Examples with ActorSystem akka.actor.ActorSystem used on opensource projects

Example 56 with ActorSystem

use of akka.actor.ActorSystem in project flink by apache.

the class JobManagerHACheckpointRecoveryITCase method testCheckpointRecoveryFailure.

/**
	 * Tests that the JobManager logs failures during recovery properly.
	 *
	 * @see <a href="https://issues.apache.org/jira/browse/FLINK-3185">FLINK-3185</a>
	 */
@Test
@RetryOnFailure(times = 1)
public void testCheckpointRecoveryFailure() throws Exception {
    final Deadline testDeadline = TestTimeOut.fromNow();
    final String zooKeeperQuorum = ZooKeeper.getConnectString();
    final String fileStateBackendPath = FileStateBackendBasePath.getAbsoluteFile().toString();
    Configuration config = ZooKeeperTestUtils.createZooKeeperHAConfig(zooKeeperQuorum, fileStateBackendPath);
    config.setInteger(ConfigConstants.LOCAL_NUMBER_JOB_MANAGER, 2);
    JobManagerProcess[] jobManagerProcess = new JobManagerProcess[2];
    LeaderRetrievalService leaderRetrievalService = null;
    ActorSystem taskManagerSystem = null;
    ActorSystem testActorSystem = null;
    try {
        // Test actor system
        testActorSystem = AkkaUtils.createActorSystem(new Configuration(), new Some<>(new Tuple2<String, Object>("localhost", 0)));
        // The job managers
        jobManagerProcess[0] = new JobManagerProcess(0, config);
        jobManagerProcess[1] = new JobManagerProcess(1, config);
        jobManagerProcess[0].startProcess();
        jobManagerProcess[1].startProcess();
        // Leader listener
        TestingListener leaderListener = new TestingListener();
        leaderRetrievalService = ZooKeeperUtils.createLeaderRetrievalService(config);
        leaderRetrievalService.start(leaderListener);
        // The task manager
        taskManagerSystem = AkkaUtils.createActorSystem(config, Option.apply(new Tuple2<String, Object>("localhost", 0)));
        TaskManager.startTaskManagerComponentsAndActor(config, ResourceID.generate(), taskManagerSystem, "localhost", Option.<String>empty(), Option.<LeaderRetrievalService>empty(), false, TaskManager.class);
        // Get the leader
        leaderListener.waitForNewLeader(testDeadline.timeLeft().toMillis());
        String leaderAddress = leaderListener.getAddress();
        UUID leaderId = leaderListener.getLeaderSessionID();
        // Get the leader ref
        ActorRef leaderRef = AkkaUtils.getActorRef(leaderAddress, testActorSystem, testDeadline.timeLeft());
        ActorGateway leader = new AkkaActorGateway(leaderRef, leaderId);
        // Who's the boss?
        JobManagerProcess leadingJobManagerProcess;
        JobManagerProcess nonLeadingJobManagerProcess;
        if (jobManagerProcess[0].getJobManagerAkkaURL(testDeadline.timeLeft()).equals(leaderListener.getAddress())) {
            leadingJobManagerProcess = jobManagerProcess[0];
            nonLeadingJobManagerProcess = jobManagerProcess[1];
        } else {
            leadingJobManagerProcess = jobManagerProcess[1];
            nonLeadingJobManagerProcess = jobManagerProcess[0];
        }
        // Blocking JobGraph
        JobVertex blockingVertex = new JobVertex("Blocking vertex");
        blockingVertex.setInvokableClass(BlockingNoOpInvokable.class);
        JobGraph jobGraph = new JobGraph(blockingVertex);
        // Submit the job in detached mode
        leader.tell(new SubmitJob(jobGraph, ListeningBehaviour.DETACHED));
        // Wait for the job to be running
        JobManagerActorTestUtils.waitForJobStatus(jobGraph.getJobID(), JobStatus.RUNNING, leader, testDeadline.timeLeft());
        // Remove all files
        FileUtils.deleteDirectory(FileStateBackendBasePath);
        // Kill the leader
        leadingJobManagerProcess.destroy();
        // Verify that the job manager logs the failed recovery. We can not
        // do more at this point. :(
        boolean success = false;
        while (testDeadline.hasTimeLeft()) {
            String output = nonLeadingJobManagerProcess.getProcessOutput();
            if (output != null) {
                if (output.contains("Failed to recover job") && output.contains("java.io.FileNotFoundException")) {
                    success = true;
                    break;
                }
            } else {
                log.warn("No process output available.");
            }
            Thread.sleep(500);
        }
        assertTrue("Did not find expected output in logs.", success);
    } catch (Throwable t) {
        // Print early (in some situations the process logs get too big
        // for Travis and the root problem is not shown)
        t.printStackTrace();
        // In case of an error, print the job manager process logs.
        if (jobManagerProcess[0] != null) {
            jobManagerProcess[0].printProcessLog();
        }
        if (jobManagerProcess[1] != null) {
            jobManagerProcess[1].printProcessLog();
        }
        throw t;
    } finally {
        if (jobManagerProcess[0] != null) {
            jobManagerProcess[0].destroy();
        }
        if (jobManagerProcess[1] != null) {
            jobManagerProcess[1].destroy();
        }
        if (leaderRetrievalService != null) {
            leaderRetrievalService.stop();
        }
        if (taskManagerSystem != null) {
            taskManagerSystem.shutdown();
        }
        if (testActorSystem != null) {
            testActorSystem.shutdown();
        }
    }
}

