474 測定値

Apache SeaTunnel Zeta エンジンのソースコード分析 (パート 3): サーバー側タスクの送信

に William Guo115m2024/09/20

長すぎる; 読むには

今回は、サーバー側のタスク送信プロセスについて説明します。

featured image - Apache SeaTunnel Zeta エンジンのソースコード分析 (パート 3): サーバー側タスクの送信

これは、Apache SeaTunnel Zeta エンジンソースコードを分析するシリーズの最後の記事です。全体像を把握するには、以前のシリーズを確認してください。

サーバーの起動後に実行されるコンポーネントを確認しましょう。

CoordinatorService : マスター/スタンバイノードでのみ有効になり、クラスターのステータスをリッスンし、マスター/スタンバイフェイルオーバーを処理します。
SlotService : ワーカーノードで有効になり、定期的にそのステータスをマスターに報告します。
TaskExecutionService : ワーカーノードで有効になり、タスクメトリックを定期的に IMAP に更新します。

クラスターがタスクを受信しない場合、これらのコンポーネントは実行されます。ただし、クライアントがSeaTunnelSubmitJobCodecメッセージをサーバーに送信すると、サーバーはそれをどのように処理するのでしょうか。

メッセージ受信

クライアントとサーバーは別のマシン上にあるため、メソッド呼び出しは使用できません。代わりに、メッセージパッシングが使用されます。サーバーはメッセージを受信すると、それをどのように処理するのでしょうか。

まず、クライアントはSeaTunnelSubmitJobCodecタイプのメッセージを送信します。

 // Client code ClientMessage request = SeaTunnelSubmitJobCodec.encodeRequest( jobImmutableInformation.getJobId(), seaTunnelHazelcastClient .getSerializationService() .toData(jobImmutableInformation), jobImmutableInformation.isStartWithSavePoint()); PassiveCompletableFuture<Void> submitJobFuture = seaTunnelHazelcastClient.requestOnMasterAndGetCompletableFuture(request);

SeaTunnelSubmitJobCodecクラスでは、メッセージタイプをMessageTaskクラスにマップするSeaTunnelMessageTaskFactoryProviderクラスに関連付けられています。 SeaTunnelSubmitJobCodecの場合、 SubmitJobTaskクラスにマップされます。

 private final Int2ObjectHashMap<MessageTaskFactory> factories = new Int2ObjectHashMap<>(60); private void initFactories() { factories.put( SeaTunnelPrintMessageCodec.REQUEST_MESSAGE_TYPE, (clientMessage, connection) -> new PrintMessageTask(clientMessage, node, connection)); factories.put( SeaTunnelSubmitJobCodec.REQUEST_MESSAGE_TYPE, (clientMessage, connection) -> new SubmitJobTask(clientMessage, node, connection)); ..... }

SubmitJobTaskクラスを調べると、 SubmitJobOperationクラスが呼び出されます。

 @Override protected Operation prepareOperation() { return new SubmitJobOperation( parameters.jobId, parameters.jobImmutableInformation, parameters.isStartWithSavePoint); }

SubmitJobOperationクラスでは、タスク情報は、 submitJobメソッドを介してCoordinatorServiceコンポーネントに渡されます。

 @Override protected PassiveCompletableFuture<?> doRun() throws Exception { SeaTunnelServer seaTunnelServer = getService(); return seaTunnelServer .getCoordinatorService() .submitJob(jobId, jobImmutableInformation, isStartWithSavePoint); }

この時点で、クライアントメッセージはメソッド呼び出しのためにサーバーに効果的に渡されます。他の種類の操作も同様の方法で追跡できます。

コーディネーターサービス

次に、 CoordinatorServiceジョブの送信をどのように処理するかを見てみましょう。

 public PassiveCompletableFuture<Void> submitJob( long jobId, Data jobImmutableInformation, boolean isStartWithSavePoint) { CompletableFuture<Void> jobSubmitFuture = new CompletableFuture<>(); // First, check if a job with the same ID already exists if (getJobMaster(jobId) != null) { logger.warning( String.format( "The job %s is currently running; no need to submit again.", jobId)); jobSubmitFuture.complete(null); return new PassiveCompletableFuture<>(jobSubmitFuture); } // Initialize JobMaster object JobMaster jobMaster = new JobMaster( jobImmutableInformation, this.nodeEngine, executorService, getResourceManager(), getJobHistoryService(), runningJobStateIMap, runningJobStateTimestampsIMap, ownedSlotProfilesIMap, runningJobInfoIMap, metricsImap, engineConfig, seaTunnelServer); executorService.submit( () -> { try { // Ensure no duplicate tasks with the same ID if (!isStartWithSavePoint && getJobHistoryService().getJobMetrics(jobId) != null) { throw new JobException( String.format( "The job id %s has already been submitted and is not starting with a savepoint.", jobId)); } // Add task info to IMAP runningJobInfoIMap.put( jobId, new JobInfo(System.currentTimeMillis(), jobImmutableInformation)); runningJobMasterMap.put(jobId, jobMaster); // Initialize JobMaster jobMaster.init( runningJobInfoIMap.get(jobId).getInitializationTimestamp(), false); // Task creation successful jobSubmitFuture.complete(null); } catch (Throwable e) { String errorMsg = ExceptionUtils.getMessage(e); logger.severe(String.format("submit job %s error %s ", jobId, errorMsg)); jobSubmitFuture.completeExceptionally(new JobException(errorMsg)); } if (!jobSubmitFuture.isCompletedExceptionally()) { // Start job execution try { jobMaster.run(); } finally { // Remove jobMaster from map if not cancelled if (!jobMaster.getJobMasterCompleteFuture().isCancelled()) { runningJobMasterMap.remove(jobId); } } } else { runningJobInfoIMap.remove(jobId); runningJobMasterMap.remove(jobId); } }); return new PassiveCompletableFuture<>(jobSubmitFuture); }

