va.sloTtft=$SLO_TTFT \
         --set hpa.enabled=$DEPLOY_HPA \
+        --set hpa.minReplicas=$HPA_MIN_REPLICAS \
         --set hpa.behavior.scaleUp.stabilizationWindowSeconds=$HPA_STABILIZATION_SECONDS \
         --set hpa.behavior.scaleDown.stabilizationWindowSeconds=$HPA_STABILIZATION_SECONDS \
         --set llmd.namespace=$LLMD_NS \

deploy/kind-emulator/setup.sh

@@ -21,6 +21,8 @@ gpu_type="$DEFAULT_GPU_TYPE"
 gpu_model="$DEFAULT_GPU_MODEL"
 gpu_memory="$DEFAULT_GPU_MEMORY"
 k8s_version="${K8S_VERSION:-$DEFAULT_K8S_VERSION}"
+# Enable HPAScaleToZero feature gate (alpha feature for scale-to-zero HPA support)
+enable_scale_to_zero="${ENABLE_SCALE_TO_ZERO:-true}"
 
 # --------------------------------------------------------------------
 # Cleanup on exit
@@ -48,7 +50,8 @@ Options:
     -h                 Show this help message
 
 Environment Variables:
-    K8S_VERSION        Kubernetes version to use (default: $DEFAULT_K8S_VERSION)
+    K8S_VERSION           Kubernetes version to use (default: $DEFAULT_K8S_VERSION)
+    ENABLE_SCALE_TO_ZERO  Enable HPAScaleToZero feature gate (default: true)
 EOF
 }
 
@@ -83,6 +86,13 @@ done
 # --------------------------------------------------------------------
 echo "[1/6] Creating Kind cluster: ${cluster_name} with ${nodes} nodes and ${gpus_per_node} GPUS each..."
 
+# Build feature gates string
+feature_gates=""
+if [ "$enable_scale_to_zero" = "true" ]; then
+    feature_gates="HPAScaleToZero=true"
+    echo "  HPAScaleToZero feature gate: enabled"
+fi
+
 cat <<EOF > kind-config.yaml
 kind: Cluster
 apiVersion: kind.x-k8s.io/v1alpha4
@@ -91,6 +101,21 @@ nodes:
   image: kindest/node:${k8s_version}
 EOF
 
+# Add kubeadmConfigPatches for feature gates if any are enabled
+if [ -n "$feature_gates" ]; then
+    cat <<EOF >> kind-config.yaml
+  kubeadmConfigPatches:
+  - |
+    kind: ClusterConfiguration
+    apiServer:
+      extraArgs:
+        feature-gates: ${feature_gates}
+    controllerManager:
+      extraArgs:
+        feature-gates: ${feature_gates}
+EOF
+fi
+
 for ((i=1; i<nodes; i++)); do
     echo "- role: worker" >> kind-config.yaml
     echo "  image: kindest/node:${k8s_version}" >> kind-config.yaml

docs/integrations/hpa-integration.md

@@ -265,6 +265,66 @@ kubectl -n kube-system get pod -l component=kube-controller-manager -o yaml | gr
 
 7. Specify the `minReplicas: 0` field in the `yaml` snippet for HPA and apply it following the integration steps
 
+### Automatic Setup with Kind Cluster Script
+
+When using the `deploy/kind-emulator/setup.sh` script, the HPAScaleToZero feature gate is **enabled by default**. You can disable it if needed:
+
+```sh
+# Default: HPAScaleToZero enabled
+./deploy/kind-emulator/setup.sh
+
+# Disable HPAScaleToZero feature gate
+ENABLE_SCALE_TO_ZERO=false ./deploy/kind-emulator/setup.sh
+```
+
+### OpenShift Cluster Configuration
+
+On OpenShift, the HPAScaleToZero feature gate must be enabled at the cluster level by a cluster administrator. This requires modifying the FeatureGate custom resource:
+
+> **Warning**: Modifying FeatureGate on OpenShift may impact cluster stability as HPAScaleToZero is a TechPreview/alpha feature. Consult your cluster administrator before making these changes.
+
+1. Check current FeatureGate configuration:
+
+```sh
+oc get featuregate cluster -o yaml
+```
+
+2. Enable HPAScaleToZero feature gate:
+
+```sh
+oc patch featuregate cluster --type=merge -p '{
+  "spec": {
+    "featureSet": "CustomNoUpgrade",
+    "customNoUpgrade": {
+      "enabled": ["HPAScaleToZero"]
+    }
+  }
+}'
+```
+
+3. Wait for the cluster to apply the changes (this may take several minutes as nodes are updated):
+
+```sh
+oc get clusterversion -w
+```
+
+4. Verify the feature gate is enabled:
+
+```sh
+oc get featuregate cluster -o jsonpath='{.spec.customNoUpgrade.enabled}'
+```
+
+5. Deploy HPA with `minReplicas: 0`:
+
+```sh
+# Using install script
+HPA_MIN_REPLICAS=0 ./deploy/install.sh -e openshift
+
+# Or via Helm
+helm upgrade workload-variant-autoscaler ./charts/workload-variant-autoscaler \
+  --set hpa.minReplicas=0
+```
+
 ### Note on possible timing issues
 
 For this discussion, please refer to the [community doc](https://docs.google.com/document/d/15z1u2HIH7qoxT-nxj4BnZ_TyqHPqIn0FcCPTnIMn7bs/edit?tab=t.0).