use of org.apache.flink.runtime.jobgraph.JobVertex in project flink by apache.
the class JobManagerHARecoveryTest method testJobRecoveryWhenLosingLeadership.
/**
* Tests that the persisted job is not removed from the SubmittedJobGraphStore if the JobManager
* loses its leadership. Furthermore, it tests that the job manager can recover the job from
* the SubmittedJobGraphStore and checkpoint state is recovered as well.
*/
@Test
public void testJobRecoveryWhenLosingLeadership() throws Exception {
FiniteDuration timeout = new FiniteDuration(30, TimeUnit.SECONDS);
FiniteDuration jobRecoveryTimeout = new FiniteDuration(3, TimeUnit.SECONDS);
Deadline deadline = new FiniteDuration(2, TimeUnit.MINUTES).fromNow();
Configuration flinkConfiguration = new Configuration();
UUID leaderSessionID = UUID.randomUUID();
UUID newLeaderSessionID = UUID.randomUUID();
int slots = 2;
ActorRef archive = null;
ActorRef jobManager = null;
ActorRef taskManager = null;
flinkConfiguration.setString(HighAvailabilityOptions.HA_MODE, "zookeeper");
flinkConfiguration.setString(HighAvailabilityOptions.HA_STORAGE_PATH, temporaryFolder.newFolder().toString());
flinkConfiguration.setInteger(ConfigConstants.TASK_MANAGER_NUM_TASK_SLOTS, slots);
try {
Scheduler scheduler = new Scheduler(TestingUtils.defaultExecutionContext());
MySubmittedJobGraphStore mySubmittedJobGraphStore = new MySubmittedJobGraphStore();
MyCheckpointStore checkpointStore = new MyCheckpointStore();
CheckpointIDCounter checkpointCounter = new StandaloneCheckpointIDCounter();
CheckpointRecoveryFactory checkpointStateFactory = new MyCheckpointRecoveryFactory(checkpointStore, checkpointCounter);
TestingLeaderElectionService myLeaderElectionService = new TestingLeaderElectionService();
TestingLeaderRetrievalService myLeaderRetrievalService = new TestingLeaderRetrievalService();
InstanceManager instanceManager = new InstanceManager();
instanceManager.addInstanceListener(scheduler);
archive = system.actorOf(Props.create(MemoryArchivist.class, 10));
Props jobManagerProps = Props.create(TestingJobManager.class, flinkConfiguration, TestingUtils.defaultExecutor(), TestingUtils.defaultExecutor(), instanceManager, scheduler, new BlobLibraryCacheManager(new BlobServer(flinkConfiguration), 3600000), archive, new FixedDelayRestartStrategy.FixedDelayRestartStrategyFactory(Int.MaxValue(), 100), timeout, myLeaderElectionService, mySubmittedJobGraphStore, checkpointStateFactory, jobRecoveryTimeout, Option.apply(null));
jobManager = system.actorOf(jobManagerProps);
ActorGateway gateway = new AkkaActorGateway(jobManager, leaderSessionID);
taskManager = TaskManager.startTaskManagerComponentsAndActor(flinkConfiguration, ResourceID.generate(), system, "localhost", Option.apply("taskmanager"), Option.apply((LeaderRetrievalService) myLeaderRetrievalService), true, TestingTaskManager.class);
ActorGateway tmGateway = new AkkaActorGateway(taskManager, leaderSessionID);
Future<Object> tmAlive = tmGateway.ask(TestingMessages.getAlive(), deadline.timeLeft());
Await.ready(tmAlive, deadline.timeLeft());
JobVertex sourceJobVertex = new JobVertex("Source");
sourceJobVertex.setInvokableClass(BlockingStatefulInvokable.class);
sourceJobVertex.setParallelism(slots);
JobGraph jobGraph = new JobGraph("TestingJob", sourceJobVertex);
List<JobVertexID> vertexId = Collections.singletonList(sourceJobVertex.getID());
