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();
}
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();
}
}
use of org.apache.hadoop.mapreduce.v2.app.job.TaskAttempt in project hadoop by apache.
the class TestRMContainerAllocator method testExcludeSchedReducesFromHeadroom.
/**
* Tests whether scheduled reducers are excluded from headroom while
* calculating headroom.
*/
@Test
public void testExcludeSchedReducesFromHeadroom() throws Exception {
LOG.info("Running testExcludeSchedReducesFromHeadroom");
Configuration conf = new Configuration();
conf.setInt(MRJobConfig.MR_JOB_REDUCER_UNCONDITIONAL_PREEMPT_DELAY_SEC, -1);
MyResourceManager rm = new MyResourceManager(conf);
rm.start();
DrainDispatcher dispatcher = (DrainDispatcher) rm.getRMContext().getDispatcher();
// Submit the application
RMApp app = rm.submitApp(1024);
dispatcher.await();
MockNM amNodeManager = rm.registerNode("amNM:1234", 1260);
amNodeManager.nodeHeartbeat(true);
dispatcher.await();
ApplicationAttemptId appAttemptId = app.getCurrentAppAttempt().getAppAttemptId();
rm.sendAMLaunched(appAttemptId);
dispatcher.await();
JobId jobId = MRBuilderUtils.newJobId(appAttemptId.getApplicationId(), 0);
Job mockJob = mock(Job.class);
when(mockJob.getReport()).thenReturn(MRBuilderUtils.newJobReport(jobId, "job", "user", JobState.RUNNING, 0, 0, 0, 0, 0, 0, 0, "jobfile", null, false, ""));
Task mockTask = mock(Task.class);
TaskAttempt mockTaskAttempt = mock(TaskAttempt.class);
when(mockJob.getTask((TaskId) any())).thenReturn(mockTask);
when(mockTask.getAttempt((TaskAttemptId) any())).thenReturn(mockTaskAttempt);
when(mockTaskAttempt.getProgress()).thenReturn(0.01f);
MyContainerAllocator allocator = new MyContainerAllocator(rm, conf, appAttemptId, mockJob);
MockNM nodeManager = rm.registerNode("h1:1234", 4096);
dispatcher.await();
// Register nodes to RM.
MockNM nodeManager2 = rm.registerNode("h2:1234", 1024);
dispatcher.await();
// Request 2 maps and 1 reducer(sone on nodes which are not registered).
ContainerRequestEvent event1 = createReq(jobId, 1, 1024, new String[] { "h1" });
allocator.sendRequest(event1);
ContainerRequestEvent event2 = createReq(jobId, 2, 1024, new String[] { "h2" });
allocator.sendRequest(event2);
ContainerRequestEvent event3 = createReq(jobId, 3, 1024, new String[] { "h1" }, false, true);
allocator.sendRequest(event3);
// This will tell the scheduler about the requests but there will be no
// allocations as nodes are not added.
allocator.schedule();
dispatcher.await();
// Request for another reducer on h3 which has not registered.
ContainerRequestEvent event4 = createReq(jobId, 4, 1024, new String[] { "h3" }, false, true);
allocator.sendRequest(event4);
allocator.schedule();
dispatcher.await();
// Update resources in scheduler through node heartbeat from h1.
nodeManager.nodeHeartbeat(true);
dispatcher.await();
rm.getMyFifoScheduler().forceResourceLimit(Resource.newInstance(3072, 3));
allocator.schedule();
dispatcher.await();
// Two maps are assigned.
Assert.assertEquals(2, allocator.getAssignedRequests().maps.size());
// Send deallocate request for map so that no maps are assigned after this.
ContainerAllocatorEvent deallocate1 = createDeallocateEvent(jobId, 1, false);
allocator.sendDeallocate(deallocate1);
ContainerAllocatorEvent deallocate2 = createDeallocateEvent(jobId, 2, false);
allocator.sendDeallocate(deallocate2);
// No map should be assigned.
Assert.assertEquals(0, allocator.getAssignedRequests().maps.size());
nodeManager.nodeHeartbeat(true);
dispatcher.await();
rm.getMyFifoScheduler().forceResourceLimit(Resource.newInstance(1024, 1));
allocator.schedule();
dispatcher.await();
// h2 heartbeats.
nodeManager2.nodeHeartbeat(true);
dispatcher.await();
// Send request for one more mapper.
