diff --git a/core/src/main/scala/org/apache/spark/scheduler/cluster/StandaloneSchedulerBackend.scala b/core/src/main/scala/org/apache/spark/scheduler/cluster/StandaloneSchedulerBackend.scala
index fd8e64454bf7..a4e2a7434128 100644
--- a/core/src/main/scala/org/apache/spark/scheduler/cluster/StandaloneSchedulerBackend.scala
+++ b/core/src/main/scala/org/apache/spark/scheduler/cluster/StandaloneSchedulerBackend.scala
@@ -58,7 +58,13 @@ private[spark] class StandaloneSchedulerBackend(
 
   override def start() {
     super.start()
-    launcherBackend.connect()
+
+    // SPARK-21159. The scheduler backend should only try to connect to the launcher when in client
+    // mode. In cluster mode, the code that submits the application to the Master needs to connect
+    // to the launcher instead.
+    if (sc.deployMode == "client") {
+      launcherBackend.connect()
+    }
 
     // The endpoint for executors to talk to us
     val driverUrl = RpcEndpointAddress(