サーバーでは、個々のタスクを管理するためにJobMasterオブジェクトが作成されます。JobMaster JobMaster作成中に、 getResourceManager()を介してリソースマネージャーを取得し、 getJobHistoryService()を介してジョブ履歴情報を取得しますjobHistoryServiceは起動時に初期化されますが、 ResourceManager最初のタスク送信時に遅延ロードされます。

 /** Lazy load for resource manager */ public ResourceManager getResourceManager() { if (resourceManager == null) { synchronized (this) { if (resourceManager == null) { ResourceManager manager = new ResourceManagerFactory(nodeEngine, engineConfig) .getResourceManager(); manager.init(); resourceManager = manager; } } } return resourceManager; }

リソースマネージャ

現在、SeaTunnel はスタンドアロンデプロイメントのみをサポートしています。ResourceManager ResourceManager初期化すると、すべてのクラスターノードが収集され、 SyncWorkerProfileOperationが送信されてノード情報が取得され、内部のregisterWorker状態が更新されます。

 @Override public void init() { log.info("Init ResourceManager"); initWorker(); } private void initWorker() { log.info("initWorker... "); List<Address> aliveNode = nodeEngine.getClusterService().getMembers().stream() .map(Member::getAddress) .collect(Collectors.toList()); log.info("init live nodes: {}", aliveNode); List<CompletableFuture<Void>> futures = aliveNode.stream() .map( node -> sendToMember(new SyncWorkerProfileOperation(), node) .thenAccept( p -> { if (p != null) { registerWorker.put( node, (WorkerProfile) p); log.info( "received new worker register: " + ((WorkerProfile) p) .getAddress()); } })) .collect(Collectors.toList()); futures.forEach(CompletableFuture::join); log.info("registerWorker: {}", registerWorker); }

以前、 SlotService各ノードからマスターにハートビートメッセージを定期的に送信していることを確認しました。これらのハートビートを受信すると、 ResourceManager内部状態のノードステータスを更新します。

 @Override public void heartbeat(WorkerProfile workerProfile) { if (!registerWorker.containsKey(workerProfile.getAddress())) { log.info("received new worker register: " + workerProfile.getAddress()); sendToMember(new ResetResourceOperation(), workerProfile.getAddress()).join(); } else { log.debug("received worker heartbeat from: " + workerProfile.getAddress()); } registerWorker.put(workerProfile.getAddress(), workerProfile); }

ジョブマスター

CoordinatorServiceでは、 JobMasterインスタンスが作成され、そのinitメソッドが呼び出されます。 initメソッドが完了すると、タスクの作成が成功したとみなされます。次に、 runメソッドが呼び出され、タスクが正式に実行されます。

初期化とinitメソッドを見てみましょう。

 public JobMaster( @NonNull Data jobImmutableInformationData, @NonNull NodeEngine nodeEngine, @NonNull ExecutorService executorService, @NonNull ResourceManager resourceManager, @NonNull JobHistoryService jobHistoryService, @NonNull IMap runningJobStateIMap, @NonNull IMap runningJobStateTimestampsIMap, @NonNull IMap ownedSlotProfilesIMap, @NonNull IMap<Long, JobInfo> runningJobInfoIMap, @NonNull IMap<Long, HashMap<TaskLocation, SeaTunnelMetricsContext>> metricsImap, EngineConfig engineConfig, SeaTunnelServer seaTunnelServer) { this.jobImmutableInformationData = jobImmutableInformationData; this.nodeEngine = nodeEngine; this.executorService = executorService; flakeIdGenerator = this.nodeEngine .getHazelcastInstance() .getFlakeIdGenerator(Constant.SEATUNNEL_ID_GENERATOR_NAME); this.ownedSlotProfilesIMap = ownedSlotProfilesIMap; this.resourceManager = resourceManager; this.jobHistoryService = jobHistoryService; this.runningJobStateIMap = runningJobStateIMap; this.runningJobStateTimestampsIMap = runningJobStateTimestampsIMap; this.runningJobInfoIMap = runningJobInfoIMap; this.engineConfig = engineConfig; this.metricsImap = metricsImap; this.seaTunnelServer = seaTunnelServer; this.releasedSlotWhenTaskGroupFinished = new ConcurrentHashMap<>(); }

