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

Example 21 with TaskAttempt

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

the class AMWebServices method getJobTaskAttemptState.

@GET
@Path("/jobs/{jobid}/tasks/{taskid}/attempts/{attemptid}/state")
@Produces({ MediaType.APPLICATION_JSON + "; " + JettyUtils.UTF_8, MediaType.APPLICATION_XML + "; " + JettyUtils.UTF_8 })
public JobTaskAttemptState getJobTaskAttemptState(@Context HttpServletRequest hsr, @PathParam("jobid") String jid, @PathParam("taskid") String tid, @PathParam("attemptid") String attId) throws IOException, InterruptedException {
    init();
    Job job = getJobFromJobIdString(jid, appCtx);
    checkAccess(job, hsr);
    Task task = getTaskFromTaskIdString(tid, job);
    TaskAttempt ta = getTaskAttemptFromTaskAttemptString(attId, task);
    return new JobTaskAttemptState(ta.getState().toString());
}

Example 22 with TaskAttempt

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

the class AMWebServices method getJobTaskAttemptId.

@GET
@Path("/jobs/{jobid}/tasks/{taskid}/attempts/{attemptid}")
@Produces({ MediaType.APPLICATION_JSON + "; " + JettyUtils.UTF_8, MediaType.APPLICATION_XML + "; " + JettyUtils.UTF_8 })
public TaskAttemptInfo getJobTaskAttemptId(@Context HttpServletRequest hsr, @PathParam("jobid") String jid, @PathParam("taskid") String tid, @PathParam("attemptid") String attId) {
    init();
    Job job = getJobFromJobIdString(jid, appCtx);
    checkAccess(job, hsr);
    Task task = getTaskFromTaskIdString(tid, job);
    TaskAttempt ta = getTaskAttemptFromTaskAttemptString(attId, task);
    if (task.getType() == TaskType.REDUCE) {
        return new ReduceTaskAttemptInfo(ta, task.getType());
    } else {
        return new TaskAttemptInfo(ta, task.getType(), true);
    }
}

Example 23 with TaskAttempt

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

the class TestRuntimeEstimators method coreTestEstimator.

private void coreTestEstimator(TaskRuntimeEstimator testedEstimator, int expectedSpeculations) {
    estimator = testedEstimator;
    clock = new ControlledClock();
    dispatcher = new AsyncDispatcher();
    myJob = null;
    slotsInUse.set(0);
    completedMaps.set(0);
    completedReduces.set(0);
    successfulSpeculations.set(0);
    taskTimeSavedBySpeculation.set(0);
    clock.tickMsec(1000);
    Configuration conf = new Configuration();
    myAppContext = new MyAppContext(MAP_TASKS, REDUCE_TASKS);
    myJob = myAppContext.getAllJobs().values().iterator().next();
    estimator.contextualize(conf, myAppContext);
    conf.setLong(MRJobConfig.SPECULATIVE_RETRY_AFTER_NO_SPECULATE, 500L);
    conf.setLong(MRJobConfig.SPECULATIVE_RETRY_AFTER_SPECULATE, 5000L);
    conf.setDouble(MRJobConfig.SPECULATIVECAP_RUNNING_TASKS, 0.1);
    conf.setDouble(MRJobConfig.SPECULATIVECAP_TOTAL_TASKS, 0.001);
    conf.setInt(MRJobConfig.SPECULATIVE_MINIMUM_ALLOWED_TASKS, 5);
    speculator = new DefaultSpeculator(conf, myAppContext, estimator, clock);
    Assert.assertEquals("wrong SPECULATIVE_RETRY_AFTER_NO_SPECULATE value", 500L, speculator.getSoonestRetryAfterNoSpeculate());
    Assert.assertEquals("wrong SPECULATIVE_RETRY_AFTER_SPECULATE value", 5000L, speculator.getSoonestRetryAfterSpeculate());
    Assert.assertEquals(speculator.getProportionRunningTasksSpeculatable(), 0.1, 0.00001);
    Assert.assertEquals(speculator.getProportionTotalTasksSpeculatable(), 0.001, 0.00001);
    Assert.assertEquals("wrong SPECULATIVE_MINIMUM_ALLOWED_TASKS value", 5, speculator.getMinimumAllowedSpeculativeTasks());
    dispatcher.register(Speculator.EventType.class, speculator);
    dispatcher.register(TaskEventType.class, new SpeculationRequestEventHandler());
    dispatcher.init(conf);
    dispatcher.start();
    speculator.init(conf);
    speculator.start();
    // Now that the plumbing is hooked up, we do the following:
    //  do until all tasks are finished, ...
    //  1: If we have spare capacity, assign as many map tasks as we can, then
    //     assign as many reduce tasks as we can.  Note that an odd reduce
    //     task might be started while there are still map tasks, because
    //     map tasks take 3 slots and reduce tasks 2 slots.
    //  2: Send a speculation event for every task attempt that's running
    //  note that new attempts might get started by the speculator
    // discover undone tasks
    int undoneMaps = MAP_TASKS;
    int undoneReduces = REDUCE_TASKS;
    // build a task sequence where all the maps precede any of the reduces
    List<Task> allTasksSequence = new LinkedList<Task>();
    allTasksSequence.addAll(myJob.getTasks(TaskType.MAP).values());
    allTasksSequence.addAll(myJob.getTasks(TaskType.REDUCE).values());
    while (undoneMaps + undoneReduces > 0) {
        undoneMaps = 0;
        undoneReduces = 0;
        // start all attempts which are new but for which there is enough slots
        for (Task task : allTasksSequence) {
            if (!task.isFinished()) {
                if (task.getType() == TaskType.MAP) {
                    ++undoneMaps;
                } else {
                    ++undoneReduces;
                }
            }
            for (TaskAttempt attempt : task.getAttempts().values()) {
                if (attempt.getState() == TaskAttemptState.NEW && INITIAL_NUMBER_FREE_SLOTS - slotsInUse.get() >= taskTypeSlots(task.getType())) {
                    MyTaskAttemptImpl attemptImpl = (MyTaskAttemptImpl) attempt;
                    SpeculatorEvent event = new SpeculatorEvent(attempt.getID(), false, clock.getTime());
                    speculator.handle(event);
                    attemptImpl.startUp();
                } else {
                    // If a task attempt is in progress we should send the news to
                    // the Speculator.
                    TaskAttemptStatus status = new TaskAttemptStatus();
                    status.id = attempt.getID();
                    status.progress = attempt.getProgress();
                    status.stateString = attempt.getState().name();
                    status.taskState = attempt.getState();
                    SpeculatorEvent event = new SpeculatorEvent(status, clock.getTime());
                    speculator.handle(event);
                }
            }
        }
        long startTime = System.currentTimeMillis();
        // drain the speculator event queue
        while (!speculator.eventQueueEmpty()) {
            Thread.yield();
            if (System.currentTimeMillis() > startTime + 130000) {
                return;
            }
        }
        clock.tickMsec(1000L);
        if (clock.getTime() % 10000L == 0L) {
            speculator.scanForSpeculations();
        }
    }
    Assert.assertEquals("We got the wrong number of successful speculations.", expectedSpeculations, successfulSpeculations.get());
}

