[Minor Improvements] Vllmruntime Autoscaling in Operator

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

Author: aeon-x

operator/api/v1alpha1/vllmruntime_types.go

@@ -328,6 +328,15 @@ type VLLMRuntimeStatus struct {
 	// Current replica count (used by scale subresource)
 	Replicas int32 `json:"replicas,omitempty"`
 
+	// Number of pods that are ready and serving traffic
+	AvailableReplicas int32 `json:"availableReplicas,omitempty"`
+
+	// Number of pods running the latest pod template
+	UpdatedReplicas int32 `json:"updatedReplicas,omitempty"`
+
+	// Number of pods that are not yet available
+	UnavailableReplicas int32 `json:"unavailableReplicas,omitempty"`
+
 	// Label selector for pods (used by scale subresource for HPA AverageValue)
 	Selector string `json:"selector,omitempty"`
 }

operator/config/crd/bases/production-stack.vllm.ai_vllmruntimes.yaml

@@ -546,6 +546,10 @@ spec:
           status:
             description: VLLMRuntimeStatus defines the observed state of VLLMRuntime
             properties:
+              availableReplicas:
+                description: Number of pods that are ready and serving traffic
+                format: int32
+                type: integer
               lastUpdated:
                 description: Last updated timestamp
                 format: date-time
@@ -561,6 +565,14 @@ spec:
                 description: Label selector for pods (used by scale subresource for
                   HPA AverageValue)
                 type: string
+              unavailableReplicas:
+                description: Number of pods that are not yet available
+                format: int32
+                type: integer
+              updatedReplicas:
+                description: Number of pods running the latest pod template
+                format: int32
+                type: integer
             type: object
         type: object
     served: true

operator/internal/controller/vllmruntime_controller.go

@@ -21,6 +21,7 @@ import (
 	"fmt"
 	"maps"
 	"reflect"
+	"strings"
 
 	appsv1 "k8s.io/api/apps/v1"
 	corev1 "k8s.io/api/core/v1"
@@ -1066,6 +1067,11 @@ func (r *VLLMRuntimeReconciler) updateStatus(
 		latestVR.Status.Replicas = dep.Status.Replicas
 		latestVR.Status.Selector = metav1.FormatLabelSelector(dep.Spec.Selector)
 
+		// Expose additional deployment status fields
+		latestVR.Status.AvailableReplicas = dep.Status.AvailableReplicas
+		latestVR.Status.UpdatedReplicas = dep.Status.UpdatedReplicas
+		latestVR.Status.UnavailableReplicas = dep.Status.UnavailableReplicas
+
 		return r.Status().Update(ctx, latestVR)
 	})
 }
@@ -1089,6 +1095,13 @@ func (r *VLLMRuntimeReconciler) reconcileScaledObject(
 
 	prometheusAddr := cfg.Triggers.PrometheusAddress
 	servedModelName := vllmRuntime.Spec.Model.ModelURL
+	// Check extraArgs for --served-model-name override
+	for _, arg := range vllmRuntime.Spec.VLLMConfig.ExtraArgs {
+		if strings.HasPrefix(arg, "--served-model-name=") {
+			servedModelName = strings.TrimPrefix(arg, "--served-model-name=")
+			break
+		}
+	}
 
 	spec := map[string]interface{}{
 		"scaleTargetRef": map[string]interface{}{
@@ -1125,7 +1138,7 @@ func (r *VLLMRuntimeReconciler) reconcileScaledObject(
 				"metadata": map[string]string{
 					"serverAddress": prometheusAddr,
 					"metricName":    "vllm_incoming_keepalive",
-					"query":         fmt.Sprintf(`sum(rate(vllm:num_incoming_requests_total{namespace="%s", model="%s"}[1m]) > bool 0)`, vllmRuntime.Namespace, servedModelName),
+					"query":         fmt.Sprintf(`sum(rate(vllm:num_incoming_requests_total{namespace="%s", model="%s"}[2m]) > bool 0)`, vllmRuntime.Namespace, servedModelName),
 					"threshold":     "1",
 				},
 			},