private void killUnfinishedAttempt(TaskAttempt attempt, String logMsg) {
if (commitAttempt != null && commitAttempt.equals(attempt)) {
LOG.info("Removing commit attempt: " + commitAttempt);
commitAttempt = null;
}
if (attempt != null && !attempt.isFinished()) {
eventHandler.handle(new TaskAttemptEventKillRequest(attempt.getID(), logMsg));
}
}
@Override
public float getProgress() {
readLock.lock();
try {
TaskAttempt bestAttempt = selectBestAttempt();
if (bestAttempt == null) {
return 0f;
}
return bestAttempt.getProgress();
} finally {
readLock.unlock();
}
}
// this is always called in read/write lock
// TODO Verify behaviour is Task is killed (no finished attempt)
private long getFinishTime() {
if (!isFinished()) {
return 0;
}
long finishTime = 0;
for (TaskAttempt at : attempts.values()) {
// select the max finish time of all attempts
// FIXME shouldnt this not count attempts killed after an attempt succeeds
if (finishTime < at.getFinishTime()) {
finishTime = at.getFinishTime();
}
}
return finishTime;
}
@Override
public boolean canCommit(TezTaskAttemptID taskAttemptID) {
writeLock.lock();
try {
if (getState() != TaskState.RUNNING) {
LOG.info("Task not running. Issuing kill to bad commit attempt " + taskAttemptID);
eventHandler.handle(
new TaskAttemptEventKillRequest(taskAttemptID, "Task not running. Bad attempt."));
return false;
}
if (commitAttempt == null) {
TaskAttempt ta = getAttempt(taskAttemptID);
if (ta == null) {
throw new TezUncheckedException("Unknown task for commit: " + taskAttemptID);
}
// Its ok to get a non-locked state snapshot since we handle changes of
// state in the task attempt. Dont want to deadlock here.
TaskAttemptState taState = ta.getStateNoLock();
if (taState == TaskAttemptState.RUNNING) {
commitAttempt = taskAttemptID;
LOG.info(taskAttemptID + " given a go for committing the task output.");
return true;
} else {
LOG.info(
taskAttemptID
+ " with state: "
+ taState
+ " given a no-go for commit because its not running.");
return false;
}
} else {
if (commitAttempt.equals(taskAttemptID)) {
LOG.info(taskAttemptID + " given a go for committing the task output.");
return true;
}
// Don't think this can be a pluggable decision, so simply raise an
// event for the TaskAttempt to delete its output.
// Wait for commit attempt to succeed. Dont kill this. If commit
// attempt fails then choose a different committer. When commit attempt
// succeeds then this and others will be killed
LOG.info(commitAttempt + " is current committer. Commit waiting for: " + taskAttemptID);
return false;
}
} finally {
writeLock.unlock();
}
}
@Override
public TezCounters getCounters() {
TezCounters counters = null;
readLock.lock();
try {
TaskAttempt bestAttempt = selectBestAttempt();
if (bestAttempt != null) {
counters = bestAttempt.getCounters();
} else {
counters = TaskAttemptImpl.EMPTY_COUNTERS;
// counters.groups = new HashMap<CharSequence, CounterGroup>();
}
return counters;
} finally {
readLock.unlock();
}
}
@Override
public void transition(TaskImpl task, TaskEvent event) {
TezTaskAttemptID successTaId = ((TaskEventTAUpdate) event).getTaskAttemptID();
if (task.commitAttempt != null && !task.commitAttempt.equals(successTaId)) {
// The succeeded attempt is not the one that was selected to commit
// This is impossible and has to be a bug
throw new TezUncheckedException(
"TA: "
+ successTaId
+ " succeeded but TA: "
+ task.commitAttempt
+ " was expected to commit and succeed");
}
task.handleTaskAttemptCompletion(successTaId, TaskAttemptStateInternal.SUCCEEDED);
task.finishedAttempts++;
--task.numberUncompletedAttempts;
task.successfulAttempt = successTaId;
task.eventHandler.handle(new VertexEventTaskCompleted(task.taskId, TaskState.SUCCEEDED));
LOG.info("Task succeeded with attempt " + task.successfulAttempt);
if (task.historyTaskStartGenerated) {
task.logJobHistoryTaskFinishedEvent();
}
// issue kill to all other attempts
for (TaskAttempt attempt : task.attempts.values()) {
if (attempt.getID() != task.successfulAttempt
&&
// This is okay because it can only talk us out of sending a
// TA_KILL message to an attempt that doesn't need one for
// other reasons.
