diff --git a/hack/test-utils.sh b/hack/test-utils.sh
index 7bdbed267..1a802a94d 100644
--- a/hack/test-utils.sh
+++ b/hack/test-utils.sh
@@ -18,9 +18,14 @@ ROOT="$( cd "$( dirname "${BASH_SOURCE[0]}" )" && pwd )"/..
 . ${ROOT}/hack/versions
 # CRI_CONTAINERD_FLAGS are the extra flags to use when start cri-containerd.
 CRI_CONTAINERD_FLAGS=${CRI_CONTAINERD_FLAGS:-""}
+# RESTART_WAIT_PERIOD is the period to wait before restarting cri-containerd/containerd.
+RESTART_WAIT_PERIOD=${RESTART_WAIT_PERIOD:-10}
 
 CRICONTAINERD_SOCK=/var/run/cri-containerd.sock
 
+cri_containerd_pid=
+containerd_pid=
+
 # test_setup starts containerd and cri-containerd.
 test_setup() {
   local report_dir=$1 
@@ -35,22 +40,42 @@ test_setup() {
     exit 1
   fi
   sudo pkill containerd
-  sudo containerd &> ${report_dir}/containerd.log &
+  keepalive "sudo containerd" ${RESTART_WAIT_PERIOD} &> ${report_dir}/containerd.log &
+  containerd_pid=$!
   # Wait for containerd to be running by using the containerd client ctr to check the version
   # of the containerd server. Wait an increasing amount of time after each of five attempts
   readiness_check "sudo ctr version"
 
   # Start cri-containerd
-  sudo ${ROOT}/_output/cri-containerd --alsologtostderr --v 4 ${CRI_CONTAINERD_FLAGS} \
-	  &> ${report_dir}/cri-containerd.log &
+  keepalive "sudo ${ROOT}/_output/cri-containerd --alsologtostderr --v 4 ${CRI_CONTAINERD_FLAGS}" \
+	  ${RESTART_WAIT_PERIOD} &> ${report_dir}/cri-containerd.log &
+  cri_containerd_pid=$!
   readiness_check "sudo ${GOPATH}/bin/crictl --runtime-endpoint=${CRICONTAINERD_SOCK} info"
 }
 
 # test_teardown kills containerd and cri-containerd.
 test_teardown() {
+  if [ -n "${containerd_pid}" ]; then
+    kill ${containerd_pid}
+  fi
+  if [ -n "${cri_containerd_pid}" ]; then
+    kill ${cri_containerd_pid}
+  fi
   sudo pkill containerd
 }
 
