Examples with BlobServer org.apache.flink.runtime.blob.BlobServer used on opensource projects

Example 1 with BlobServer

use of org.apache.flink.runtime.blob.BlobServer 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());
        }
    }
}

Example 2 with BlobServer

use of org.apache.flink.runtime.blob.BlobServer in project flink by apache.

the class JobManagerLeaderElectionTest method createJobManagerProps.

private Props createJobManagerProps(Configuration configuration) throws Exception {
    LeaderElectionService leaderElectionService;
    if (HighAvailabilityMode.fromConfig(configuration) == HighAvailabilityMode.NONE) {
        leaderElectionService = new StandaloneLeaderElectionService();
    } else {
        CuratorFramework client = ZooKeeperUtils.startCuratorFramework(configuration);
        leaderElectionService = ZooKeeperUtils.createLeaderElectionService(client, configuration);
    }
    // We don't need recovery in this test
    SubmittedJobGraphStore submittedJobGraphStore = new StandaloneSubmittedJobGraphStore();
    CheckpointRecoveryFactory checkpointRecoveryFactory = new StandaloneCheckpointRecoveryFactory();
    return Props.create(TestingJobManager.class, configuration, TestingUtils.defaultExecutor(), TestingUtils.defaultExecutor(), new InstanceManager(), new Scheduler(TestingUtils.defaultExecutionContext()), new BlobLibraryCacheManager(new BlobServer(configuration), 10L), ActorRef.noSender(), new NoRestartStrategy.NoRestartStrategyFactory(), AkkaUtils.getDefaultTimeoutAsFiniteDuration(), leaderElectionService, submittedJobGraphStore, checkpointRecoveryFactory, AkkaUtils.getDefaultTimeoutAsFiniteDuration(), Option.apply(null));
}

Example 3 with BlobServer

use of org.apache.flink.runtime.blob.BlobServer in project flink by apache.

the class MiniDispatcherTest method setupClass.

@BeforeClass
public static void setupClass() throws IOException {
    jobGraph = JobGraphTestUtils.singleNoOpJobGraph();
    executionGraphInfo = new ExecutionGraphInfo(new ArchivedExecutionGraphBuilder().setJobID(jobGraph.getJobID()).setState(JobStatus.FINISHED).build());
    rpcService = new TestingRpcService();
    configuration = new Configuration();
    blobServer = new BlobServer(configuration, temporaryFolder.newFolder(), new VoidBlobStore());
}

Example 4 with BlobServer

use of org.apache.flink.runtime.blob.BlobServer in project flink by apache.

the class BlobLibraryCacheManagerTest method testLibraryCacheManagerCleanup.

