@SuppressWarnings("unchecked")
public void sendTezEventToSourceTasks(TezEvent tezEvent) {
if (!bufferEvents.get()) {
switch (tezEvent.getEventType()) {
case INPUT_READ_ERROR_EVENT:
InputReadErrorEvent event = (InputReadErrorEvent) tezEvent.getEvent();
TezTaskAttemptID destAttemptId = tezEvent.getSourceInfo().getTaskAttemptID();
int destTaskIndex = destAttemptId.getTaskID().getId();
int srcTaskIndex = edgeManager.routeEventToSourceTasks(destTaskIndex, event);
int numConsumers =
edgeManager.getDestinationConsumerTaskNumber(
srcTaskIndex, destinationVertex.getTotalTasks());
TezTaskID srcTaskId = sourceVertex.getTask(srcTaskIndex).getTaskId();
int taskAttemptIndex = event.getVersion();
TezTaskAttemptID srcTaskAttemptId = new TezTaskAttemptID(srcTaskId, taskAttemptIndex);
eventHandler.handle(
new TaskAttemptEventOutputFailed(srcTaskAttemptId, tezEvent, numConsumers));
break;
default:
throw new TezUncheckedException("Unhandled tez event type: " + tezEvent.getEventType());
}
} else {
sourceEventBuffer.add(tezEvent);
}
}
public InputSpec getDestinationSpec(int destinationTaskIndex) {
return new InputSpec(
sourceVertex.getName(),
edgeProperty.getEdgeDestination(),
edgeManager.getNumDestinationTaskInputs(
sourceVertex.getTotalTasks(), destinationTaskIndex));
}
// TODO remove hacky name lookup
@Override
public boolean needsWaitAfterOutputConsumable() {
Vertex vertex = getVertex();
ProcessorDescriptor processorDescriptor = vertex.getProcessorDescriptor();
if (processorDescriptor != null
&& processorDescriptor.getClassName().contains("InitialTaskWithInMemSort")) {
return true;
} else {
return false;
}
}
@Test(timeout = 10000)
public void testBasicSpeculationPerVertexConf() throws Exception {
DAG dag = DAG.create("test");
String vNameNoSpec = "A";
String vNameSpec = "B";
Vertex vA = Vertex.create(vNameNoSpec, ProcessorDescriptor.create("Proc.class"), 5);
Vertex vB = Vertex.create(vNameSpec, ProcessorDescriptor.create("Proc.class"), 5);
vA.setConf(TezConfiguration.TEZ_AM_SPECULATION_ENABLED, "false");
dag.addVertex(vA);
dag.addVertex(vB);
// min/max src fraction is set to 1. So vertices will run sequentially
dag.addEdge(
Edge.create(
vA,
vB,
EdgeProperty.create(
DataMovementType.SCATTER_GATHER,
DataSourceType.PERSISTED,
SchedulingType.SEQUENTIAL,
OutputDescriptor.create("O"),
InputDescriptor.create("I"))));
MockTezClient tezClient = createTezSession();
DAGClient dagClient = tezClient.submitDAG(dag);
DAGImpl dagImpl = (DAGImpl) mockApp.getContext().getCurrentDAG();
TezVertexID vertexId = dagImpl.getVertex(vNameSpec).getVertexId();
TezVertexID vertexIdNoSpec = dagImpl.getVertex(vNameNoSpec).getVertexId();
// original attempt is killed and speculative one is successful
TezTaskAttemptID killedTaId =
TezTaskAttemptID.getInstance(TezTaskID.getInstance(vertexId, 0), 0);
TezTaskAttemptID noSpecTaId =
TezTaskAttemptID.getInstance(TezTaskID.getInstance(vertexIdNoSpec, 0), 0);
// cause speculation trigger for both
mockLauncher.setStatusUpdatesForTask(killedTaId, 100);
mockLauncher.setStatusUpdatesForTask(noSpecTaId, 100);
mockLauncher.startScheduling(true);
dagClient.waitForCompletion();
Assert.assertEquals(DAGStatus.State.SUCCEEDED, dagClient.getDAGStatus(null).getState());
org.apache.tez.dag.app.dag.Vertex vSpec = dagImpl.getVertex(vertexId);
org.apache.tez.dag.app.dag.Vertex vNoSpec = dagImpl.getVertex(vertexIdNoSpec);
// speculation for vA but not for vB
Assert.assertTrue(
vSpec.getAllCounters().findCounter(TaskCounter.NUM_SPECULATIONS).getValue() > 0);
Assert.assertEquals(
0, vNoSpec.getAllCounters().findCounter(TaskCounter.NUM_SPECULATIONS).getValue());
tezClient.stop();
}
public void sendTezEventToDestinationTasks(TezEvent tezEvent) {
if (!bufferEvents.get()) {
