diff --git a/cmd/integration/integration.go b/cmd/integration/integration.go
index 41511b13180..aa7c6e191fb 100644
--- a/cmd/integration/integration.go
+++ b/cmd/integration/integration.go
@@ -71,8 +71,8 @@ var (
 
 type fakeKubeletClient struct{}
 
-func (fakeKubeletClient) GetPodStatus(host, podNamespace, podID string) (api.PodStatusResult, error) {
-	glog.V(3).Infof("Trying to get container info for %v/%v/%v", host, podNamespace, podID)
+func (fakeKubeletClient) GetPodStatus(host, podNamespace, podName string) (api.PodStatusResult, error) {
+	glog.V(3).Infof("Trying to get container info for %v/%v/%v", host, podNamespace, podName)
 	// This is a horrible hack to get around the fact that we can't provide
 	// different port numbers per kubelet...
 	var c client.PodInfoGetter
@@ -88,9 +88,9 @@ func (fakeKubeletClient) GetPodStatus(host, podNamespace, podID string) (api.Pod
 			Port:   10251,
 		}
 	default:
-		glog.Fatalf("Can't get info for: '%v', '%v - %v'", host, podNamespace, podID)
+		glog.Fatalf("Can't get info for: '%v', '%v - %v'", host, podNamespace, podName)
 	}
-	r, err := c.GetPodStatus("127.0.0.1", podNamespace, podID)
+	r, err := c.GetPodStatus("127.0.0.1", podNamespace, podName)
 	if err != nil {
 		return r, err
 	}
@@ -304,29 +304,29 @@ func countEndpoints(eps *api.Endpoints) int {
 	return count
 }
 
-func podExists(c *client.Client, podNamespace string, podID string) wait.ConditionFunc {
+func podExists(c *client.Client, podNamespace string, podName string) wait.ConditionFunc {
 	return func() (bool, error) {
-		_, err := c.Pods(podNamespace).Get(podID)
+		_, err := c.Pods(podNamespace).Get(podName)
 		return err == nil, nil
 	}
 }
 
-func podNotFound(c *client.Client, podNamespace string, podID string) wait.ConditionFunc {
+func podNotFound(c *client.Client, podNamespace string, podName string) wait.ConditionFunc {
 	return func() (bool, error) {
-		_, err := c.Pods(podNamespace).Get(podID)
+		_, err := c.Pods(podNamespace).Get(podName)
 		return apierrors.IsNotFound(err), nil
 	}
 }
 
-func podRunning(c *client.Client, podNamespace string, podID string) wait.ConditionFunc {
+func podRunning(c *client.Client, podNamespace string, podName string) wait.ConditionFunc {
 	return func() (bool, error) {
-		pod, err := c.Pods(podNamespace).Get(podID)
+		pod, err := c.Pods(podNamespace).Get(podName)
 		if apierrors.IsNotFound(err) {
 			return false, nil
 		}
 		if err != nil {
 			// This could be a connection error so we want to retry, but log the error.
-			glog.Errorf("Error when reading pod %q: %v", podID, err)
+			glog.Errorf("Error when reading pod %q: %v", podName, err)
 			return false, nil
 		}
 		if pod.Status.Phase != api.PodRunning {
@@ -814,6 +814,61 @@ func runServiceTest(client *client.Client) {
 	glog.Info("Service test passed.")
 }
 
+func runSchedulerNoPhantomPodsTest(client *client.Client) {
+	pod := &api.Pod{
+		Spec: api.PodSpec{
+			Containers: []api.Container{
+				{
+					Name:  "c1",
+					Image: "kubernetes/pause",
+					Ports: []api.ContainerPort{
+						{ContainerPort: 1234, HostPort: 9999},
+					},
+					ImagePullPolicy: api.PullIfNotPresent,
+				},
+			},
+		},
+	}
+
+	// Assuming we only have two kublets, the third pod here won't schedule
+	// if the scheduler doesn't correctly handle the delete for the second
+	// pod.
+	pod.ObjectMeta.Name = "phantom.foo"
+	foo, err := client.Pods(api.NamespaceDefault).Create(pod)
+	if err != nil {
+		glog.Fatalf("Failed to create pod: %v, %v", pod, err)
+	}
+	if err := wait.Poll(time.Second, time.Second*10, podRunning(client, foo.Namespace, foo.Name)); err != nil {
+		glog.Fatalf("FAILED: pod never started running %v", err)
+	}
+
+	pod.ObjectMeta.Name = "phantom.bar"
+	bar, err := client.Pods(api.NamespaceDefault).Create(pod)
+	if err != nil {
+		glog.Fatalf("Failed to create pod: %v, %v", pod, err)
+	}
+	if err := wait.Poll(time.Second, time.Second*10, podRunning(client, bar.Namespace, bar.Name)); err != nil {
+		glog.Fatalf("FAILED: pod never started running %v", err)
+	}
+
+	// Delete a pod to free up room.
+	err = client.Pods(api.NamespaceDefault).Delete(bar.Name)
+	if err != nil {
+		glog.Fatalf("FAILED: couldn't delete pod %q: %v", bar.Name, err)
+	}
+
+	pod.ObjectMeta.Name = "phantom.baz"
+	baz, err := client.Pods(api.NamespaceDefault).Create(pod)
+	if err != nil {
+		glog.Fatalf("Failed to create pod: %v, %v", pod, err)
+	}
+	if err := wait.Poll(time.Second, time.Second*10, podRunning(client, baz.Namespace, baz.Name)); err != nil {
+		glog.Fatalf("FAILED: (Scheduler probably didn't process deletion of 'phantom.bar') Pod never started running: %v", err)
+	}
+
+	glog.Info("Scheduler doesn't make phantom pods: test passed.")
+}
+
 type testFunc func(*client.Client)
 