jobGraph.setSnapshotSettings(new JobSnapshottingSettings(vertexId, vertexId, vertexId, 100, 10 * 60 * 1000, 0, 1, ExternalizedCheckpointSettings.none(), null, true));
BlockingStatefulInvokable.initializeStaticHelpers(slots);
Future<Object> isLeader = gateway.ask(TestingJobManagerMessages.getNotifyWhenLeader(), deadline.timeLeft());
Future<Object> isConnectedToJobManager = tmGateway.ask(new TestingTaskManagerMessages.NotifyWhenRegisteredAtJobManager(jobManager), deadline.timeLeft());
// tell jobManager that he's the leader
myLeaderElectionService.isLeader(leaderSessionID);
// tell taskManager who's the leader
myLeaderRetrievalService.notifyListener(gateway.path(), leaderSessionID);
Await.ready(isLeader, deadline.timeLeft());
Await.ready(isConnectedToJobManager, deadline.timeLeft());
// submit blocking job
Future<Object> jobSubmitted = gateway.ask(new JobManagerMessages.SubmitJob(jobGraph, ListeningBehaviour.DETACHED), deadline.timeLeft());
Await.ready(jobSubmitted, deadline.timeLeft());
// Wait for some checkpoints to complete
BlockingStatefulInvokable.awaitCompletedCheckpoints();
Future<Object> jobRemoved = gateway.ask(new TestingJobManagerMessages.NotifyWhenJobRemoved(jobGraph.getJobID()), deadline.timeLeft());
// Revoke leadership
myLeaderElectionService.notLeader();
// check that the job gets removed from the JobManager
Await.ready(jobRemoved, deadline.timeLeft());
// but stays in the submitted job graph store
assertTrue(mySubmittedJobGraphStore.contains(jobGraph.getJobID()));
Future<Object> jobRunning = gateway.ask(new TestingJobManagerMessages.NotifyWhenJobStatus(jobGraph.getJobID(), JobStatus.RUNNING), deadline.timeLeft());
// Make JobManager again a leader
myLeaderElectionService.isLeader(newLeaderSessionID);
// tell the TaskManager about it
myLeaderRetrievalService.notifyListener(gateway.path(), newLeaderSessionID);
// wait that the job is recovered and reaches state RUNNING
Await.ready(jobRunning, deadline.timeLeft());
Future<Object> jobFinished = gateway.ask(new TestingJobManagerMessages.NotifyWhenJobRemoved(jobGraph.getJobID()), deadline.timeLeft());
BlockingInvokable.unblock();
// wait til the job has finished
Await.ready(jobFinished, deadline.timeLeft());
// check that the job has been removed from the submitted job graph store
assertFalse(mySubmittedJobGraphStore.contains(jobGraph.getJobID()));
// Check that state has been recovered
long[] recoveredStates = BlockingStatefulInvokable.getRecoveredStates();
for (long state : recoveredStates) {
boolean isExpected = state >= BlockingStatefulInvokable.NUM_CHECKPOINTS_TO_COMPLETE;
assertTrue("Did not recover checkpoint state correctly, expecting >= " + BlockingStatefulInvokable.NUM_CHECKPOINTS_TO_COMPLETE + ", but state was " + state, isExpected);
}
} finally {
if (archive != null) {
archive.tell(PoisonPill.getInstance(), ActorRef.noSender());
}
if (jobManager != null) {
jobManager.tell(PoisonPill.getInstance(), ActorRef.noSender());
}
if (taskManager != null) {
taskManager.tell(PoisonPill.getInstance(), ActorRef.noSender());
}
}
}
use of org.apache.flink.runtime.jobgraph.JobVertex in project flink by apache.
the class JobManagerTest method testKvStateMessages.
/**
* Tests that the JobManager handles {@link org.apache.flink.runtime.query.KvStateMessage}
* instances as expected.