+# keepalive runs a command and keeps it alive.
+# keepalive process is eventually killed in test_teardown.
+keepalive() {
+  local command=$1
+  echo ${command}
+  local wait_period=$2
+  while true; do
+    ${command}
+    sleep ${wait_period}
+  done
+}
+
 # readiness_check checks readiness of a daemon with specified command.
 readiness_check() {
   local command=$1
diff --git a/integration/restart_test.go b/integration/restart_test.go
new file mode 100644
index 000000000..768b45fc9
--- /dev/null
+++ b/integration/restart_test.go
@@ -0,0 +1,200 @@
+/*
+Copyright 2017 The Kubernetes Authors.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+*/
+
+package integration
+
+import (
+	"os"
+	"path/filepath"
+	"testing"
+	"time"
+
+	"github.com/containerd/containerd"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+	"golang.org/x/net/context"
+	"golang.org/x/sys/unix"
+	"k8s.io/kubernetes/pkg/kubelet/apis/cri/v1alpha1/runtime"
+)
+
+// Restart test must run sequentially.
+
+func TestSandboxAcrossCRIContainerdRestart(t *testing.T) {
+	ctx := context.Background()
+	sandboxNS := "sandbox-restart-cri-containerd"
+	sandboxes := []struct {
+		name            string
+		id              string
+		stateBeforeExit runtime.PodSandboxState
+		actionAfterExit string
+		expectedState   runtime.PodSandboxState
+	}{
+		{
+			name:            "task-always-ready",
+			stateBeforeExit: runtime.PodSandboxState_SANDBOX_READY,
+			expectedState:   runtime.PodSandboxState_SANDBOX_READY,
+		},
+		{
+			name:            "task-always-not-ready",
+			stateBeforeExit: runtime.PodSandboxState_SANDBOX_NOTREADY,
+			expectedState:   runtime.PodSandboxState_SANDBOX_NOTREADY,
+		},
+		{
+			name:            "task-exit-before-restart",
+			stateBeforeExit: runtime.PodSandboxState_SANDBOX_READY,
+			actionAfterExit: "kill",
+			expectedState:   runtime.PodSandboxState_SANDBOX_NOTREADY,
+		},
+		{
+			name:            "task-deleted-before-restart",
+			stateBeforeExit: runtime.PodSandboxState_SANDBOX_READY,
+			actionAfterExit: "delete",
+			expectedState:   runtime.PodSandboxState_SANDBOX_NOTREADY,
+		},
+	}
+	t.Logf("Make sure no sandbox is running before test")
+	existingSandboxes, err := runtimeService.ListPodSandbox(&runtime.PodSandboxFilter{})
+	require.NoError(t, err)
+	require.Empty(t, existingSandboxes)
+
+	t.Logf("Start test sandboxes")
+	for i := range sandboxes {
+		s := &sandboxes[i]
+		cfg := PodSandboxConfig(s.name, sandboxNS)
+		sb, err := runtimeService.RunPodSandbox(cfg)
+		require.NoError(t, err)
+		defer func() {
+			// Make sure the sandbox is cleaned up in any case.
+			runtimeService.StopPodSandbox(sb)
+			runtimeService.RemovePodSandbox(sb)
+		}()
+		s.id = sb
+		if s.stateBeforeExit == runtime.PodSandboxState_SANDBOX_NOTREADY {
+			require.NoError(t, runtimeService.StopPodSandbox(sb))
+		}
+	}
+
+	t.Logf("Kill cri-containerd")
+	require.NoError(t, KillProcess("cri-containerd"))
+	defer func() {
+		assert.NoError(t, Eventually(func() (bool, error) {
+			return ConnectDaemons() == nil, nil
+		}, time.Second, 30*time.Second), "make sure cri-containerd is running before test finish")
+	}()
+
+	t.Logf("Change sandbox state, must finish before cri-containerd is restarted")
+	for _, s := range sandboxes {
+		if s.actionAfterExit == "" {
+			continue
+		}
+		cntr, err := containerdClient.LoadContainer(ctx, s.id)
+		require.NoError(t, err)
+		task, err := cntr.Task(ctx, nil)
+		require.NoError(t, err)
+		switch s.actionAfterExit {
+		case "kill":
+			require.NoError(t, task.Kill(ctx, unix.SIGKILL, containerd.WithKillAll))
+		case "delete":
+			_, err := task.Delete(ctx, containerd.WithProcessKill)
+			require.NoError(t, err)
