Examples with TaskAttempt org.apache.hadoop.mapreduce.v2.app.job.TaskAttempt used on opensource projects

Example 31 with TaskAttempt

use of org.apache.hadoop.mapreduce.v2.app.job.TaskAttempt in project hadoop by apache.

the class TestRecovery method testCrashed.

/**
   * AM with 2 maps and 1 reduce. For 1st map, one attempt fails, one attempt
   * completely disappears because of failed launch, one attempt gets killed and
   * one attempt succeeds. AM crashes after the first tasks finishes and
   * recovers completely and succeeds in the second generation.
   * 
   * @throws Exception
   */
@Test
public void testCrashed() throws Exception {
    int runCount = 0;
    long am1StartTimeEst = System.currentTimeMillis();
    MRApp app = new MRAppWithHistory(2, 1, false, this.getClass().getName(), true, ++runCount);
    Configuration conf = new Configuration();
    conf.setBoolean("mapred.mapper.new-api", true);
    conf.setBoolean("mapred.reducer.new-api", true);
    conf.setBoolean(MRJobConfig.JOB_UBERTASK_ENABLE, false);
    conf.set(FileOutputFormat.OUTDIR, outputDir.toString());
    Job job = app.submit(conf);
    app.waitForState(job, JobState.RUNNING);
    long jobStartTime = job.getReport().getStartTime();
    //all maps would be running
    Assert.assertEquals("No of tasks not correct", 3, job.getTasks().size());
    Iterator<Task> it = job.getTasks().values().iterator();
    Task mapTask1 = it.next();
    Task mapTask2 = it.next();
    Task reduceTask = it.next();
    // all maps must be running
    app.waitForState(mapTask1, TaskState.RUNNING);
    app.waitForState(mapTask2, TaskState.RUNNING);
    TaskAttempt task1Attempt1 = mapTask1.getAttempts().values().iterator().next();
    TaskAttempt task2Attempt = mapTask2.getAttempts().values().iterator().next();
    //before sending the TA_DONE, event make sure attempt has come to 
    //RUNNING state
    app.waitForState(task1Attempt1, TaskAttemptState.RUNNING);
    app.waitForState(task2Attempt, TaskAttemptState.RUNNING);
    app.waitForState(reduceTask, TaskState.RUNNING);
    /////////// Play some games with the TaskAttempts of the first task //////
    //send the fail signal to the 1st map task attempt
    app.getContext().getEventHandler().handle(new TaskAttemptEvent(task1Attempt1.getID(), TaskAttemptEventType.TA_FAILMSG));
    app.waitForState(task1Attempt1, TaskAttemptState.FAILED);
    int timeOut = 0;
    while (mapTask1.getAttempts().size() != 2 && timeOut++ < 10) {
        Thread.sleep(2000);
        LOG.info("Waiting for next attempt to start");
    }
    Assert.assertEquals(2, mapTask1.getAttempts().size());
    Iterator<TaskAttempt> itr = mapTask1.getAttempts().values().iterator();
    itr.next();
    TaskAttempt task1Attempt2 = itr.next();
    // wait for the second task attempt to be assigned.
    waitForContainerAssignment(task1Attempt2);
    // This attempt will automatically fail because of the way ContainerLauncher
    // is setup
    // This attempt 'disappears' from JobHistory and so causes MAPREDUCE-3846
    app.getContext().getEventHandler().handle(new TaskAttemptEvent(task1Attempt2.getID(), TaskAttemptEventType.TA_CONTAINER_LAUNCH_FAILED));
    app.waitForState(task1Attempt2, TaskAttemptState.FAILED);
    timeOut = 0;
    while (mapTask1.getAttempts().size() != 3 && timeOut++ < 10) {
        Thread.sleep(2000);
        LOG.info("Waiting for next attempt to start");
    }
    Assert.assertEquals(3, mapTask1.getAttempts().size());
    itr = mapTask1.getAttempts().values().iterator();
    itr.next();
    itr.next();
    TaskAttempt task1Attempt3 = itr.next();
    app.waitForState(task1Attempt3, TaskAttemptState.RUNNING);
    //send the kill signal to the 1st map 3rd attempt
    app.getContext().getEventHandler().handle(new TaskAttemptEvent(task1Attempt3.getID(), TaskAttemptEventType.TA_KILL));
    app.waitForState(task1Attempt3, TaskAttemptState.KILLED);
    timeOut = 0;
    while (mapTask1.getAttempts().size() != 4 && timeOut++ < 10) {
        Thread.sleep(2000);
        LOG.info("Waiting for next attempt to start");
