@Test
public void testDisposeSavepointSuccess() throws Exception {
replaceStdOutAndStdErr();
try {
String savepointPath = "expectedSavepointPath";
ActorGateway jobManager = mock(ActorGateway.class);
Promise<Object> triggerResponse = new scala.concurrent.impl.Promise.DefaultPromise<>();
when(jobManager.ask(
Mockito.eq(new JobManagerMessages.DisposeSavepoint(savepointPath)),
Mockito.any(FiniteDuration.class)))
.thenReturn(triggerResponse.future());
triggerResponse.success(JobManagerMessages.getDisposeSavepointSuccess());
CliFrontend frontend = new MockCliFrontend(CliFrontendTestUtils.getConfigDir(), jobManager);
String[] parameters = {"-d", savepointPath};
int returnCode = frontend.savepoint(parameters);
assertEquals(0, returnCode);
verify(jobManager, times(1))
.ask(
Mockito.eq(new JobManagerMessages.DisposeSavepoint(savepointPath)),
Mockito.any(FiniteDuration.class));
String outMsg = buffer.toString();
assertTrue(outMsg.contains(savepointPath));
assertTrue(outMsg.contains("disposed"));
} finally {
restoreStdOutAndStdErr();
}
}
/**
* Tests that a job is properly canceled in the case of a leader change. However, this time only
* the JMs are notified about the leader change and the TMs still believe the old leader to have
* leadership.
*/
@Test
public void testStateCleanupAfterNewLeaderElection() throws Exception {
UUID leaderSessionID = UUID.randomUUID();
UUID newLeaderSessionID = UUID.randomUUID();
cluster.grantLeadership(0, leaderSessionID);
cluster.notifyRetrievalListeners(0, leaderSessionID);
cluster.waitForTaskManagersToBeRegistered();
// submit blocking job so that we can test job clean up
cluster.submitJobDetached(job);
ActorGateway jm = cluster.getLeaderGateway(timeout);
Future<Object> wait =
jm.ask(new WaitForAllVerticesToBeRunningOrFinished(job.getJobID()), timeout);
Await.ready(wait, timeout);
Future<Object> jobRemoval = jm.ask(new NotifyWhenJobRemoved(job.getJobID()), timeout);
// only notify the JMs about the new leader JM(1)
cluster.grantLeadership(1, newLeaderSessionID);
// job should be removed anyway
Await.ready(jobRemoval, timeout);
}
@Test
public void testTriggerSavepointFailure() throws Exception {
replaceStdOutAndStdErr();
try {
JobID jobId = new JobID();
ActorGateway jobManager = mock(ActorGateway.class);
Promise<Object> triggerResponse = new scala.concurrent.impl.Promise.DefaultPromise<>();
when(jobManager.ask(
Mockito.eq(new JobManagerMessages.TriggerSavepoint(jobId)),
Mockito.any(FiniteDuration.class)))
.thenReturn(triggerResponse.future());
Exception testException = new Exception("expectedTestException");
triggerResponse.success(new JobManagerMessages.TriggerSavepointFailure(jobId, testException));
CliFrontend frontend = new MockCliFrontend(CliFrontendTestUtils.getConfigDir(), jobManager);
String[] parameters = {jobId.toString()};
int returnCode = frontend.savepoint(parameters);
assertTrue(returnCode != 0);
verify(jobManager, times(1))
.ask(
Mockito.eq(new JobManagerMessages.TriggerSavepoint(jobId)),
Mockito.any(FiniteDuration.class));
assertTrue(buffer.toString().contains("expectedTestException"));
} finally {
restoreStdOutAndStdErr();
}
}
/**
* Tests that a job is properly canceled in the event of a leader change. However, this time only
* the TMs are notified about the changing leader. This should be enough to cancel the currently
* running job, though.
*/
@Test
public void testStateCleanupAfterListenerNotification() throws Exception {
UUID leaderSessionID = UUID.randomUUID();
UUID newLeaderSessionID = UUID.randomUUID();
cluster.grantLeadership(0, leaderSessionID);
cluster.notifyRetrievalListeners(0, leaderSessionID);
cluster.waitForTaskManagersToBeRegistered();
// submit blocking job
cluster.submitJobDetached(job);
ActorGateway jm = cluster.getLeaderGateway(timeout);
Future<Object> wait =
jm.ask(new WaitForAllVerticesToBeRunningOrFinished(job.getJobID()), timeout);
Await.ready(wait, timeout);
Future<Object> jobRemoval = jm.ask(new NotifyWhenJobRemoved(job.getJobID()), timeout);
// notify listeners (TMs) about the leader change
cluster.notifyRetrievalListeners(1, newLeaderSessionID);
Await.ready(jobRemoval, timeout);
}
/**
* This method sends a CancelTask message to the instance of the assigned slot.
