MAPREDUCE-5924. Changed TaskAttemptImpl to ignore TA_COMMIT_PENDING event at COMMIT_PENDING state. Contributed by Zhijie Shen

git-svn-id: https://svn.apache.org/repos/asf/hadoop/common/trunk@1603338 13f79535-47bb-0310-9956-ffa450edef68

hadoop-mapreduce-project/CHANGES.txt

@@ -264,6 +264,9 @@ Release 2.5.0 - UNRELEASED
 
     MAPREDUCE-5920. Add Xattr option in DistCp docs. (Yi Liu via cnauroth)
 
+    MAPREDUCE-5924. Changed TaskAttemptImpl to ignore TA_COMMIT_PENDING event
+    at COMMIT_PENDING state. (Zhijie Shen via jianhe)
+
 Release 2.4.1 - 2014-06-23 
 
   INCOMPATIBLE CHANGES

hadoop-mapreduce-project/hadoop-mapreduce-client/hadoop-mapreduce-client-app/src/main/java/org/apache/hadoop/mapreduce/v2/app/job/impl/TaskAttemptImpl.java

@@ -335,6 +335,15 @@ public abstract class TaskAttemptImpl implements
      .addTransition(TaskAttemptStateInternal.COMMIT_PENDING,
          TaskAttemptStateInternal.FAIL_CONTAINER_CLEANUP,
          TaskAttemptEventType.TA_TIMED_OUT, CLEANUP_CONTAINER_TRANSITION)
+     // AM is likely to receive duplicate TA_COMMIT_PENDINGs as the task attempt
+     // will re-send the commit message until it doesn't encounter any
+     // IOException and succeeds in delivering the commit message.
+     // Ignoring the duplicate commit message is a short-term fix. In long term,
+     // we need to make use of retry cache to help this and other MR protocol
+     // APIs that can be considered as @AtMostOnce.
+     .addTransition(TaskAttemptStateInternal.COMMIT_PENDING,
+         TaskAttemptStateInternal.COMMIT_PENDING,
+         TaskAttemptEventType.TA_COMMIT_PENDING)
 
      // Transitions from SUCCESS_CONTAINER_CLEANUP state
      // kill and cleanup the container

hadoop-mapreduce-project/hadoop-mapreduce-client/hadoop-mapreduce-client-app/src/test/java/org/apache/hadoop/mapreduce/v2/app/TestMRApp.java

@@ -112,6 +112,15 @@ public class TestMRApp {
     //wait for first attempt to commit pending
     app.waitForState(attempt, TaskAttemptState.COMMIT_PENDING);
 
+    //re-send the commit pending signal to the task
+    app.getContext().getEventHandler().handle(
+        new TaskAttemptEvent(
+            attempt.getID(),
+            TaskAttemptEventType.TA_COMMIT_PENDING));
+
+    //the task attempt should be still at COMMIT_PENDING
+    app.waitForState(attempt, TaskAttemptState.COMMIT_PENDING);
+
     //send the done signal to the task
     app.getContext().getEventHandler().handle(
         new TaskAttemptEvent(