Examples with TaskExecutorPartitionTracker org.apache.flink.runtime.io.network.partition.TaskExecutorPartitionTracker used on opensource projects

Example 1 with TaskExecutorPartitionTracker

use of org.apache.flink.runtime.io.network.partition.TaskExecutorPartitionTracker in project flink by apache.

the class TaskExecutorPartitionLifecycleTest method internalTestPartitionRelease.

private void internalTestPartitionRelease(TaskExecutorPartitionTracker partitionTracker, ShuffleEnvironment<?, ?> shuffleEnvironment, CompletableFuture<ResultPartitionID> startTrackingFuture, TestAction testAction) throws Exception {
    final ResultPartitionDeploymentDescriptor taskResultPartitionDescriptor = PartitionTestUtils.createPartitionDeploymentDescriptor(ResultPartitionType.BLOCKING);
    final ExecutionAttemptID eid1 = taskResultPartitionDescriptor.getShuffleDescriptor().getResultPartitionID().getProducerId();
    final TaskDeploymentDescriptor taskDeploymentDescriptor = TaskExecutorSubmissionTest.createTaskDeploymentDescriptor(jobId, "job", eid1, new SerializedValue<>(new ExecutionConfig()), "Sender", 1, 0, 1, 0, new Configuration(), new Configuration(), TestingInvokable.class.getName(), Collections.singletonList(taskResultPartitionDescriptor), Collections.emptyList(), Collections.emptyList(), Collections.emptyList());
    final TaskSlotTable<Task> taskSlotTable = createTaskSlotTable();
    final TaskExecutorLocalStateStoresManager localStateStoresManager = new TaskExecutorLocalStateStoresManager(false, Reference.owned(new File[] { tmp.newFolder() }), Executors.directExecutor());
    final TaskManagerServices taskManagerServices = new TaskManagerServicesBuilder().setTaskSlotTable(taskSlotTable).setTaskStateManager(localStateStoresManager).setShuffleEnvironment(shuffleEnvironment).build();
    final CompletableFuture<Void> taskFinishedFuture = new CompletableFuture<>();
    final OneShotLatch slotOfferedLatch = new OneShotLatch();
    final TestingJobMasterGateway jobMasterGateway = new TestingJobMasterGatewayBuilder().setRegisterTaskManagerFunction((ignoredJobId, ignoredTaskManagerRegistrationInformation) -> CompletableFuture.completedFuture(new JMTMRegistrationSuccess(ResourceID.generate()))).setOfferSlotsFunction((resourceID, slotOffers) -> {
        slotOfferedLatch.trigger();
        return CompletableFuture.completedFuture(slotOffers);
    }).setUpdateTaskExecutionStateFunction(taskExecutionState -> {
        if (taskExecutionState.getExecutionState() == ExecutionState.FINISHED) {
            taskFinishedFuture.complete(null);
        }
        return CompletableFuture.completedFuture(Acknowledge.get());
    }).build();
    final TestingTaskExecutor taskExecutor = createTestingTaskExecutor(taskManagerServices, partitionTracker);
    final CompletableFuture<SlotReport> initialSlotReportFuture = new CompletableFuture<>();
    final TestingResourceManagerGateway testingResourceManagerGateway = new TestingResourceManagerGateway();
    testingResourceManagerGateway.setSendSlotReportFunction(resourceIDInstanceIDSlotReportTuple3 -> {
        initialSlotReportFuture.complete(resourceIDInstanceIDSlotReportTuple3.f2);