*
* <p>The sending is tried up to NUM_CANCEL_CALL_TRIES times.
*/
private void sendCancelRpcCall() {
final SimpleSlot slot = this.assignedResource;
if (slot != null) {
final ActorGateway gateway = slot.getInstance().getActorGateway();
Future<Object> cancelResult =
gateway.retry(
new CancelTask(attemptId), NUM_CANCEL_CALL_TRIES, timeout, executionContext);
cancelResult.onComplete(
new OnComplete<Object>() {
@Override
public void onComplete(Throwable failure, Object success) throws Throwable {
if (failure != null) {
fail(new Exception("Task could not be canceled.", failure));
} else {
TaskOperationResult result = (TaskOperationResult) success;
if (!result.success()) {
LOG.debug(
"Cancel task call did not find task. Probably akka message call" + " race.");
}
}
}
},
executionContext);
}
}
@Override
public void run() {
try {
int returnValue;
switch (type) {
case YARN_SESSION:
yCli = new FlinkYarnSessionCli("", "", false);
returnValue = yCli.run(args);
break;
case CLI_FRONTEND:
TestingCLI cli;
try {
cli = new TestingCLI();
returnValue = cli.parseParameters(args);
} catch (Exception e) {
throw new RuntimeException(
"Failed to execute the following args with CliFrontend: " + Arrays.toString(args),
e);
}
final ClusterClient client = cli.getClusterClient();
try {
// check if the JobManager is still alive after running the job
final FiniteDuration finiteDuration = new FiniteDuration(10, TimeUnit.SECONDS);
ActorGateway jobManagerGateway = client.getJobManagerGateway();
Await.ready(
jobManagerGateway.ask(new Identify(true), finiteDuration), finiteDuration);
} catch (Exception e) {
throw new RuntimeException(
"It seems like the JobManager died although it should still be alive");
}
// verify we would have shut down anyways and then shutdown
Mockito.verify(cli.getSpiedClusterClient()).shutdown();
client.shutdown();
break;
default:
throw new RuntimeException("Unknown type " + type);
}
if (returnValue != this.expectedReturnValue) {
Assert.fail(
"The YARN session returned with unexpected value="
+ returnValue
+ " expected="
+ expectedReturnValue);
}
} catch (Throwable t) {
LOG.info("Runner stopped with exception", t);
// save error.
this.runnerError = t;
}
}
private void sendFailIntermediateResultPartitionsRpcCall() {
final SimpleSlot slot = this.assignedResource;
if (slot != null) {
final Instance instance = slot.getInstance();
if (instance.isAlive()) {
final ActorGateway gateway = instance.getActorGateway();
// TODO For some tests this could be a problem when querying too early if all resources were
// released
gateway.tell(new FailIntermediateResultPartitions(attemptId));
}
}
}
/**
* Verifies a correct error message when vertices with master initialization (input formats /
* output formats) fail.
*/
@Test
public void testFailureWhenInitializeOnMasterFails() {
try {
// create a simple job graph
JobVertex jobVertex =
new JobVertex("Vertex that fails in initializeOnMaster") {
@Override
public void initializeOnMaster(ClassLoader loader) throws Exception {
throw new RuntimeException("test exception");
}
};
jobVertex.setInvokableClass(Tasks.NoOpInvokable.class);
JobGraph jg = new JobGraph("test job", jobVertex);
// submit the job
Future<Object> submitFuture =
jmGateway.ask(
new JobManagerMessages.SubmitJob(jg, ListeningBehaviour.EXECUTION_RESULT), timeout);
try {
Await.result(submitFuture, timeout);
} catch (JobExecutionException e) {
// that is what we expect
// test that the exception nesting is not too deep
assertTrue(e.getCause() instanceof RuntimeException);
} catch (Exception e) {
fail("Wrong exception type");
}
} catch (Exception e) {
e.printStackTrace();
fail(e.getMessage());
}
}
/**
* Tests that a job is properly canceled in the case of a leader change. In such an event all
* TaskManagers have to disconnect from the previous leader and connect to the newly elected
* leader.
