diff --git a/pkg/kubelet/container/helpers.go b/pkg/kubelet/container/helpers.go
index b8d841894bd..f64aeec72e2 100644
--- a/pkg/kubelet/container/helpers.go
+++ b/pkg/kubelet/container/helpers.go
@@ -72,8 +72,8 @@ func ShouldContainerBeRestarted(container *v1.Container, pod *v1.Pod, podStatus
 	if status.State == ContainerStateRunning {
 		return false
 	}
-	// Always restart container in unknown state now
-	if status.State == ContainerStateUnknown {
+	// Always restart container in the unknown, or in the created state.
+	if status.State == ContainerStateUnknown || status.State == ContainerStateCreated {
 		return true
 	}
 	// Check RestartPolicy for dead container
diff --git a/pkg/kubelet/kubelet_pods.go b/pkg/kubelet/kubelet_pods.go
index c662f1a2ce0..0fc4fe9106b 100644
--- a/pkg/kubelet/kubelet_pods.go
+++ b/pkg/kubelet/kubelet_pods.go
@@ -1210,6 +1210,13 @@ func (kl *Kubelet) convertToAPIContainerStatuses(pod *v1.Pod, podStatus *kubecon
 		switch cs.State {
 		case kubecontainer.ContainerStateRunning:
 			status.State.Running = &v1.ContainerStateRunning{StartedAt: metav1.NewTime(cs.StartedAt)}
+		case kubecontainer.ContainerStateCreated:
+			// Treat containers in the "created" state as if they are exited.
+			// The pod workers are supposed start all containers it creates in
+			// one sync (syncPod) iteration. There should not be any normal
+			// "created" containers when the pod worker generates the status at
+			// the beginning of a sync iteration.
+			fallthrough
 		case kubecontainer.ContainerStateExited:
 			status.State.Terminated = &v1.ContainerStateTerminated{
 				ExitCode:    int32(cs.ExitCode),