ContainerRequestEvent event5 = createReq(jobId, 5, 1024, new String[] { "h1" });
allocator.sendRequest(event5);
rm.getMyFifoScheduler().forceResourceLimit(Resource.newInstance(2048, 2));
allocator.schedule();
dispatcher.await();
// One reducer is assigned and one map is scheduled
Assert.assertEquals(1, allocator.getScheduledRequests().maps.size());
Assert.assertEquals(1, allocator.getAssignedRequests().reduces.size());
// Headroom enough to run a mapper if headroom is taken as it is but wont be
// enough if scheduled reducers resources are deducted.
rm.getMyFifoScheduler().forceResourceLimit(Resource.newInstance(1260, 2));
allocator.schedule();
dispatcher.await();
// After allocate response, the one assigned reducer is preempted and killed
Assert.assertEquals(1, MyContainerAllocator.getTaskAttemptKillEvents().size());
Assert.assertEquals(RMContainerAllocator.RAMPDOWN_DIAGNOSTIC, MyContainerAllocator.getTaskAttemptKillEvents().get(0).getMessage());
Assert.assertEquals(1, allocator.getNumOfPendingReduces());
}
use of org.apache.hadoop.mapreduce.v2.app.job.TaskAttempt in project hadoop by apache.
the class TestRMContainerAllocator method testUpdatedNodes.
@Test
public void testUpdatedNodes() throws Exception {
Configuration conf = new Configuration();
MyResourceManager rm = new MyResourceManager(conf);
rm.start();
DrainDispatcher dispatcher = (DrainDispatcher) rm.getRMContext().getDispatcher();
// Submit the application
RMApp app = rm.submitApp(1024);
dispatcher.await();
MockNM amNodeManager = rm.registerNode("amNM:1234", 2048);
amNodeManager.nodeHeartbeat(true);
dispatcher.await();
ApplicationAttemptId appAttemptId = app.getCurrentAppAttempt().getAppAttemptId();
rm.sendAMLaunched(appAttemptId);
dispatcher.await();
JobId jobId = MRBuilderUtils.newJobId(appAttemptId.getApplicationId(), 0);
Job mockJob = mock(Job.class);
MyContainerAllocator allocator = new MyContainerAllocator(rm, conf, appAttemptId, mockJob);
// add resources to scheduler
MockNM nm1 = rm.registerNode("h1:1234", 10240);
MockNM nm2 = rm.registerNode("h2:1234", 10240);
dispatcher.await();
// create the map container request
ContainerRequestEvent event = createReq(jobId, 1, 1024, new String[] { "h1" });
allocator.sendRequest(event);
TaskAttemptId attemptId = event.getAttemptID();
TaskAttempt mockTaskAttempt = mock(TaskAttempt.class);
when(mockTaskAttempt.getNodeId()).thenReturn(nm1.getNodeId());
Task mockTask = mock(Task.class);
when(mockTask.getAttempt(attemptId)).thenReturn(mockTaskAttempt);
when(mockJob.getTask(attemptId.getTaskId())).thenReturn(mockTask);
// this tells the scheduler about the requests
List<TaskAttemptContainerAssignedEvent> assigned = allocator.schedule();
dispatcher.await();
nm1.nodeHeartbeat(true);
dispatcher.await();
Assert.assertEquals(1, allocator.getJobUpdatedNodeEvents().size());
Assert.assertEquals(3, allocator.getJobUpdatedNodeEvents().get(0).getUpdatedNodes().size());
allocator.getJobUpdatedNodeEvents().clear();
// get the assignment
assigned = allocator.schedule();
dispatcher.await();
Assert.assertEquals(1, assigned.size());
Assert.assertEquals(nm1.getNodeId(), assigned.get(0).getContainer().getNodeId());
// no updated nodes reported
Assert.assertTrue(allocator.getJobUpdatedNodeEvents().isEmpty());
Assert.assertTrue(allocator.getTaskAttemptKillEvents().isEmpty());
// mark nodes bad
nm1.nodeHeartbeat(false);
nm2.nodeHeartbeat(false);
dispatcher.await();
// schedule response returns updated nodes
assigned = allocator.schedule();
dispatcher.await();
Assert.assertEquals(0, assigned.size());
// updated nodes are reported
Assert.assertEquals(1, allocator.getJobUpdatedNodeEvents().size());
Assert.assertEquals(1, allocator.getTaskAttemptKillEvents().size());
Assert.assertEquals(2, allocator.getJobUpdatedNodeEvents().get(0).getUpdatedNodes().size());
Assert.assertEquals(attemptId, allocator.getTaskAttemptKillEvents().get(0).getTaskAttemptID());
allocator.getJobUpdatedNodeEvents().clear();
allocator.getTaskAttemptKillEvents().clear();
assigned = allocator.schedule();
dispatcher.await();
Assert.assertEquals(0, assigned.size());
// no updated nodes reported
Assert.assertTrue(allocator.getJobUpdatedNodeEvents().isEmpty());
Assert.assertTrue(allocator.getTaskAttemptKillEvents().isEmpty());
}
use of org.apache.hadoop.mapreduce.v2.app.job.TaskAttempt in project hadoop by apache.