初期化中は、重要な操作は行われず、単純な変数の割り当てのみが実行されます。init initに注目する必要があります。

 public synchronized void init(long initializationTimestamp, boolean restart) throws Exception { // The server receives a binary object from the client, // which is first converted to a JobImmutableInformation object, the same object sent by the client jobImmutableInformation = nodeEngine.getSerializationService().toObject(jobImmutableInformationData); // Get the checkpoint configuration, such as the interval, timeout, etc. jobCheckpointConfig = createJobCheckpointConfig( engineConfig.getCheckpointConfig(), jobImmutableInformation.getJobConfig()); LOGGER.info( String.format( "Init JobMaster for Job %s (%s) ", jobImmutableInformation.getJobConfig().getName(), jobImmutableInformation.getJobId())); LOGGER.info( String.format( "Job %s (%s) needed jar urls %s", jobImmutableInformation.getJobConfig().getName(), jobImmutableInformation.getJobId(), jobImmutableInformation.getPluginJarsUrls())); ClassLoader appClassLoader = Thread.currentThread().getContextClassLoader(); // Get the ClassLoader ClassLoader classLoader = seaTunnelServer .getClassLoaderService() .getClassLoader( jobImmutableInformation.getJobId(), jobImmutableInformation.getPluginJarsUrls()); // Deserialize the logical DAG from the client-provided information logicalDag = CustomClassLoadedObject.deserializeWithCustomClassLoader( nodeEngine.getSerializationService(), classLoader, jobImmutableInformation.getLogicalDag()); try { Thread.currentThread().setContextClassLoader(classLoader); // Execute save mode functionality, such as table creation and deletion if (!restart && !logicalDag.isStartWithSavePoint() && ReadonlyConfig.fromMap(logicalDag.getJobConfig().getEnvOptions()) .get(EnvCommonOptions.SAVEMODE_EXECUTE_LOCATION) .equals(SaveModeExecuteLocation.CLUSTER)) { logicalDag.getLogicalVertexMap().values().stream() .map(LogicalVertex::getAction) .filter(action -> action instanceof SinkAction) .map(sink -> ((SinkAction<?, ?, ?, ?>) sink).getSink()) .forEach(JobMaster::handleSaveMode); } // Parse the logical plan into a physical plan final Tuple2<PhysicalPlan, Map<Integer, CheckpointPlan>> planTuple = PlanUtils.fromLogicalDAG( logicalDag, nodeEngine, jobImmutableInformation, initializationTimestamp, executorService, flakeIdGenerator, runningJobStateIMap, runningJobStateTimestampsIMap, engineConfig.getQueueType(), engineConfig); this.physicalPlan = planTuple.f0(); this.physicalPlan.setJobMaster(this); this.checkpointPlanMap = planTuple.f1(); } finally { // Reset the current thread's ClassLoader and release the created classLoader Thread.currentThread().setContextClassLoader(appClassLoader); seaTunnelServer .getClassLoaderService() .releaseClassLoader( jobImmutableInformation.getJobId(), jobImmutableInformation.getPluginJarsUrls()); } Exception initException = null; try { // Initialize the checkpoint manager this.initCheckPointManager(restart); } catch (Exception e) { initException = e; } // Add some callback functions for job state listening this.initStateFuture(); if (initException != null) { if (restart) { cancelJob(); } throw initException; } }

最後に、 runメソッドを見てみましょう。

 public void run() { try { physicalPlan.startJob(); } catch (Throwable e) { LOGGER.severe( String.format( "Job %s (%s) run error with: %s", physicalPlan.getJobImmutableInformation().getJobConfig().getName(), physicalPlan.getJobImmutableInformation().getJobId(), ExceptionUtils.getMessage(e))); } finally { jobMasterCompleteFuture.join(); if (engineConfig.getConnectorJarStorageConfig().getEnable()) { List<ConnectorJarIdentifier> pluginJarIdentifiers = jobImmutableInformation.getPluginJarIdentifiers(); seaTunnelServer .getConnectorPackageService() .cleanUpWhenJobFinished( jobImmutableInformation.getJobId(), pluginJarIdentifiers); } } }

この方法は比較的簡単で、生成された物理プランを実行するためにphysicalPlan.startJob()を呼び出します。

上記のコードから、サーバーがクライアントのタスク送信要求を受信した後、論理プランから物理プランを生成するJobMasterクラスを初期化し、物理プランを実行することがわかります。

次に、論理プランが物理プランに変換される方法を詳しく調べる必要があります。

論理プランから物理プランへの変換

物理計画の生成は、 JobMasterで次のメソッドを呼び出すことによって行われます。

 final Tuple2<PhysicalPlan, Map<Integer, CheckpointPlan>> planTuple = PlanUtils.fromLogicalDAG( logicalDag, nodeEngine, jobImmutableInformation, initializationTimestamp, executorService, flakeIdGenerator, runningJobStateIMap, runningJobStateTimestampsIMap, engineConfig.getQueueType(), engineConfig);

物理プランを生成する方法は、まず論理プランを実行プランに変換し、次に実行プランを物理プランに変換する。

 public static Tuple2<PhysicalPlan, Map<Integer, CheckpointPlan>> fromLogicalDAG( @NonNull LogicalDag logicalDag, @NonNull NodeEngine nodeEngine, @NonNull JobImmutableInformation jobImmutableInformation, long initializationTimestamp, @NonNull ExecutorService executorService, @NonNull FlakeIdGenerator flakeIdGenerator, @NonNull IMap runningJobStateIMap, @NonNull IMap runningJobStateTimestampsIMap, @NonNull QueueType queueType, @NonNull EngineConfig engineConfig) { return new PhysicalPlanGenerator( new ExecutionPlanGenerator( logicalDag, jobImmutableInformation, engineConfig) .generate(), nodeEngine, jobImmutableInformation, initializationTimestamp, executorService, flakeIdGenerator, runningJobStateIMap, runningJobStateTimestampsIMap, queueType) .generate(); }

実行計画の生成

public ExecutionPlanGenerator( @NonNull LogicalDag logicalPlan, @NonNull JobImmutableInformation jobImmutableInformation, @NonNull EngineConfig engineConfig) { checkArgument( logicalPlan.getEdges().size() > 0, "ExecutionPlan Builder must have LogicalPlan."); this.logicalPlan = logicalPlan; this.jobImmutableInformation = jobImmutableInformation; this.engineConfig = engineConfig; } public ExecutionPlan generate() { log.debug("Generate execution plan using logical plan:"); Set<ExecutionEdge> executionEdges = generateExecutionEdges(logicalPlan.getEdges()); log.debug("Phase 1: generate execution edge list {}", executionEdges); executionEdges = generateShuffleEdges(executionEdges); log.debug("Phase 2: generate shuffle edge list {}", executionEdges); executionEdges = generateTransformChainEdges(executionEdges); log.debug("Phase 3: generate transform chain edge list {}", executionEdges); List<Pipeline> pipelines = generatePipelines(executionEdges); log.debug("Phase 4: generate pipeline list {}", pipelines); ExecutionPlan executionPlan = new ExecutionPlan(pipelines, jobImmutableInformation); log.debug("Phase 5 : generate execution plan {}", executionPlan); return executionPlan; }