        return CompletableFuture.completedFuture(Acknowledge.get());
    });
    testingResourceManagerGateway.setRegisterTaskExecutorFunction(input -> CompletableFuture.completedFuture(new TaskExecutorRegistrationSuccess(new InstanceID(), testingResourceManagerGateway.getOwnResourceId(), new ClusterInformation("blobServerHost", 55555))));
    try {
        taskExecutor.start();
        taskExecutor.waitUntilStarted();
        final TaskExecutorGateway taskExecutorGateway = taskExecutor.getSelfGateway(TaskExecutorGateway.class);
        final String jobMasterAddress = "jm";
        rpc.registerGateway(jobMasterAddress, jobMasterGateway);
        rpc.registerGateway(testingResourceManagerGateway.getAddress(), testingResourceManagerGateway);
        // inform the task manager about the job leader
        taskManagerServices.getJobLeaderService().addJob(jobId, jobMasterAddress);
        jobManagerLeaderRetriever.notifyListener(jobMasterAddress, UUID.randomUUID());
        resourceManagerLeaderRetriever.notifyListener(testingResourceManagerGateway.getAddress(), testingResourceManagerGateway.getFencingToken().toUUID());
        final Optional<SlotStatus> slotStatusOptional = StreamSupport.stream(initialSlotReportFuture.get().spliterator(), false).findAny();
        assertTrue(slotStatusOptional.isPresent());
        final SlotStatus slotStatus = slotStatusOptional.get();
        while (true) {
            try {
                taskExecutorGateway.requestSlot(slotStatus.getSlotID(), jobId, taskDeploymentDescriptor.getAllocationId(), ResourceProfile.ZERO, jobMasterAddress, testingResourceManagerGateway.getFencingToken(), timeout).get();
                break;
            } catch (Exception e) {
                // the proper establishment of the RM connection is tracked
                // asynchronously, so we have to poll here until it went through
                // until then, slot requests will fail with an exception
                Thread.sleep(50);
            }
        }
        TestingInvokable.sync = new BlockerSync();
        // Wait till the slot has been successfully offered before submitting the task.
        // This ensures TM has been successfully registered to JM.
        slotOfferedLatch.await();
        taskExecutorGateway.submitTask(taskDeploymentDescriptor, jobMasterGateway.getFencingToken(), timeout).get();
        TestingInvokable.sync.awaitBlocker();
        // the task is still running => the partition is in in-progress and should be tracked
        assertThat(startTrackingFuture.get(), equalTo(taskResultPartitionDescriptor.getShuffleDescriptor().getResultPartitionID()));
        TestingInvokable.sync.releaseBlocker();
        taskFinishedFuture.get(timeout.getSize(), timeout.getUnit());
        testAction.accept(jobId, taskResultPartitionDescriptor, taskExecutor, taskExecutorGateway);
    } finally {
        RpcUtils.terminateRpcEndpoint(taskExecutor, timeout);
    }
    // the shutdown of the backing shuffle environment releases all partitions
    // the book-keeping is not aware of this
    assertTrue(shuffleEnvironment.getPartitionsOccupyingLocalResources().isEmpty());
}