Example 57 with ActorSystem

use of akka.actor.ActorSystem in project flink by apache.

the class JobManagerHAJobGraphRecoveryITCase method testClientNonDetachedListeningBehaviour.

/**
	 * Tests that clients receive updates after recovery by a new leader.
	 */
@Test
public void testClientNonDetachedListeningBehaviour() throws Exception {
    Configuration config = ZooKeeperTestUtils.createZooKeeperHAConfig(ZooKeeper.getConnectString(), FileStateBackendBasePath.getPath());
    // Test actor system
    ActorSystem testSystem = null;
    // JobManager setup. Start the job managers as separate processes in order to not run the
    // actors postStop, which cleans up all running jobs.
    JobManagerProcess[] jobManagerProcess = new JobManagerProcess[2];
    LeaderRetrievalService leaderRetrievalService = null;
    ActorSystem taskManagerSystem = null;
    try {
        final Deadline deadline = TestTimeOut.fromNow();
        // Test actor system
        testSystem = AkkaUtils.createActorSystem(new Configuration(), new Some<>(new Tuple2<String, Object>("localhost", 0)));
        // The job managers
        jobManagerProcess[0] = new JobManagerProcess(0, config);
        jobManagerProcess[1] = new JobManagerProcess(1, config);
        jobManagerProcess[0].startProcess();
        jobManagerProcess[1].startProcess();
        // Leader listener
        TestingListener leaderListener = new TestingListener();
        leaderRetrievalService = ZooKeeperUtils.createLeaderRetrievalService(config);
        leaderRetrievalService.start(leaderListener);
        // The task manager
        taskManagerSystem = AkkaUtils.createActorSystem(AkkaUtils.getDefaultAkkaConfig());
        TaskManager.startTaskManagerComponentsAndActor(config, ResourceID.generate(), taskManagerSystem, "localhost", Option.<String>empty(), Option.<LeaderRetrievalService>empty(), false, TaskManager.class);
        // Client test actor
        TestActorRef<RecordingTestClient> clientRef = TestActorRef.create(testSystem, Props.create(RecordingTestClient.class));
        JobGraph jobGraph = createBlockingJobGraph();
        {
            // Initial submission
            leaderListener.waitForNewLeader(deadline.timeLeft().toMillis());
            String leaderAddress = leaderListener.getAddress();
            UUID leaderId = leaderListener.getLeaderSessionID();
            // The client
            AkkaActorGateway client = new AkkaActorGateway(clientRef, leaderId);
            // Get the leader ref
            ActorRef leaderRef = AkkaUtils.getActorRef(leaderAddress, testSystem, deadline.timeLeft());
            ActorGateway leader = new AkkaActorGateway(leaderRef, leaderId);
            int numSlots = 0;
            while (numSlots == 0) {
                Future<?> slotsFuture = leader.ask(JobManagerMessages.getRequestTotalNumberOfSlots(), deadline.timeLeft());
                numSlots = (Integer) Await.result(slotsFuture, deadline.timeLeft());
            }
            // Submit the job in non-detached mode
            leader.tell(new SubmitJob(jobGraph, ListeningBehaviour.EXECUTION_RESULT_AND_STATE_CHANGES), client);