!attempt.isFinished()) {
LOG.info("Issuing kill to other attempt " + attempt.getID());
task.eventHandler.handle(
new TaskAttemptEventKillRequest(attempt.getID(), "Alternate attempt succeeded"));
}
}
// send notification to DAG scheduler
task.eventHandler.handle(
new DAGEventSchedulerUpdate(
DAGEventSchedulerUpdate.UpdateType.TA_SUCCEEDED,
task.attempts.get(task.successfulAttempt)));
task.finished(TaskStateInternal.SUCCEEDED);
}
public void testBasicSpeculation(boolean withProgress) throws Exception {
DAG dag = DAG.create("test");
Vertex vA = Vertex.create("A", ProcessorDescriptor.create("Proc.class"), 5);
dag.addVertex(vA);
MockTezClient tezClient = createTezSession();
DAGClient dagClient = tezClient.submitDAG(dag);
DAGImpl dagImpl = (DAGImpl) mockApp.getContext().getCurrentDAG();
TezVertexID vertexId = TezVertexID.getInstance(dagImpl.getID(), 0);
// original attempt is killed and speculative one is successful
TezTaskAttemptID killedTaId =
TezTaskAttemptID.getInstance(TezTaskID.getInstance(vertexId, 0), 0);
TezTaskAttemptID successTaId =
TezTaskAttemptID.getInstance(TezTaskID.getInstance(vertexId, 0), 1);
mockLauncher.updateProgress(withProgress);
// cause speculation trigger
mockLauncher.setStatusUpdatesForTask(killedTaId, 100);
mockLauncher.startScheduling(true);
dagClient.waitForCompletion();
Assert.assertEquals(DAGStatus.State.SUCCEEDED, dagClient.getDAGStatus(null).getState());
Task task = dagImpl.getTask(killedTaId.getTaskID());
Assert.assertEquals(2, task.getAttempts().size());
Assert.assertEquals(successTaId, task.getSuccessfulAttempt().getID());
TaskAttempt killedAttempt = task.getAttempt(killedTaId);
Joiner.on(",").join(killedAttempt.getDiagnostics()).contains("Killed as speculative attempt");
Assert.assertEquals(
TaskAttemptTerminationCause.TERMINATED_EFFECTIVE_SPECULATION,
killedAttempt.getTerminationCause());
if (withProgress) {
// without progress updates occasionally more than 1 task speculates
Assert.assertEquals(
1, task.getCounters().findCounter(TaskCounter.NUM_SPECULATIONS).getValue());
Assert.assertEquals(
1, dagImpl.getAllCounters().findCounter(TaskCounter.NUM_SPECULATIONS).getValue());
org.apache.tez.dag.app.dag.Vertex v = dagImpl.getVertex(killedTaId.getTaskID().getVertexID());
Assert.assertEquals(
1, v.getAllCounters().findCounter(TaskCounter.NUM_SPECULATIONS).getValue());
}
tezClient.stop();
}
// this is always called in read/write lock
private long getLaunchTime() {
long taskLaunchTime = 0;
boolean launchTimeSet = false;
for (TaskAttempt at : attempts.values()) {
// select the least launch time of all attempts
long attemptLaunchTime = at.getLaunchTime();
if (attemptLaunchTime != 0 && !launchTimeSet) {
// For the first non-zero launch time
launchTimeSet = true;
taskLaunchTime = attemptLaunchTime;
} else if (attemptLaunchTime != 0 && taskLaunchTime > attemptLaunchTime) {
taskLaunchTime = attemptLaunchTime;
}
}
if (!launchTimeSet) {
return this.scheduledTime;
}
return taskLaunchTime;
}
// This is always called in the Write Lock
private void addAndScheduleAttempt() {
TaskAttempt attempt = createAttempt(attempts.size());
if (LOG.isDebugEnabled()) {
LOG.debug("Created attempt " + attempt.getID());
}
switch (attempts.size()) {
case 0:
attempts = Collections.singletonMap(attempt.getID(), attempt);
break;
case 1:
Map<TezTaskAttemptID, TaskAttempt> newAttempts =