*/
@Test
public void testStateCleanupAfterNewLeaderElectionAndListenerNotification() throws Exception {
UUID leaderSessionID1 = UUID.randomUUID();
UUID leaderSessionID2 = UUID.randomUUID();
// first make JM(0) the leader
cluster.grantLeadership(0, leaderSessionID1);
// notify all listeners
cluster.notifyRetrievalListeners(0, leaderSessionID1);
cluster.waitForTaskManagersToBeRegistered();
// submit blocking job so that it is not finished when we cancel it
cluster.submitJobDetached(job);
ActorGateway jm = cluster.getLeaderGateway(timeout);
Future<Object> wait =
jm.ask(new WaitForAllVerticesToBeRunningOrFinished(job.getJobID()), timeout);
Await.ready(wait, timeout);
Future<Object> jobRemoval = jm.ask(new NotifyWhenJobRemoved(job.getJobID()), timeout);
// make the JM(1) the new leader
cluster.grantLeadership(1, leaderSessionID2);
// notify all listeners about the event
cluster.notifyRetrievalListeners(1, leaderSessionID2);
Await.ready(jobRemoval, timeout);
cluster.waitForTaskManagersToBeRegistered();
ActorGateway jm2 = cluster.getLeaderGateway(timeout);
Future<Object> futureNumberSlots =
jm2.ask(JobManagerMessages.getRequestTotalNumberOfSlots(), timeout);
// check that all TMs have registered at the new leader
int numberSlots = (Integer) Await.result(futureNumberSlots, timeout);
assertEquals(parallelism, numberSlots);
// try to resubmit now the non-blocking job, it should complete successfully
Tasks.BlockingOnceReceiver$.MODULE$.blocking_$eq(false);
cluster.submitJobAndWait(job, false, timeout);
}
/**
* Tests that the same JobManager can be reelected as the leader. Even though, the same JM is
* elected as the next leader, all currently running jobs should be canceled properly and all TMs
* should disconnect from the leader and then reconnect to it.
*/
@Test
public void testReelectionOfSameJobManager() throws Exception {
UUID leaderSessionID = UUID.randomUUID();
UUID newLeaderSessionID = UUID.randomUUID();
FiniteDuration shortTimeout = new FiniteDuration(20, TimeUnit.SECONDS);
cluster.grantLeadership(0, leaderSessionID);
cluster.notifyRetrievalListeners(0, leaderSessionID);
cluster.waitForTaskManagersToBeRegistered();
// submit blocking job
cluster.submitJobDetached(job);
ActorGateway jm = cluster.getLeaderGateway(timeout);
Future<Object> wait =
jm.ask(new WaitForAllVerticesToBeRunningOrFinished(job.getJobID()), timeout);
Await.ready(wait, timeout);
Future<Object> jobRemoval = jm.ask(new NotifyWhenJobRemoved(job.getJobID()), timeout);
// make JM(0) again the leader --> this implies first a leadership revokal
cluster.grantLeadership(0, newLeaderSessionID);
Await.ready(jobRemoval, timeout);
// The TMs should not be able to reconnect since they don't know the current leader
// session ID
try {
cluster.waitForTaskManagersToBeRegistered(shortTimeout);
fail("TaskManager should not be able to register at JobManager.");
} catch (TimeoutException e) {
// expected exception since the TMs have still the old leader session ID
}
// notify the TMs about the new (old) leader
cluster.notifyRetrievalListeners(0, newLeaderSessionID);
cluster.waitForTaskManagersToBeRegistered();
// try to resubmit now the non-blocking job, it should complete successfully
Tasks.BlockingOnceReceiver$.MODULE$.blocking_$eq(false);
cluster.submitJobAndWait(job, false, timeout);
}
@Test
public void testFailureWhenJarBlobsMissing() {
try {
// create a simple job graph
JobVertex jobVertex = new JobVertex("Test Vertex");
jobVertex.setInvokableClass(Tasks.NoOpInvokable.class);
JobGraph jg = new JobGraph("test job", jobVertex);
// request the blob port from the job manager
Future<Object> future =
jmGateway.ask(JobManagerMessages.getRequestBlobManagerPort(), timeout);
int blobPort = (Integer) Await.result(future, timeout);
// upload two dummy bytes and add their keys to the job graph as dependencies
BlobKey key1, key2;
BlobClient bc = new BlobClient(new InetSocketAddress("localhost", blobPort));
try {
key1 = bc.put(new byte[10]);
key2 = bc.put(new byte[10]);
// delete one of the blobs to make sure that the startup failed
bc.delete(key2);
} finally {
bc.close();
}
jg.addBlob(key1);
jg.addBlob(key2);
// submit the job
Future<Object> submitFuture =
jmGateway.ask(
new JobManagerMessages.SubmitJob(jg, ListeningBehaviour.EXECUTION_RESULT), timeout);
try {
Await.result(submitFuture, timeout);
} catch (JobExecutionException e) {
// that is what we expect
assertTrue(e.getCause() instanceof IOException);
} catch (Exception e) {
fail("Wrong exception type");
}
} catch (Exception e) {
e.printStackTrace();
fail(e.getMessage());
}
}
/**
* Sends an UpdatePartitionInfo message to the instance of the consumerSlot.