ExecutionPlanGeneratorクラスは論理プランを受け取り、実行エッジ、シャッフルエッジ、変換チェーンエッジ、そしてパイプラインの生成を含む一連の手順を通じて実行プランを生成します。

物理計画の作成

PhysicalPlanGeneratorクラスは、実行プランを物理プランに変換します。

 public PhysicalPlanGenerator( @NonNull ExecutionPlan executionPlan, @NonNull NodeEngine nodeEngine, @NonNull JobImmutableInformation jobImmutableInformation, long initializationTimestamp, @NonNull ExecutorService executorService, @NonNull FlakeIdGenerator flakeIdGenerator, @NonNull IMap runningJobStateIMap, @NonNull IMap runningJobStateTimestampsIMap, @NonNull QueueType queueType) { this.executionPlan = executionPlan; this.nodeEngine = nodeEngine; this.jobImmutableInformation = jobImmutableInformation; this.initializationTimestamp = initializationTimestamp; this.executorService = executorService; this.flakeIdGenerator = flakeIdGenerator; this.runningJobStateIMap = runningJobStateIMap; this.runningJobStateTimestampsIMap = runningJobStateTimestampsIMap; this.queueType = queueType; } public PhysicalPlan generate() { List<PhysicalVertex> physicalVertices = generatePhysicalVertices(executionPlan); List<PhysicalEdge> physicalEdges = generatePhysicalEdges(executionPlan); PhysicalPlan physicalPlan = new PhysicalPlan(physicalVertices, physicalEdges); log.debug("Generate physical plan: {}", physicalPlan); return physicalPlan; }

これら 2 つのクラスの内容を調べてみましょう。

 public class ExecutionPlan { private final List<Pipeline> pipelines; private final JobImmutableInformation jobImmutableInformation; } public class Pipeline { /** The ID of the pipeline. */ private final Integer id; private final List<ExecutionEdge> edges; private final Map<Long, ExecutionVertex> vertexes; } public class ExecutionEdge { private ExecutionVertex leftVertex; private ExecutionVertex rightVertex; } public class ExecutionVertex { private Long vertexId; private Action action; private int parallelism; }

これを論理的な計画と比較してみましょう。

 public class LogicalDag implements IdentifiedDataSerializable { @Getter private JobConfig jobConfig; private final Set<LogicalEdge> edges = new LinkedHashSet<>(); private final Map<Long, LogicalVertex> logicalVertexMap = new LinkedHashMap<>(); private IdGenerator idGenerator; private boolean isStartWithSavePoint = false; } public class LogicalEdge implements IdentifiedDataSerializable { private LogicalVertex inputVertex; private LogicalVertex targetVertex; private Long inputVertexId; private Long targetVertexId; } public class LogicalVertex implements IdentifiedDataSerializable { private Long vertexId; private Action action; private int parallelism; }

各パイプラインは論理プランに似ているようです。なぜこの変換ステップが必要なのでしょうか? 論理プランを生成するプロセスを詳しく見てみましょう。

上で示したように、実行プランの生成には 5 つのステップが含まれており、これらを 1 つずつ確認していきます。

ステップ1: 論理プランを実行プランに変換する

// Input is a set of logical plan edges, where each edge stores upstream and downstream nodes private Set<ExecutionEdge> generateExecutionEdges(Set<LogicalEdge> logicalEdges) { Set<ExecutionEdge> executionEdges = new LinkedHashSet<>(); Map<Long, ExecutionVertex> logicalVertexIdToExecutionVertexMap = new HashMap(); // Sort in order: first by input node, then by output node List<LogicalEdge> sortedLogicalEdges = new ArrayList<>(logicalEdges); Collections.sort( sortedLogicalEdges, (o1, o2) -> { if (o1.getInputVertexId() != o2.getInputVertexId()) { return o1.getInputVertexId() > o2.getInputVertexId() ? 1 : -1; } if (o1.getTargetVertexId() != o2.getTargetVertexId()) { return o1.getTargetVertexId() > o2.getTargetVertexId() ? 1 : -1; } return 0; }); // Loop to convert each logical plan edge to an execution plan edge for (LogicalEdge logicalEdge : sortedLogicalEdges) { LogicalVertex logicalInputVertex = logicalEdge.getInputVertex(); ExecutionVertex executionInputVertex = logicalVertexIdToExecutionVertexMap.computeIfAbsent( logicalInputVertex.getVertexId(), vertexId -> { long newId = idGenerator.getNextId(); // Recreate Action for each logical plan node Action newLogicalInputAction = recreateAction( logicalInputVertex.getAction(), newId, logicalInputVertex.getParallelism()); // Convert to execution plan node return new ExecutionVertex( newId, newLogicalInputAction, logicalInputVertex.getParallelism()); }); // Similarly, recreate execution plan nodes for target nodes LogicalVertex logicalTargetVertex = logicalEdge.getTargetVertex(); ExecutionVertex executionTargetVertex = logicalVertexIdToExecutionVertexMap.computeIfAbsent( logicalTargetVertex.getVertexId(), vertexId -> { long newId = idGenerator.getNextId(); Action newLogicalTargetAction = recreateAction( logicalTargetVertex.getAction(), newId, logicalTargetVertex.getParallelism()); return new ExecutionVertex( newId, newLogicalTargetAction, logicalTargetVertex.getParallelism()); }); // Generate execution plan edge ExecutionEdge executionEdge = new ExecutionEdge(executionInputVertex, executionTargetVertex); executionEdges.add(executionEdge); } return executionEdges; }