            JobManagerActorTestUtils.waitForJobStatus(jobGraph.getJobID(), JobStatus.RUNNING, leader, deadline.timeLeft());
        }
        // Who's the boss?
        JobManagerProcess leadingJobManagerProcess;
        if (jobManagerProcess[0].getJobManagerAkkaURL(deadline.timeLeft()).equals(leaderListener.getAddress())) {
            leadingJobManagerProcess = jobManagerProcess[0];
        } else {
            leadingJobManagerProcess = jobManagerProcess[1];
        }
        // Kill the leading job manager process
        leadingJobManagerProcess.destroy();
        {
            // Recovery by the standby JobManager
            leaderListener.waitForNewLeader(deadline.timeLeft().toMillis());
            String leaderAddress = leaderListener.getAddress();
            UUID leaderId = leaderListener.getLeaderSessionID();
            ActorRef leaderRef = AkkaUtils.getActorRef(leaderAddress, testSystem, deadline.timeLeft());
            ActorGateway leader = new AkkaActorGateway(leaderRef, leaderId);
            JobManagerActorTestUtils.waitForJobStatus(jobGraph.getJobID(), JobStatus.RUNNING, leader, deadline.timeLeft());
            // Cancel the job
            leader.tell(new JobManagerMessages.CancelJob(jobGraph.getJobID()));
        }
        // Wait for the execution result
        clientRef.underlyingActor().awaitJobResult(deadline.timeLeft().toMillis());
        int jobSubmitSuccessMessages = 0;
        for (Object msg : clientRef.underlyingActor().getMessages()) {
            if (msg instanceof JobManagerMessages.JobSubmitSuccess) {
                jobSubmitSuccessMessages++;
            }
        }
        // At least two submissions should be ack-ed (initial and recovery). This is quite
        // conservative, but it is still possible that these messages are overtaken by the
        // final message.
        assertEquals(2, jobSubmitSuccessMessages);
    } catch (Throwable t) {
        // Print early (in some situations the process logs get too big
        // for Travis and the root problem is not shown)
        t.printStackTrace();
        // In case of an error, print the job manager process logs.
        if (jobManagerProcess[0] != null) {
            jobManagerProcess[0].printProcessLog();
        }
        if (jobManagerProcess[1] != null) {
            jobManagerProcess[1].printProcessLog();
        }
        throw t;
    } finally {
        if (jobManagerProcess[0] != null) {
            jobManagerProcess[0].destroy();
        }
        if (jobManagerProcess[1] != null) {
            jobManagerProcess[1].destroy();
        }
        if (leaderRetrievalService != null) {
            leaderRetrievalService.stop();
        }
        if (taskManagerSystem != null) {
            taskManagerSystem.shutdown();
        }
        if (testSystem != null) {
            testSystem.shutdown();
        }
    }
}

Example 58 with ActorSystem

use of akka.actor.ActorSystem in project flink by apache.

the class JobManagerHAProcessFailureBatchRecoveryITCase method testJobManagerProcessFailure.

