Examples with TaskAttemptCompletionEvent org.apache.hadoop.mapreduce.v2.api.records.TaskAttemptCompletionEvent used on opensource projects

Example 1 with TaskAttemptCompletionEvent

use of org.apache.hadoop.mapreduce.v2.api.records.TaskAttemptCompletionEvent in project hadoop by apache.

the class TestJobHistoryEntities method testGetTaskAttemptCompletionEvent.

/**
   * Simple test of some methods of CompletedJob
   * @throws Exception
   */
@Test(timeout = 30000)
public void testGetTaskAttemptCompletionEvent() throws Exception {
    HistoryFileInfo info = mock(HistoryFileInfo.class);
    when(info.getConfFile()).thenReturn(fullConfPath);
    completedJob = new CompletedJob(conf, jobId, fullHistoryPath, loadTasks, "user", info, jobAclsManager);
    TaskCompletionEvent[] events = completedJob.getMapAttemptCompletionEvents(0, 1000);
    assertEquals(10, completedJob.getMapAttemptCompletionEvents(0, 10).length);
    int currentEventId = 0;
    for (TaskCompletionEvent taskAttemptCompletionEvent : events) {
        int eventId = taskAttemptCompletionEvent.getEventId();
        assertTrue(eventId >= currentEventId);
        currentEventId = eventId;
    }
    assertNull(completedJob.loadConfFile());
    // job name
    assertEquals("Sleep job", completedJob.getName());
    // queue name
    assertEquals("default", completedJob.getQueueName());
    // progress
    assertEquals(1.0, completedJob.getProgress(), 0.001);
    // 12 rows in answer
    assertEquals(12, completedJob.getTaskAttemptCompletionEvents(0, 1000).length);
    // select first 10 rows
    assertEquals(10, completedJob.getTaskAttemptCompletionEvents(0, 10).length);
    // select 5-10 rows include 5th
    assertEquals(7, completedJob.getTaskAttemptCompletionEvents(5, 10).length);
    // without errors
    assertEquals(1, completedJob.getDiagnostics().size());
    assertEquals("", completedJob.getDiagnostics().get(0));
    assertEquals(0, completedJob.getJobACLs().size());
}

Example 2 with TaskAttemptCompletionEvent

use of org.apache.hadoop.mapreduce.v2.api.records.TaskAttemptCompletionEvent in project hadoop by apache.

the class TestFetchFailure method testFetchFailure.

@Test
public void testFetchFailure() throws Exception {
    MRApp app = new MRApp(1, 1, false, this.getClass().getName(), true);
    Configuration conf = new Configuration();
    // map -> reduce -> fetch-failure -> map retry is incompatible with
    // sequential, single-task-attempt approach in uber-AM, so disable:
    conf.setBoolean(MRJobConfig.JOB_UBERTASK_ENABLE, false);
    Job job = app.submit(conf);
    app.waitForState(job, JobState.RUNNING);
    //all maps would be running
    Assert.assertEquals("Num tasks not correct", 2, job.getTasks().size());
    Iterator<Task> it = job.getTasks().values().iterator();
    Task mapTask = it.next();
    Task reduceTask = it.next();
    //wait for Task state move to RUNNING
    app.waitForState(mapTask, TaskState.RUNNING);
    TaskAttempt mapAttempt1 = mapTask.getAttempts().values().iterator().next();
    app.waitForState(mapAttempt1, TaskAttemptState.RUNNING);
    //send the done signal to the map attempt
    app.getContext().getEventHandler().handle(new TaskAttemptEvent(mapAttempt1.getID(), TaskAttemptEventType.TA_DONE));
    // wait for map success
    app.waitForState(mapTask, TaskState.SUCCEEDED);
    final int checkIntervalMillis = 10;
    final int waitForMillis = 800;
    GenericTestUtils.waitFor(new Supplier<Boolean>() {