+		}
+	}
+
+	t.Logf("Wait until cri-containerd is restarted")
+	require.NoError(t, Eventually(func() (bool, error) {
+		return ConnectDaemons() == nil, nil
+	}, time.Second, 30*time.Second), "wait for cri-containerd to be restarted")
+
+	t.Logf("Check sandbox state after restart")
+	loadedSandboxes, err := runtimeService.ListPodSandbox(&runtime.PodSandboxFilter{})
+	require.NoError(t, err)
+	assert.Len(t, loadedSandboxes, len(sandboxes))
+	for _, s := range sandboxes {
+		for _, loaded := range loadedSandboxes {
+			if s.id == loaded.Id {
+				assert.Equal(t, s.expectedState, loaded.State)
+				break
+			}
+		}
+	}
+
+	t.Logf("Should be able to stop and remove sandbox after restart")
+	for _, s := range sandboxes {
+		// Properly stop the sandbox if it's ready before restart.
+		if s.stateBeforeExit == runtime.PodSandboxState_SANDBOX_READY {
+			assert.NoError(t, runtimeService.StopPodSandbox(s.id))
+		}
+		assert.NoError(t, runtimeService.RemovePodSandbox(s.id))
+	}
+}
+
+// TestSandboxDeletionAcrossCRIContainerdRestart tests the case that sandbox container
+// is deleted from containerd during cri-containerd is down. This should not happen.
+// However, if this really happens, cri-containerd should not load such sandbox and
+// should do best effort cleanup of the sandbox root directory. Note that in this case,
+// cri-containerd loses the network namespace of the sandbox, so it won't be able to
+// teardown the network properly.
+// This test uses host network sandbox to avoid resource leakage.
+func TestSandboxDeletionAcrossCRIContainerdRestart(t *testing.T) {
+	ctx := context.Background()
+	sandboxNS := "sandbox-delete-restart-cri-containerd"
+	t.Logf("Make sure no sandbox is running before test")
+	existingSandboxes, err := runtimeService.ListPodSandbox(&runtime.PodSandboxFilter{})
+	require.NoError(t, err)
+	require.Empty(t, existingSandboxes)
+
+	t.Logf("Start test sandboxes")
+	cfg := PodSandboxConfig("sandbox", sandboxNS, WithHostNetwork)
+	sb, err := runtimeService.RunPodSandbox(cfg)
+	require.NoError(t, err)
+	defer func() {
+		// Make sure the sandbox is cleaned up in any case.
+		runtimeService.StopPodSandbox(sb)
+		runtimeService.RemovePodSandbox(sb)
+	}()
+
+	t.Logf("Kill cri-containerd")
+	require.NoError(t, KillProcess("cri-containerd"))
+	defer func() {
+		assert.NoError(t, Eventually(func() (bool, error) {
+			return ConnectDaemons() == nil, nil
+		}, time.Second, 30*time.Second), "make sure cri-containerd is running before test finish")
+	}()
+
+	t.Logf("Delete sandbox container from containerd")
+	cntr, err := containerdClient.LoadContainer(ctx, sb)
+	require.NoError(t, err)
+	task, err := cntr.Task(ctx, nil)
+	require.NoError(t, err)
+	_, err = task.Delete(ctx, containerd.WithProcessKill)
+	require.NoError(t, err)
+	require.NoError(t, cntr.Delete(ctx, containerd.WithSnapshotCleanup))
+
+	t.Logf("Wait until cri-containerd is restarted")
+	require.NoError(t, Eventually(func() (bool, error) {
+		return ConnectDaemons() == nil, nil
+	}, time.Second, 30*time.Second), "wait for cri-containerd to be restarted")
+
+	t.Logf("Check sandbox state after restart")
+	loadedSandboxes, err := runtimeService.ListPodSandbox(&runtime.PodSandboxFilter{})
+	require.NoError(t, err)
+	assert.Empty(t, loadedSandboxes)
+
+	t.Logf("Make sure sandbox root is removed")
+	sandboxRoot := filepath.Join(criContainerdRoot, "sandboxes", sb)
+	_, err = os.Stat(sandboxRoot)
+	assert.True(t, os.IsNotExist(err))
+}
diff --git a/integration/test_utils.go b/integration/test_utils.go
index a58047ba4..574e0ded7 100644
--- a/integration/test_utils.go
+++ b/integration/test_utils.go
@@ -18,6 +18,8 @@ package integration
 