@Test
public void testJobManagerProcessFailure() throws Exception {
    // Config
    final int numberOfJobManagers = 2;
    final int numberOfTaskManagers = 2;
    final int numberOfSlotsPerTaskManager = 2;
    assertEquals(PARALLELISM, numberOfTaskManagers * numberOfSlotsPerTaskManager);
    // Setup
    // Test actor system
    ActorSystem testActorSystem;
    // Job managers
    final JobManagerProcess[] jmProcess = new JobManagerProcess[numberOfJobManagers];
    // Task managers
    final ActorSystem[] tmActorSystem = new ActorSystem[numberOfTaskManagers];
    // Leader election service
    LeaderRetrievalService leaderRetrievalService = null;
    // Coordination between the processes goes through a directory
    File coordinateTempDir = null;
    try {
        final Deadline deadline = TestTimeOut.fromNow();
        // Coordination directory
        coordinateTempDir = createTempDirectory();
        // Job Managers
        Configuration config = ZooKeeperTestUtils.createZooKeeperHAConfig(ZooKeeper.getConnectString(), FileStateBackendBasePath.getPath());
        // Start first process
        jmProcess[0] = new JobManagerProcess(0, config);
        jmProcess[0].startProcess();
        // Task manager configuration
        config.setInteger(ConfigConstants.TASK_MANAGER_MEMORY_SIZE_KEY, 4);
        config.setInteger(ConfigConstants.TASK_MANAGER_NETWORK_NUM_BUFFERS_KEY, 100);
        config.setInteger(ConfigConstants.TASK_MANAGER_NUM_TASK_SLOTS, 2);
        // Start the task manager process
        for (int i = 0; i < numberOfTaskManagers; i++) {
            tmActorSystem[i] = AkkaUtils.createActorSystem(AkkaUtils.getDefaultAkkaConfig());
            TaskManager.startTaskManagerComponentsAndActor(config, ResourceID.generate(), tmActorSystem[i], "localhost", Option.<String>empty(), Option.<LeaderRetrievalService>empty(), false, TaskManager.class);
        }
        // Test actor system
        testActorSystem = AkkaUtils.createActorSystem(AkkaUtils.getDefaultAkkaConfig());
        jmProcess[0].getActorRef(testActorSystem, deadline.timeLeft());
        // Leader listener
        TestingListener leaderListener = new TestingListener();
        leaderRetrievalService = ZooKeeperUtils.createLeaderRetrievalService(config);
        leaderRetrievalService.start(leaderListener);
        // Initial submission
        leaderListener.waitForNewLeader(deadline.timeLeft().toMillis());
        String leaderAddress = leaderListener.getAddress();
        UUID leaderId = leaderListener.getLeaderSessionID();
        // Get the leader ref
        ActorRef leaderRef = AkkaUtils.getActorRef(leaderAddress, testActorSystem, deadline.timeLeft());
        ActorGateway leaderGateway = new AkkaActorGateway(leaderRef, leaderId);
        // Wait for all task managers to connect to the leading job manager
        JobManagerActorTestUtils.waitForTaskManagers(numberOfTaskManagers, leaderGateway, deadline.timeLeft());
        final File coordinateDirClosure = coordinateTempDir;
        final Throwable[] errorRef = new Throwable[1];
        // we trigger program execution in a separate thread
        Thread programTrigger = new Thread("Program Trigger") {