        @Override
        public Boolean get() {
            TaskAttemptCompletionEvent[] events = job.getTaskAttemptCompletionEvents(0, 100);
            return events.length >= 1;
        }
    }, checkIntervalMillis, waitForMillis);
    TaskAttemptCompletionEvent[] events = job.getTaskAttemptCompletionEvents(0, 100);
    Assert.assertEquals("Num completion events not correct", 1, events.length);
    Assert.assertEquals("Event status not correct", TaskAttemptCompletionEventStatus.SUCCEEDED, events[0].getStatus());
    // wait for reduce to start running
    app.waitForState(reduceTask, TaskState.RUNNING);
    TaskAttempt reduceAttempt = reduceTask.getAttempts().values().iterator().next();
    app.waitForState(reduceAttempt, TaskAttemptState.RUNNING);
    //send 3 fetch failures from reduce to trigger map re execution
    sendFetchFailure(app, reduceAttempt, mapAttempt1, "host");
    sendFetchFailure(app, reduceAttempt, mapAttempt1, "host");
    sendFetchFailure(app, reduceAttempt, mapAttempt1, "host");
    //wait for map Task state move back to RUNNING
    app.waitForState(mapTask, TaskState.RUNNING);
    //map attempt must have become FAILED
    Assert.assertEquals("Map TaskAttempt state not correct", TaskAttemptState.FAILED, mapAttempt1.getState());
    Assert.assertEquals("Num attempts in Map Task not correct", 2, mapTask.getAttempts().size());
    Iterator<TaskAttempt> atIt = mapTask.getAttempts().values().iterator();
    atIt.next();
    TaskAttempt mapAttempt2 = atIt.next();
    app.waitForState(mapAttempt2, TaskAttemptState.RUNNING);
    //send the done signal to the second map attempt
    app.getContext().getEventHandler().handle(new TaskAttemptEvent(mapAttempt2.getID(), TaskAttemptEventType.TA_DONE));
    // wait for map success
    app.waitForState(mapTask, TaskState.SUCCEEDED);
    //send done to reduce
    app.getContext().getEventHandler().handle(new TaskAttemptEvent(reduceAttempt.getID(), TaskAttemptEventType.TA_DONE));
    app.waitForState(job, JobState.SUCCEEDED);
    //previous completion event now becomes obsolete
    Assert.assertEquals("Event status not correct", TaskAttemptCompletionEventStatus.OBSOLETE, events[0].getStatus());
    events = job.getTaskAttemptCompletionEvents(0, 100);
    Assert.assertEquals("Num completion events not correct", 4, events.length);
    Assert.assertEquals("Event map attempt id not correct", mapAttempt1.getID(), events[0].getAttemptId());
    Assert.assertEquals("Event map attempt id not correct", mapAttempt1.getID(), events[1].getAttemptId());
    Assert.assertEquals("Event map attempt id not correct", mapAttempt2.getID(), events[2].getAttemptId());
    Assert.assertEquals("Event redude attempt id not correct", reduceAttempt.getID(), events[3].getAttemptId());
    Assert.assertEquals("Event status not correct for map attempt1", TaskAttemptCompletionEventStatus.OBSOLETE, events[0].getStatus());
    Assert.assertEquals("Event status not correct for map attempt1", TaskAttemptCompletionEventStatus.FAILED, events[1].getStatus());
    Assert.assertEquals("Event status not correct for map attempt2", TaskAttemptCompletionEventStatus.SUCCEEDED, events[2].getStatus());
    Assert.assertEquals("Event status not correct for reduce attempt1", TaskAttemptCompletionEventStatus.SUCCEEDED, events[3].getStatus());
    TaskCompletionEvent[] mapEvents = job.getMapAttemptCompletionEvents(0, 2);
    TaskCompletionEvent[] convertedEvents = TypeConverter.fromYarn(events);
    Assert.assertEquals("Incorrect number of map events", 2, mapEvents.length);
    Assert.assertArrayEquals("Unexpected map events", Arrays.copyOfRange(convertedEvents, 0, 2), mapEvents);
    mapEvents = job.getMapAttemptCompletionEvents(2, 200);
    Assert.assertEquals("Incorrect number of map events", 1, mapEvents.length);
    Assert.assertEquals("Unexpected map event", convertedEvents[2], mapEvents[0]);
}

Example 3 with TaskAttemptCompletionEvent

use of org.apache.hadoop.mapreduce.v2.api.records.TaskAttemptCompletionEvent in project hadoop by apache.

the class TestFetchFailure method testFetchFailureWithRecovery.