    }
    Assert.assertEquals(4, mapTask1.getAttempts().size());
    itr = mapTask1.getAttempts().values().iterator();
    itr.next();
    itr.next();
    itr.next();
    TaskAttempt task1Attempt4 = itr.next();
    app.waitForState(task1Attempt4, TaskAttemptState.RUNNING);
    //send the done signal to the 1st map 4th attempt
    app.getContext().getEventHandler().handle(new TaskAttemptEvent(task1Attempt4.getID(), TaskAttemptEventType.TA_DONE));
    /////////// End of games with the TaskAttempts of the first task //////
    //wait for first map task to complete
    app.waitForState(mapTask1, TaskState.SUCCEEDED);
    long task1StartTime = mapTask1.getReport().getStartTime();
    long task1FinishTime = mapTask1.getReport().getFinishTime();
    //stop the app
    app.stop();
    //rerun
    //in rerun the 1st map will be recovered from previous run
    long am2StartTimeEst = System.currentTimeMillis();
    app = new MRAppWithHistory(2, 1, false, this.getClass().getName(), false, ++runCount);
    conf = new Configuration();
    conf.setBoolean(MRJobConfig.MR_AM_JOB_RECOVERY_ENABLE, true);
    conf.setBoolean("mapred.mapper.new-api", true);
    conf.setBoolean("mapred.reducer.new-api", true);
    conf.set(FileOutputFormat.OUTDIR, outputDir.toString());
    conf.setBoolean(MRJobConfig.JOB_UBERTASK_ENABLE, false);
    job = app.submit(conf);
    app.waitForState(job, JobState.RUNNING);
    //all maps would be running
    Assert.assertEquals("No of tasks not correct", 3, job.getTasks().size());
    it = job.getTasks().values().iterator();
    mapTask1 = it.next();
    mapTask2 = it.next();
    reduceTask = it.next();
    // first map will be recovered, no need to send done
    app.waitForState(mapTask1, TaskState.SUCCEEDED);
    app.waitForState(mapTask2, TaskState.RUNNING);
    task2Attempt = mapTask2.getAttempts().values().iterator().next();
    //before sending the TA_DONE, event make sure attempt has come to 
    //RUNNING state
    app.waitForState(task2Attempt, TaskAttemptState.RUNNING);
    //send the done signal to the 2nd map task
    app.getContext().getEventHandler().handle(new TaskAttemptEvent(mapTask2.getAttempts().values().iterator().next().getID(), TaskAttemptEventType.TA_DONE));
    //wait to get it completed
    app.waitForState(mapTask2, TaskState.SUCCEEDED);
    //wait for reduce to be running before sending done
    app.waitForState(reduceTask, TaskState.RUNNING);
    //send the done signal to the reduce
    app.getContext().getEventHandler().handle(new TaskAttemptEvent(reduceTask.getAttempts().values().iterator().next().getID(), TaskAttemptEventType.TA_DONE));
    app.waitForState(job, JobState.SUCCEEDED);
    app.verifyCompleted();
    Assert.assertEquals("Job Start time not correct", jobStartTime, job.getReport().getStartTime());
    Assert.assertEquals("Task Start time not correct", task1StartTime, mapTask1.getReport().getStartTime());
    Assert.assertEquals("Task Finish time not correct", task1FinishTime, mapTask1.getReport().getFinishTime());
    Assert.assertEquals(2, job.getAMInfos().size());
    int attemptNum = 1;
    // Verify AMInfo
    for (AMInfo amInfo : job.getAMInfos()) {
        Assert.assertEquals(attemptNum++, amInfo.getAppAttemptId().getAttemptId());
        Assert.assertEquals(amInfo.getAppAttemptId(), amInfo.getContainerId().getApplicationAttemptId());
        Assert.assertEquals(MRApp.NM_HOST, amInfo.getNodeManagerHost());
        Assert.assertEquals(MRApp.NM_PORT, amInfo.getNodeManagerPort());
        Assert.assertEquals(MRApp.NM_HTTP_PORT, amInfo.getNodeManagerHttpPort());
    }
    long am1StartTimeReal = job.getAMInfos().get(0).getStartTime();
    long am2StartTimeReal = job.getAMInfos().get(1).getStartTime();
    Assert.assertTrue(am1StartTimeReal >= am1StartTimeEst && am1StartTimeReal <= am2StartTimeEst);
    Assert.assertTrue(am2StartTimeReal >= am2StartTimeEst && am2StartTimeReal <= System.currentTimeMillis());
// TODO Add verification of additional data from jobHistory - whatever was
// available in the failed attempt should be available here
}