Example 2 with TaskExecutorPartitionTracker

use of org.apache.flink.runtime.io.network.partition.TaskExecutorPartitionTracker in project flink by apache.

the class TaskExecutorTest method testHeartbeatReporting.

/**
 * Tests that the correct partition/slot report is sent as part of the heartbeat response.
 */
@Test
public void testHeartbeatReporting() throws Exception {
    final String rmAddress = "rm";
    final UUID rmLeaderId = UUID.randomUUID();
    // register the mock resource manager gateway
    final TestingResourceManagerGateway rmGateway = new TestingResourceManagerGateway();
    final CompletableFuture<ResourceID> taskExecutorRegistrationFuture = new CompletableFuture<>();
    final ResourceID rmResourceId = rmGateway.getOwnResourceId();
    final CompletableFuture<RegistrationResponse> registrationResponse = CompletableFuture.completedFuture(new TaskExecutorRegistrationSuccess(new InstanceID(), rmResourceId, new ClusterInformation("localhost", 1234)));
    rmGateway.setRegisterTaskExecutorFunction(taskExecutorRegistration -> {
        taskExecutorRegistrationFuture.complete(taskExecutorRegistration.getResourceId());
        return registrationResponse;
    });
    final CompletableFuture<SlotReport> initialSlotReportFuture = new CompletableFuture<>();
    rmGateway.setSendSlotReportFunction(resourceIDInstanceIDSlotReportTuple3 -> {
        initialSlotReportFuture.complete(resourceIDInstanceIDSlotReportTuple3.f2);
        return CompletableFuture.completedFuture(Acknowledge.get());
    });
    final CompletableFuture<TaskExecutorHeartbeatPayload> heartbeatPayloadCompletableFuture = new CompletableFuture<>();
    rmGateway.setTaskExecutorHeartbeatFunction((resourceID, heartbeatPayload) -> {
        heartbeatPayloadCompletableFuture.complete(heartbeatPayload);
        return FutureUtils.completedVoidFuture();
    });
    rpc.registerGateway(rmAddress, rmGateway);
    final SlotID slotId = buildSlotID(0);
    final ResourceProfile resourceProfile = ResourceProfile.fromResources(1.0, 1);
    final SlotReport slotReport1 = new SlotReport(new SlotStatus(slotId, resourceProfile));
    final SlotReport slotReport2 = new SlotReport(new SlotStatus(slotId, resourceProfile, new JobID(), new AllocationID()));
    final Queue<SlotReport> reports = new ArrayDeque<>(Arrays.asList(slotReport1, slotReport2));
    final TaskSlotTable<Task> taskSlotTable = TestingTaskSlotTable.<Task>newBuilder().createSlotReportSupplier(reports::poll).closeAsyncReturns(CompletableFuture.completedFuture(null)).build();
    final TaskExecutorLocalStateStoresManager localStateStoresManager = createTaskExecutorLocalStateStoresManager();
    final TaskManagerServices taskManagerServices = new TaskManagerServicesBuilder().setUnresolvedTaskManagerLocation(unresolvedTaskManagerLocation).setTaskSlotTable(taskSlotTable).setTaskStateManager(localStateStoresManager).build();
    final TaskExecutorPartitionTracker partitionTracker = createPartitionTrackerWithFixedPartitionReport(taskManagerServices.getShuffleEnvironment());
    final TaskExecutor taskManager = createTaskExecutor(taskManagerServices, HEARTBEAT_SERVICES, partitionTracker);
    try {
        taskManager.start();
        // define a leader and see that a registration happens
        resourceManagerLeaderRetriever.notifyListener(rmAddress, rmLeaderId);
        // register resource manager success will trigger monitoring heartbeat target between tm
        // and rm
        assertThat(taskExecutorRegistrationFuture.get(), equalTo(unresolvedTaskManagerLocation.getResourceID()));
        assertThat(initialSlotReportFuture.get(), equalTo(slotReport1));
        TaskExecutorGateway taskExecutorGateway = taskManager.getSelfGateway(TaskExecutorGateway.class);
        // trigger the heartbeat asynchronously
        taskExecutorGateway.heartbeatFromResourceManager(rmResourceId);
        // wait for heartbeat response
        SlotReport actualSlotReport = heartbeatPayloadCompletableFuture.get().getSlotReport();
        // the new slot report should be reported
        assertEquals(slotReport2, actualSlotReport);
        ClusterPartitionReport actualClusterPartitionReport = heartbeatPayloadCompletableFuture.get().getClusterPartitionReport();
        assertEquals(partitionTracker.createClusterPartitionReport(), actualClusterPartitionReport);
    } finally {
        RpcUtils.terminateRpcEndpoint(taskManager, timeout);
    }
}

Example 3 with TaskExecutorPartitionTracker

use of org.apache.flink.runtime.io.network.partition.TaskExecutorPartitionTracker in project flink by apache.

the class TaskExecutorTest method testReleaseOfJobResourcesIfJobMasterIsNotCorrect.