 func addFlags(fs *pflag.FlagSet) {
@@ -904,6 +959,11 @@ func main() {
 		glog.Fatalf("Expected 16 containers; got %v\n\nlist of created containers:\n\n%#v\n\nDocker 1 Created:\n\n%#v\n\nDocker 2 Created:\n\n%#v\n\n", len(createdConts), createdConts.List(), fakeDocker1.Created, fakeDocker2.Created)
 	}
 	glog.Infof("OK - found created containers: %#v", createdConts.List())
+
+	// This test doesn't run with the others because it can't run in
+	// parallel and also it schedules extra pods which would change the
+	// above pod counting logic.
+	runSchedulerNoPhantomPodsTest(kubeClient)
 }
 
 // ServeCachedManifestFile serves a file for kubelet to read.
diff --git a/plugin/pkg/scheduler/factory/factory.go b/plugin/pkg/scheduler/factory/factory.go
index 4fc10b398ad..30b5dd60f5a 100644
--- a/plugin/pkg/scheduler/factory/factory.go
+++ b/plugin/pkg/scheduler/factory/factory.go
@@ -103,6 +103,21 @@ func (f *ConfigFactory) CreateFromConfig(policy schedulerapi.Policy) (*scheduler
 	return f.CreateFromKeys(predicateKeys, priorityKeys)
 }
 
+// ReflectorDeletionHook passes all operations through to Store, but calls
+// OnDelete in a goroutine if there is a deletion.
+type ReflectorDeletionHook struct {
+	cache.Store
+	OnDelete func(obj interface{})
+}
+
+func (r ReflectorDeletionHook) Delete(obj interface{}) error {
+	go func() {
+		defer util.HandleCrash()
+		r.OnDelete(obj)
+	}()
+	return r.Store.Delete(obj)
+}
+
 // Creates a scheduler from a set of registered fit predicate keys and priority keys.
 func (f *ConfigFactory) CreateFromKeys(predicateKeys, priorityKeys util.StringSet) (*scheduler.Config, error) {
 	glog.V(2).Infof("creating scheduler with fit predicates '%v' and priority functions '%v", predicateKeys, priorityKeys)
@@ -125,9 +140,22 @@ func (f *ConfigFactory) CreateFromKeys(predicateKeys, priorityKeys util.StringSe
 	// Watch and queue pods that need scheduling.
 	cache.NewReflector(f.createUnassignedPodLW(), &api.Pod{}, f.PodQueue, 0).Run()
 
+	// Pass through all events to the scheduled pod store, but on a deletion,
+	// also remove from the assumed pods.
+	assumedPodDeleter := ReflectorDeletionHook{
+		Store: f.ScheduledPodLister.Store,
+		OnDelete: func(obj interface{}) {
+			if pod, ok := obj.(*api.Pod); ok {
+				f.modeler.LockedAction(func() {
+					f.modeler.ForgetPod(pod)
+				})
+			}
+		},
+	}
+
 	// Watch and cache all running pods. Scheduler needs to find all pods
 	// so it knows where it's safe to place a pod. Cache this locally.
-	cache.NewReflector(f.createAssignedPodLW(), &api.Pod{}, f.ScheduledPodLister.Store, 0).Run()
+	cache.NewReflector(f.createAssignedPodLW(), &api.Pod{}, assumedPodDeleter, 0).Run()
 
 	// Watch minions.
 	// Minions may be listed frequently, so provide a local up-to-date cache.
diff --git a/plugin/pkg/scheduler/modeler.go b/plugin/pkg/scheduler/modeler.go
index 97e0907475b..7833b5b94b6 100644
--- a/plugin/pkg/scheduler/modeler.go
+++ b/plugin/pkg/scheduler/modeler.go
@@ -19,6 +19,7 @@ package scheduler
 import (
 	"fmt"
 	"strings"
+	"sync"
 