*/
@Test
public void testKvStateMessages() throws Exception {
Deadline deadline = new FiniteDuration(100, TimeUnit.SECONDS).fromNow();
Configuration config = new Configuration();
config.setString(ConfigConstants.AKKA_ASK_TIMEOUT, "100ms");
UUID leaderSessionId = null;
ActorGateway jobManager = new AkkaActorGateway(JobManager.startJobManagerActors(config, system, TestingUtils.defaultExecutor(), TestingUtils.defaultExecutor(), TestingJobManager.class, MemoryArchivist.class)._1(), leaderSessionId);
LeaderRetrievalService leaderRetrievalService = new StandaloneLeaderRetrievalService(AkkaUtils.getAkkaURL(system, jobManager.actor()));
Configuration tmConfig = new Configuration();
tmConfig.setInteger(ConfigConstants.TASK_MANAGER_MEMORY_SIZE_KEY, 4);
tmConfig.setInteger(ConfigConstants.TASK_MANAGER_NUM_TASK_SLOTS, 8);
ActorRef taskManager = TaskManager.startTaskManagerComponentsAndActor(tmConfig, ResourceID.generate(), system, "localhost", scala.Option.<String>empty(), scala.Option.apply(leaderRetrievalService), true, TestingTaskManager.class);
Future<Object> registrationFuture = jobManager.ask(new NotifyWhenAtLeastNumTaskManagerAreRegistered(1), deadline.timeLeft());
Await.ready(registrationFuture, deadline.timeLeft());
//
// Location lookup
//
LookupKvStateLocation lookupNonExistingJob = new LookupKvStateLocation(new JobID(), "any-name");
Future<KvStateLocation> lookupFuture = jobManager.ask(lookupNonExistingJob, deadline.timeLeft()).mapTo(ClassTag$.MODULE$.<KvStateLocation>apply(KvStateLocation.class));
try {
Await.result(lookupFuture, deadline.timeLeft());
fail("Did not throw expected Exception");
} catch (IllegalStateException ignored) {
// Expected
}
JobGraph jobGraph = new JobGraph("croissant");
JobVertex jobVertex1 = new JobVertex("cappuccino");
jobVertex1.setParallelism(4);
jobVertex1.setMaxParallelism(16);
jobVertex1.setInvokableClass(BlockingNoOpInvokable.class);
JobVertex jobVertex2 = new JobVertex("americano");
jobVertex2.setParallelism(4);
jobVertex2.setMaxParallelism(16);
jobVertex2.setInvokableClass(BlockingNoOpInvokable.class);
jobGraph.addVertex(jobVertex1);
jobGraph.addVertex(jobVertex2);
Future<JobSubmitSuccess> submitFuture = jobManager.ask(new SubmitJob(jobGraph, ListeningBehaviour.DETACHED), deadline.timeLeft()).mapTo(ClassTag$.MODULE$.<JobSubmitSuccess>apply(JobSubmitSuccess.class));
Await.result(submitFuture, deadline.timeLeft());
Object lookupUnknownRegistrationName = new LookupKvStateLocation(jobGraph.getJobID(), "unknown");
lookupFuture = jobManager.ask(lookupUnknownRegistrationName, deadline.timeLeft()).mapTo(ClassTag$.MODULE$.<KvStateLocation>apply(KvStateLocation.class));
try {
Await.result(lookupFuture, deadline.timeLeft());
fail("Did not throw expected Exception");
} catch (UnknownKvStateLocation ignored) {
// Expected
}
//
// Registration
//
NotifyKvStateRegistered registerNonExistingJob = new NotifyKvStateRegistered(new JobID(), new JobVertexID(), new KeyGroupRange(0, 0), "any-name", new KvStateID(), new KvStateServerAddress(InetAddress.getLocalHost(), 1233));
jobManager.tell(registerNonExistingJob);
LookupKvStateLocation lookupAfterRegistration = new LookupKvStateLocation(registerNonExistingJob.getJobId(), registerNonExistingJob.getRegistrationName());
lookupFuture = jobManager.ask(lookupAfterRegistration, deadline.timeLeft()).mapTo(ClassTag$.MODULE$.<KvStateLocation>apply(KvStateLocation.class));