ステップ2

 private Set<ExecutionEdge> generateShuffleEdges(Set<ExecutionEdge> executionEdges) { // Map of upstream node ID to list of downstream nodes Map<Long, List<ExecutionVertex>> targetVerticesMap = new LinkedHashMap<>(); // Store only nodes of type Source Set<ExecutionVertex> sourceExecutionVertices = new HashSet<>(); executionEdges.forEach( edge -> { ExecutionVertex leftVertex = edge.getLeftVertex(); ExecutionVertex rightVertex = edge.getRightVertex(); if (leftVertex.getAction() instanceof SourceAction) { sourceExecutionVertices.add(leftVertex); } targetVerticesMap .computeIfAbsent(leftVertex.getVertexId(), id -> new ArrayList<>()) .add(rightVertex); }); if (sourceExecutionVertices.size() != 1) { return executionEdges; } ExecutionVertex sourceExecutionVertex = sourceExecutionVertices.stream().findFirst().get(); Action sourceAction = sourceExecutionVertex.getAction(); List<CatalogTable> producedCatalogTables = new ArrayList<>(); if (sourceAction instanceof SourceAction) { try { producedCatalogTables = ((SourceAction<?, ?, ?>) sourceAction) .getSource() .getProducedCatalogTables(); } catch (UnsupportedOperationException e) { } } else if (sourceAction instanceof TransformChainAction) { return executionEdges; } else { throw new SeaTunnelException( "source action must be SourceAction or TransformChainAction"); } // If the source produces a single table or // the source has only one downstream output, return directly if (producedCatalogTables.size() <= 1 || targetVerticesMap.get(sourceExecutionVertex.getVertexId()).size() <= 1) { return executionEdges; } List<ExecutionVertex> sinkVertices = targetVerticesMap.get(sourceExecutionVertex.getVertexId()); // Check if there are other types of actions, currently downstream nodes should ideally have two types: Transform and Sink; here we check if only Sink type is present Optional<ExecutionVertex> hasOtherAction = sinkVertices.stream() .filter(vertex -> !(vertex.getAction() instanceof SinkAction)) .findFirst(); checkArgument(!hasOtherAction.isPresent()); // After executing the above code, the current scenario is: // There is only one data source, this source produces multiple tables, and multiple sink nodes depend on these tables // This means the task has only two types of nodes: a source node that produces multiple tables and a group of sink nodes depending on this source // A new shuffle node will be created and added between the source and sinks // Modify the dependency relationship to source -> shuffle -> multiple sinks Set<ExecutionEdge> newExecutionEdges = new LinkedHashSet<>(); // Shuffle strategy will not be explored in detail here ShuffleStrategy shuffleStrategy = ShuffleMultipleRowStrategy.builder() .jobId(jobImmutableInformation.getJobId()) .inputPartitions(sourceAction.getParallelism()) .catalogTables(producedCatalogTables) .queueEmptyQueueTtl( (int) (engineConfig.getCheckpointConfig().getCheckpointInterval() * 3)) .build(); ShuffleConfig shuffleConfig = ShuffleConfig.builder().shuffleStrategy(shuffleStrategy).build(); long shuffleVertexId = idGenerator.getNextId(); String shuffleActionName = String.format("Shuffle [%s]", sourceAction.getName()); ShuffleAction shuffleAction = new ShuffleAction(shuffleVertexId, shuffleActionName, shuffleConfig); shuffleAction.setParallelism(sourceAction.getParallelism()); ExecutionVertex shuffleVertex = new ExecutionVertex(shuffleVertexId, shuffleAction, shuffleAction.getParallelism()); ExecutionEdge sourceToShuffleEdge = new ExecutionEdge(sourceExecutionVertex, shuffleVertex); newExecutionEdges.add(sourceToShuffleEdge); // Set the parallelism of multiple sink nodes to 1 for (ExecutionVertex sinkVertex : sinkVertices) { sinkVertex.setParallelism(1); sinkVertex.getAction().setParallelism(1); ExecutionEdge shuffleToSinkEdge = new ExecutionEdge(shuffleVertex, sinkVertex); newExecutionEdges.add(shuffleToSinkEdge); } return newExecutionEdges; }

シャッフルステップは、ソースが複数のテーブルの読み取りをサポートし、このソースに依存するシンクノードが複数ある特定のシナリオに対処します。このような場合、間にシャッフルノードが追加されます。