/**
   * This tests that if a map attempt was failed (say due to fetch failures),
   * then it gets re-run. When the next map attempt is running, if the AM dies,
   * then, on AM re-run, the AM does not incorrectly remember the first failed
   * attempt. Currently recovery does not recover running tasks. Effectively,
   * the AM re-runs the maps from scratch.
   */
@Test
public void testFetchFailureWithRecovery() throws Exception {
    int runCount = 0;
    MRApp app = new MRAppWithHistory(1, 1, false, this.getClass().getName(), true, ++runCount);
    Configuration conf = new Configuration();
    // map -> reduce -> fetch-failure -> map retry is incompatible with
    // sequential, single-task-attempt approach in uber-AM, so disable:
    conf.setBoolean(MRJobConfig.JOB_UBERTASK_ENABLE, false);
    Job job = app.submit(conf);
    app.waitForState(job, JobState.RUNNING);
    //all maps would be running
    Assert.assertEquals("Num tasks not correct", 2, job.getTasks().size());
    Iterator<Task> it = job.getTasks().values().iterator();
    Task mapTask = it.next();
    Task reduceTask = it.next();
    //wait for Task state move to RUNNING
    app.waitForState(mapTask, TaskState.RUNNING);
    TaskAttempt mapAttempt1 = mapTask.getAttempts().values().iterator().next();
    app.waitForState(mapAttempt1, TaskAttemptState.RUNNING);
    //send the done signal to the map attempt
    app.getContext().getEventHandler().handle(new TaskAttemptEvent(mapAttempt1.getID(), TaskAttemptEventType.TA_DONE));
    // wait for map success
    app.waitForState(mapTask, TaskState.SUCCEEDED);
    TaskAttemptCompletionEvent[] events = job.getTaskAttemptCompletionEvents(0, 100);
    Assert.assertEquals("Num completion events not correct", 1, events.length);
    Assert.assertEquals("Event status not correct", TaskAttemptCompletionEventStatus.SUCCEEDED, events[0].getStatus());
    // wait for reduce to start running
    app.waitForState(reduceTask, TaskState.RUNNING);
    TaskAttempt reduceAttempt = reduceTask.getAttempts().values().iterator().next();
    app.waitForState(reduceAttempt, TaskAttemptState.RUNNING);
    //send 3 fetch failures from reduce to trigger map re execution
    sendFetchFailure(app, reduceAttempt, mapAttempt1, "host");
    sendFetchFailure(app, reduceAttempt, mapAttempt1, "host");
    sendFetchFailure(app, reduceAttempt, mapAttempt1, "host");
    //wait for map Task state move back to RUNNING
    app.waitForState(mapTask, TaskState.RUNNING);
    // Crash the app again.
    app.stop();
    //rerun
    app = new MRAppWithHistory(1, 1, false, this.getClass().getName(), false, ++runCount);
    conf = new Configuration();
    conf.setBoolean(MRJobConfig.MR_AM_JOB_RECOVERY_ENABLE, true);
    conf.setBoolean(MRJobConfig.JOB_UBERTASK_ENABLE, false);
    job = app.submit(conf);
    app.waitForState(job, JobState.RUNNING);
    //all maps would be running
    Assert.assertEquals("Num tasks not correct", 2, job.getTasks().size());
    it = job.getTasks().values().iterator();
    mapTask = it.next();
    reduceTask = it.next();
    // the map is not in a SUCCEEDED state after restart of AM
    app.waitForState(mapTask, TaskState.RUNNING);
    mapAttempt1 = mapTask.getAttempts().values().iterator().next();
    app.waitForState(mapAttempt1, TaskAttemptState.RUNNING);
    //send the done signal to the map attempt
    app.getContext().getEventHandler().handle(new TaskAttemptEvent(mapAttempt1.getID(), TaskAttemptEventType.TA_DONE));
    // wait for map success
    app.waitForState(mapTask, TaskState.SUCCEEDED);
    reduceAttempt = reduceTask.getAttempts().values().iterator().next();
    //send done to reduce
    app.getContext().getEventHandler().handle(new TaskAttemptEvent(reduceAttempt.getID(), TaskAttemptEventType.TA_DONE));
    app.waitForState(job, JobState.SUCCEEDED);
    events = job.getTaskAttemptCompletionEvents(0, 100);
    Assert.assertEquals("Num completion events not correct", 2, events.length);
}

Example 4 with TaskAttemptCompletionEvent

use of org.apache.hadoop.mapreduce.v2.api.records.TaskAttemptCompletionEvent in project hadoop by apache.

the class TestTaskAttemptListenerImpl method createTce.

private static TaskAttemptCompletionEvent createTce(int eventId, boolean isMap, TaskAttemptCompletionEventStatus status) {
    JobId jid = MRBuilderUtils.newJobId(12345, 1, 1);
    TaskId tid = MRBuilderUtils.newTaskId(jid, 0, isMap ? org.apache.hadoop.mapreduce.v2.api.records.TaskType.MAP : org.apache.hadoop.mapreduce.v2.api.records.TaskType.REDUCE);
    TaskAttemptId attemptId = MRBuilderUtils.newTaskAttemptId(tid, 0);
    RecordFactory recordFactory = RecordFactoryProvider.getRecordFactory(null);
    TaskAttemptCompletionEvent tce = recordFactory.newRecordInstance(TaskAttemptCompletionEvent.class);
    tce.setEventId(eventId);
    tce.setAttemptId(attemptId);
    tce.setStatus(status);
    return tce;
}

Example 5 with TaskAttemptCompletionEvent

use of org.apache.hadoop.mapreduce.v2.api.records.TaskAttemptCompletionEvent in project hadoop by apache.

the class TestJobImpl method testUnusableNodeTransition.

