fix scaling restriciton for pending state

Author: dittops

api/scaler/v1alpha1/budaiscaler_types.go

@@ -467,15 +467,15 @@ type StartingPodsConfig struct {
 	StartingPodWeight *string `json:"startingPodWeight,omitempty"`
 
 	// MaxStartingPods is the maximum number of starting pods allowed before
-	// gating further scale-up operations. Set to 0 to disable the gate.
-	// Default: 0 (disabled)
+	// gating further scale-up operations. Set to 0 to disable this gate.
+	// Default: 0 (disabled, uses MaxStartingPodPercent instead)
 	// +optional
 	// +kubebuilder:validation:Minimum=0
 	MaxStartingPods *int32 `json:"maxStartingPods,omitempty"`
 
 	// MaxStartingPodPercent is the maximum percentage of total pods that can
-	// be in starting state before gating scale-up.
-	// Set to 0 to disable. Default: 0 (disabled)
+	// be in starting state before gating scale-up. Set to 0 to disable.
+	// Default: 50 (gate scale-up if more than 50% of pods are starting)
 	// +optional
 	// +kubebuilder:validation:Minimum=0
 	// +kubebuilder:validation:Maximum=100

pkg/context/context.go

@@ -38,6 +38,12 @@ const (
 	DefaultPanicWindow              = 60 * time.Second
 	DefaultStableWindow             = 180 * time.Second
 	DefaultActivationScale          = int32(1)
+
+	// Starting pods defaults - enabled by default for LLM workloads with long cold starts
+	DefaultStartingPodWeight     = 0.5 // Count starting pods as 50% capacity
+	DefaultMaxStartingPods       = 0   // Disabled (use percent-based instead)
+	DefaultMaxStartingPodPercent = 50  // Gate scale-up if >50% of pods are starting
+	DefaultBypassGateOnPanic     = false
 )
 
 // ScalingContext provides configuration for scaling algorithms.
@@ -204,11 +210,11 @@ func NewBaseScalingContext() ScalingContext {
 		scaleDownPolicies:        nil,
 		scaleUpSelectPolicy:      scalerv1alpha1.MaxChangePolicySelect,
 		scaleDownSelectPolicy:    scalerv1alpha1.MinChangePolicySelect,
-		// Starting pods defaults
-		startingPodWeight:     0.5,   // Count starting pods as 50% capacity
-		maxStartingPods:       0,     // Disabled by default
-		maxStartingPodPercent: 0,     // Disabled by default
-		bypassGateOnPanic:     false, // Don't bypass gate on panic (important for LLM workloads)
+		// Starting pods defaults - enabled by default for LLM workloads
+		startingPodWeight:     DefaultStartingPodWeight,
+		maxStartingPods:       DefaultMaxStartingPods,
+		maxStartingPodPercent: DefaultMaxStartingPodPercent,
+		bypassGateOnPanic:     DefaultBypassGateOnPanic
 	}
 }
 

pkg/controller/budaiscaler/autoscaler.go

@@ -132,7 +132,18 @@ func (a *AutoScaler) Scale(ctx context.Context, scaler *scalerv1alpha1.BudAIScal
 
 	// Build scaling request
 	readyPodCount := a.countReadyPods(pods)
-	startingPodCount := a.countStartingPods(pods)
+	runningNotReadyCount := a.countStartingPods(pods)
+
+	// Count pending pods - these are also "starting" as they represent
+	// capacity that's been requested but not yet available
+	pendingPodCount, err := a.countPendingPods(ctx, scaler, scale)
+	if err != nil {
+		klog.V(4).InfoS("Failed to count pending pods", "scaler", scaler.Name, "error", err)
+		pendingPodCount = 0
+	}
+
+	// Total starting pods = running-but-not-ready + pending
+	startingPodCount := runningNotReadyCount + pendingPodCount
 
 	request := algorithm.ScalingRequest{
 		Scaler:           scaler,
@@ -145,7 +156,8 @@ func (a *AutoScaler) Scale(ctx context.Context, scaler *scalerv1alpha1.BudAIScal
 	}
 
 	klog.V(4).InfoS("Pod counts", "scaler", scaler.Name,
-		"readyPods", readyPodCount, "startingPods", startingPodCount,
+		"readyPods", readyPodCount, "runningNotReady", runningNotReadyCount,
+		"pendingPods", pendingPodCount, "totalStarting", startingPodCount,
 		"totalRunning", len(pods), "desiredReplicas", scale.Spec.Replicas)
 
 	// Get last scale time from status
@@ -279,6 +291,35 @@ func (a *AutoScaler) countStartingPods(pods []corev1.Pod) int32 {
 	return count
 }
 
+// countPendingPods counts pods that are in Pending state.
+// These are pods that have been requested but not yet scheduled or started.
+// They represent capacity that's been committed but not yet available.
+func (a *AutoScaler) countPendingPods(ctx context.Context, scaler *scalerv1alpha1.BudAIScaler, scale *Scale) (int32, error) {
+	// Parse selector from scale
+	selector, err := labels.Parse(scale.Status.Selector)
+	if err != nil {
+		return 0, fmt.Errorf("failed to parse selector: %w", err)
+	}
+
+	// List all pods matching selector
+	podList := &corev1.PodList{}
+	err = a.client.List(ctx, podList, &client.ListOptions{
+		Namespace:     scaler.Namespace,
+		LabelSelector: selector,
+	})
+	if err != nil {
+		return 0, fmt.Errorf("failed to list pods: %w", err)
+	}
+
+	var count int32
+	for _, pod := range podList.Items {
+		if pod.Status.Phase == corev1.PodPending && pod.DeletionTimestamp == nil {
+			count++
+		}
+	}
+	return count, nil
+}
+
 // isPodReady checks if a pod is ready.
 func isPodReady(pod *corev1.Pod) bool {
 	for _, condition := range pod.Status.Conditions {