*
* @param consumerSlot Slot to whose instance the message will be sent
* @param updatePartitionInfo UpdatePartitionInfo message
*/
private void sendUpdatePartitionInfoRpcCall(
final SimpleSlot consumerSlot, final UpdatePartitionInfo updatePartitionInfo) {
if (consumerSlot != null) {
final Instance instance = consumerSlot.getInstance();
final ActorGateway gateway = instance.getActorGateway();
Future<Object> futureUpdate = gateway.ask(updatePartitionInfo, timeout);
futureUpdate.onFailure(
new OnFailure() {
@Override
public void onFailure(Throwable failure) throws Throwable {
fail(
new IllegalStateException(
"Update task on instance " + instance + " failed due to:", failure));
}
},
executionContext);
}
}
/**
* Shuts down the checkpoint coordinator.
*
* <p>After this method has been called, the coordinator does not accept and further messages and
* cannot trigger any further checkpoints.
*/
public void shutdown() throws Exception {
synchronized (lock) {
try {
if (!shutdown) {
shutdown = true;
LOG.info("Stopping checkpoint coordinator for job " + job);
periodicScheduling = false;
triggerRequestQueued = false;
// shut down the thread that handles the timeouts and pending triggers
timer.cancel();
// make sure that the actor does not linger
if (jobStatusListener != null) {
jobStatusListener.tell(PoisonPill.getInstance());
jobStatusListener = null;
}
checkpointIdCounter.stop();
// clear and discard all pending checkpoints
for (PendingCheckpoint pending : pendingCheckpoints.values()) {
pending.discard(userClassLoader);
}
pendingCheckpoints.clear();
// clean and discard all successful checkpoints
completedCheckpointStore.discardAllCheckpoints();
onShutdown();
}
} finally {
// Remove shutdown hook to prevent resource leaks, unless this is invoked by the
// shutdown hook itself.
if (shutdownHook != null && shutdownHook != Thread.currentThread()) {
try {
Runtime.getRuntime().removeShutdownHook(shutdownHook);
} catch (IllegalStateException ignored) {
// race, JVM is in shutdown already, we can safely ignore this
} catch (Throwable t) {
LOG.warn("Error unregistering checkpoint coordinator shutdown hook.", t);
}
}
}
}
}
public ExecutionGraph getExecutionGraph(JobID jid) {
ExecutionGraph cached = cache.get(jid);
if (cached != null) {
return cached;
}
try {
Future<Object> future = source.ask(new JobManagerMessages.RequestJob(jid), timeout);
Object result = Await.result(future, timeout);
if (result instanceof JobManagerMessages.JobNotFound) {
return null;
} else if (result instanceof JobManagerMessages.JobFound) {
ExecutionGraph eg = ((JobManagerMessages.JobFound) result).executionGraph();
cache.put(jid, eg);
return eg;
} else {
throw new RuntimeException("Unknown response from JobManager / Archive: " + result);
}
} catch (Exception e) {
throw new RuntimeException("Error requesting execution graph", e);
}
}
/**
* Tests that a job cannot be restarted from a savepoint with a different parallelism if the
* rescaled operator has non-partitioned state.
