Reconcile on all private service events in KPA. (knative#7514) (#445)

We used to have a metrics service which the KPA watched and reacted to. We got rid of that and now have a race condition if we specify minScale but the SKS becomes ready with less than minScale pods. Nothing after that will ever kick the KPA to reconcile again.

This fixes that by watching the private services to make sure the KPA sees every change in deployment size. It itself keys off of the private services too to calculate the current replicas, closing the loop here.

pkg/reconciler/autoscaling/kpa/controller.go

@@ -38,6 +38,8 @@ import (
 	"knative.dev/pkg/controller"
 	pkgreconciler "knative.dev/pkg/reconciler"
 	"knative.dev/serving/pkg/apis/autoscaling"
+	"knative.dev/serving/pkg/apis/networking"
+	"knative.dev/serving/pkg/apis/serving"
 	autoscalerconfig "knative.dev/serving/pkg/autoscaler/config"
 	servingreconciler "knative.dev/serving/pkg/reconciler"
 	areconciler "knative.dev/serving/pkg/reconciler/autoscaling"
@@ -114,16 +116,12 @@ func NewController(
 		},
 	})
 
+	// Watch all the private service endpoints.
 	endpointsInformer.Informer().AddEventHandler(cache.FilteringResourceEventHandler{
-		FilterFunc: pkgreconciler.LabelExistsFilterFunc(autoscaling.KPALabelKey),
-		Handler:    controller.HandleAll(impl.EnqueueLabelOfNamespaceScopedResource("", autoscaling.KPALabelKey)),
+		FilterFunc: pkgreconciler.LabelFilterFunc(networking.ServiceTypeKey, string(networking.ServiceTypePrivate), false),
+		Handler:    controller.HandleAll(impl.EnqueueLabelOfNamespaceScopedResource("", serving.RevisionLabelKey)),
 	})
 
-	// Watch all the services that we have created.
-	serviceInformer.Informer().AddEventHandler(cache.FilteringResourceEventHandler{
-		FilterFunc: onlyKpaClass,
-		Handler:    controller.HandleAll(impl.EnqueueControllerOf),
-	})
 	sksInformer.Informer().AddEventHandler(cache.FilteringResourceEventHandler{
 		FilterFunc: onlyKpaClass,
 		Handler:    controller.HandleAll(impl.EnqueueControllerOf),