try {
Await.result(lookupFuture, deadline.timeLeft());
fail("Did not throw expected Exception");
} catch (IllegalStateException ignored) {
// Expected
}
NotifyKvStateRegistered registerForExistingJob = new NotifyKvStateRegistered(jobGraph.getJobID(), jobVertex1.getID(), new KeyGroupRange(0, 0), "register-me", new KvStateID(), new KvStateServerAddress(InetAddress.getLocalHost(), 1293));
jobManager.tell(registerForExistingJob);
lookupAfterRegistration = new LookupKvStateLocation(registerForExistingJob.getJobId(), registerForExistingJob.getRegistrationName());
lookupFuture = jobManager.ask(lookupAfterRegistration, deadline.timeLeft()).mapTo(ClassTag$.MODULE$.<KvStateLocation>apply(KvStateLocation.class));
KvStateLocation location = Await.result(lookupFuture, deadline.timeLeft());
assertNotNull(location);
assertEquals(jobGraph.getJobID(), location.getJobId());
assertEquals(jobVertex1.getID(), location.getJobVertexId());
assertEquals(jobVertex1.getMaxParallelism(), location.getNumKeyGroups());
assertEquals(1, location.getNumRegisteredKeyGroups());
KeyGroupRange keyGroupRange = registerForExistingJob.getKeyGroupRange();
assertEquals(1, keyGroupRange.getNumberOfKeyGroups());
assertEquals(registerForExistingJob.getKvStateId(), location.getKvStateID(keyGroupRange.getStartKeyGroup()));
assertEquals(registerForExistingJob.getKvStateServerAddress(), location.getKvStateServerAddress(keyGroupRange.getStartKeyGroup()));
//
// Unregistration
//
NotifyKvStateUnregistered unregister = new NotifyKvStateUnregistered(registerForExistingJob.getJobId(), registerForExistingJob.getJobVertexId(), registerForExistingJob.getKeyGroupRange(), registerForExistingJob.getRegistrationName());
jobManager.tell(unregister);
lookupFuture = jobManager.ask(lookupAfterRegistration, deadline.timeLeft()).mapTo(ClassTag$.MODULE$.<KvStateLocation>apply(KvStateLocation.class));
try {
Await.result(lookupFuture, deadline.timeLeft());
fail("Did not throw expected Exception");
} catch (UnknownKvStateLocation ignored) {
// Expected
}
//
// Duplicate registration fails task
//
NotifyKvStateRegistered register = new NotifyKvStateRegistered(jobGraph.getJobID(), jobVertex1.getID(), new KeyGroupRange(0, 0), "duplicate-me", new KvStateID(), new KvStateServerAddress(InetAddress.getLocalHost(), 1293));
NotifyKvStateRegistered duplicate = new NotifyKvStateRegistered(jobGraph.getJobID(), // <--- different operator, but...
jobVertex2.getID(), new KeyGroupRange(0, 0), // ...same name
"duplicate-me", new KvStateID(), new KvStateServerAddress(InetAddress.getLocalHost(), 1293));
Future<TestingJobManagerMessages.JobStatusIs> failedFuture = jobManager.ask(new NotifyWhenJobStatus(jobGraph.getJobID(), JobStatus.FAILED), deadline.timeLeft()).mapTo(ClassTag$.MODULE$.<JobStatusIs>apply(JobStatusIs.class));
jobManager.tell(register);
jobManager.tell(duplicate);
// Wait for failure
JobStatusIs jobStatus = Await.result(failedFuture, deadline.timeLeft());
assertEquals(JobStatus.FAILED, jobStatus.state());
}
use of org.apache.flink.runtime.jobgraph.JobVertex in project flink by apache.
the class JobManagerTest method testRequestPartitionStateUnregisteredExecution.
/**
* Tests the JobManager response when the execution is not registered with
* the ExecutionGraph.