/**
 * Tests that the TaskExecutor releases all of its job resources if the JobMaster is not running
 * the specified job. See FLINK-21606.
 */
@Test
public void testReleaseOfJobResourcesIfJobMasterIsNotCorrect() throws Exception {
    final TaskManagerServices taskManagerServices = new TaskManagerServicesBuilder().setTaskSlotTable(TaskSlotUtils.createTaskSlotTable(1)).build();
    final TestingTaskExecutorPartitionTracker taskExecutorPartitionTracker = new TestingTaskExecutorPartitionTracker();
    final CompletableFuture<JobID> jobPartitionsReleaseFuture = new CompletableFuture<>();
    // simulate that we have some partitions tracked
    taskExecutorPartitionTracker.setIsTrackingPartitionsForFunction(ignored -> true);
    taskExecutorPartitionTracker.setStopTrackingAndReleaseAllPartitionsConsumer(jobPartitionsReleaseFuture::complete);
    final TaskExecutor taskExecutor = createTaskExecutor(taskManagerServices, HEARTBEAT_SERVICES, taskExecutorPartitionTracker);
    final TestingJobMasterGateway jobMasterGateway = new TestingJobMasterGatewayBuilder().setRegisterTaskManagerFunction((ignoredJobId, ignoredTaskManagerRegistrationInformation) -> CompletableFuture.completedFuture(new JMTMRegistrationRejection("foobar"))).build();
    rpc.registerGateway(jobMasterGateway.getAddress(), jobMasterGateway);
    final InstanceID registrationId = new InstanceID();
    final OneShotLatch taskExecutorIsRegistered = new OneShotLatch();
    final CompletableFuture<Tuple3<InstanceID, SlotID, AllocationID>> availableSlotFuture = new CompletableFuture<>();
    final TestingResourceManagerGateway resourceManagerGateway = createRmWithTmRegisterAndNotifySlotHooks(registrationId, taskExecutorIsRegistered, availableSlotFuture);
    rpc.registerGateway(resourceManagerGateway.getAddress(), resourceManagerGateway);
    resourceManagerLeaderRetriever.notifyListener(resourceManagerGateway.getAddress(), resourceManagerGateway.getFencingToken().toUUID());
    try {
        taskExecutor.start();
        final TaskExecutorGateway taskExecutorGateway = taskExecutor.getSelfGateway(TaskExecutorGateway.class);
        taskExecutorIsRegistered.await();
        final AllocationID allocationId = new AllocationID();
        final SlotID slotId = new SlotID(taskExecutor.getResourceID(), 0);
        requestSlot(taskExecutorGateway, jobId, allocationId, slotId, ResourceProfile.UNKNOWN, jobMasterGateway.getAddress(), resourceManagerGateway.getFencingToken());
        // The JobManager should reject the registration which should release all job resources
        // on the TaskExecutor
        jobManagerLeaderRetriever.notifyListener(jobMasterGateway.getAddress(), jobMasterGateway.getFencingToken().toUUID());
        // the slot should be freed
        assertThat(availableSlotFuture.get().f1, is(slotId));
        assertThat(availableSlotFuture.get().f2, is(allocationId));
        // all job partitions should be released
        assertThat(jobPartitionsReleaseFuture.get(), is(jobId));
    } finally {
        RpcUtils.terminateRpcEndpoint(taskExecutor, timeout);
    }
}

Example 4 with TaskExecutorPartitionTracker

use of org.apache.flink.runtime.io.network.partition.TaskExecutorPartitionTracker in project flink by apache.

the class TaskExecutorPartitionLifecycleTest method testBlockingLocalPartitionReleaseDoesNotBlockTaskExecutor.

@Test
public void testBlockingLocalPartitionReleaseDoesNotBlockTaskExecutor() throws Exception {
    BlockerSync sync = new BlockerSync();
    ResultPartitionManager blockingResultPartitionManager = new ResultPartitionManager() {

        @Override
        public void releasePartition(ResultPartitionID partitionId, Throwable cause) {
            sync.blockNonInterruptible();
            super.releasePartition(partitionId, cause);
        }
    };
    NettyShuffleEnvironment shuffleEnvironment = new NettyShuffleEnvironmentBuilder().setResultPartitionManager(blockingResultPartitionManager).setIoExecutor(TEST_EXECUTOR_SERVICE_RESOURCE.getExecutor()).build();
    final CompletableFuture<ResultPartitionID> startTrackingFuture = new CompletableFuture<>();
    final TaskExecutorPartitionTracker partitionTracker = new TaskExecutorPartitionTrackerImpl(shuffleEnvironment) {

        @Override
        public void startTrackingPartition(JobID producingJobId, TaskExecutorPartitionInfo partitionInfo) {
            super.startTrackingPartition(producingJobId, partitionInfo);
            startTrackingFuture.complete(partitionInfo.getResultPartitionId());
        }
    };
    try {
        internalTestPartitionRelease(partitionTracker, shuffleEnvironment, startTrackingFuture, (jobId, resultPartitionDeploymentDescriptor, taskExecutor, taskExecutorGateway) -> {
            final IntermediateDataSetID dataSetId = resultPartitionDeploymentDescriptor.getResultId();
            taskExecutorGateway.releaseClusterPartitions(Collections.singleton(dataSetId), timeout);
            // execute some operation to check whether the TaskExecutor is blocked
            taskExecutorGateway.canBeReleased().get(5, TimeUnit.SECONDS);
        });
    } finally {
        sync.releaseBlocker();
    }
}