@Test(timeout = 20000)
public void testUnusableNodeTransition() throws Exception {
    Configuration conf = new Configuration();
    conf.set(MRJobConfig.MR_AM_STAGING_DIR, stagingDir);
    conf.setInt(MRJobConfig.NUM_REDUCES, 1);
    DrainDispatcher dispatcher = new DrainDispatcher();
    dispatcher.init(conf);
    dispatcher.start();
    CyclicBarrier syncBarrier = new CyclicBarrier(2);
    OutputCommitter committer = new TestingOutputCommitter(syncBarrier, true);
    CommitterEventHandler commitHandler = createCommitterEventHandler(dispatcher, committer);
    commitHandler.init(conf);
    commitHandler.start();
    final JobImpl job = createRunningStubbedJob(conf, dispatcher, 2, null);
    // add a special task event handler to put the task back to running in case
    // of task rescheduling/killing
    EventHandler<TaskAttemptEvent> taskAttemptEventHandler = new EventHandler<TaskAttemptEvent>() {

        @Override
        public void handle(TaskAttemptEvent event) {
            if (event.getType() == TaskAttemptEventType.TA_KILL) {
                job.decrementSucceededMapperCount();
            }
        }
    };
    dispatcher.register(TaskAttemptEventType.class, taskAttemptEventHandler);
    // replace the tasks with spied versions to return the right attempts
    Map<TaskId, Task> spiedTasks = new HashMap<TaskId, Task>();
    List<NodeReport> nodeReports = new ArrayList<NodeReport>();
    Map<NodeReport, TaskId> nodeReportsToTaskIds = new HashMap<NodeReport, TaskId>();
    for (Map.Entry<TaskId, Task> e : job.tasks.entrySet()) {
        TaskId taskId = e.getKey();
        Task task = e.getValue();
        if (taskId.getTaskType() == TaskType.MAP) {
            // add an attempt to the task to simulate nodes
            NodeId nodeId = mock(NodeId.class);
            TaskAttempt attempt = mock(TaskAttempt.class);
            when(attempt.getNodeId()).thenReturn(nodeId);
            TaskAttemptId attemptId = MRBuilderUtils.newTaskAttemptId(taskId, 0);
            when(attempt.getID()).thenReturn(attemptId);
            // create a spied task
            Task spied = spy(task);
            doReturn(attempt).when(spied).getAttempt(any(TaskAttemptId.class));
            spiedTasks.put(taskId, spied);
            // create a NodeReport based on the node id
            NodeReport report = mock(NodeReport.class);
            when(report.getNodeState()).thenReturn(NodeState.UNHEALTHY);
            when(report.getNodeId()).thenReturn(nodeId);
            nodeReports.add(report);
            nodeReportsToTaskIds.put(report, taskId);
        }
    }
    // replace the tasks with the spied tasks
    job.tasks.putAll(spiedTasks);
    // complete all mappers first
    for (TaskId taskId : job.tasks.keySet()) {
        if (taskId.getTaskType() == TaskType.MAP) {
            // generate a task attempt completed event first to populate the
            // nodes-to-succeeded-attempts map
            TaskAttemptCompletionEvent tce = Records.newRecord(TaskAttemptCompletionEvent.class);
            TaskAttemptId attemptId = MRBuilderUtils.newTaskAttemptId(taskId, 0);
            tce.setAttemptId(attemptId);
            tce.setStatus(TaskAttemptCompletionEventStatus.SUCCEEDED);
            job.handle(new JobTaskAttemptCompletedEvent(tce));
            // complete the task itself
            job.handle(new JobTaskEvent(taskId, TaskState.SUCCEEDED));
            Assert.assertEquals(JobState.RUNNING, job.getState());
        }
    }
    // add an event for a node transition
    NodeReport firstMapperNodeReport = nodeReports.get(0);
    NodeReport secondMapperNodeReport = nodeReports.get(1);
    job.handle(new JobUpdatedNodesEvent(job.getID(), Collections.singletonList(firstMapperNodeReport)));
    dispatcher.await();
    // complete the reducer
    for (TaskId taskId : job.tasks.keySet()) {
        if (taskId.getTaskType() == TaskType.REDUCE) {
            job.handle(new JobTaskEvent(taskId, TaskState.SUCCEEDED));
        }
    }
    // add another event for a node transition for the other mapper
    // this should not trigger rescheduling
    job.handle(new JobUpdatedNodesEvent(job.getID(), Collections.singletonList(secondMapperNodeReport)));
    // complete the first mapper that was rescheduled
    TaskId firstMapper = nodeReportsToTaskIds.get(firstMapperNodeReport);
    job.handle(new JobTaskEvent(firstMapper, TaskState.SUCCEEDED));
    // verify the state is moving to committing
    assertJobState(job, JobStateInternal.COMMITTING);
    // let the committer complete and verify the job succeeds
    syncBarrier.await();
    assertJobState(job, JobStateInternal.SUCCEEDED);
    dispatcher.stop();
    commitHandler.stop();
}