use of org.apache.flink.runtime.jobgraph.JobGraph in project flink by apache.
the class JobClientActorRecoveryITCase method testJobClientRecovery.
/**
* Tests wether the JobClientActor can connect to a newly elected leading job manager to obtain
* the JobExecutionResult. The submitted job blocks for the first execution attempt. The
* leading job manager will be killed so that the second job manager will be elected as the
* leader. The newly elected leader has to retrieve the checkpointed job from ZooKeeper
* and continue its execution. This time, the job does not block and, thus, can be finished.
* The execution result should be sent to the JobClientActor which originally submitted the
* job.
*
* @throws Exception
*/
@Test
public void testJobClientRecovery() throws Exception {
File rootFolder = tempFolder.getRoot();
Configuration config = ZooKeeperTestUtils.createZooKeeperHAConfig(zkServer.getConnectString(), rootFolder.getPath());
config.setInteger(ConfigConstants.LOCAL_NUMBER_JOB_MANAGER, 2);
config.setInteger(ConfigConstants.LOCAL_NUMBER_TASK_MANAGER, 1);
final TestingCluster cluster = new TestingCluster(config);
cluster.start();
JobVertex blockingVertex = new JobVertex("Blocking Vertex");
blockingVertex.setInvokableClass(BlockingTask.class);
blockingVertex.setParallelism(1);
final JobGraph jobGraph = new JobGraph("Blocking Test Job", blockingVertex);
final Promise<JobExecutionResult> promise = new scala.concurrent.impl.Promise.DefaultPromise<>();
Deadline deadline = new FiniteDuration(2, TimeUnit.MINUTES).fromNow();
try {
Thread submitter = new Thread(new Runnable() {
@Override
public void run() {
try {
JobExecutionResult result = cluster.submitJobAndWait(jobGraph, false);
promise.success(result);
} catch (Exception e) {
promise.failure(e);
}
}
});
submitter.start();
synchronized (BlockingTask.waitLock) {
while (BlockingTask.HasBlockedExecution < 1 && deadline.hasTimeLeft()) {
BlockingTask.waitLock.wait(deadline.timeLeft().toMillis());
}
}
if (deadline.isOverdue()) {
Assert.fail("The job has not blocked within the given deadline.");
}
ActorGateway gateway = cluster.getLeaderGateway(deadline.timeLeft());
gateway.tell(TestingJobManagerMessages.getDisablePostStop());
gateway.tell(PoisonPill.getInstance());
// if the job fails then an exception is thrown here
Await.result(promise.future(), deadline.timeLeft());
} finally {
cluster.shutdown();
}
}
use of org.apache.flink.runtime.jobgraph.JobGraph in project flink by apache.
the class BackPressureStatsTrackerITCase method testBackPressuredProducer.
/**
* Tests a simple fake-back pressured task. Back pressure is assumed when
* sampled stack traces are in blocking buffer requests.
*/
@Test
public void testBackPressuredProducer() throws Exception {
new JavaTestKit(testActorSystem) {
{
final FiniteDuration deadline = new FiniteDuration(60, TimeUnit.SECONDS);
// The JobGraph
final JobGraph jobGraph = new JobGraph();
final int parallelism = 4;
final JobVertex task = new JobVertex("Task");
task.setInvokableClass(BackPressuredTask.class);
task.setParallelism(parallelism);
jobGraph.addVertex(task);
ActorGateway jobManger = null;
ActorGateway taskManager = null;
//
// 1) Consume all buffers at first (no buffers for the test task)
//
testBufferPool = networkBufferPool.createBufferPool(1, Integer.MAX_VALUE);
final List<Buffer> buffers = new ArrayList<>();
while (true) {
Buffer buffer = testBufferPool.requestBuffer();
if (buffer != null) {
buffers.add(buffer);
} else {
break;
}
}
try {
jobManger = TestingUtils.createJobManager(testActorSystem, TestingUtils.defaultExecutor(), TestingUtils.defaultExecutor(), new Configuration());
final Configuration config = new Configuration();
config.setInteger(ConfigConstants.TASK_MANAGER_NUM_TASK_SLOTS, parallelism);
taskManager = TestingUtils.createTaskManager(testActorSystem, jobManger, config, true, true);
final ActorGateway jm = jobManger;
new Within(deadline) {
@Override
protected void run() {
try {
ActorGateway testActor = new AkkaActorGateway(getTestActor(), null);
// Submit the job and wait until it is running
JobClient.submitJobDetached(jm, config, jobGraph, deadline, ClassLoader.getSystemClassLoader());
jm.tell(new WaitForAllVerticesToBeRunning(jobGraph.getJobID()), testActor);
expectMsgEquals(new AllVerticesRunning(jobGraph.getJobID()));
// Get the ExecutionGraph
jm.tell(new RequestExecutionGraph(jobGraph.getJobID()), testActor);
ExecutionGraphFound executionGraphResponse = expectMsgClass(ExecutionGraphFound.class);
ExecutionGraph executionGraph = (ExecutionGraph) executionGraphResponse.executionGraph();
ExecutionJobVertex vertex = executionGraph.getJobVertex(task.getID());
StackTraceSampleCoordinator coordinator = new StackTraceSampleCoordinator(testActorSystem.dispatcher(), 60000);
// Verify back pressure (clean up interval can be ignored)
BackPressureStatsTracker statsTracker = new BackPressureStatsTracker(coordinator, 100 * 1000, 20, Time.milliseconds(10L));
int numAttempts = 10;
int nextSampleId = 0;
// the buffer.
