diff --git a/integration/cluster.go b/integration/cluster.go
index d2f5de39f861..aa693698c2ee 100644
--- a/integration/cluster.go
+++ b/integration/cluster.go
@@ -465,6 +465,8 @@ type member struct {
 
 	// serverClient is a clientv3 that directly calls the etcdserver.
 	serverClient *clientv3.Client
+
+	keepDataDirStop bool
 }
 
 func (m *member) GRPCAddr() string { return m.grpcAddr }
@@ -769,8 +771,10 @@ func (m *member) Restart(t *testing.T) error {
 func (m *member) Terminate(t *testing.T) {
 	plog.Printf("terminating %s (%s)", m.Name, m.grpcAddr)
 	m.Close()
-	if err := os.RemoveAll(m.ServerConfig.DataDir); err != nil {
-		t.Fatal(err)
+	if !m.keepDataDirStop {
+		if err := os.RemoveAll(m.ServerConfig.DataDir); err != nil {
+			t.Fatal(err)
+		}
 	}
 	plog.Printf("terminated %s (%s)", m.Name, m.grpcAddr)
 }
diff --git a/integration/cluster_test.go b/integration/cluster_test.go
index 1bf5fc56c485..9c97b935963c 100644
--- a/integration/cluster_test.go
+++ b/integration/cluster_test.go
@@ -441,6 +441,50 @@ func TestRejectUnhealthyRemove(t *testing.T) {
 	}
 }
 
+// TestRestartRemoved ensures that restarting removed member must exit
+// if 'initial-cluster-state' is set 'new' and old data directory still exists
+// (see https://github.com/coreos/etcd/issues/7512 for more).
+func TestRestartRemoved(t *testing.T) {
+	defer testutil.AfterTest(t)
+
+	// 1. start single-member cluster
+	c := NewCluster(t, 1)
+	for _, m := range c.Members {
+		m.ServerConfig.StrictReconfigCheck = true
+	}
+	c.Launch(t)
+	defer c.Terminate(t)
+
+	// 2. add a new member
+	c.AddMember(t)
+	c.WaitLeader(t)
+
+	oldm := c.Members[0]
+	oldm.keepDataDirStop = true
+
+	// 3. remove first member, shut down without deleting data
+	if err := c.removeMember(t, uint64(c.Members[0].s.ID())); err != nil {
+		t.Fatalf("expected to remove member, got error %v", err)
+	}
+	c.WaitLeader(t)
+
+	// 4. restart first member with 'initial-cluster-state=new'
+	// wrong config, expects exit within ReqTimeout
+	oldm.ServerConfig.NewCluster = false
+	if err := oldm.Restart(t); err != nil {
+		t.Fatalf("unexpected ForceRestart error: %v", err)
+	}
+	defer func() {
+		oldm.Close()
+		os.RemoveAll(oldm.ServerConfig.DataDir)
+	}()
+	select {
+	case <-oldm.s.StopNotify():
+	case <-time.After(30 * time.Second):
+		t.Fatalf("removed member didn't return after %v", 30*time.Second)
+	}
+}
+
 // clusterMustProgress ensures that cluster can make progress. It creates
 // a random key first, and check the new key could be got from all client urls
 // of the cluster.