ステップ3

 private Set<ExecutionEdge> generateTransformChainEdges(Set<ExecutionEdge> executionEdges) { // Uses three structures: stores all Source nodes and the input/output nodes for each // inputVerticesMap stores all upstream input nodes by downstream node id as the key // targetVerticesMap stores all downstream output nodes by upstream node id as the key Map<Long, List<ExecutionVertex>> inputVerticesMap = new HashMap<>(); Map<Long, List<ExecutionVertex>> targetVerticesMap = new HashMap<>(); Set<ExecutionVertex> sourceExecutionVertices = new HashSet<>(); executionEdges.forEach( edge -> { ExecutionVertex leftVertex = edge.getLeftVertex(); ExecutionVertex rightVertex = edge.getRightVertex(); if (leftVertex.getAction() instanceof SourceAction) { sourceExecutionVertices.add(leftVertex); } inputVerticesMap .computeIfAbsent(rightVertex.getVertexId(), id -> new ArrayList<>()) .add(leftVertex); targetVerticesMap .computeIfAbsent(leftVertex.getVertexId(), id -> new ArrayList<>()) .add(rightVertex); }); Map<Long, ExecutionVertex> transformChainVertexMap = new HashMap<>(); Map<Long, Long> chainedTransformVerticesMapping = new HashMap<>(); // Loop over each source, starting with all head nodes in the DAG for (ExecutionVertex sourceVertex : sourceExecutionVertices) { List<ExecutionVertex> vertices = new ArrayList<>(); vertices.add(sourceVertex); for (int index = 0; index < vertices.size(); index++) { ExecutionVertex vertex = vertices.get(index); fillChainedTransformExecutionVertex( vertex, chainedTransformVerticesMapping, transformChainVertexMap, executionEdges, Collections.unmodifiableMap(inputVerticesMap), Collections.unmodifiableMap(targetVerticesMap)); // If the current node has downstream nodes, add all downstream nodes to the list // The second loop will recalculate the newly added downstream nodes, which could be Transform nodes or Sink nodes if (targetVerticesMap.containsKey(vertex.getVertexId())) { vertices.addAll(targetVerticesMap.get(vertex.getVertexId())); } } } // After looping, chained Transform nodes will be chained, and the chainable edges will be removed from the execution plan // Therefore, the logical plan at this point cannot form the graph relationship and needs to be rebuilt Set<ExecutionEdge> transformChainEdges = new LinkedHashSet<>(); // Loop over existing relationships for (ExecutionEdge executionEdge : executionEdges) { ExecutionVertex leftVertex = executionEdge.getLeftVertex(); ExecutionVertex rightVertex = executionEdge.getRightVertex(); boolean needRebuild = false; // Check if the input or output nodes of the current edge are in the chain mapping // If so, the node has been chained, and we need to find the chained node in the mapping // and rebuild the DAG if (chainedTransformVerticesMapping.containsKey(leftVertex.getVertexId())) { needRebuild = true; leftVertex = transformChainVertexMap.get( chainedTransformVerticesMapping.get(leftVertex.getVertexId())); } if (chainedTransformVerticesMapping.containsKey(rightVertex.getVertexId())) { needRebuild = true; rightVertex = transformChainVertexMap.get( chainedTransformVerticesMapping.get(rightVertex.getVertexId())); } if (needRebuild) { executionEdge = new ExecutionEdge(leftVertex, rightVertex); } transformChainEdges.add(executionEdge); } return transformChainEdges; } private void fillChainedTransformExecutionVertex( ExecutionVertex currentVertex, Map<Long, Long> chainedTransformVerticesMapping, Map<Long, ExecutionVertex> transformChainVertexMap, Set<ExecutionEdge> executionEdges, Map<Long, List<ExecutionVertex>> inputVerticesMap, Map<Long, List<ExecutionVertex>> targetVerticesMap) { // Exit if the map already contains the current node if (chainedTransformVerticesMapping.containsKey(currentVertex.getVertexId())) { return; } List<ExecutionVertex> transformChainedVertices = new ArrayList<>(); collectChainedVertices( currentVertex, transformChainedVertices, executionEdges, inputVerticesMap, targetVerticesMap); // If the list is not empty, it means the Transform nodes in the list can be merged into one if (transformChainedVertices.size() > 0) { long newVertexId = idGenerator.getNextId(); List<SeaTunnelTransform> transforms = new ArrayList<>(transformChainedVertices.size()); List<String> names = new ArrayList<>(transformChainedVertices.size()); Set<URL> jars = new HashSet<>(); Set<ConnectorJarIdentifier> identifiers = new HashSet<>(); transformChainedVertices.stream() .peek( // Add all historical node IDs and new node IDs to the mapping vertex -> chainedTransformVerticesMapping.put( vertex.getVertexId(), newVertexId)) .map(ExecutionVertex::getAction) .map(action -> (TransformAction) action) .forEach( action -> { transforms.add(action.getTransform()); jars.addAll(action.getJarUrls()); identifiers.addAll(action.getConnectorJarIdentifiers()); names.add(action.getName()); }); String transformChainActionName = String.format("TransformChain[%s]", String.join("->", names)); // Merge multiple TransformActions into one TransformChainAction TransformChainAction transformChainAction = new TransformChainAction( newVertexId, transformChainActionName, jars, identifiers, transforms); transformChainAction.setParallelism(currentVertex.getAction().getParallelism()); ExecutionVertex executionVertex = new ExecutionVertex( newVertexId, transformChainAction, currentVertex.getParallelism()); // Store the modified node information in the state transformChainVertexMap.put(newVertexId, executionVertex); chainedTransformVerticesMapping.put( currentVertex.getVertexId(), executionVertex.getVertexId()); } } private void collectChainedVertices( ExecutionVertex currentVertex, List<ExecutionVertex> chainedVertices, Set<ExecutionEdge> executionEdges, Map<Long, List<ExecutionVertex>> inputVerticesMap, Map<Long, List<ExecutionVertex>> targetVerticesMap) { Action action = currentVertex.getAction(); // Only merge TransformAction if (action instanceof TransformAction) { if (chainedVertices.size() == 0) { // If the list of vertices to be merged is empty, add itself to the list // The condition for entering this branch is that the current node is a TransformAction and the list to be merged is empty // There may be several scenarios: the first Transform node enters, and this Transform node has no constraints chainedVertices.add(currentVertex); } else if (inputVerticesMap.get(currentVertex.getVertexId()).size() == 1) { // When this condition is entered, it means: // The list of vertices to be merged already has at least one TransformAction // The scenario at this point is that the upstream Transform node has only one downstream node, ie, the current node. This constraint is ensured by the following judgment // Chain the current TransformAction node with the previous TransformAction node // Delete this relationship from the execution plan executionEdges.remove( new ExecutionEdge( chainedVertices.get(chainedVertices.size() - 1), currentVertex)); // Add itself to the list of nodes to be merged chainedVertices.add(currentVertex); } else { return; } } else { return; } // It cannot chain to any target vertex if it has multiple target vertices. if (targetVerticesMap.get(currentVertex.getVertexId()).size() == 1) { // If the current node has only one downstream node, try chaining again // If the current node has multiple downstream nodes, it will not chain the downstream nodes, so it can be ensured that the above chaining is a one-to-one relationship // This call occurs when the Transform node has only one downstream node collectChainedVertices( targetVerticesMap.get(currentVertex.getVertexId()).get(0), chainedVertices, executionEdges, inputVerticesMap, targetVerticesMap); } }