Example 24 with TaskAttempt

use of org.apache.hadoop.mapreduce.v2.app.job.TaskAttempt 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();
}

Example 25 with TaskAttempt

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

the class TestContainerLauncher method testSlowNM.

@Test(timeout = 15000)
public void testSlowNM() throws Exception {
    conf = new Configuration();
    int maxAttempts = 1;
    conf.setInt(MRJobConfig.MAP_MAX_ATTEMPTS, maxAttempts);
    conf.setBoolean(MRJobConfig.JOB_UBERTASK_ENABLE, false);
    // set timeout low for the test
    conf.setInt("yarn.rpc.nm-command-timeout", 3000);
    conf.set(YarnConfiguration.IPC_RPC_IMPL, HadoopYarnProtoRPC.class.getName());
    YarnRPC rpc = YarnRPC.create(conf);
    String bindAddr = "localhost:0";
    InetSocketAddress addr = NetUtils.createSocketAddr(bindAddr);
    NMTokenSecretManagerInNM tokenSecretManager = new NMTokenSecretManagerInNM();
    MasterKey masterKey = Records.newRecord(MasterKey.class);
    masterKey.setBytes(ByteBuffer.wrap("key".getBytes()));
    tokenSecretManager.setMasterKey(masterKey);
    conf.set(CommonConfigurationKeysPublic.HADOOP_SECURITY_AUTHENTICATION, "token");
    server = rpc.getServer(ContainerManagementProtocol.class, new DummyContainerManager(), addr, conf, tokenSecretManager, 1);
    server.start();
    MRApp app = new MRAppWithSlowNM(tokenSecretManager);
    try {
        Job job = app.submit(conf);
        app.waitForState(job, JobState.RUNNING);
        Map<TaskId, Task> tasks = job.getTasks();
        Assert.assertEquals("Num tasks is not correct", 1, tasks.size());
        Task task = tasks.values().iterator().next();
        app.waitForState(task, TaskState.SCHEDULED);
        Map<TaskAttemptId, TaskAttempt> attempts = tasks.values().iterator().next().getAttempts();
        Assert.assertEquals("Num attempts is not correct", maxAttempts, attempts.size());
        TaskAttempt attempt = attempts.values().iterator().next();
        app.waitForInternalState((TaskAttemptImpl) attempt, TaskAttemptStateInternal.ASSIGNED);
        app.waitForState(job, JobState.FAILED);
        String diagnostics = attempt.getDiagnostics().toString();
        LOG.info("attempt.getDiagnostics: " + diagnostics);
        Assert.assertTrue(diagnostics.contains("Container launch failed for " + "container_0_0000_01_000000 : "));
        Assert.assertTrue(diagnostics.contains("java.net.SocketTimeoutException: 3000 millis timeout while waiting for channel"));
    } finally {
        server.stop();
        app.stop();
    }
}