diff --git a/clientv3/client.go b/clientv3/client.go
index 28041ffb3..ef5e7ea24 100644
--- a/clientv3/client.go
+++ b/clientv3/client.go
@@ -384,5 +384,5 @@ func dialEndpointList(c *Client) (*grpc.ClientConn, error) {
 // progress can be made, even after reconnecting.
 func isHaltErr(ctx context.Context, err error) bool {
 	isRPCError := strings.HasPrefix(grpc.ErrorDesc(err), "etcdserver: ")
-	return isRPCError || ctx.Err() != nil
+	return isRPCError || ctx.Err() != nil || err == rpctypes.ErrConnClosed
 }
diff --git a/clientv3/integration/kv_test.go b/clientv3/integration/kv_test.go
index 391022bef..8a5050aab 100644
--- a/clientv3/integration/kv_test.go
+++ b/clientv3/integration/kv_test.go
@@ -279,6 +279,42 @@ func TestKVRange(t *testing.T) {
 	}
 }
 
+func TestKVGetErrConnClosed(t *testing.T) {
+	defer testutil.AfterTest(t)
+
+	clus := integration.NewClusterV3(t, &integration.ClusterConfig{Size: 1})
+	defer clus.Terminate(t)
+
+	cli := clus.Client(0)
+	kv := clientv3.NewKV(cli)
+
+	closed, donec := make(chan struct{}), make(chan struct{})
+	go func() {
+		select {
+		case <-time.After(3 * time.Second):
+			t.Fatal("cli.Close took too long")
+		case <-closed:
+		}
+
+		if _, err := kv.Get(context.TODO(), "foo"); err != rpctypes.ErrConnClosed {
+			t.Fatalf("expected %v, got %v", rpctypes.ErrConnClosed, err)
+		}
+		close(donec)
+	}()
+
+	if err := cli.Close(); err != nil {
+		t.Fatal(err)
+	}
+	clus.TakeClient(0)
+	close(closed)
+
+	select {
+	case <-time.After(3 * time.Second):
+		t.Fatal("kv.Get took too long")
+	case <-donec:
+	}
+}
+
 func TestKVDeleteRange(t *testing.T) {
 	defer testutil.AfterTest(t)
 
diff --git a/clientv3/kv.go b/clientv3/kv.go
index 7b4082a08..41ac5113f 100644
--- a/clientv3/kv.go
+++ b/clientv3/kv.go
@@ -189,6 +189,7 @@ func (kv *kv) do(ctx context.Context, op Op) (OpResponse, error) {
 	return OpResponse{}, err
 }
 
+// getRemote must be followed by kv.rc.release() call.
 func (kv *kv) getRemote(ctx context.Context) (pb.KVClient, error) {
 	if err := kv.rc.acquire(ctx); err != nil {
 		return nil, err
diff --git a/clientv3/remote_client.go b/clientv3/remote_client.go
index 216d7ed08..b8209b8a5 100644
--- a/clientv3/remote_client.go
+++ b/clientv3/remote_client.go
@@ -17,6 +17,8 @@ package clientv3
 import (
 	"sync"
 
+	"github.com/coreos/etcd/etcdserver/api/v3rpc/rpctypes"
+
 	"golang.org/x/net/context"
 	"google.golang.org/grpc"
 )
@@ -81,10 +83,14 @@ func (r *remoteClient) tryUpdate() bool {
 func (r *remoteClient) acquire(ctx context.Context) error {
 	for {
 		r.client.mu.RLock()
+		closed := r.client.cancel == nil
 		c := r.client.conn
 		r.mu.Lock()
 		match := r.conn == c
 		r.mu.Unlock()
+		if closed {
+			return rpctypes.ErrConnClosed
+		}
 		if match {
 			return nil
 		}
diff --git a/etcdserver/api/v3rpc/rpctypes/error.go b/etcdserver/api/v3rpc/rpctypes/error.go
index 42c8c5a13..3868eff26 100644
--- a/etcdserver/api/v3rpc/rpctypes/error.go
+++ b/etcdserver/api/v3rpc/rpctypes/error.go
@@ -104,6 +104,8 @@ var (
 
 	ErrNoLeader   = Error(ErrGRPCNoLeader)
 	ErrNotCapable = Error(ErrGRPCNotCapable)
+
+	ErrConnClosed = EtcdError{code: codes.Unavailable, desc: "clientv3: connection closed"}
 )
 
 // EtcdError defines gRPC server errors.
diff --git a/integration/cluster.go b/integration/cluster.go
index 68f6bd112..c339e37c9 100644
--- a/integration/cluster.go
+++ b/integration/cluster.go
@@ -27,6 +27,7 @@ import (
 	"sort"
 	"strconv"
 	"strings"
+	"sync"
 	"sync/atomic"
 	"testing"
 	"time"
@@ -749,6 +750,8 @@ func (p SortableMemberSliceByPeerURLs) Swap(i, j int) { p[i], p[j] = p[j], p[i]
 
 type ClusterV3 struct {
 	*cluster
+
+	mu      sync.Mutex
 	clients []*clientv3.Client
 }
 
@@ -756,7 +759,9 @@ type ClusterV3 struct {
 // for each cluster member.
 func NewClusterV3(t *testing.T, cfg *ClusterConfig) *ClusterV3 {
 	cfg.UseGRPC = true
-	clus := &ClusterV3{cluster: NewClusterByConfig(t, cfg)}
+	clus := &ClusterV3{
+		cluster: NewClusterByConfig(t, cfg),
+	}
 	for _, m := range clus.Members {
 		client, err := NewClientV3(m)
 		if err != nil {
@@ -769,12 +774,23 @@ func NewClusterV3(t *testing.T, cfg *ClusterConfig) *ClusterV3 {
 	return clus
 }
 
+func (c *ClusterV3) TakeClient(idx int) {
+	c.mu.Lock()
+	c.clients[idx] = nil
+	c.mu.Unlock()
+}
+
 func (c *ClusterV3) Terminate(t *testing.T) {
+	c.mu.Lock()
 	for _, client := range c.clients {
+		if client == nil {
+			continue
+		}
 		if err := client.Close(); err != nil {
 			t.Error(err)
 		}
 	}
+	c.mu.Unlock()
 	c.cluster.Terminate(t)
 }