new LinkedHashMap<TezTaskAttemptID, TaskAttempt>(maxAttempts);
newAttempts.putAll(attempts);
attempts = newAttempts;
attempts.put(attempt.getID(), attempt);
break;
default:
attempts.put(attempt.getID(), attempt);
break;
}
// TODO: Recovery
/*
// Update nextATtemptNumber
if (taskAttemptsFromPreviousGeneration.isEmpty()) {
++nextAttemptNumber;
} else {
// There are still some TaskAttempts from previous generation, use them
nextAttemptNumber =
taskAttemptsFromPreviousGeneration.remove(0).getAttemptId().getId();
}
*/
++numberUncompletedAttempts;
// schedule the nextAttemptNumber
// send event to DAG to assign priority and schedule the attempt with global
// picture in mind
eventHandler.handle(
new DAGEventSchedulerUpdate(DAGEventSchedulerUpdate.UpdateType.TA_SCHEDULE, attempt));
}
@Test(timeout = 10000)
public void testDAGSchedulerNaturalOrder() {
DAG mockDag = mock(DAG.class);
Vertex mockVertex = mock(Vertex.class);
TaskAttempt mockAttempt = mock(TaskAttempt.class);
when(mockDag.getVertex((TezVertexID) any())).thenReturn(mockVertex);
when(mockVertex.getDistanceFromRoot()).thenReturn(0).thenReturn(1).thenReturn(2);
when(mockAttempt.getIsRescheduled()).thenReturn(false);
DAGEventSchedulerUpdate event =
new DAGEventSchedulerUpdate(DAGEventSchedulerUpdate.UpdateType.TA_SCHEDULE, mockAttempt);
DAGScheduler scheduler = new DAGSchedulerNaturalOrder(mockDag, mockEventHandler);
scheduler.scheduleTask(event);
Assert.assertTrue(mockEventHandler.event.getPriority().getPriority() == 2);
scheduler.scheduleTask(event);
Assert.assertTrue(mockEventHandler.event.getPriority().getPriority() == 4);
scheduler.scheduleTask(event);
Assert.assertTrue(mockEventHandler.event.getPriority().getPriority() == 6);
when(mockAttempt.getIsRescheduled()).thenReturn(true);
scheduler.scheduleTask(event);
Assert.assertTrue(mockEventHandler.event.getPriority().getPriority() == 5);
}
// select the nextAttemptNumber with best progress
// always called inside the Read Lock
private TaskAttempt selectBestAttempt() {
float progress = 0f;
TaskAttempt result = null;
for (TaskAttempt at : attempts.values()) {
switch (at.getState()) {
// ignore all failed task attempts
case FAILED:
case KILLED:
continue;
default:
}
if (result == null) {
result = at; // The first time around
}
// calculate the best progress
float attemptProgress = at.getProgress();
if (attemptProgress > progress) {
result = at;
progress = attemptProgress;
}
}
return result;
}
@Override
public TaskReport getReport() {
// TODO TEZPB This is broken. Records will not work without the PBImpl, which
// is in a different package.
TaskReport report = Records.newRecord(TaskReport.class);
readLock.lock();
try {
report.setTaskId(taskId);
report.setStartTime(getLaunchTime());
report.setFinishTime(getFinishTime());
report.setTaskState(getState());
report.setProgress(getProgress());
for (TaskAttempt attempt : attempts.values()) {
if (TaskAttemptState.RUNNING.equals(attempt.getState())) {
report.addRunningAttempt(attempt.getID());
}
}
report.setSuccessfulAttempt(successfulAttempt);
for (TaskAttempt att : attempts.values()) {
String prefix = "AttemptID:" + att.getID() + " Info:";
for (CharSequence cs : att.getDiagnostics()) {
report.addDiagnostics(prefix + cs);
}
}
// Add a copy of counters as the last step so that their lifetime on heap
// is as small as possible.