/**
 * Tests that the {@link BlobLibraryCacheManager} cleans up after all class loader leases for a
 * single job a closed.
 */
@Test
public void testLibraryCacheManagerCleanup() throws Exception {
    JobID jobId = new JobID();
    List<PermanentBlobKey> keys = new ArrayList<>();
    BlobServer server = null;
    PermanentBlobCache cache = null;
    BlobLibraryCacheManager libCache = null;
    final byte[] buf = new byte[128];
    try {
        Configuration config = new Configuration();
        config.setLong(BlobServerOptions.CLEANUP_INTERVAL, 1L);
        server = new BlobServer(config, temporaryFolder.newFolder(), new VoidBlobStore());
        server.start();
        InetSocketAddress serverAddress = new InetSocketAddress("localhost", server.getPort());
        cache = new PermanentBlobCache(config, temporaryFolder.newFolder(), new VoidBlobStore(), serverAddress);
        keys.add(server.putPermanent(jobId, buf));
        buf[0] += 1;
        keys.add(server.putPermanent(jobId, buf));
        libCache = createBlobLibraryCacheManager(cache);
        cache.registerJob(jobId);
        assertEquals(0, libCache.getNumberOfManagedJobs());
        assertEquals(0, libCache.getNumberOfReferenceHolders(jobId));
        checkFileCountForJob(2, jobId, server);
        checkFileCountForJob(0, jobId, cache);
        final LibraryCacheManager.ClassLoaderLease classLoaderLease1 = libCache.registerClassLoaderLease(jobId);
        UserCodeClassLoader classLoader1 = classLoaderLease1.getOrResolveClassLoader(keys, Collections.emptyList());
        assertEquals(1, libCache.getNumberOfManagedJobs());
        assertEquals(1, libCache.getNumberOfReferenceHolders(jobId));
        assertEquals(2, checkFilesExist(jobId, keys, cache, true));
        checkFileCountForJob(2, jobId, server);
        checkFileCountForJob(2, jobId, cache);
        final LibraryCacheManager.ClassLoaderLease classLoaderLease2 = libCache.registerClassLoaderLease(jobId);
        final UserCodeClassLoader classLoader2 = classLoaderLease2.getOrResolveClassLoader(keys, Collections.emptyList());
        assertThat(classLoader1, sameInstance(classLoader2));
        try {
            classLoaderLease1.getOrResolveClassLoader(Collections.emptyList(), Collections.emptyList());
            fail("Should fail with an IllegalStateException");
        } catch (IllegalStateException e) {
        // that's what we want
        }
        try {
            classLoaderLease1.getOrResolveClassLoader(keys, Collections.singletonList(new URL("file:///tmp/does-not-exist")));
            fail("Should fail with an IllegalStateException");
        } catch (IllegalStateException e) {
        // that's what we want
        }
        assertEquals(1, libCache.getNumberOfManagedJobs());
        assertEquals(2, libCache.getNumberOfReferenceHolders(jobId));
        assertEquals(2, checkFilesExist(jobId, keys, cache, true));
        checkFileCountForJob(2, jobId, server);
        checkFileCountForJob(2, jobId, cache);
        classLoaderLease1.release();
        assertEquals(1, libCache.getNumberOfManagedJobs());
        assertEquals(1, libCache.getNumberOfReferenceHolders(jobId));
        assertEquals(2, checkFilesExist(jobId, keys, cache, true));
        checkFileCountForJob(2, jobId, server);
        checkFileCountForJob(2, jobId, cache);
        classLoaderLease2.release();
        assertEquals(0, libCache.getNumberOfManagedJobs());
        assertEquals(0, libCache.getNumberOfReferenceHolders(jobId));
        assertEquals(2, checkFilesExist(jobId, keys, cache, true));
        checkFileCountForJob(2, jobId, server);
        checkFileCountForJob(2, jobId, cache);
    // only PermanentBlobCache#releaseJob() calls clean up files (tested in
    // BlobCacheCleanupTest etc.
    } finally {
        if (libCache != null) {
            libCache.shutdown();
        }
        // should have been closed by the libraryCacheManager, but just in case
        if (cache != null) {
            cache.close();
        }
        if (server != null) {
            server.close();
        }
    }
}

Example 5 with BlobServer

use of org.apache.flink.runtime.blob.BlobServer in project flink by apache.

the class BlobLibraryCacheManagerTest method testRegisterAndDownload.

