[SPARK-14701][STREAMING] First stop the event loop, then stop the checkpoint writer in JobGenerator

Currently if we call `streamingContext.stop` (e.g. in a `StreamingListener.onBatchCompleted` callback) when a batch is about to complete, a `rejectedException` may get thrown from `checkPointWriter.executor`, since the `eventLoop` will try to process `DoCheckpoint` events even after the `checkPointWriter.executor` was stopped.

Please see [SPARK-14701](https://issues.apache.org/jira/browse/SPARK-14701) for details and stack traces.

## What changes were proposed in this pull request?

Reversed the stopping order of `event loop` and `checkpoint writer`.

## How was this patch tested?

Existing test suits.
(no dedicated test suits were added because the change is simple to reason about)

Author: Liwei Lin <lwlin7@gmail.com>

Closes #12489 from lw-lin/spark-14701.

1 files changed, 2 insertions, 2 deletions

diff --git a/streaming/src/main/scala/org/apache/spark/streaming/scheduler/JobGenerator.scala b/streaming/src/main/scala/org/apache/spark/streaming/scheduler/JobGenerator.scala
index 307ff1f7ec..8f9421fc09 100644
--- a/streaming/src/main/scala/org/apache/spark/streaming/scheduler/JobGenerator.scala
+++ b/streaming/src/main/scala/org/apache/spark/streaming/scheduler/JobGenerator.scala
@@ -154,9 +154,9 @@ class JobGenerator(jobScheduler: JobScheduler) extends Logging {
       graph.stop()
     }
 
-    // Stop the event loop and checkpoint writer
-    if (shouldCheckpoint) checkpointWriter.stop()
+    // First stop the event loop, then stop the checkpoint writer; see SPARK-14701
     eventLoop.stop()
+    if (shouldCheckpoint) checkpointWriter.stop()
     logInfo("Stopped JobGenerator")
   }