for (int attempt = 0; attempt < numAttempts; attempt++) {
try {
OperatorBackPressureStats stats = triggerStatsSample(statsTracker, vertex);
assertEquals(nextSampleId + attempt, stats.getSampleId());
assertEquals(parallelism, stats.getNumberOfSubTasks());
assertEquals(1.0, stats.getMaxBackPressureRatio(), 0.0);
for (int i = 0; i < parallelism; i++) {
assertEquals(1.0, stats.getBackPressureRatio(i), 0.0);
}
nextSampleId = stats.getSampleId() + 1;
break;
} catch (Throwable t) {
if (attempt == numAttempts - 1) {
throw t;
} else {
Thread.sleep(500);
}
}
}
//
for (Buffer buf : buffers) {
buf.recycle();
}
// grab them and then immediately release them.
while (testBufferPool.getNumberOfAvailableMemorySegments() < 100) {
Thread.sleep(100);
}
// Verify that no task is back pressured any more.
for (int attempt = 0; attempt < numAttempts; attempt++) {
try {
OperatorBackPressureStats stats = triggerStatsSample(statsTracker, vertex);
assertEquals(nextSampleId + attempt, stats.getSampleId());
assertEquals(parallelism, stats.getNumberOfSubTasks());
// Verify that no task is back pressured
for (int i = 0; i < parallelism; i++) {
assertEquals(0.0, stats.getBackPressureRatio(i), 0.0);
}
break;
} catch (Throwable t) {
if (attempt == numAttempts - 1) {
throw t;
} else {
Thread.sleep(500);
}
}
}
// Shut down
jm.tell(new TestingJobManagerMessages.NotifyWhenJobRemoved(jobGraph.getJobID()), testActor);
// Cancel job
jm.tell(new JobManagerMessages.CancelJob(jobGraph.getJobID()));
// Response to removal notification
expectMsgEquals(true);
//
// 3) Trigger stats for archived job
//
statsTracker.invalidateOperatorStatsCache();
assertFalse("Unexpected trigger", statsTracker.triggerStackTraceSample(vertex));
} catch (Exception e) {
e.printStackTrace();
fail(e.getMessage());
}
}
};
} finally {
TestingUtils.stopActor(jobManger);
TestingUtils.stopActor(taskManager);
for (Buffer buf : buffers) {
buf.recycle();
}
testBufferPool.lazyDestroy();
}
}
};
}
use of org.apache.flink.runtime.jobgraph.JobGraph in project flink by apache.
the class JarPlanHandler method handleJsonRequest.
@Override
public String handleJsonRequest(Map<String, String> pathParams, Map<String, String> queryParams, ActorGateway jobManager) throws Exception {
try {
JarActionHandlerConfig config = JarActionHandlerConfig.fromParams(pathParams, queryParams);
JobGraph graph = getJobGraphAndClassLoader(config).f0;
StringWriter writer = new StringWriter();
JsonGenerator gen = JsonFactory.jacksonFactory.createGenerator(writer);
gen.writeStartObject();
gen.writeFieldName("plan");
gen.writeRawValue(JsonPlanGenerator.generatePlan(graph));
gen.writeEndObject();
gen.close();
return writer.toString();
} catch (Exception e) {
return sendError(e);
}
}
use of org.apache.flink.runtime.jobgraph.JobGraph 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.JobGraph 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());
}
Aggregations