Example 32 with TaskAttempt

use of org.apache.hadoop.mapreduce.v2.app.job.TaskAttempt in project hadoop by apache.

the class TestRecovery method testOutputRecovery.

@Test
public void testOutputRecovery() throws Exception {
    int runCount = 0;
    MRApp app = new MRAppWithHistory(1, 2, false, this.getClass().getName(), true, ++runCount);
    Configuration conf = new Configuration();
    conf.setBoolean("mapred.mapper.new-api", true);
    conf.setBoolean("mapred.reducer.new-api", true);
    conf.setBoolean(MRJobConfig.JOB_UBERTASK_ENABLE, false);
    conf.set(FileOutputFormat.OUTDIR, outputDir.toString());
    Job job = app.submit(conf);
    app.waitForState(job, JobState.RUNNING);
    Assert.assertEquals("No of tasks not correct", 3, job.getTasks().size());
    Iterator<Task> it = job.getTasks().values().iterator();
    Task mapTask1 = it.next();
    Task reduceTask1 = it.next();
    // all maps must be running
    app.waitForState(mapTask1, TaskState.RUNNING);
    TaskAttempt task1Attempt1 = mapTask1.getAttempts().values().iterator().next();
    //before sending the TA_DONE, event make sure attempt has come to 
    //RUNNING state
    app.waitForState(task1Attempt1, TaskAttemptState.RUNNING);
    //send the done signal to the map
    app.getContext().getEventHandler().handle(new TaskAttemptEvent(task1Attempt1.getID(), TaskAttemptEventType.TA_DONE));
    //wait for map task to complete
    app.waitForState(mapTask1, TaskState.SUCCEEDED);
    // Verify the shuffle-port
    Assert.assertEquals(5467, task1Attempt1.getShufflePort());
    app.waitForState(reduceTask1, TaskState.RUNNING);
    TaskAttempt reduce1Attempt1 = reduceTask1.getAttempts().values().iterator().next();
    // write output corresponding to reduce1
    writeOutput(reduce1Attempt1, conf);
    //send the done signal to the 1st reduce
    app.getContext().getEventHandler().handle(new TaskAttemptEvent(reduce1Attempt1.getID(), TaskAttemptEventType.TA_DONE));
    //wait for first reduce task to complete
    app.waitForState(reduceTask1, TaskState.SUCCEEDED);
    //stop the app before the job completes.
    app.stop();
    //rerun
    //in rerun the map will be recovered from previous run
    app = new MRAppWithHistory(1, 2, false, this.getClass().getName(), false, ++runCount);
    conf = new Configuration();
    conf.setBoolean(MRJobConfig.MR_AM_JOB_RECOVERY_ENABLE, true);
    conf.setBoolean("mapred.mapper.new-api", true);
    conf.setBoolean("mapred.reducer.new-api", true);
    conf.set(FileOutputFormat.OUTDIR, outputDir.toString());
    conf.setBoolean(MRJobConfig.JOB_UBERTASK_ENABLE, false);
    job = app.submit(conf);
    app.waitForState(job, JobState.RUNNING);
    Assert.assertEquals("No of tasks not correct", 3, job.getTasks().size());
    it = job.getTasks().values().iterator();
    mapTask1 = it.next();
    reduceTask1 = it.next();
    Task reduceTask2 = it.next();
    // map will be recovered, no need to send done
    app.waitForState(mapTask1, TaskState.SUCCEEDED);
    // Verify the shuffle-port after recovery
    task1Attempt1 = mapTask1.getAttempts().values().iterator().next();
    Assert.assertEquals(5467, task1Attempt1.getShufflePort());
    // first reduce will be recovered, no need to send done
    app.waitForState(reduceTask1, TaskState.SUCCEEDED);
    app.waitForState(reduceTask2, TaskState.RUNNING);
    TaskAttempt reduce2Attempt = reduceTask2.getAttempts().values().iterator().next();
    //before sending the TA_DONE, event make sure attempt has come to 
    //RUNNING state
    app.waitForState(reduce2Attempt, TaskAttemptState.RUNNING);
    //send the done signal to the 2nd reduce task
    app.getContext().getEventHandler().handle(new TaskAttemptEvent(reduce2Attempt.getID(), TaskAttemptEventType.TA_DONE));
    //wait to get it completed
    app.waitForState(reduceTask2, TaskState.SUCCEEDED);
    app.waitForState(job, JobState.SUCCEEDED);
    app.verifyCompleted();
    validateOutput();
}