*/
@Test
public void testRequestPartitionStateUnregisteredExecution() throws Exception {
new JavaTestKit(system) {
{
new Within(duration("15 seconds")) {
@Override
protected void run() {
// Setup
TestingCluster cluster = null;
try {
cluster = startTestingCluster(4, 1, DEFAULT_AKKA_ASK_TIMEOUT());
final IntermediateDataSetID rid = new IntermediateDataSetID();
// Create a task
final JobVertex sender = new JobVertex("Sender");
sender.setParallelism(1);
// just finish
sender.setInvokableClass(NoOpInvokable.class);
sender.createAndAddResultDataSet(rid, PIPELINED);
final JobVertex sender2 = new JobVertex("Blocking Sender");
sender2.setParallelism(1);
// just block
sender2.setInvokableClass(BlockingNoOpInvokable.class);
sender2.createAndAddResultDataSet(new IntermediateDataSetID(), PIPELINED);
final JobGraph jobGraph = new JobGraph("Fast finishing producer test job", sender, sender2);
final JobID jid = jobGraph.getJobID();
final ActorGateway jobManagerGateway = cluster.getLeaderGateway(TestingUtils.TESTING_DURATION());
// we can set the leader session ID to None because we don't use this gateway to send messages
final ActorGateway testActorGateway = new AkkaActorGateway(getTestActor(), null);
// Submit the job and wait for all vertices to be running
jobManagerGateway.tell(new SubmitJob(jobGraph, ListeningBehaviour.EXECUTION_RESULT), testActorGateway);
expectMsgClass(JobSubmitSuccess.class);
jobManagerGateway.tell(new WaitForAllVerticesToBeRunningOrFinished(jid), testActorGateway);
expectMsgClass(AllVerticesRunning.class);
Future<Object> egFuture = jobManagerGateway.ask(new RequestExecutionGraph(jobGraph.getJobID()), remaining());
ExecutionGraphFound egFound = (ExecutionGraphFound) Await.result(egFuture, remaining());
ExecutionGraph eg = (ExecutionGraph) egFound.executionGraph();
ExecutionVertex vertex = eg.getJobVertex(sender.getID()).getTaskVertices()[0];
while (vertex.getExecutionState() != ExecutionState.FINISHED) {
Thread.sleep(1);
}
IntermediateResultPartition partition = vertex.getProducedPartitions().values().iterator().next();
ResultPartitionID partitionId = new ResultPartitionID(partition.getPartitionId(), vertex.getCurrentExecutionAttempt().getAttemptId());
// Producer finished, request state
Object request = new RequestPartitionProducerState(jid, rid, partitionId);
Future<ExecutionState> producerStateFuture = jobManagerGateway.ask(request, getRemainingTime()).mapTo(ClassTag$.MODULE$.<ExecutionState>apply(ExecutionState.class));
assertEquals(ExecutionState.FINISHED, Await.result(producerStateFuture, getRemainingTime()));
} catch (Exception e) {
e.printStackTrace();
fail(e.getMessage());
} finally {
if (cluster != null) {
cluster.shutdown();
}
}
}
};
}
};
}
use of org.apache.flink.runtime.jobgraph.JobVertex in project flink by apache.
the class JobManagerTest method testRequestPartitionState.
/**
* Tests responses to partition state requests.
*/
@Test
public void testRequestPartitionState() throws Exception {
new JavaTestKit(system) {
{
new Within(duration("15 seconds")) {
@Override
protected void run() {
// Setup
TestingCluster cluster = null;
try {
cluster = startTestingCluster(2, 1, DEFAULT_AKKA_ASK_TIMEOUT());
final IntermediateDataSetID rid = new IntermediateDataSetID();
// Create a task
final JobVertex sender = new JobVertex("Sender");
sender.setParallelism(1);
// just block
sender.setInvokableClass(BlockingNoOpInvokable.class);
sender.createAndAddResultDataSet(rid, PIPELINED);
final JobGraph jobGraph = new JobGraph("Blocking test job", sender);
final JobID jid = jobGraph.getJobID();
final ActorGateway jobManagerGateway = cluster.getLeaderGateway(TestingUtils.TESTING_DURATION());
// we can set the leader session ID to None because we don't use this gateway to send messages
final ActorGateway testActorGateway = new AkkaActorGateway(getTestActor(), null);
// Submit the job and wait for all vertices to be running
jobManagerGateway.tell(new SubmitJob(jobGraph, ListeningBehaviour.EXECUTION_RESULT), testActorGateway);