*
* @throws Exception
*/
@Test
public void testSavepointRescalingNonPartitionedStateCausesException() throws Exception {
final int parallelism = numSlots / 2;
final int parallelism2 = numSlots;
final int maxParallelism = 13;
FiniteDuration timeout = new FiniteDuration(3, TimeUnit.MINUTES);
Deadline deadline = timeout.fromNow();
JobID jobID = null;
ActorGateway jobManager = null;
try {
jobManager = cluster.getLeaderGateway(deadline.timeLeft());
JobGraph jobGraph = createJobGraphWithOperatorState(parallelism, maxParallelism, false);
jobID = jobGraph.getJobID();
cluster.submitJobDetached(jobGraph);
Object savepointResponse = null;
// wait until the operator is started
StateSourceBase.workStartedLatch.await();
while (deadline.hasTimeLeft()) {
Future<Object> savepointPathFuture =
jobManager.ask(new JobManagerMessages.TriggerSavepoint(jobID), deadline.timeLeft());
FiniteDuration waitingTime = new FiniteDuration(10, TimeUnit.SECONDS);
savepointResponse = Await.result(savepointPathFuture, waitingTime);
if (savepointResponse instanceof JobManagerMessages.TriggerSavepointSuccess) {
break;
}
}
assertTrue(savepointResponse instanceof JobManagerMessages.TriggerSavepointSuccess);
final String savepointPath =
((JobManagerMessages.TriggerSavepointSuccess) savepointResponse).savepointPath();
Future<Object> jobRemovedFuture =
jobManager.ask(
new TestingJobManagerMessages.NotifyWhenJobRemoved(jobID), deadline.timeLeft());
Future<Object> cancellationResponseFuture =
jobManager.ask(new JobManagerMessages.CancelJob(jobID), deadline.timeLeft());
Object cancellationResponse = Await.result(cancellationResponseFuture, deadline.timeLeft());
assertTrue(cancellationResponse instanceof JobManagerMessages.CancellationSuccess);
Await.ready(jobRemovedFuture, deadline.timeLeft());
// job successfully removed
jobID = null;
JobGraph scaledJobGraph =
createJobGraphWithOperatorState(parallelism2, maxParallelism, false);
scaledJobGraph.setSavepointPath(savepointPath);
jobID = scaledJobGraph.getJobID();
cluster.submitJobAndWait(scaledJobGraph, false);
jobID = null;
} catch (JobExecutionException exception) {
if (exception.getCause() instanceof SuppressRestartsException) {
SuppressRestartsException suppressRestartsException =
(SuppressRestartsException) exception.getCause();
if (suppressRestartsException.getCause() instanceof IllegalStateException) {
// we expect a IllegalStateException wrapped in a SuppressRestartsException wrapped
// in a JobExecutionException, because the job containing non-partitioned state
// is being rescaled
} else {
throw exception;
}
} else {
throw exception;
}
} finally {
// clear any left overs from a possibly failed job
if (jobID != null && jobManager != null) {
Future<Object> jobRemovedFuture =
jobManager.ask(new TestingJobManagerMessages.NotifyWhenJobRemoved(jobID), timeout);
try {
Await.ready(jobRemovedFuture, timeout);
} catch (TimeoutException | InterruptedException ie) {
fail("Failed while cleaning up the cluster.");
}
}
}
}
/**
* Tests rescaling of partitioned operator state. More specific, we test the mechanism with {@link
* ListCheckpointed} as it subsumes {@link
* org.apache.flink.streaming.api.checkpoint.CheckpointedFunction}.