report.setCounters(getCounters());
return report;
} finally {
readLock.unlock();
}
}
@Ignore
@Test(timeout = 10000)
public void testDAGSchedulerMRR() {
DAG mockDag = mock(DAG.class);
TezDAGID dagId = TezDAGID.getInstance("1", 1, 1);
TaskSchedulerEventHandler mockTaskScheduler = mock(TaskSchedulerEventHandler.class);
Vertex mockVertex1 = mock(Vertex.class);
TezVertexID mockVertexId1 = TezVertexID.getInstance(dagId, 1);
when(mockVertex1.getVertexId()).thenReturn(mockVertexId1);
when(mockVertex1.getDistanceFromRoot()).thenReturn(0);
TaskAttempt mockAttempt1 = mock(TaskAttempt.class);
when(mockAttempt1.getVertexID()).thenReturn(mockVertexId1);
when(mockAttempt1.getIsRescheduled()).thenReturn(false);
when(mockDag.getVertex(mockVertexId1)).thenReturn(mockVertex1);
Vertex mockVertex2 = mock(Vertex.class);
TezVertexID mockVertexId2 = TezVertexID.getInstance(dagId, 2);
when(mockVertex2.getVertexId()).thenReturn(mockVertexId2);
when(mockVertex2.getDistanceFromRoot()).thenReturn(1);
TaskAttempt mockAttempt2 = mock(TaskAttempt.class);
when(mockAttempt2.getVertexID()).thenReturn(mockVertexId2);
when(mockAttempt2.getIsRescheduled()).thenReturn(false);
when(mockDag.getVertex(mockVertexId2)).thenReturn(mockVertex2);
TaskAttempt mockAttempt2f = mock(TaskAttempt.class);
when(mockAttempt2f.getVertexID()).thenReturn(mockVertexId2);
when(mockAttempt2f.getIsRescheduled()).thenReturn(true);
Vertex mockVertex3 = mock(Vertex.class);
TezVertexID mockVertexId3 = TezVertexID.getInstance(dagId, 3);
when(mockVertex3.getVertexId()).thenReturn(mockVertexId3);
when(mockVertex3.getDistanceFromRoot()).thenReturn(2);
TaskAttempt mockAttempt3 = mock(TaskAttempt.class);
when(mockAttempt3.getVertexID()).thenReturn(mockVertexId3);
when(mockAttempt3.getIsRescheduled()).thenReturn(false);
when(mockDag.getVertex(mockVertexId3)).thenReturn(mockVertex3);
DAGEventSchedulerUpdate mockEvent1 = mock(DAGEventSchedulerUpdate.class);
when(mockEvent1.getAttempt()).thenReturn(mockAttempt1);
DAGEventSchedulerUpdate mockEvent2 = mock(DAGEventSchedulerUpdate.class);
when(mockEvent2.getAttempt()).thenReturn(mockAttempt2);
DAGEventSchedulerUpdate mockEvent2f = mock(DAGEventSchedulerUpdate.class);
when(mockEvent2f.getAttempt()).thenReturn(mockAttempt2f);
DAGEventSchedulerUpdate mockEvent3 = mock(DAGEventSchedulerUpdate.class);
when(mockEvent3.getAttempt()).thenReturn(mockAttempt3);
DAGScheduler scheduler =
new DAGSchedulerMRR(mockDag, mockEventHandler, mockTaskScheduler, 0.5f);
// M starts. M completes. R1 starts. R1 completes. R2 starts. R2 completes
scheduler.scheduleTask(mockEvent1); // M starts
Assert.assertTrue(mockEventHandler.event.getPriority().getPriority() == 3);
scheduler.scheduleTask(mockEvent1); // M runs another
Assert.assertTrue(mockEventHandler.event.getPriority().getPriority() == 3);
scheduler.vertexCompleted(mockVertex1); // M completes
scheduler.scheduleTask(mockEvent2); // R1 starts
Assert.assertTrue(mockEventHandler.event.getPriority().getPriority() == 6);
scheduler.scheduleTask(mockEvent2); // R1 runs another
Assert.assertTrue(mockEventHandler.event.getPriority().getPriority() == 6);
scheduler.scheduleTask(mockEvent2f); // R1 runs retry. Retry priority
Assert.assertTrue(mockEventHandler.event.getPriority().getPriority() == 4);
scheduler.vertexCompleted(mockVertex2); // R1 completes
scheduler.scheduleTask(mockEvent3); // R2 starts