            @Override
            public void run() {
                try {
                    testJobManagerFailure(ZooKeeper.getConnectString(), coordinateDirClosure);
                } catch (Throwable t) {
                    t.printStackTrace();
                    errorRef[0] = t;
                }
            }
        };
        //start the test program
        programTrigger.start();
        // wait until all marker files are in place, indicating that all tasks have started
        AbstractTaskManagerProcessFailureRecoveryTest.waitForMarkerFiles(coordinateTempDir, READY_MARKER_FILE_PREFIX, PARALLELISM, deadline.timeLeft().toMillis());
        // Kill one of the job managers and trigger recovery
        jmProcess[0].destroy();
        jmProcess[1] = new JobManagerProcess(1, config);
        jmProcess[1].startProcess();
        jmProcess[1].getActorRef(testActorSystem, deadline.timeLeft());
        // we create the marker file which signals the program functions tasks that they can complete
        AbstractTaskManagerProcessFailureRecoveryTest.touchFile(new File(coordinateTempDir, PROCEED_MARKER_FILE));
        programTrigger.join(deadline.timeLeft().toMillis());
        // We wait for the finish marker file. We don't wait for the program trigger, because
        // we submit in detached mode.
        AbstractTaskManagerProcessFailureRecoveryTest.waitForMarkerFiles(coordinateTempDir, FINISH_MARKER_FILE_PREFIX, 1, deadline.timeLeft().toMillis());
        // check that the program really finished
        assertFalse("The program did not finish in time", programTrigger.isAlive());
        // check whether the program encountered an error
        if (errorRef[0] != null) {
            Throwable error = errorRef[0];
            error.printStackTrace();
            fail("The program encountered a " + error.getClass().getSimpleName() + " : " + error.getMessage());
        }
    } catch (Throwable t) {
        // Print early (in some situations the process logs get too big
        // for Travis and the root problem is not shown)
        t.printStackTrace();
        for (JobManagerProcess p : jmProcess) {
            if (p != null) {
                p.printProcessLog();
            }
        }
        throw t;
    } finally {
        for (int i = 0; i < numberOfTaskManagers; i++) {
            if (tmActorSystem[i] != null) {
                tmActorSystem[i].shutdown();
            }
        }
        if (leaderRetrievalService != null) {
            leaderRetrievalService.stop();
        }
        for (JobManagerProcess jmProces : jmProcess) {
            if (jmProces != null) {
                jmProces.destroy();
            }
        }
        // Delete coordination directory
        if (coordinateTempDir != null) {
            try {
                FileUtils.deleteDirectory(coordinateTempDir);
            } catch (Throwable ignored) {
            }
        }
    }
}

Example 59 with ActorSystem

use of akka.actor.ActorSystem in project flink by apache.

the class ChaosMonkeyITCase method testChaosMonkey.