Example 33 with TaskAttempt

use of org.apache.hadoop.mapreduce.v2.app.job.TaskAttempt in project hadoop by apache.

the class TestRecovery method testRecoveryFailsUsingCustomOutputCommitter.

/**
   * This test case primarily verifies if the recovery is controlled through config
   * property. In this case, recover is turned OFF. AM with 3 maps and 0 reduce.
   * AM crashes after the first two tasks finishes and recovery fails and have
   * to rerun fully in the second generation and succeeds.
   * 
   * @throws Exception
   */
@Test
public void testRecoveryFailsUsingCustomOutputCommitter() throws Exception {
    int runCount = 0;
    MRApp app = new MRAppWithHistory(3, 0, false, this.getClass().getName(), true, ++runCount);
    Configuration conf = new Configuration();
    conf.setClass("mapred.output.committer.class", TestFileOutputCommitter.class, org.apache.hadoop.mapred.OutputCommitter.class);
    conf.set(FileOutputFormat.OUTDIR, outputDir.toString());
    conf.setBoolean("want.am.recovery", false);
    Job job = app.submit(conf);
    app.waitForState(job, JobState.RUNNING);
    // all maps would be running
    Assert.assertEquals("No of tasks not correct", 3, job.getTasks().size());
    Iterator<Task> it = job.getTasks().values().iterator();
    Task mapTask1 = it.next();
    Task mapTask2 = it.next();
    Task mapTask3 = it.next();
    // all maps must be running
    app.waitForState(mapTask1, TaskState.RUNNING);
    app.waitForState(mapTask2, TaskState.RUNNING);
    app.waitForState(mapTask3, TaskState.RUNNING);
    TaskAttempt task1Attempt = mapTask1.getAttempts().values().iterator().next();
    TaskAttempt task2Attempt = mapTask2.getAttempts().values().iterator().next();
    TaskAttempt task3Attempt = mapTask3.getAttempts().values().iterator().next();
    // before sending the TA_DONE, event make sure attempt has come to
    // RUNNING state
    app.waitForState(task1Attempt, TaskAttemptState.RUNNING);
    app.waitForState(task2Attempt, TaskAttemptState.RUNNING);
    app.waitForState(task3Attempt, TaskAttemptState.RUNNING);
    // send the done signal to the 1st two maps
    app.getContext().getEventHandler().handle(new TaskAttemptEvent(task1Attempt.getID(), TaskAttemptEventType.TA_DONE));
    app.getContext().getEventHandler().handle(new TaskAttemptEvent(task2Attempt.getID(), TaskAttemptEventType.TA_DONE));
    // wait for first two map task to complete
    app.waitForState(mapTask1, TaskState.SUCCEEDED);
