MAPREDUCE-6190. If a task stucks before its first heartbeat, it never timeouts and the MR job becomes stuck. Contributed by Zhaohui Xin.

hadoop-mapreduce-project/hadoop-mapreduce-client/hadoop-mapreduce-client-app/src/main/java/org/apache/hadoop/mapreduce/v2/app/TaskHeartbeatHandler.java

@@ -22,6 +22,7 @@
 import java.util.Map;
 import java.util.concurrent.ConcurrentHashMap;
 import java.util.concurrent.ConcurrentMap;
+import java.util.concurrent.atomic.AtomicBoolean;
 
 import com.google.common.annotations.VisibleForTesting;
 import org.apache.hadoop.conf.Configuration;
@@ -47,11 +48,13 @@
 @SuppressWarnings({"unchecked", "rawtypes"})
 public class TaskHeartbeatHandler extends AbstractService {
 
-  private static class ReportTime {
+  static class ReportTime {
     private long lastProgress;
+    private final AtomicBoolean reported;
 
     public ReportTime(long time) {
       setLastProgress(time);
+      reported = new AtomicBoolean(false);
     }
     
     public synchronized void setLastProgress(long time) {
@@ -61,6 +64,10 @@ public synchronized void setLastProgress(long time) {
     public synchronized long getLastProgress() {
       return lastProgress;
     }
+
+    public boolean isReported(){
+      return reported.get();
+    }
   }
   
   private static final Logger LOG =
@@ -72,6 +79,7 @@ public synchronized long getLastProgress() {
   private volatile boolean stopped;
   private long taskTimeOut;
   private long unregisterTimeOut;
+  private long taskStuckTimeOut;
   private int taskTimeOutCheckInterval = 30 * 1000; // 30 seconds.
 
   private final EventHandler eventHandler;
@@ -98,6 +106,8 @@ protected void serviceInit(Configuration conf) throws Exception {
         MRJobConfig.TASK_TIMEOUT, MRJobConfig.DEFAULT_TASK_TIMEOUT_MILLIS);
     unregisterTimeOut = conf.getLong(MRJobConfig.TASK_EXIT_TIMEOUT,
         MRJobConfig.TASK_EXIT_TIMEOUT_DEFAULT);
+    taskStuckTimeOut = conf.getLong(MRJobConfig.TASK_STUCK_TIMEOUT_MS,
+        MRJobConfig.DEFAULT_TASK_STUCK_TIMEOUT_MS);
 
     // enforce task timeout is at least twice as long as task report interval
     long taskProgressReportIntervalMillis = MRJobConfUtil.
@@ -135,6 +145,7 @@ public void progressing(TaskAttemptId attemptID) {
     //TODO throw an exception if the task isn't registered.
     ReportTime time = runningAttempts.get(attemptID);
     if(time != null) {
+      time.reported.compareAndSet(false, true);
       time.setLastProgress(clock.getTime());
     }
   }
@@ -179,13 +190,21 @@ private void checkRunning(long currentTime) {
         Map.Entry<TaskAttemptId, ReportTime> entry = iterator.next();
         boolean taskTimedOut = (taskTimeOut > 0) &&
             (currentTime > (entry.getValue().getLastProgress() + taskTimeOut));
+        // when container in NM not started in a long time,
+        // we think the taskAttempt is stuck
+        boolean taskStuck = (!entry.getValue().isReported()) &&
+            (currentTime >
+                (entry.getValue().getLastProgress() + taskStuckTimeOut));
 
-        if(taskTimedOut) {
+        if(taskTimedOut || taskStuck) {
           // task is lost, remove from the list and raise lost event
           iterator.remove();
           eventHandler.handle(new TaskAttemptDiagnosticsUpdateEvent(entry
               .getKey(), "AttemptID:" + entry.getKey().toString()
-              + " Timed out after " + taskTimeOut / 1000 + " secs"));
+              + " task timeout set: " + taskTimeOut / 1000 + "s,"
+              + " taskTimedOut: " + taskTimedOut + ";"
+              + " task stuck timeout set: " + taskStuckTimeOut / 1000 + "s,"
+              + " taskStuck: " + taskStuck));
           eventHandler.handle(new TaskAttemptEvent(entry.getKey(),
               TaskAttemptEventType.TA_TIMED_OUT));
         }
@@ -205,6 +224,11 @@ private void checkRecentlyUnregistered(long currentTime) {
     }
   }
 
+  @VisibleForTesting
+  ConcurrentMap getRunningAttempts(){
+    return runningAttempts;
+  }
+
   @VisibleForTesting
   public long getTaskTimeOut() {
     return taskTimeOut;

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

@@ -18,6 +18,7 @@
 
 package org.apache.hadoop.mapreduce.v2.app;
 
+import static org.junit.Assert.assertFalse;
 import static org.mockito.Matchers.any;
 import static org.mockito.Mockito.mock;
 import static org.mockito.Mockito.times;
@@ -41,12 +42,15 @@
 import org.junit.Assert;
 import org.junit.Test;
 