*/
public void testSavepointRescalingPartitionedOperatorState(boolean scaleOut) throws Exception {
final int parallelism = scaleOut ? numSlots : numSlots / 2;
final int parallelism2 = scaleOut ? numSlots / 2 : numSlots;
final int maxParallelism = 13;
FiniteDuration timeout = new FiniteDuration(3, TimeUnit.MINUTES);
Deadline deadline = timeout.fromNow();
JobID jobID = null;
ActorGateway jobManager = null;
int counterSize = Math.max(parallelism, parallelism2);
PartitionedStateSource.CHECK_CORRECT_SNAPSHOT = new int[counterSize];
PartitionedStateSource.CHECK_CORRECT_RESTORE = new int[counterSize];
try {
jobManager = cluster.getLeaderGateway(deadline.timeLeft());
JobGraph jobGraph = createJobGraphWithOperatorState(parallelism, maxParallelism, true);
jobID = jobGraph.getJobID();
cluster.submitJobDetached(jobGraph);
Object savepointResponse = null;
// wait until the operator is started
StateSourceBase.workStartedLatch.await();
while (deadline.hasTimeLeft()) {
Future<Object> savepointPathFuture =
jobManager.ask(new JobManagerMessages.TriggerSavepoint(jobID), deadline.timeLeft());
FiniteDuration waitingTime = new FiniteDuration(10, TimeUnit.SECONDS);
savepointResponse = Await.result(savepointPathFuture, waitingTime);
if (savepointResponse instanceof JobManagerMessages.TriggerSavepointSuccess) {
break;
}
}
assertTrue(savepointResponse instanceof JobManagerMessages.TriggerSavepointSuccess);
final String savepointPath =
((JobManagerMessages.TriggerSavepointSuccess) savepointResponse).savepointPath();
Future<Object> jobRemovedFuture =
jobManager.ask(
new TestingJobManagerMessages.NotifyWhenJobRemoved(jobID), deadline.timeLeft());
Future<Object> cancellationResponseFuture =
jobManager.ask(new JobManagerMessages.CancelJob(jobID), deadline.timeLeft());
Object cancellationResponse = Await.result(cancellationResponseFuture, deadline.timeLeft());
assertTrue(cancellationResponse instanceof JobManagerMessages.CancellationSuccess);
Await.ready(jobRemovedFuture, deadline.timeLeft());
// job successfully removed
jobID = null;
JobGraph scaledJobGraph = createJobGraphWithOperatorState(parallelism2, maxParallelism, true);
scaledJobGraph.setSavepointPath(savepointPath);
jobID = scaledJobGraph.getJobID();
cluster.submitJobAndWait(scaledJobGraph, false);
int sumExp = 0;
int sumAct = 0;
for (int c : PartitionedStateSource.CHECK_CORRECT_SNAPSHOT) {
sumExp += c;
}
for (int c : PartitionedStateSource.CHECK_CORRECT_RESTORE) {
sumAct += c;
}
assertEquals(sumExp, sumAct);
jobID = null;
} finally {
// clear any left overs from a possibly failed job
if (jobID != null && jobManager != null) {
Future<Object> jobRemovedFuture =
jobManager.ask(new TestingJobManagerMessages.NotifyWhenJobRemoved(jobID), timeout);
try {
Await.ready(jobRemovedFuture, timeout);
} catch (TimeoutException | InterruptedException ie) {
fail("Failed while cleaning up the cluster.");
}
}
}
}
@Override
public String handleJsonRequest(
Map<String, String> pathParams, Map<String, String> queryParams, ActorGateway jobManager)
throws Exception {
try {
if (jobManager != null) {
// whether one task manager's metrics are requested, or all task manager, we
// return them in an array. This avoids unnecessary code complexity.
// If only one task manager is requested, we only fetch one task manager metrics.
final List<Instance> instances = new ArrayList<>();
if (pathParams.containsKey(TASK_MANAGER_ID_KEY)) {
try {
InstanceID instanceID =
new InstanceID(StringUtils.hexStringToByte(pathParams.get(TASK_MANAGER_ID_KEY)));
Future<Object> future =
jobManager.ask(
new JobManagerMessages.RequestTaskManagerInstance(instanceID), timeout);
TaskManagerInstance instance = (TaskManagerInstance) Await.result(future, timeout);
if (instance.instance().nonEmpty()) {
instances.add(instance.instance().get());
}
}
// this means the id string was invalid. Keep the list empty.
catch (IllegalArgumentException e) {
// do nothing.