    app.waitForState(mapTask2, TaskState.SUCCEEDED);
    // stop the app
    app.stop();
    // rerun
    // in rerun the 1st two map will be recovered from previous run
    app = new MRAppWithHistory(2, 1, false, this.getClass().getName(), false, ++runCount);
    conf = new Configuration();
    conf.setClass("mapred.output.committer.class", TestFileOutputCommitter.class, org.apache.hadoop.mapred.OutputCommitter.class);
    conf.setBoolean("want.am.recovery", false);
    conf.set(FileOutputFormat.OUTDIR, outputDir.toString());
    // Set num-reduces explicitly in conf as recovery logic depends on it.
    conf.setInt(MRJobConfig.NUM_REDUCES, 0);
    conf.setBoolean(MRJobConfig.JOB_UBERTASK_ENABLE, false);
    job = app.submit(conf);
    app.waitForState(job, JobState.RUNNING);
    Assert.assertEquals("No of tasks not correct", 3, job.getTasks().size());
    it = job.getTasks().values().iterator();
    mapTask1 = it.next();
    mapTask2 = it.next();
    mapTask3 = it.next();
    // first two maps will NOT  be recovered, need to send done from them
    app.waitForState(mapTask1, TaskState.RUNNING);
    app.waitForState(mapTask2, TaskState.RUNNING);
    app.waitForState(mapTask3, TaskState.RUNNING);
    task3Attempt = mapTask3.getAttempts().values().iterator().next();
    // before sending the TA_DONE, event make sure attempt has come to
    // RUNNING state
    app.waitForState(task3Attempt, TaskAttemptState.RUNNING);
    // send the done signal to all 3 tasks map task
    app.getContext().getEventHandler().handle(new TaskAttemptEvent(mapTask1.getAttempts().values().iterator().next().getID(), TaskAttemptEventType.TA_DONE));
    app.getContext().getEventHandler().handle(new TaskAttemptEvent(mapTask2.getAttempts().values().iterator().next().getID(), TaskAttemptEventType.TA_DONE));
    app.getContext().getEventHandler().handle(new TaskAttemptEvent(mapTask3.getAttempts().values().iterator().next().getID(), TaskAttemptEventType.TA_DONE));
    // wait to get it completed
    app.waitForState(mapTask3, TaskState.SUCCEEDED);
    app.waitForState(job, JobState.SUCCEEDED);
    app.verifyCompleted();
}

Example 34 with TaskAttempt

use of org.apache.hadoop.mapreduce.v2.app.job.TaskAttempt in project hadoop by apache.

the class TestKill method testKillTaskWait.

@Test
public void testKillTaskWait() throws Exception {
    final Dispatcher dispatcher = new AsyncDispatcher() {

        private TaskAttemptEvent cachedKillEvent;