ステップ4

 private List<Pipeline> generatePipelines(Set<ExecutionEdge> executionEdges) { // Stores each execution plan node Set<ExecutionVertex> executionVertices = new LinkedHashSet<>(); for (ExecutionEdge edge : executionEdges) { executionVertices.add(edge.getLeftVertex()); executionVertices.add(edge.getRightVertex()); } // Calls the Pipeline generator to convert the execution plan into Pipelines PipelineGenerator pipelineGenerator = new PipelineGenerator(executionVertices, new ArrayList<>(executionEdges)); List<Pipeline> pipelines = pipelineGenerator.generatePipelines(); Set<String> duplicatedActionNames = new HashSet<>(); Set<String> actionNames = new HashSet<>(); for (Pipeline pipeline : pipelines) { Integer pipelineId = pipeline.getId(); for (ExecutionVertex vertex : pipeline.getVertexes().values()) { // Get each execution node of the current Pipeline, reset the Action name, and add the pipeline name Action action = vertex.getAction(); String actionName = String.format("pipeline-%s [%s]", pipelineId, action.getName()); action.setName(actionName); if (actionNames.contains(actionName)) { duplicatedActionNames.add(actionName); } actionNames.add(action Name); } } if (duplicatedActionNames.size() > 0) { throw new RuntimeException( String.format( "Duplicated Action names found: %s", duplicatedActionNames)); } return pipelines; } public PipelineGenerator(Collection<ExecutionVertex> vertices, List<ExecutionEdge> edges) { this.vertices = vertices; this.edges = edges; } public List<Pipeline> generatePipelines() { List<ExecutionEdge> executionEdges = expandEdgeByParallelism(edges); // Split the execution plan into unrelated execution plans based on their relationships // Divide into several unrelated execution plans List<List<ExecutionEdge>> edgesList = splitUnrelatedEdges(executionEdges); edgesList = edgesList.stream() .flatMap(e -> this.splitUnionEdge(e).stream()) .collect(Collectors.toList()); // Just convert execution plan to pipeline at now. We should split it to multi pipeline with // cache in the future IdGenerator idGenerator = new IdGenerator(); // Convert execution plan graph to Pipeline return edgesList.stream() .map( e -> { Map<Long, ExecutionVertex> vertexes = new HashMap<>(); List<ExecutionEdge> pipelineEdges = e.stream() .map( edge -> { if (!vertexes.containsKey( edge.getLeftVertexId())) { vertexes.put( edge.getLeftVertexId(), edge.getLeftVertex()); } ExecutionVertex source = vertexes.get( edge.getLeftVertexId()); if (!vertexes.containsKey( edge.getRightVertexId())) { vertexes.put( edge.getRightVertexId(), edge.getRightVertex()); } ExecutionVertex destination = vertexes.get( edge.getRightVertexId()); return new ExecutionEdge( source, destination); }) .collect(Collectors.toList()); return new Pipeline( (int) idGenerator.getNextId(), pipelineEdges, vertexes); }) .collect(Collectors.toList()); }

ステップ5

ステップ 5 では、ステップ 4 で生成されたパイプラインパラメータを渡して、実行プランインスタンスを生成します。

まとめ：

実行プランは、論理プランに対して次のタスクを実行します。

ソースが複数のテーブルを生成し、複数のシンクのノードがこのソースに依存する場合、間にシャッフルノードが追加されます。
複数の変換ノードを 1 つのノードに結合する、変換ノードのチェーンマージを試行します。
タスクを分割し、 configuration file/LogicalDag List<Pipeline>として表されるいくつかの無関係なタスクに分割します。

物理プラン生成

物理プランの生成について詳しく説明する前に、まず、生成された物理プランに含まれる情報を確認し、その内部コンポーネントを調べてみましょう。

 public class PhysicalPlan { private final List<SubPlan> pipelineList; private final AtomicInteger finishedPipelineNum = new AtomicInteger(0); private final AtomicInteger canceledPipelineNum = new AtomicInteger(0); private final AtomicInteger failedPipelineNum = new AtomicInteger(0); private final JobImmutableInformation jobImmutableInformation; private final IMap<Object, Object> runningJobStateIMap; private final IMap<Object, Long[]> runningJobStateTimestampsIMap; private CompletableFuture<JobResult> jobEndFuture; private final AtomicReference<String> errorBySubPlan = new AtomicReference<>(); private final String jobFullName; private final long jobId; private JobMaster jobMaster; private boolean makeJobEndWhenPipelineEnded = true; private volatile boolean isRunning = false; }

このクラスでは、キーフィールドはpipelineListで、これはSubPlanインスタンスのリストです。

 public class SubPlan { private final int pipelineMaxRestoreNum; private final int pipelineRestoreIntervalSeconds; private final List<PhysicalVertex> physicalVertexList; private final List<PhysicalVertex> coordinatorVertexList; private final int pipelineId; private final AtomicInteger finishedTaskNum = new AtomicInteger(0); private final AtomicInteger canceledTaskNum = new AtomicInteger(0); private final AtomicInteger failedTaskNum = new AtomicInteger(0); private final String pipelineFullName; private final IMap<Object, Object> runningJobStateIMap; private final Map<String, String> tags; private final IMap<Object, Long[]> runningJobStateTimestampsIMap; private CompletableFuture<PipelineExecutionState> pipelineFuture; private final PipelineLocation pipelineLocation; private AtomicReference<String> errorByPhysicalVertex = new AtomicReference<>(); private final ExecutorService executorService; private JobMaster jobMaster; private PassiveCompletableFuture<Void> reSchedulerPipelineFuture; private Integer pipelineRestoreNum; private final Object restoreLock = new Object(); private volatile PipelineStatus currPipelineStatus; public volatile boolean isRunning = false; private Map<TaskGroupLocation, SlotProfile> slotProfiles; }