 	"github.com/GoogleCloudPlatform/kubernetes/pkg/api"
 	"github.com/GoogleCloudPlatform/kubernetes/pkg/client/cache"
@@ -40,9 +41,24 @@ type ExtendedPodLister interface {
 	Exists(pod *api.Pod) (bool, error)
 }
 
+// actionLocker implements lockedAction (so the fake and SimpleModeler can both
+// use it)
+type actionLocker struct {
+	sync.Mutex
+}
+
+// LockedAction serializes calls of whatever is passed as 'do'.
+func (a *actionLocker) LockedAction(do func()) {
+	a.Lock()
+	defer a.Unlock()
+	do()
+}
+
 // FakeModeler implements the SystemModeler interface.
 type FakeModeler struct {
 	AssumePodFunc func(pod *api.Pod)
+	ForgetPodFunc func(pod *api.Pod)
+	actionLocker
 }
 
 // AssumePod calls the function variable if it is not nil.
@@ -52,6 +68,13 @@ func (f *FakeModeler) AssumePod(pod *api.Pod) {
 	}
 }
 
+// ForgetPod calls the function variable if it is not nil.
+func (f *FakeModeler) ForgetPod(pod *api.Pod) {
+	if f.ForgetPodFunc != nil {
+		f.ForgetPodFunc(pod)
+	}
+}
+
 // SimpleModeler implements the SystemModeler interface with a timed pod cache.
 type SimpleModeler struct {
 	queuedPods    ExtendedPodLister
@@ -61,6 +84,8 @@ type SimpleModeler struct {
 	// haven't yet shown up in the scheduledPods variable.
 	// TODO: periodically clear this.
 	assumedPods *cache.StoreToPodLister
+
+	actionLocker
 }
 
 // NewSimpleModeler returns a new SimpleModeler.
@@ -78,6 +103,10 @@ func (s *SimpleModeler) AssumePod(pod *api.Pod) {
 	s.assumedPods.Add(pod)
 }
 
+func (s *SimpleModeler) ForgetPod(pod *api.Pod) {
+	s.assumedPods.Delete(pod)
+}
+
 // Extract names for readable logging.
 func podNames(pods []api.Pod) []string {
 	out := make([]string, len(pods))
diff --git a/plugin/pkg/scheduler/scheduler.go b/plugin/pkg/scheduler/scheduler.go
index 86b45f4322b..3dcedb8e1d8 100644
--- a/plugin/pkg/scheduler/scheduler.go
+++ b/plugin/pkg/scheduler/scheduler.go
@@ -44,6 +44,16 @@ type SystemModeler interface {
 	// The assumtion should last until the system confirms the
 	// assumtion or disconfirms it.
 	AssumePod(pod *api.Pod)
+	// ForgetPod removes a pod assumtion. (It won't make the model
+	// show the absence of the given pod if the pod is in the scheduled
+	// pods list!)
+	ForgetPod(pod *api.Pod)
+
+	// For serializing calls to Assume/ForgetPod: imagine you want to add
+	// a pod iff a bind succeeds, but also remove a pod if it is deleted.
+	// TODO: if SystemModeler begins modeling things other than pods, this
+	// should probably be parameterized or specialized for pods.
+	LockedAction(f func())
 }
 
 // Scheduler watches for new unscheduled pods. It attempts to find
@@ -104,16 +114,21 @@ func (s *Scheduler) scheduleOne() {
 			Name: dest,
 		},
 	}
-	if err := s.config.Binder.Bind(b); err != nil {
-		glog.V(1).Infof("Failed to bind pod: %v", err)
-		s.config.Recorder.Eventf(pod, "failedScheduling", "Binding rejected: %v", err)
-		s.config.Error(pod, err)
-		return
-	}
-	s.config.Recorder.Eventf(pod, "scheduled", "Successfully assigned %v to %v", pod.Name, dest)
-	// tell the model to assume that this binding took effect.
-	assumed := *pod
-	assumed.Spec.Host = dest
-	assumed.Status.Host = dest
-	s.config.Modeler.AssumePod(&assumed)
+
+	// We want to add the pod to the model iff the bind succeeds, but we don't want to race
+	// with any deletions, which happen asyncronously.
+	s.config.Modeler.LockedAction(func() {
+		if err := s.config.Binder.Bind(b); err != nil {
+			glog.V(1).Infof("Failed to bind pod: %v", err)
+			s.config.Recorder.Eventf(pod, "failedScheduling", "Binding rejected: %v", err)
+			s.config.Error(pod, err)
+			return
+		}
+		s.config.Recorder.Eventf(pod, "scheduled", "Successfully assigned %v to %v", pod.Name, dest)
+		// tell the model to assume that this binding took effect.
+		assumed := *pod
+		assumed.Spec.Host = dest
+		assumed.Status.Host = dest
+		s.config.Modeler.AssumePod(&assumed)
+	})
 }