@Test
public void testChaosMonkey() throws Exception {
    // Test config
    final int numberOfJobManagers = 3;
    final int numberOfTaskManagers = 3;
    final int numberOfSlotsPerTaskManager = 2;
    // The final count each source is counting to: 1...n
    final int n = 5000;
    // Parallelism for the program
    final int parallelism = numberOfTaskManagers * numberOfSlotsPerTaskManager;
    // The test should not run longer than this
    final FiniteDuration testDuration = new FiniteDuration(10, TimeUnit.MINUTES);
    // Every x seconds a random job or task manager is killed
    //
    // The job will will be running for $killEvery seconds and then a random Job/TaskManager
    // will be killed. On recovery (which takes some time to bring up the new process etc.),
    // this test will wait for task managers to reconnect before starting the next count down.
    // Therefore the delay between retries is not important in this setup.
    final FiniteDuration killEvery = new FiniteDuration(5, TimeUnit.SECONDS);
    // Trigger a checkpoint every
    final int checkpointingIntervalMs = 1000;
    // Total number of kills
    final int totalNumberOfKills = 10;
    // -----------------------------------------------------------------------------------------
    // Setup
    Configuration config = ZooKeeperTestUtils.createZooKeeperHAConfig(ZooKeeper.getConnectString(), FileStateBackendBasePath.toURI().toString());
    // Akka and restart timeouts
    config.setString(ConfigConstants.AKKA_WATCH_HEARTBEAT_INTERVAL, "1000 ms");
    config.setString(ConfigConstants.AKKA_WATCH_HEARTBEAT_PAUSE, "6 s");
    config.setInteger(ConfigConstants.AKKA_WATCH_THRESHOLD, 9);
    if (checkpointingIntervalMs >= killEvery.toMillis()) {
        throw new IllegalArgumentException("Relax! You want to kill processes every " + killEvery + ", but the checkpointing interval is " + checkpointingIntervalMs / 1000 + " seconds. Either decrease the interval or " + "increase the kill interval. Otherwise, the program will not complete any " + "checkpoint.");
    }
    // Task manager
    config.setInteger(ConfigConstants.TASK_MANAGER_NUM_TASK_SLOTS, numberOfSlotsPerTaskManager);
    ActorSystem testActorSystem = null;
    LeaderRetrievalService leaderRetrievalService = null;
    List<JobManagerProcess> jobManagerProcesses = new ArrayList<>();
    List<TaskManagerProcess> taskManagerProcesses = new ArrayList<>();
    try {
        // Initial state
        for (int i = 0; i < numberOfJobManagers; i++) {
            jobManagerProcesses.add(createAndStartJobManagerProcess(config));
        }
        for (int i = 0; i < numberOfTaskManagers; i++) {
            taskManagerProcesses.add(createAndStartTaskManagerProcess(config));
        }
        testActorSystem = AkkaUtils.createDefaultActorSystem();
        // Leader listener
        leaderRetrievalService = ZooKeeperUtils.createLeaderRetrievalService(config);
        TestingListener leaderListener = new TestingListener();
        leaderRetrievalService.start(leaderListener);
        Deadline deadline = testDuration.fromNow();
        // Wait for the new leader
        int leaderIndex = waitForNewLeader(leaderListener, jobManagerProcesses, deadline.timeLeft());
        // Wait for the task managers to connect
        waitForTaskManagers(numberOfTaskManagers, jobManagerProcesses.get(leaderIndex), testActorSystem, deadline.timeLeft());
        // The job
        JobGraph jobGraph = createJobGraph(n, CheckpointCompletedCoordination.getPath(), ProceedCoordination.getPath(), parallelism, checkpointingIntervalMs);
        LOG.info("Submitting job {}", jobGraph.getJobID());
        submitJobGraph(jobGraph, jobManagerProcesses.get(leaderIndex), leaderListener, testActorSystem, deadline.timeLeft());
        LOG.info("Waiting for a checkpoint to complete before kicking off chaos");
        // Wait for a checkpoint to complete
        TestJvmProcess.waitForMarkerFiles(FileStateBackendBasePath, COMPLETED_PREFIX, parallelism, deadline.timeLeft().toMillis());
        LOG.info("Checkpoint completed... ready for chaos");
        int currentKillNumber = 1;
        int currentJobManagerKills = 0;
        int currentTaskManagerKills = 0;
        for (int i = 0; i < totalNumberOfKills; i++) {
            LOG.info("Waiting for {} before next kill ({}/{})", killEvery, currentKillNumber++, totalNumberOfKills);
            Thread.sleep(killEvery.toMillis());
            LOG.info("Checking job status...");
            JobStatus jobStatus = requestJobStatus(jobGraph.getJobID(), jobManagerProcesses.get(leaderIndex), testActorSystem, deadline.timeLeft());