the class TestTaskImpl method testFailedTransitions.
@Test
public void testFailedTransitions() {
mockTask = new MockTaskImpl(jobId, partition, dispatcher.getEventHandler(), remoteJobConfFile, conf, taskAttemptListener, jobToken, credentials, clock, startCount, metrics, appContext, TaskType.MAP) {
@Override
protected int getMaxAttempts() {
return 1;
}
};
TaskId taskId = getNewTaskID();
scheduleTaskAttempt(taskId);
launchTaskAttempt(getLastAttempt().getAttemptId());
// add three more speculative attempts
mockTask.handle(new TaskTAttemptEvent(getLastAttempt().getAttemptId(), TaskEventType.T_ADD_SPEC_ATTEMPT));
launchTaskAttempt(getLastAttempt().getAttemptId());
mockTask.handle(new TaskTAttemptEvent(getLastAttempt().getAttemptId(), TaskEventType.T_ADD_SPEC_ATTEMPT));
launchTaskAttempt(getLastAttempt().getAttemptId());
mockTask.handle(new TaskTAttemptEvent(getLastAttempt().getAttemptId(), TaskEventType.T_ADD_SPEC_ATTEMPT));
launchTaskAttempt(getLastAttempt().getAttemptId());
assertEquals(4, taskAttempts.size());
// have the first attempt fail, verify task failed due to no retries
MockTaskAttemptImpl taskAttempt = taskAttempts.get(0);
taskAttempt.setState(TaskAttemptState.FAILED);
mockTask.handle(new TaskTAttemptEvent(taskAttempt.getAttemptId(), TaskEventType.T_ATTEMPT_FAILED));
assertEquals(TaskState.FAILED, mockTask.getState());
// verify task can no longer be killed
mockTask.handle(new TaskEvent(taskId, TaskEventType.T_KILL));
assertEquals(TaskState.FAILED, mockTask.getState());
// verify speculative doesn't launch new tasks
mockTask.handle(new TaskTAttemptEvent(getLastAttempt().getAttemptId(), TaskEventType.T_ADD_SPEC_ATTEMPT));
mockTask.handle(new TaskTAttemptEvent(getLastAttempt().getAttemptId(), TaskEventType.T_ATTEMPT_LAUNCHED));
assertEquals(TaskState.FAILED, mockTask.getState());
assertEquals(4, taskAttempts.size());
// verify attempt events from active tasks don't knock task out of FAILED
taskAttempt = taskAttempts.get(1);
taskAttempt.setState(TaskAttemptState.COMMIT_PENDING);
mockTask.handle(new TaskTAttemptEvent(taskAttempt.getAttemptId(), TaskEventType.T_ATTEMPT_COMMIT_PENDING));
assertEquals(TaskState.FAILED, mockTask.getState());
taskAttempt.setState(TaskAttemptState.FAILED);
mockTask.handle(new TaskTAttemptEvent(taskAttempt.getAttemptId(), TaskEventType.T_ATTEMPT_FAILED));
assertEquals(TaskState.FAILED, mockTask.getState());
taskAttempt = taskAttempts.get(2);
taskAttempt.setState(TaskAttemptState.SUCCEEDED);
mockTask.handle(new TaskTAttemptEvent(taskAttempt.getAttemptId(), TaskEventType.T_ATTEMPT_SUCCEEDED));
assertEquals(TaskState.FAILED, mockTask.getState());
taskAttempt = taskAttempts.get(3);
taskAttempt.setState(TaskAttemptState.KILLED);
mockTask.handle(new TaskTAttemptKilledEvent(taskAttempt.getAttemptId(), false));
assertEquals(TaskState.FAILED, mockTask.getState());
}
Aggregations