}
} else {
Future<Object> future =
jobManager.ask(JobManagerMessages.getRequestRegisteredTaskManagers(), timeout);
RegisteredTaskManagers taskManagers =
(RegisteredTaskManagers) Await.result(future, timeout);
instances.addAll(taskManagers.asJavaCollection());
}
StringWriter writer = new StringWriter();
JsonGenerator gen = JsonFactory.jacksonFactory.createGenerator(writer);
gen.writeStartObject();
gen.writeArrayFieldStart("taskmanagers");
for (Instance instance : instances) {
gen.writeStartObject();
gen.writeStringField("id", instance.getId().toString());
gen.writeStringField("path", instance.getActorGateway().path());
gen.writeNumberField("dataPort", instance.getTaskManagerLocation().dataPort());
gen.writeNumberField("timeSinceLastHeartbeat", instance.getLastHeartBeat());
gen.writeNumberField("slotsNumber", instance.getTotalNumberOfSlots());
gen.writeNumberField("freeSlots", instance.getNumberOfAvailableSlots());
gen.writeNumberField("cpuCores", instance.getResources().getNumberOfCPUCores());
gen.writeNumberField("physicalMemory", instance.getResources().getSizeOfPhysicalMemory());
gen.writeNumberField("freeMemory", instance.getResources().getSizeOfJvmHeap());
gen.writeNumberField("managedMemory", instance.getResources().getSizeOfManagedMemory());
// only send metrics when only one task manager requests them.
if (pathParams.containsKey(TASK_MANAGER_ID_KEY)) {
byte[] report = instance.getLastMetricsReport();
if (report != null) {
gen.writeFieldName("metrics");
gen.writeRawValue(new String(report, "utf-8"));
}
}
gen.writeEndObject();
}
gen.writeEndArray();
gen.writeEndObject();
gen.close();
return writer.toString();
} else {
throw new Exception("No connection to the leading JobManager.");
}
} catch (Exception e) {
throw new RuntimeException("Failed to fetch list of all task managers: " + e.getMessage(), e);
}
}
/**
* Tests that a job with non partitioned state can be restarted from a savepoint with a different
* parallelism if the operator with non-partitioned state are not rescaled.
*
* @throws Exception
*/
@Test
public void testSavepointRescalingWithKeyedAndNonPartitionedState() throws Exception {
int numberKeys = 42;
int numberElements = 1000;
int numberElements2 = 500;
int parallelism = numSlots / 2;
int parallelism2 = numSlots;
int maxParallelism = 13;
FiniteDuration timeout = new FiniteDuration(3, TimeUnit.MINUTES);
Deadline deadline = timeout.fromNow();
ActorGateway jobManager = null;
JobID jobID = null;
try {
jobManager = cluster.getLeaderGateway(deadline.timeLeft());
JobGraph jobGraph =
createJobGraphWithKeyedAndNonPartitionedOperatorState(
parallelism, maxParallelism, parallelism, numberKeys, numberElements, false, 100);
jobID = jobGraph.getJobID();
cluster.submitJobDetached(jobGraph);
// wait til the sources have emitted numberElements for each key and completed a checkpoint
SubtaskIndexFlatMapper.workCompletedLatch.await(
deadline.timeLeft().toMillis(), TimeUnit.MILLISECONDS);
// verify the current state
Set<Tuple2<Integer, Integer>> actualResult = CollectionSink.getElementsSet();
Set<Tuple2<Integer, Integer>> expectedResult = new HashSet<>();
for (int key = 0; key < numberKeys; key++) {
int keyGroupIndex = KeyGroupRangeAssignment.assignToKeyGroup(key, maxParallelism);
expectedResult.add(
Tuple2.of(
KeyGroupRangeAssignment.computeOperatorIndexForKeyGroup(
maxParallelism, parallelism, keyGroupIndex),
numberElements * key));
}
assertEquals(expectedResult, actualResult);
// clear the CollectionSink set for the restarted job
CollectionSink.clearElementsSet();
Future<Object> savepointPathFuture =
jobManager.ask(new JobManagerMessages.TriggerSavepoint(jobID), deadline.timeLeft());
final String savepointPath =
((JobManagerMessages.TriggerSavepointSuccess)
Await.result(savepointPathFuture, deadline.timeLeft()))
.savepointPath();
Future<Object> jobRemovedFuture =
jobManager.ask(
new TestingJobManagerMessages.NotifyWhenJobRemoved(jobID), deadline.timeLeft());
Future<Object> cancellationResponseFuture =
jobManager.ask(new JobManagerMessages.CancelJob(jobID), deadline.timeLeft());
Object cancellationResponse = Await.result(cancellationResponseFuture, deadline.timeLeft());