+import java.util.Map;
+import java.util.concurrent.ConcurrentMap;
+
 
 public class TestTaskHeartbeatHandler {
   
   @SuppressWarnings({ "rawtypes", "unchecked" })
   @Test
-  public void testTimeout() throws InterruptedException {
+  public void testTaskTimeout() throws InterruptedException {
     EventHandler mockHandler = mock(EventHandler.class);
     Clock clock = SystemClock.getInstance();
     TaskHeartbeatHandler hb = new TaskHeartbeatHandler(mockHandler, clock, 1);
@@ -62,11 +66,13 @@ public void testTimeout() throws InterruptedException {
     hb.init(conf);
     hb.start();
     try {
-      ApplicationId appId = ApplicationId.newInstance(0l, 5);
+      ApplicationId appId = ApplicationId.newInstance(0L, 5);
       JobId jobId = MRBuilderUtils.newJobId(appId, 4);
       TaskId tid = MRBuilderUtils.newTaskId(jobId, 3, TaskType.MAP);
       TaskAttemptId taid = MRBuilderUtils.newTaskAttemptId(tid, 2);
       hb.register(taid);
+      // Task heartbeat once to avoid stuck
+      hb.progressing(taid);
       Thread.sleep(100);
       //Events only happen when the task is canceled
       verify(mockHandler, times(2)).handle(any(Event.class));
@@ -75,6 +81,47 @@ public void testTimeout() throws InterruptedException {
     }
   }
 
+  @SuppressWarnings("unchecked")
+  @Test
+  public void testTaskStuck() throws InterruptedException {
+    EventHandler mockHandler = mock(EventHandler.class);
+    Clock clock = SystemClock.getInstance();
+    TaskHeartbeatHandler hb = new TaskHeartbeatHandler(mockHandler, clock, 1);
+
+
+    Configuration conf = new Configuration();
+    conf.setLong(MRJobConfig.TASK_STUCK_TIMEOUT_MS, 10); // 10ms
+    conf.setInt(MRJobConfig.TASK_TIMEOUT, 1000); //1000 ms
+    // set TASK_PROGRESS_REPORT_INTERVAL to a value smaller than TASK_TIMEOUT
+    // so that TASK_TIMEOUT is not overridden
+    conf.setLong(MRJobConfig.TASK_PROGRESS_REPORT_INTERVAL, 5);
+    conf.setInt(MRJobConfig.TASK_TIMEOUT_CHECK_INTERVAL_MS, 10); //10 ms
+
+    hb.init(conf);
+    hb.start();
+    try {
+      ApplicationId appId = ApplicationId.newInstance(0L, 5);
+      JobId jobId = MRBuilderUtils.newJobId(appId, 4);
+      TaskId tid = MRBuilderUtils.newTaskId(jobId, 3, TaskType.MAP);
+      TaskAttemptId taid = MRBuilderUtils.newTaskAttemptId(tid, 2);
+      hb.register(taid);
+
+      ConcurrentMap<TaskAttemptId, TaskHeartbeatHandler.ReportTime>
+          runningAttempts = hb.getRunningAttempts();
+      for (Map.Entry<TaskAttemptId, TaskHeartbeatHandler.ReportTime> entry
+          : runningAttempts.entrySet()) {
+        assertFalse(entry.getValue().isReported());
+      }
+
+      Thread.sleep(100);
+
+      //Events only happen when the task is canceled
+      verify(mockHandler, times(2)).handle(any(Event.class));
+    } finally {
+      hb.stop();
+    }
+  }
+
   /**
    * Test if the final heartbeat timeout is set correctly when task progress
    * report interval is set bigger than the task timeout in the configuration.
@@ -120,7 +167,7 @@ public void testTaskUnregistered() throws Exception {
     hb.init(conf);
     hb.start();
     try {
-      ApplicationId appId = ApplicationId.newInstance(0l, 5);
+      ApplicationId appId = ApplicationId.newInstance(0L, 5);
       JobId jobId = MRBuilderUtils.newJobId(appId, 4);
       TaskId tid = MRBuilderUtils.newTaskId(jobId, 3, TaskType.MAP);
       final TaskAttemptId taid = MRBuilderUtils.newTaskAttemptId(tid, 2);

hadoop-mapreduce-project/hadoop-mapreduce-client/hadoop-mapreduce-client-core/src/main/java/org/apache/hadoop/mapreduce/MRJobConfig.java

@@ -353,6 +353,14 @@ public interface MRJobConfig {
   public static final String TASK_TIMEOUT = "mapreduce.task.timeout";
   long DEFAULT_TASK_TIMEOUT_MILLIS = 5 * 60 * 1000L;
 
+  /**
+   * The max timeout before receiving remote task's first heartbeat.
+   * This parameter is in order to avoid waiting for the container
+   * to start indefinitely, which made task stuck in the NEW state.
+   */
+  String TASK_STUCK_TIMEOUT_MS = "mapreduce.task.stuck.timeout-ms";
+  long DEFAULT_TASK_STUCK_TIMEOUT_MS = 10 * 60 * 1000L;
+
   String TASK_PROGRESS_REPORT_INTERVAL =
       "mapreduce.task.progress-report.interval";
 

hadoop-mapreduce-project/hadoop-mapreduce-client/hadoop-mapreduce-client-core/src/main/resources/mapred-default.xml

@@ -276,6 +276,15 @@
   </description>
 </property>
 
+<property>
+  <name>mapreduce.task.stuck.timeout-ms</name>
+  <value>600000</value>
+  <description>The max timeout before receiving remote task's first heartbeat.
+    This parameter is in order to avoid waiting for the container
+    to start indefinitely, which made task stuck in the NEW state.
+  </description>
+</property>
+
 <property>
   <name>mapreduce.map.memory.mb</name>
   <value>-1</value>