expectMsgClass(JobSubmitSuccess.class);
jobManagerGateway.tell(new WaitForAllVerticesToBeRunningOrFinished(jid), testActorGateway);
expectMsgClass(AllVerticesRunning.class);
// This is the mock execution ID of the task requesting the state of the partition
final ExecutionAttemptID receiver = new ExecutionAttemptID();
// Request the execution graph to get the runtime info
jobManagerGateway.tell(new RequestExecutionGraph(jid), testActorGateway);
final ExecutionGraph eg = (ExecutionGraph) expectMsgClass(ExecutionGraphFound.class).executionGraph();
final ExecutionVertex vertex = eg.getJobVertex(sender.getID()).getTaskVertices()[0];
final IntermediateResultPartition partition = vertex.getProducedPartitions().values().iterator().next();
final ResultPartitionID partitionId = new ResultPartitionID(partition.getPartitionId(), vertex.getCurrentExecutionAttempt().getAttemptId());
// - The test ----------------------------------------------------------------------
// 1. All execution states
RequestPartitionProducerState request = new RequestPartitionProducerState(jid, rid, partitionId);
for (ExecutionState state : ExecutionState.values()) {
ExecutionGraphTestUtils.setVertexState(vertex, state);
Future<ExecutionState> futurePartitionState = jobManagerGateway.ask(request, getRemainingTime()).mapTo(ClassTag$.MODULE$.<ExecutionState>apply(ExecutionState.class));
ExecutionState resp = Await.result(futurePartitionState, getRemainingTime());
assertEquals(state, resp);
}
// 2. Non-existing execution
request = new RequestPartitionProducerState(jid, rid, new ResultPartitionID());
Future<?> futurePartitionState = jobManagerGateway.ask(request, getRemainingTime());
try {
Await.result(futurePartitionState, getRemainingTime());
fail("Did not fail with expected RuntimeException");
} catch (RuntimeException e) {
assertEquals(IllegalArgumentException.class, e.getCause().getClass());
}
// 3. Non-existing job
request = new RequestPartitionProducerState(new JobID(), rid, new ResultPartitionID());
futurePartitionState = jobManagerGateway.ask(request, getRemainingTime());
try {
Await.result(futurePartitionState, getRemainingTime());
fail("Did not fail with expected IllegalArgumentException");
} catch (IllegalArgumentException ignored) {
}
} catch (Exception e) {
e.printStackTrace();
fail(e.getMessage());
} finally {
if (cluster != null) {
cluster.shutdown();
}
}
}
};
}
};
}
use of org.apache.flink.runtime.jobgraph.JobVertex in project flink by apache.
the class JobManagerTest method testStopSignal.
@Test
public void testStopSignal() throws Exception {
new JavaTestKit(system) {
{
new Within(duration("15 seconds")) {
@Override
protected void run() {
// Setup
TestingCluster cluster = null;
try {
cluster = startTestingCluster(2, 1, DEFAULT_AKKA_ASK_TIMEOUT());
// Create a task
final JobVertex sender = new JobVertex("Sender");
sender.setParallelism(2);
sender.setInvokableClass(StoppableInvokable.class);
final JobGraph jobGraph = new JobGraph("Stoppable streaming test job", sender);
final JobID jid = jobGraph.getJobID();
final ActorGateway jobManagerGateway = cluster.getLeaderGateway(TestingUtils.TESTING_DURATION());
// we can set the leader session ID to None because we don't use this gateway to send messages
final ActorGateway testActorGateway = new AkkaActorGateway(getTestActor(), null);
// Submit the job and wait for all vertices to be running
jobManagerGateway.tell(new SubmitJob(jobGraph, ListeningBehaviour.EXECUTION_RESULT), testActorGateway);
expectMsgClass(JobSubmitSuccess.class);
jobManagerGateway.tell(new WaitForAllVerticesToBeRunning(jid), testActorGateway);
expectMsgClass(AllVerticesRunning.class);
jobManagerGateway.tell(new StopJob(jid), testActorGateway);
// - The test ----------------------------------------------------------------------
expectMsgClass(StoppingSuccess.class);
expectMsgClass(JobResultSuccess.class);
} finally {
if (cluster != null) {
cluster.shutdown();
}
}
}
};
}
};
}
Aggregations