[Feat] Support KEDA Auto Scaling in Production Stack Operator (#903)

* init

Signed-off-by: aeon-x <talexcao@gmail.com>

* update proto

Signed-off-by: aeon-x <talexcao@gmail.com>

* fix review comments

Signed-off-by: aeon-x <talexcao@gmail.com>

* fix review comments for autoscaling

Signed-off-by: aeon-x <talexcao@gmail.com>

* fix lint

Signed-off-by: aeon-x <talexcao@gmail.com>

---------

Signed-off-by: aeon-x <talexcao@gmail.com>
Co-authored-by: Rui Zhang <51696593+ruizhang0101@users.noreply.github.com>

Author: aeon-x

operator/api/v1alpha1/vllmruntime_types.go

@@ -51,6 +51,96 @@ type DeploymentConfig struct {
 	SidecarConfig SidecarConfig `json:"sidecarConfig,omitempty"`
 }
 
+// ScaleUpPolicy defines the HPA scaling behavior when scaling up
+type ScaleUpPolicy struct {
+	// StabilizationWindowSeconds is the HPA stabilization window for scaling up
+	// +kubebuilder:default=0
+	StabilizationWindowSeconds *int32 `json:"stabilizationWindowSeconds,omitempty"`
+
+	// PodValue is the max pods to add per period
+	// +kubebuilder:default=1
+	// +kubebuilder:validation:Minimum=1
+	PodValue *int32 `json:"podValue,omitempty"`
+
+	// PeriodSeconds is the period for the scale up policy
+	// +kubebuilder:default=60
+	// +kubebuilder:validation:Minimum=1
+	PeriodSeconds *int32 `json:"periodSeconds,omitempty"`
+}
+
+// ScaleDownPolicy defines the HPA scaling behavior when scaling down
+type ScaleDownPolicy struct {
+	// StabilizationWindowSeconds is the HPA stabilization window for scaling down
+	// +kubebuilder:default=300
+	StabilizationWindowSeconds *int32 `json:"stabilizationWindowSeconds,omitempty"`
+
+	// PodValue is the max pods to remove per period
+	// +kubebuilder:default=1
+	// +kubebuilder:validation:Minimum=1
+	PodValue *int32 `json:"podValue,omitempty"`
+
+	// PeriodSeconds is the period for the scale down policy
+	// +kubebuilder:default=60
+	// +kubebuilder:validation:Minimum=1
+	PeriodSeconds *int32 `json:"periodSeconds,omitempty"`
+
+	// ScaleToZeroDelaySeconds is the wait time before scaling to zero (seconds).
+	// Only applicable when minReplicas is set to 0.
+	// +kubebuilder:default=1800
+	ScaleToZeroDelaySeconds *int32 `json:"scaleToZeroDelaySeconds,omitempty"`
+}
+
+// TriggerConfig defines the metric thresholds for autoscaling triggers
+type TriggerConfig struct {
+	// PrometheusAddress is the Prometheus server address for metric queries
+	// +kubebuilder:default="http://kube-prom-stack-kube-prome-prometheus.monitoring.svc:9090"
+	PrometheusAddress string `json:"prometheusAddress,omitempty"`
+
+	// RequestsRunningThreshold is the per-pod concurrent requests threshold for scaling
+	// +kubebuilder:default=5
+	// +kubebuilder:validation:Minimum=1
+	RequestsRunningThreshold *int32 `json:"requestsRunningThreshold,omitempty"`
+
+	// GenerationTokensThreshold is the per-pod generation tokens/s threshold for scaling
+	// +kubebuilder:default=100
+	// +kubebuilder:validation:Minimum=1
+	GenerationTokensThreshold *int32 `json:"generationTokensThreshold,omitempty"`
+
+	// PromptTokensThreshold is the per-pod prompt tokens/s threshold for scaling
+	// +kubebuilder:default=100
+	// +kubebuilder:validation:Minimum=1
+	PromptTokensThreshold *int32 `json:"promptTokensThreshold,omitempty"`
+}
+
+// AutoscalingConfig defines the KEDA autoscaling configuration
+type AutoscalingConfig struct {
+	// Enabled enables autoscaling
+	Enabled bool `json:"enabled,omitempty"`
+
+	// MinReplicas is the minimum number of replicas (defaults to 1, set to 0 to enable scale-to-zero)
+	// +kubebuilder:default=1
+	// +kubebuilder:validation:Minimum=0
+	MinReplicas *int32 `json:"minReplicas,omitempty"`
+
+	// MaxReplicas is the maximum number of replicas
+	// +kubebuilder:validation:Required
+	// +kubebuilder:validation:Minimum=1
+	MaxReplicas int32 `json:"maxReplicas"`
+
+	// PollingInterval is how often KEDA checks metrics (seconds)
+	// +kubebuilder:default=15
+	PollingInterval *int32 `json:"pollingInterval,omitempty"`
+
+	// ScaleUpPolicy defines the scaling behavior when scaling up
+	ScaleUpPolicy ScaleUpPolicy `json:"scaleUpPolicy,omitempty"`
+
+	// ScaleDownPolicy defines the scaling behavior when scaling down
+	ScaleDownPolicy ScaleDownPolicy `json:"scaleDownPolicy,omitempty"`
+
+	// Triggers defines the metric thresholds for autoscaling
+	Triggers TriggerConfig `json:"triggers,omitempty"`
+}
+
 // VLLMRuntimeSpec defines the desired state of VLLMRuntime
 type VLLMRuntimeSpec struct {
 	// Model configuration
@@ -67,6 +157,9 @@ type VLLMRuntimeSpec struct {
 
 	// Deployment configuration
 	DeploymentConfig DeploymentConfig `json:"deploymentConfig"`
+
+	// Autoscaling configuration
+	AutoscalingConfig *AutoscalingConfig `json:"autoscalingConfig,omitempty"`
 }
 
 // VLLMConfig defines the vLLM server configuration
@@ -231,10 +324,17 @@ type VLLMRuntimeStatus struct {
 
 	// Last updated timestamp
 	LastUpdated metav1.Time `json:"lastUpdated,omitempty"`
+
+	// Current replica count (used by scale subresource)
+	Replicas int32 `json:"replicas,omitempty"`
+
+	// Label selector for pods (used by scale subresource for HPA AverageValue)
+	Selector string `json:"selector,omitempty"`
 }
 
 // +kubebuilder:object:root=true
 // +kubebuilder:subresource:status
+// +kubebuilder:subresource:scale:specpath=.spec.deploymentConfig.replicas,statuspath=.status.replicas,selectorpath=.status.selector
 // +kubebuilder:resource:shortName=vr
 
 // VLLMRuntime is the Schema for the vllmruntimes API