 import (
 	"errors"
+	"fmt"
+	"os/exec"
 	"time"
 
 	"github.com/containerd/containerd"
@@ -38,6 +40,7 @@ const (
 	k8sNamespace          = "k8s.io"                             // This is the same with server.k8sContainerdNamespace.
 	containerdEndpoint    = "/run/containerd/containerd.sock"
 	criContainerdEndpoint = "/var/run/cri-containerd.sock"
+	criContainerdRoot     = "/var/lib/cri-containerd"
 )
 
 var (
@@ -48,28 +51,61 @@ var (
 )
 
 func init() {
+	if err := ConnectDaemons(); err != nil {
+		glog.Exitf("Failed to connect daemons: %v", err)
+	}
+}
+
+// ConnectDaemons connect cri-containerd and containerd, and initialize the clients.
+func ConnectDaemons() error {
 	var err error
 	runtimeService, err = remote.NewRemoteRuntimeService(sock, timeout)
 	if err != nil {
-		glog.Exitf("Failed to create runtime service: %v", err)
+		return fmt.Errorf("failed to create runtime service: %v", err)
 	}
 	imageService, err = remote.NewRemoteImageService(sock, timeout)
 	if err != nil {
-		glog.Exitf("Failed to create image service: %v", err)
+		return fmt.Errorf("failed to create image service: %v", err)
+	}
+	// Since CRI grpc client doesn't have `WithBlock` specified, we
+	// need to check whether it is actually connected.
+	// TODO(random-liu): Extend cri remote client to accept extra grpc options.
+	_, err = runtimeService.ListContainers(&runtime.ContainerFilter{})
+	if err != nil {
+		return fmt.Errorf("failed to list containers: %v", err)
+	}
+	_, err = imageService.ListImages(&runtime.ImageFilter{})
+	if err != nil {
+		return fmt.Errorf("failed to list images: %v", err)
 	}
 	containerdClient, err = containerd.New(containerdEndpoint, containerd.WithDefaultNamespace(k8sNamespace))
 	if err != nil {
-		glog.Exitf("Failed to connect containerd: %v", err)
+		return fmt.Errorf("failed to connect containerd: %v", err)
 	}
 	criContainerdClient, err = client.NewCRIContainerdClient(criContainerdEndpoint, timeout)
 	if err != nil {
-		glog.Exitf("Failed to connect cri-containerd: %v", err)
+		return fmt.Errorf("failed to connect cri-containerd: %v", err)
 	}
+	return nil
 }
 
 // Opts sets specific information in pod sandbox config.
 type PodSandboxOpts func(*runtime.PodSandboxConfig)
 
+func WithHostNetwork(p *runtime.PodSandboxConfig) {
+	if p.Linux == nil {
+		p.Linux = &runtime.LinuxPodSandboxConfig{}
+	}
+	if p.Linux.SecurityContext == nil {
+		p.Linux.SecurityContext = &runtime.LinuxSandboxSecurityContext{}
+	}
+	if p.Linux.SecurityContext.NamespaceOptions == nil {
+		p.Linux.SecurityContext.NamespaceOptions = &runtime.NamespaceOption{
+			HostNetwork: true,
+		}
+	}
+}
+
 // PodSandboxConfig generates a pod sandbox config for test.
 func PodSandboxConfig(name, ns string, opts ...PodSandboxOpts) *runtime.PodSandboxConfig {
 	config := &runtime.PodSandboxConfig{
@@ -164,3 +200,12 @@ func Eventually(f CheckFunc, period, timeout time.Duration) error {
 func Randomize(str string) string {
 	return str + "-" + util.GenerateID()
 }
+
+// KillProcess kills the process by name. pkill is used.
+func KillProcess(name string) error {
+	output, err := exec.Command("pkill", fmt.Sprintf("^%s$", name)).CombinedOutput()
+	if err != nil {
+		return fmt.Errorf("failed to kill %q - error: %v, output: %q", name, err, output)
+	}
+	return nil
+}
diff --git a/pkg/client/client.go b/pkg/client/client.go
index a7d1cbed5..276186153 100644
--- a/pkg/client/client.go
+++ b/pkg/client/client.go
@@ -34,7 +34,9 @@ func NewCRIContainerdClient(endpoint string, timeout time.Duration) (api.CRICont
 		return nil, fmt.Errorf("failed to get dialer: %v", err)
 	}
 	conn, err := grpc.Dial(addr,
+		grpc.WithBlock(),
 		grpc.WithInsecure(),
+		// TODO(random-liu): WithTimeout is being deprecated, use context instead.
 		grpc.WithTimeout(timeout),
 		grpc.WithDialer(dialer),
 	)