            if (jobStatus != JobStatus.RUNNING && jobStatus != JobStatus.FINISHED) {
                // Wait for it to run
                LOG.info("Waiting for job status {}", JobStatus.RUNNING);
                waitForJobRunning(jobGraph.getJobID(), jobManagerProcesses.get(leaderIndex), testActorSystem, deadline.timeLeft());
            } else if (jobStatus == JobStatus.FINISHED) {
                // Early finish
                LOG.info("Job finished");
                return;
            } else {
                LOG.info("Job status is {}", jobStatus);
            }
            if (rand.nextBoolean()) {
                LOG.info("Killing the leading JobManager");
                JobManagerProcess newJobManager = createAndStartJobManagerProcess(config);
                JobManagerProcess leader = jobManagerProcesses.remove(leaderIndex);
                leader.destroy();
                currentJobManagerKills++;
                LOG.info("Killed {}", leader);
                // Make sure to add the new job manager before looking for a new leader
                jobManagerProcesses.add(newJobManager);
                // Wait for the new leader
                leaderIndex = waitForNewLeader(leaderListener, jobManagerProcesses, deadline.timeLeft());
                // Wait for the task managers to connect
                waitForTaskManagers(numberOfTaskManagers, jobManagerProcesses.get(leaderIndex), testActorSystem, deadline.timeLeft());
            } else {
                LOG.info("Killing a random TaskManager");
                TaskManagerProcess newTaskManager = createAndStartTaskManagerProcess(config);
                // Wait for this new task manager to be connected
                waitForTaskManagers(numberOfTaskManagers + 1, jobManagerProcesses.get(leaderIndex), testActorSystem, deadline.timeLeft());
                // Now it's safe to kill a process
                int next = rand.nextInt(numberOfTaskManagers);
                TaskManagerProcess taskManager = taskManagerProcesses.remove(next);
                LOG.info("{} has been chosen. Killing process...", taskManager);
                taskManager.destroy();
                currentTaskManagerKills++;
                // Add the new task manager after killing an old one
                taskManagerProcesses.add(newTaskManager);
            }
        }
        LOG.info("Chaos is over. Total kills: {} ({} job manager + {} task managers). " + "Checking job status...", totalNumberOfKills, currentJobManagerKills, currentTaskManagerKills);
        // Signal the job to speed up (if it is not done yet)
        TestJvmProcess.touchFile(ProceedCoordination);
        // Wait for the job to finish
        LOG.info("Waiting for job status {}", JobStatus.FINISHED);
        waitForJobFinished(jobGraph.getJobID(), jobManagerProcesses.get(leaderIndex), testActorSystem, deadline.timeLeft());
        LOG.info("Job finished");
        LOG.info("Waiting for job removal");
        waitForJobRemoved(jobGraph.getJobID(), jobManagerProcesses.get(leaderIndex), testActorSystem, deadline.timeLeft());
        LOG.info("Job removed");
        LOG.info("Checking clean recovery state...");
        checkCleanRecoveryState(config);
        LOG.info("Recovery state clean");
    } catch (Throwable t) {
        // Print early (in some situations the process logs get too big
        // for Travis and the root problem is not shown)
        t.printStackTrace();
        System.out.println("#################################################");
        System.out.println(" TASK MANAGERS");
        System.out.println("#################################################");
        for (TaskManagerProcess taskManagerProcess : taskManagerProcesses) {
            taskManagerProcess.printProcessLog();
        }
        System.out.println("#################################################");
        System.out.println(" JOB MANAGERS");
        System.out.println("#################################################");
        for (JobManagerProcess jobManagerProcess : jobManagerProcesses) {
            jobManagerProcess.printProcessLog();
        }
        throw t;
    } finally {
        for (JobManagerProcess jobManagerProcess : jobManagerProcesses) {
            if (jobManagerProcess != null) {
                jobManagerProcess.destroy();
            }
        }
        if (leaderRetrievalService != null) {
            leaderRetrievalService.stop();
        }
        if (testActorSystem != null) {
            testActorSystem.shutdown();
        }
    }
}