        @Override
        protected void dispatch(Event event) {
            if (event instanceof TaskAttemptEvent) {
                TaskAttemptEvent killEvent = (TaskAttemptEvent) event;
                if (killEvent.getType() == TaskAttemptEventType.TA_KILL) {
                    TaskAttemptId taID = killEvent.getTaskAttemptID();
                    if (taID.getTaskId().getTaskType() == TaskType.REDUCE && taID.getTaskId().getId() == 0 && taID.getId() == 0) {
                        // Task is asking the reduce TA to kill itself. 'Create' a race
                        // condition. Make the task succeed and then inform the task that
                        // TA has succeeded. Once Task gets the TA succeeded event at
                        // KILL_WAIT, then relay the actual kill signal to TA
                        super.dispatch(new TaskAttemptEvent(taID, TaskAttemptEventType.TA_DONE));
                        super.dispatch(new TaskAttemptEvent(taID, TaskAttemptEventType.TA_CONTAINER_COMPLETED));
                        super.dispatch(new TaskTAttemptEvent(taID, TaskEventType.T_ATTEMPT_SUCCEEDED));
                        this.cachedKillEvent = killEvent;
                        return;
                    }
                }
            } else if (event instanceof TaskEvent) {
                TaskEvent taskEvent = (TaskEvent) event;
                if (taskEvent.getType() == TaskEventType.T_ATTEMPT_SUCCEEDED && this.cachedKillEvent != null) {
                    // When the TA comes and reports that it is done, send the
                    // cachedKillEvent
                    super.dispatch(this.cachedKillEvent);
                    return;
                }
            }
            super.dispatch(event);
        }
    };
    MRApp app = new MRApp(1, 1, false, this.getClass().getName(), true) {

        @Override
        public Dispatcher createDispatcher() {
            return dispatcher;
        }
    };
    Job job = app.submit(new Configuration());
    JobId jobId = app.getJobId();
    app.waitForState(job, JobState.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();
    app.waitForState(mapTask, TaskState.RUNNING);
    app.waitForState(reduceTask, TaskState.RUNNING);
    TaskAttempt mapAttempt = mapTask.getAttempts().values().iterator().next();
    app.waitForState(mapAttempt, TaskAttemptState.RUNNING);
    TaskAttempt reduceAttempt = reduceTask.getAttempts().values().iterator().next();
    app.waitForState(reduceAttempt, TaskAttemptState.RUNNING);
    // Finish map
    app.getContext().getEventHandler().handle(new TaskAttemptEvent(mapAttempt.getID(), TaskAttemptEventType.TA_DONE));
    app.waitForState(mapTask, TaskState.SUCCEEDED);
    // Now kill the job
    app.getContext().getEventHandler().handle(new JobEvent(jobId, JobEventType.JOB_KILL));
    app.waitForInternalState((JobImpl) job, JobStateInternal.KILLED);
}

Example 35 with TaskAttempt

use of org.apache.hadoop.mapreduce.v2.app.job.TaskAttempt in project hadoop by apache.

the class TestKill method testKillJob.

@Test
public void testKillJob() throws Exception {
    final CountDownLatch latch = new CountDownLatch(1);
    MRApp app = new BlockingMRApp(1, 0, latch);
    //this will start the job but job won't complete as task is
    //blocked
    Job job = app.submit(new Configuration());
    //wait and vailidate for Job to become RUNNING
    app.waitForInternalState((JobImpl) job, JobStateInternal.RUNNING);
    //send the kill signal to Job
    app.getContext().getEventHandler().handle(new JobEvent(job.getID(), JobEventType.JOB_KILL));
    //unblock Task
    latch.countDown();
    //wait and validate for Job to be KILLED
    app.waitForState(job, JobState.KILLED);
    // make sure all events are processed. The AM is stopped
    // only when all tasks and task attempts have been killed
    app.waitForState(Service.STATE.STOPPED);
    Map<TaskId, Task> tasks = job.getTasks();
    Assert.assertEquals("No of tasks is not correct", 1, tasks.size());
    Task task = tasks.values().iterator().next();
    Assert.assertEquals("Task state not correct", TaskState.KILLED, task.getReport().getTaskState());
    Map<TaskAttemptId, TaskAttempt> attempts = tasks.values().iterator().next().getAttempts();
    Assert.assertEquals("No of attempts is not correct", 1, attempts.size());
    Iterator<TaskAttempt> it = attempts.values().iterator();
    Assert.assertEquals("Attempt state not correct", TaskAttemptState.KILLED, it.next().getReport().getTaskAttemptState());
}