Assert.assertTrue(mockEventHandler.event.getPriority().getPriority() == 9);
scheduler.scheduleTask(mockEvent3); // R2 runs another
Assert.assertTrue(mockEventHandler.event.getPriority().getPriority() == 9);
scheduler.vertexCompleted(mockVertex3); // R2 completes
// M starts. R1 starts. M completes. R2 starts. R1 completes. R2 completes
scheduler.scheduleTask(mockEvent1); // M starts
Assert.assertTrue(mockEventHandler.event.getPriority().getPriority() == 3);
scheduler.scheduleTask(mockEvent2); // R1 starts. Reordered priority
Assert.assertTrue(mockEventHandler.event.getPriority().getPriority() == 2);
scheduler.scheduleTask(mockEvent1); // M runs another
Assert.assertTrue(mockEventHandler.event.getPriority().getPriority() == 3);
scheduler.scheduleTask(mockEvent2); // R1 runs another. Reordered priority
Assert.assertTrue(mockEventHandler.event.getPriority().getPriority() == 2);
scheduler.scheduleTask(mockEvent2f); // R1 runs retry. Reordered priority
Assert.assertTrue(mockEventHandler.event.getPriority().getPriority() == 2);
scheduler.vertexCompleted(mockVertex1); // M completes
scheduler.scheduleTask(mockEvent3); // R2 starts. Reordered priority
Assert.assertTrue(mockEventHandler.event.getPriority().getPriority() == 5);
scheduler.scheduleTask(mockEvent2); // R1 runs another. Normal priority
Assert.assertTrue(mockEventHandler.event.getPriority().getPriority() == 6);
scheduler.scheduleTask(mockEvent2f); // R1 runs retry. Retry priority
Assert.assertTrue(mockEventHandler.event.getPriority().getPriority() == 4);
scheduler.scheduleTask(mockEvent3); // R2 runs another. Reordered priority
Assert.assertTrue(mockEventHandler.event.getPriority().getPriority() == 5);
scheduler.vertexCompleted(mockVertex2); // R1 completes
scheduler.vertexCompleted(mockVertex3); // R2 completes
// M starts. M completes. R1 starts. R2 starts. R1 completes. R2 completes
scheduler.scheduleTask(mockEvent1); // M starts
Assert.assertTrue(mockEventHandler.event.getPriority().getPriority() == 3);
scheduler.vertexCompleted(mockVertex1); // M completes
scheduler.scheduleTask(mockEvent2); // R1 starts
Assert.assertTrue(mockEventHandler.event.getPriority().getPriority() == 6);
scheduler.scheduleTask(mockEvent3); // R2 starts. Reordered priority
Assert.assertTrue(mockEventHandler.event.getPriority().getPriority() == 5);
scheduler.scheduleTask(mockEvent2); // R1 runs another
Assert.assertTrue(mockEventHandler.event.getPriority().getPriority() == 6);
scheduler.vertexCompleted(mockVertex2); // R1 completes
scheduler.vertexCompleted(mockVertex3); // R2 completes
// M starts. R1 starts. M completes. R1 completes. R2 starts. R2 completes
scheduler.scheduleTask(mockEvent1); // M starts
Assert.assertTrue(mockEventHandler.event.getPriority().getPriority() == 3);
scheduler.scheduleTask(mockEvent2); // R1 starts. Reordered priority
Assert.assertTrue(mockEventHandler.event.getPriority().getPriority() == 2);
scheduler.vertexCompleted(mockVertex1); // M completes
scheduler.scheduleTask(mockEvent2); // R1 starts. Normal priority
Assert.assertTrue(mockEventHandler.event.getPriority().getPriority() == 6);
scheduler.vertexCompleted(mockVertex2); // R1 completes
scheduler.scheduleTask(mockEvent3); // R2 starts
Assert.assertTrue(mockEventHandler.event.getPriority().getPriority() == 9);
scheduler.vertexCompleted(mockVertex3); // R2 completes
}