Example 60 with ActorSystem

use of akka.actor.ActorSystem in project flink by apache.

the class ZooKeeperLeaderElectionITCase method testJobExecutionOnClusterWithLeaderReelection.

/**
	 * Tests that a job can be executed after a new leader has been elected. For all except for the
	 * last leader, the job is blocking. The JobManager will be terminated while executing the
	 * blocking job. Once only one JobManager is left, it is checked that a non-blocking can be
	 * successfully executed.
	 */
@Test
public void testJobExecutionOnClusterWithLeaderReelection() throws Exception {
    int numJMs = 10;
    int numTMs = 2;
    int numSlotsPerTM = 3;
    int parallelism = numTMs * numSlotsPerTM;
    File rootFolder = tempFolder.getRoot();
    Configuration configuration = ZooKeeperTestUtils.createZooKeeperHAConfig(zkServer.getConnectString(), rootFolder.getPath());
    configuration.setInteger(ConfigConstants.LOCAL_NUMBER_JOB_MANAGER, numJMs);
    configuration.setInteger(ConfigConstants.LOCAL_NUMBER_TASK_MANAGER, numTMs);
    configuration.setInteger(ConfigConstants.TASK_MANAGER_NUM_TASK_SLOTS, numSlotsPerTM);
    // we "effectively" disable the automatic RecoverAllJobs message and sent it manually to make
    // sure that all TMs have registered to the JM prior to issueing the RecoverAllJobs message
    configuration.setString(ConfigConstants.AKKA_ASK_TIMEOUT, AkkaUtils.INF_TIMEOUT().toString());
    Tasks.BlockingOnceReceiver$.MODULE$.blocking_$eq(true);
    JobVertex sender = new JobVertex("sender");
    JobVertex receiver = new JobVertex("receiver");
    sender.setInvokableClass(Tasks.Sender.class);
    receiver.setInvokableClass(Tasks.BlockingOnceReceiver.class);
    sender.setParallelism(parallelism);
    receiver.setParallelism(parallelism);
    receiver.connectNewDataSetAsInput(sender, DistributionPattern.POINTWISE, ResultPartitionType.PIPELINED);
    SlotSharingGroup slotSharingGroup = new SlotSharingGroup();
    sender.setSlotSharingGroup(slotSharingGroup);
    receiver.setSlotSharingGroup(slotSharingGroup);
    final JobGraph graph = new JobGraph("Blocking test job", sender, receiver);
    final TestingCluster cluster = new TestingCluster(configuration);
    ActorSystem clientActorSystem = null;
    Thread thread = null;
    JobSubmitterRunnable jobSubmission = null;
    try {
        cluster.start();
        clientActorSystem = cluster.startJobClientActorSystem(graph.getJobID());
        final ActorSystem clientAS = clientActorSystem;
        jobSubmission = new JobSubmitterRunnable(clientAS, cluster, graph);
        thread = new Thread(jobSubmission);
        thread.start();
        Deadline deadline = timeout.$times(3).fromNow();
        // Kill all JobManager except for two
        for (int i = 0; i < numJMs; i++) {
            ActorGateway jm = cluster.getLeaderGateway(deadline.timeLeft());
            cluster.waitForTaskManagersToBeRegisteredAtJobManager(jm.actor());
            // recover all jobs, sent manually
            log.info("Sent recover all jobs manually to job manager {}.", jm.path());
            jm.tell(JobManagerMessages.getRecoverAllJobs());
            if (i < numJMs - 1) {
                Future<Object> future = jm.ask(new WaitForAllVerticesToBeRunningOrFinished(graph.getJobID()), deadline.timeLeft());
                Await.ready(future, deadline.timeLeft());
                cluster.clearLeader();
                if (i == numJMs - 2) {
                    Tasks.BlockingOnceReceiver$.MODULE$.blocking_$eq(false);
                }
                log.info("Kill job manager {}.", jm.path());
                jm.tell(TestingJobManagerMessages.getDisablePostStop());
                jm.tell(Kill.getInstance());
            }
        }
        log.info("Waiting for submitter thread to terminate.");
        thread.join(deadline.timeLeft().toMillis());
        log.info("Submitter thread has terminated.");
        if (thread.isAlive()) {
            fail("The job submission thread did not stop (meaning it did not succeeded in" + "executing the test job.");
        }
        Await.result(jobSubmission.resultPromise.future(), deadline.timeLeft());
    } finally {
        if (clientActorSystem != null) {
            cluster.shutdownJobClientActorSystem(clientActorSystem);
        }
        if (thread != null && thread.isAlive()) {
            jobSubmission.finished = true;
        }
        cluster.stop();
    }
}