Examples with WaitForAllVerticesToBeRunningOrFinished org.apache.flink.runtime.testingUtils.TestingJobManagerMessages.WaitForAllVerticesToBeRunningOrFinished used on opensource projects

Example 6 with WaitForAllVerticesToBeRunningOrFinished

use of org.apache.flink.runtime.testingUtils.TestingJobManagerMessages.WaitForAllVerticesToBeRunningOrFinished in project flink by apache.

the class LeaderChangeStateCleanupTest method testReelectionOfSameJobManager.

/**
	 * 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(10, TimeUnit.SECONDS);
    cluster.grantLeadership(0, leaderSessionID);
    cluster.notifyRetrievalListeners(0, leaderSessionID);
    cluster.waitForTaskManagersToBeRegistered(timeout);
    // 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);
    LOG.info("Make JM(0) again the leader. This should first revoke the leadership.");
    // make JM(0) again the leader --> this implies first a leadership revocation
    cluster.grantLeadership(0, newLeaderSessionID);
    Await.ready(jobRemoval, timeout);
    LOG.info("Job removed.");
    // 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
    }
    LOG.info("Notify TMs about the new (old) leader.");
    // notify the TMs about the new (old) leader
    cluster.notifyRetrievalListeners(0, newLeaderSessionID);
    cluster.waitForTaskManagersToBeRegistered(timeout);
    ActorGateway leaderGateway = cluster.getLeaderGateway(timeout);
    // try to resubmit now the non-blocking job, it should complete successfully
    Tasks.BlockingOnceReceiver$.MODULE$.blocking_$eq(false);
    cluster.submitJobAndWait(job, false, timeout, new TestingLeaderRetrievalService(leaderGateway.path(), leaderGateway.leaderSessionID()));
}

Example 7 with WaitForAllVerticesToBeRunningOrFinished

use of org.apache.flink.runtime.testingUtils.TestingJobManagerMessages.WaitForAllVerticesToBeRunningOrFinished in project flink by apache.

the class LeaderChangeStateCleanupTest method testStateCleanupAfterNewLeaderElectionAndListenerNotification.

/**
	 * 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(timeout);
    // 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(timeout);
    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, new TestingLeaderRetrievalService(jm2.path(), jm2.leaderSessionID()));
}

Example 8 with WaitForAllVerticesToBeRunningOrFinished

use of org.apache.flink.runtime.testingUtils.TestingJobManagerMessages.WaitForAllVerticesToBeRunningOrFinished 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();
    }
}