List<Integer> destTaskIndices = new ArrayList<Integer>();
boolean isDataMovementEvent = true;
switch (tezEvent.getEventType()) {
case INPUT_FAILED_EVENT:
isDataMovementEvent = false;
case DATA_MOVEMENT_EVENT:
Event event = tezEvent.getEvent();
TezTaskAttemptID sourceAttemptId = tezEvent.getSourceInfo().getTaskAttemptID();
int sourceTaskIndex = sourceAttemptId.getTaskID().getId();
if (isDataMovementEvent) {
edgeManager.routeEventToDestinationTasks(
(DataMovementEvent) event,
sourceTaskIndex,
destinationVertex.getTotalTasks(),
destTaskIndices);
} else {
edgeManager.routeEventToDestinationTasks(
(InputFailedEvent) event,
sourceTaskIndex,
destinationVertex.getTotalTasks(),
destTaskIndices);
}
for (Integer destTaskIndex : destTaskIndices) {
EventMetaData destMeta =
new EventMetaData(
EventProducerConsumerType.INPUT,
destinationVertex.getName(),
sourceVertex.getName(),
null); // will be filled by Task when sending the event. Is it needed?
if (isDataMovementEvent) {
destMeta.setIndex(((DataMovementEvent) event).getTargetIndex());
} else {
destMeta.setIndex(((InputFailedEvent) event).getTargetIndex());
}
tezEvent.setDestinationInfo(destMeta);
TezTaskID destTaskId = destinationVertex.getTask(destTaskIndex).getTaskId();
sendEventToTask(destTaskId, tezEvent);
}
break;
default:
throw new TezUncheckedException("Unhandled tez event type: " + tezEvent.getEventType());
}
} else {
destinationEventBuffer.add(tezEvent);
}
}
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();
}
@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);
}
@Before
public void setUp() {
dispatcher = new DrainDispatcher();
dispatcher.register(DAGEventType.class, mock(EventHandler.class));
dispatcher.register(VertexEventType.class, mock(EventHandler.class));
dispatcher.register(TaskEventType.class, new TaskEventHandler());
dispatcher.register(TaskAttemptEventType.class, taEventHandler);
dispatcher.init(new Configuration());
dispatcher.start();
vertex = mock(Vertex.class, RETURNS_DEEP_STUBS);
when(vertex.getProcessorDescriptor().getClassName()).thenReturn("");
mockAppContext = mock(AppContext.class, RETURNS_DEEP_STUBS);
when(mockAppContext.getCurrentDAG().getVertex(any(TezVertexID.class))).thenReturn(vertex);
mockHistoryEventHandler = new MockHistoryEventHandler(mockAppContext);
when(mockAppContext.getHistoryHandler()).thenReturn(mockHistoryEventHandler);
task =
new TaskImpl(
vertexId,
0,
dispatcher.getEventHandler(),
new Configuration(),
mock(TaskCommunicatorManagerInterface.class),
new SystemClock(),
mock(TaskHeartbeatHandler.class),
mockAppContext,
false,
Resource.newInstance(1, 1),
mock(ContainerContext.class),
mock(StateChangeNotifier.class),
vertex);
Map<String, OutputCommitter> committers = new HashMap<String, OutputCommitter>();
committers.put(
"out1", new TestOutputCommitter(mock(OutputCommitterContext.class), true, false));
when(task.getVertex().getOutputCommitters()).thenReturn(committers);
}
@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
}
public void setDestinationVertex(Vertex destinationVertex) {
if (this.destinationVertex != null && this.destinationVertex != destinationVertex) {
throw new TezUncheckedException("Destination vertex exists: " + destinationVertex.getName());
}
this.destinationVertex = destinationVertex;
}
public void setSourceVertex(Vertex sourceVertex) {
if (this.sourceVertex != null && this.sourceVertex != sourceVertex) {
throw new TezUncheckedException("Source vertex exists: " + sourceVertex.getName());
}
this.sourceVertex = sourceVertex;
}
public OutputSpec getSourceSpec(int sourceTaskIndex) {
return new OutputSpec(
destinationVertex.getName(),
edgeProperty.getEdgeSource(),
edgeManager.getNumSourceTaskOutputs(destinationVertex.getTotalTasks(), sourceTaskIndex));
}