diff --git a/pkg/kubelet/kubelet.go b/pkg/kubelet/kubelet.go
index 0ed51bad2df..6c11f12a3f0 100644
--- a/pkg/kubelet/kubelet.go
+++ b/pkg/kubelet/kubelet.go
@@ -1725,13 +1725,41 @@ func (kl *Kubelet) getPullSecretsForPod(pod *api.Pod) ([]api.Secret, error) {
 	return pullSecrets, nil
 }
 
+// Return name of a volume. When the volume is a PersistentVolumeClaim,
+// it returns name of the real PersistentVolume bound to the claim.
+// It returns errror when the clam is not bound yet.
+func (kl *Kubelet) resolveVolumeName(pod *api.Pod, volume *api.Volume) (string, error) {
+	claimSource := volume.VolumeSource.PersistentVolumeClaim
+	if claimSource != nil {
+		// resolve real volume behind the claim
+		claim, err := kl.kubeClient.Legacy().PersistentVolumeClaims(pod.Namespace).Get(claimSource.ClaimName)
+		if err != nil {
+			return "", fmt.Errorf("Cannot find claim %s/%s for volume %s", pod.Namespace, claimSource.ClaimName, volume.Name)
+		}
+		if claim.Status.Phase != api.ClaimBound {
+			return "", fmt.Errorf("Claim for volume %s/%s is not bound yet", pod.Namespace, claimSource.ClaimName)
+		}
+		// Use the real bound volume instead of PersistentVolume.Name
+		return claim.Spec.VolumeName, nil
+	}
+	return volume.Name, nil
+}
+
 // Stores all volumes defined by the set of pods into a map.
+// It stores real volumes there, i.e. persistent volume claims are resolved
+// to volumes that are bound to them.
 // Keys for each entry are in the format (POD_ID)/(VOLUME_NAME)
-func getDesiredVolumes(pods []*api.Pod) map[string]api.Volume {
+func (kl *Kubelet) getDesiredVolumes(pods []*api.Pod) map[string]api.Volume {
 	desiredVolumes := make(map[string]api.Volume)
 	for _, pod := range pods {
 		for _, volume := range pod.Spec.Volumes {
-			identifier := path.Join(string(pod.UID), volume.Name)
+			volumeName, err := kl.resolveVolumeName(pod, &volume)
+			if err != nil {
+				glog.V(3).Infof("%v", err)
+				// Ignore the error and hope it's resolved next time
+				continue
+			}
+			identifier := path.Join(string(pod.UID), volumeName)
 			desiredVolumes[identifier] = volume
 		}
 	}
@@ -1815,8 +1843,11 @@ func (kl *Kubelet) cleanupBandwidthLimits(allPods []*api.Pod) error {
 
 // Compares the map of current volumes to the map of desired volumes.
 // If an active volume does not have a respective desired volume, clean it up.
+// This method is blocking:
+// 1) it talks to API server to find volumes bound to persistent volume claims
+// 2) it talks to cloud to detach volumes
 func (kl *Kubelet) cleanupOrphanedVolumes(pods []*api.Pod, runningPods []*kubecontainer.Pod) error {
-	desiredVolumes := getDesiredVolumes(pods)
+	desiredVolumes := kl.getDesiredVolumes(pods)
 	currentVolumes := kl.getPodVolumesFromDisk()
 
 	runningSet := sets.String{}
diff --git a/pkg/kubelet/kubelet_test.go b/pkg/kubelet/kubelet_test.go
index 9fee58d8874..eccb6b959d6 100644
--- a/pkg/kubelet/kubelet_test.go
+++ b/pkg/kubelet/kubelet_test.go
@@ -551,6 +551,106 @@ func TestGetPodVolumesFromDisk(t *testing.T) {
 	}
 }
 
