[SPARK-15262] Synchronize block manager / scheduler executor state

## What changes were proposed in this pull request?

If an executor is still alive even after the scheduler has removed its metadata, we may receive a heartbeat from that executor and tell its block manager to reregister itself. If that happens, the block manager master will know about the executor, but the scheduler will not.

That is a dangerous situation, because when the executor does get disconnected later, the scheduler will not ask the block manager to also remove metadata for that executor. Later, when we try to clean up an RDD or a broadcast variable, we may try to send a message to that executor, triggering an exception.

## How was this patch tested?

Jenkins.

Author: Andrew Or <[email protected]>

Closes #13055 from andrewor14/block-manager-remove.

core/src/main/scala/org/apache/spark/scheduler/cluster/CoarseGrainedSchedulerBackend.scala

@@ -289,7 +289,14 @@ class CoarseGrainedSchedulerBackend(scheduler: TaskSchedulerImpl, val rpcEnv: Rp
           scheduler.executorLost(executorId, if (killed) ExecutorKilled else reason)
           listenerBus.post(
             SparkListenerExecutorRemoved(System.currentTimeMillis(), executorId, reason.toString))
-        case None => logInfo(s"Asked to remove non-existent executor $executorId")
+        case None =>
+          // SPARK-15262: If an executor is still alive even after the scheduler has removed
+          // its metadata, we may receive a heartbeat from that executor and tell its block
+          // manager to reregister itself. If that happens, the block manager master will know
+          // about the executor, but the scheduler will not. Therefore, we should remove the
+          // executor from the block manager when we hit this case.
+          scheduler.sc.env.blockManager.master.removeExecutor(executorId)
+          logInfo(s"Asked to remove non-existent executor $executorId")
       }
     }