assertTrue(cancellationResponse instanceof JobManagerMessages.CancellationSuccess);
Await.ready(jobRemovedFuture, deadline.timeLeft());
jobID = null;
JobGraph scaledJobGraph =
createJobGraphWithKeyedAndNonPartitionedOperatorState(
parallelism2,
maxParallelism,
parallelism,
numberKeys,
numberElements + numberElements2,
true,
100);
scaledJobGraph.setSavepointPath(savepointPath);
jobID = scaledJobGraph.getJobID();
cluster.submitJobAndWait(scaledJobGraph, false);
jobID = null;
Set<Tuple2<Integer, Integer>> actualResult2 = CollectionSink.getElementsSet();
Set<Tuple2<Integer, Integer>> expectedResult2 = new HashSet<>();
for (int key = 0; key < numberKeys; key++) {
int keyGroupIndex = KeyGroupRangeAssignment.assignToKeyGroup(key, maxParallelism);
expectedResult2.add(
Tuple2.of(
KeyGroupRangeAssignment.computeOperatorIndexForKeyGroup(
maxParallelism, parallelism2, keyGroupIndex),
key * (numberElements + numberElements2)));
}
assertEquals(expectedResult2, actualResult2);
} finally {
// clear the CollectionSink set for the restarted job
CollectionSink.clearElementsSet();
// clear any left overs from a possibly failed job
if (jobID != null && jobManager != null) {
Future<Object> jobRemovedFuture =
jobManager.ask(new TestingJobManagerMessages.NotifyWhenJobRemoved(jobID), timeout);
try {
Await.ready(jobRemovedFuture, timeout);
} catch (TimeoutException | InterruptedException ie) {
fail("Failed while cleaning up the cluster.");
}
}
}
}
public void deployToSlot(final SimpleSlot slot) throws JobException {
// sanity checks
if (slot == null) {
throw new NullPointerException();
}
if (!slot.isAlive()) {
throw new JobException("Target slot for deployment is not alive.");
}
// make sure exactly one deployment call happens from the correct state
// note: the transition from CREATED to DEPLOYING is for testing purposes only
ExecutionState previous = this.state;
if (previous == SCHEDULED || previous == CREATED) {
if (!transitionState(previous, DEPLOYING)) {
// race condition, someone else beat us to the deploying call.
// this should actually not happen and indicates a race somewhere else
throw new IllegalStateException("Cannot deploy task: Concurrent deployment call race.");
}
} else {
// vertex may have been cancelled, or it was already scheduled
throw new IllegalStateException(
"The vertex must be in CREATED or SCHEDULED state to be deployed. Found state "
+ previous);
}
try {
// good, we are allowed to deploy
if (!slot.setExecutedVertex(this)) {
throw new JobException("Could not assign the ExecutionVertex to the slot " + slot);
}
this.assignedResource = slot;
this.assignedResourceLocation = slot.getInstance().getInstanceConnectionInfo();
// race double check, did we fail/cancel and do we need to release the slot?
if (this.state != DEPLOYING) {
slot.releaseSlot();
return;
}
if (LOG.isInfoEnabled()) {
LOG.info(
String.format(
"Deploying %s (attempt #%d) to %s",
vertex.getSimpleName(),
attemptNumber,
slot.getInstance().getInstanceConnectionInfo().getHostname()));
}
final TaskDeploymentDescriptor deployment =
vertex.createDeploymentDescriptor(
attemptId, slot, operatorState, recoveryTimestamp, attemptNumber);
// register this execution at the execution graph, to receive call backs
vertex.getExecutionGraph().registerExecution(this);
final Instance instance = slot.getInstance();
final ActorGateway gateway = instance.getActorGateway();
final Future<Object> deployAction = gateway.ask(new SubmitTask(deployment), timeout);
deployAction.onComplete(
new OnComplete<Object>() {
@Override
public void onComplete(Throwable failure, Object success) throws Throwable {
if (failure != null) {
if (failure instanceof TimeoutException) {
String taskname =
deployment.getTaskInfo().getTaskNameWithSubtasks() + " (" + attemptId + ')';
markFailed(
new Exception(
"Cannot deploy task "
+ taskname
+ " - TaskManager ("
+ instance
+ ") not responding after a timeout of "
+ timeout,
failure));
} else {
markFailed(failure);
}
} else {
if (!(success.equals(Messages.getAcknowledge()))) {
markFailed(
new Exception(
"Failed to deploy the task to slot "
+ slot
+ ": Response was not of type Acknowledge"));
}
}
}
},
executionContext);
} catch (Throwable t) {
markFailed(t);
ExceptionUtils.rethrow(t);
}
}