+// Test for https://github.com/kubernetes/kubernetes/pull/19600
+func TestCleanupOrphanedVolumes(t *testing.T) {
+	testKubelet := newTestKubelet(t)
+	kubelet := testKubelet.kubelet
+	kubeClient := testKubelet.fakeKubeClient
+	plug := &volume.FakeVolumePlugin{PluginName: "fake", Host: nil}
+	kubelet.volumePluginMgr.InitPlugins([]volume.VolumePlugin{plug}, &volumeHost{kubelet})
+
+	// create a volume "on disk"
+	volsOnDisk := []struct {
+		podUID  types.UID
+		volName string
+	}{
+		{"podUID", "myrealvol"},
+	}
+
+	pathsOnDisk := []string{}
+	for i := range volsOnDisk {
+		fv := volume.FakeVolume{PodUID: volsOnDisk[i].podUID, VolName: volsOnDisk[i].volName, Plugin: plug}
+		fv.SetUp(nil)
+		pathsOnDisk = append(pathsOnDisk, fv.GetPath())
+	}
+
+	// store the claim in fake kubelet database
+	claim := api.PersistentVolumeClaim{
+		ObjectMeta: api.ObjectMeta{
+			Name:      "myclaim",
+			Namespace: "test",
+		},
+		Spec: api.PersistentVolumeClaimSpec{
+			VolumeName: "myrealvol",
+		},
+		Status: api.PersistentVolumeClaimStatus{
+			Phase: api.ClaimBound,
+		},
+	}
+	kubeClient.ReactionChain = fake.NewSimpleClientset(&api.PersistentVolumeClaimList{Items: []api.PersistentVolumeClaim{
+		claim,
+	}}).ReactionChain
+
+	// Create a pod referencing the volume via a PersistentVolumeClaim
+	pod := api.Pod{
+		ObjectMeta: api.ObjectMeta{
+			UID:       "podUID",
+			Name:      "pod",
+			Namespace: "test",
+		},
+		Spec: api.PodSpec{
+			Volumes: []api.Volume{
+				{
+					Name: "myvolumeclaim",
+					VolumeSource: api.VolumeSource{
+						PersistentVolumeClaim: &api.PersistentVolumeClaimVolumeSource{
+							ClaimName: "myclaim",
+						},
+					},
+				},
+			},
+		},
+	}
+
+	// The pod is pending and not running yet. Test that cleanupOrphanedVolumes
+	// won't remove the volume from disk if the volume is referenced only
+	// indirectly by a claim.
+	err := kubelet.cleanupOrphanedVolumes([]*api.Pod{&pod}, []*kubecontainer.Pod{})
+	if err != nil {
+		t.Errorf("cleanupOrphanedVolumes failed: %v", err)
+	}
+
+	volumesFound := kubelet.getPodVolumesFromDisk()
+	if len(volumesFound) != len(pathsOnDisk) {
+		t.Errorf("Expected to find %d cleaners, got %d", len(pathsOnDisk), len(volumesFound))
+	}
+	for _, ep := range pathsOnDisk {
+		found := false
+		for _, cl := range volumesFound {
+			if ep == cl.GetPath() {
+				found = true
+				break
+			}
+		}
+		if !found {
+			t.Errorf("Could not find a volume with path %s", ep)
+		}
+	}
+
+	// The pod is deleted -> kubelet should delete the volume
+	err = kubelet.cleanupOrphanedVolumes([]*api.Pod{}, []*kubecontainer.Pod{})
+	if err != nil {
+		t.Errorf("cleanupOrphanedVolumes failed: %v", err)
+	}
+	volumesFound = kubelet.getPodVolumesFromDisk()
+	if len(volumesFound) != 0 {
+		t.Errorf("Expected to find 0 cleaners, got %d", len(volumesFound))
+	}
+	for _, cl := range volumesFound {
+		t.Errorf("Found unexpected volume %s", cl.GetPath())
+	}
+}
+
 type stubVolume struct {
 	path string
 	volume.MetricsNil