@Test
public void testRegisterAndDownload() throws IOException {
    // setWritable doesn't work on Windows.
    assumeTrue(!OperatingSystem.isWindows());
    JobID jobId = new JobID();
    BlobServer server = null;
    PermanentBlobCache cache = null;
    BlobLibraryCacheManager libCache = null;
    File cacheDir = null;
    try {
        // create the blob transfer services
        Configuration config = new Configuration();
        config.setLong(BlobServerOptions.CLEANUP_INTERVAL, 1_000_000L);
        server = new BlobServer(config, temporaryFolder.newFolder(), new VoidBlobStore());
        server.start();
        InetSocketAddress serverAddress = new InetSocketAddress("localhost", server.getPort());
        cache = new PermanentBlobCache(config, temporaryFolder.newFolder(), new VoidBlobStore(), serverAddress);
        // upload some meaningless data to the server
        PermanentBlobKey dataKey1 = server.putPermanent(jobId, new byte[] { 1, 2, 3, 4, 5, 6, 7, 8 });
        PermanentBlobKey dataKey2 = server.putPermanent(jobId, new byte[] { 11, 12, 13, 14, 15, 16, 17, 18 });
        libCache = createBlobLibraryCacheManager(cache);
        assertEquals(0, libCache.getNumberOfManagedJobs());
        checkFileCountForJob(2, jobId, server);
        checkFileCountForJob(0, jobId, cache);
        // first try to access a non-existing entry
        assertEquals(0, libCache.getNumberOfReferenceHolders(new JobID()));
        // register some BLOBs as libraries
        {
            Collection<PermanentBlobKey> keys = Collections.singleton(dataKey1);
            cache.registerJob(jobId);
            final LibraryCacheManager.ClassLoaderLease classLoaderLease1 = libCache.registerClassLoaderLease(jobId);
            final UserCodeClassLoader classLoader1 = classLoaderLease1.getOrResolveClassLoader(keys, Collections.emptyList());
            assertEquals(1, libCache.getNumberOfManagedJobs());
            assertEquals(1, libCache.getNumberOfReferenceHolders(jobId));
            assertEquals(1, checkFilesExist(jobId, keys, cache, true));
            checkFileCountForJob(2, jobId, server);
            checkFileCountForJob(1, jobId, cache);
            final LibraryCacheManager.ClassLoaderLease classLoaderLease2 = libCache.registerClassLoaderLease(jobId);
            final UserCodeClassLoader classLoader2 = classLoaderLease2.getOrResolveClassLoader(keys, Collections.emptyList());
            assertThat(classLoader1, sameInstance(classLoader2));
            assertEquals(1, libCache.getNumberOfManagedJobs());
            assertEquals(2, libCache.getNumberOfReferenceHolders(jobId));
            assertEquals(1, checkFilesExist(jobId, keys, cache, true));
            checkFileCountForJob(2, jobId, server);
            checkFileCountForJob(1, jobId, cache);
            // un-register the job
            classLoaderLease1.release();
            // still one task
            assertEquals(1, libCache.getNumberOfManagedJobs());
            assertEquals(1, libCache.getNumberOfReferenceHolders(jobId));
            assertEquals(1, checkFilesExist(jobId, keys, cache, true));
            checkFileCountForJob(2, jobId, server);
            checkFileCountForJob(1, jobId, cache);
            // unregister the task registration
            classLoaderLease2.release();
            assertEquals(0, libCache.getNumberOfManagedJobs());
            assertEquals(0, libCache.getNumberOfReferenceHolders(jobId));
            // changing the libCache registration does not influence the BLOB stores...
            checkFileCountForJob(2, jobId, server);
            checkFileCountForJob(1, jobId, cache);
            cache.releaseJob(jobId);
            // library is still cached (but not associated with job any more)
            checkFileCountForJob(2, jobId, server);
            checkFileCountForJob(1, jobId, cache);
        }
        // see BlobUtils for the directory layout
        cacheDir = cache.getStorageLocation(jobId, new PermanentBlobKey()).getParentFile();
        assertTrue(cacheDir.exists());
        // make sure no further blobs can be downloaded by removing the write
        // permissions from the directory
        assertTrue("Could not remove write permissions from cache directory", cacheDir.setWritable(false, false));
        // since we cannot download this library any more, this call should fail
        try {
            cache.registerJob(jobId);
            final LibraryCacheManager.ClassLoaderLease classLoaderLease = libCache.registerClassLoaderLease(jobId);
            classLoaderLease.getOrResolveClassLoader(Collections.singleton(dataKey2), Collections.emptyList());
            fail("This should fail with an IOException");
        } catch (IOException e) {
            // splendid!
            cache.releaseJob(jobId);
        }
    } finally {
        if (cacheDir != null) {
            if (!cacheDir.setWritable(true, false)) {
                System.err.println("Could not re-add write permissions to cache directory.");
            }
        }
        if (cache != null) {
            cache.close();
        }
        if (libCache != null) {
            libCache.shutdown();
        }
        if (server != null) {
            server.close();
        }
    }
}