SubPlanクラスは、物理プランノードとコーディネーターノードに分割されたPhysicalVertexインスタンスのリストを保持します。

 public class PhysicalVertex { private final TaskGroupLocation taskGroupLocation; private final String taskFullName; private final TaskGroupDefaultImpl taskGroup; private final ExecutorService executorService; private final FlakeIdGenerator flakeIdGenerator; private final Set<URL> pluginJarsUrls; private final Set<ConnectorJarIdentifier> connectorJarIdentifiers; private final IMap<Object, Object> runningJobStateIMap; private CompletableFuture<TaskExecutionState> taskFuture; private final IMap<Object, Long[]> runningJobStateTimestampsIMap; private final NodeEngine nodeEngine; private JobMaster jobMaster; private volatile ExecutionState currExecutionState = ExecutionState.CREATED; public volatile boolean isRunning = false; private AtomicReference<String> errorByPhysicalVertex = new AtomicReference<>(); }

 public class TaskGroupDefaultImpl implements TaskGroup { private final TaskGroupLocation taskGroupLocation; private final String taskGroupName; // Stores the tasks that the physical node needs to execute // Each task could be for reading data, writing data, data splitting, checkpoint tasks, etc. private final Map<Long, Task> tasks; }

PhysicalPlanGenerator実行プランをSeaTunnelTask に変換し、実行中にデータ分割、データコミット、チェックポイントタスクなどのさまざまな調整タスクを追加する役割を担います。

 public PhysicalPlanGenerator( @NonNull ExecutionPlan executionPlan, @NonNull NodeEngine nodeEngine, @NonNull JobImmutableInformation jobImmutableInformation, long initializationTimestamp, @NonNull ExecutorService executorService, @NonNull FlakeIdGenerator flakeIdGenerator, @NonNull IMap runningJobStateIMap, @NonNull IMap runningJobStateTimestampsIMap, @NonNull QueueType queueType) { this.pipelines = executionPlan.getPipelines(); this.nodeEngine = nodeEngine; this.jobImmutableInformation = jobImmutableInformation; this.initializationTimestamp = initializationTimestamp; this.executorService = executorService; this.flakeIdGenerator = flakeIdGenerator; // the checkpoint of a pipeline this.pipelineTasks = new HashSet<>(); this.startingTasks = new HashSet<>(); this.subtaskActions = new HashMap<>(); this.runningJobStateIMap = runningJobStateIMap; this.runningJobStateTimestampsIMap = runningJobStateTimestampsIMap; this.queueType = queueType; } public Tuple2<PhysicalPlan, Map<Integer, CheckpointPlan>> generate() { // Get the node filter conditions from user configuration to select the nodes where tasks will run Map<String, String> tagFilter = (Map<String, String>) jobImmutableInformation .getJobConfig() .getEnvOptions() .get(EnvCommonOptions.NODE_TAG_FILTER.key()); // TODO Determine which tasks do not need to be restored according to state CopyOnWriteArrayList<PassiveCompletableFuture<PipelineStatus>> waitForCompleteBySubPlanList = new CopyOnWriteArrayList<>(); Map<Integer, CheckpointPlan> checkpointPlans = new HashMap<>(); final int totalPipelineNum = pipelines.size(); Stream<SubPlan> subPlanStream = pipelines.stream() .map( pipeline -> { // Clear the state each time this.pipelineTasks.clear(); this.startingTasks.clear(); this.subtaskActions.clear(); final int pipelineId = pipeline.getId(); // Get current task information final List<ExecutionEdge> edges = pipeline.getEdges(); // Get all SourceActions List<SourceAction<?, ?, ?>> sources = findSourceAction(edges); // Generate Source data slice tasks, ie, SourceSplitEnumeratorTask // This task calls the SourceSplitEnumerator class in the connector if supported List<PhysicalVertex> coordinatorVertexList = getEnumeratorTask( sources, pipelineId, totalPipelineNum); // Generate Sink commit tasks, ie, SinkAggregatedCommitterTask // This task calls the SinkAggregatedCommitter class in the connector if supported // These two tasks are executed as coordination tasks coordinatorVertexList.addAll( getCommitterTask(edges, pipelineId, totalPipelineNum)); List<PhysicalVertex> physicalVertexList = getSourceTask( edges, sources, pipelineId, totalPipelineNum); // physicalVertexList.addAll( getShuffleTask(edges, pipelineId, totalPipelineNum)); CompletableFuture<PipelineStatus> pipelineFuture = new CompletableFuture<>(); waitForCompleteBySubPlanList.add( new PassiveCompletableFuture<>(pipelineFuture)); // Add checkpoint tasks checkpointPlans.put( pipelineId, CheckpointPlan.builder() .pipelineId(pipelineId) .pipelineSubtasks(pipelineTasks) .startingSubtasks(startingTasks) .pipelineActions(pipeline.getActions()) .subtaskActions(subtaskActions) .build()); return new SubPlan( pipelineId, totalPipelineNum, initializationTimestamp, physicalVertexList, coordinatorVertexList, jobImmutableInformation, executorService, runningJobStateIMap, runningJobStateTimestampsIMap, tagFilter); }); PhysicalPlan physicalPlan = new PhysicalPlan( subPlanStream.collect(Collectors.toList()), executorService, jobImmutableInformation, initializationTimestamp, runningJobStateIMap, runningJobStateTimestampsIMap); return Tuple2.tuple2(physicalPlan, checkpointPlans); }

物理プランを生成するプロセスには、実行プランをSeaTunnelTask に変換し、データ分割タスク、データコミットタスク、チェックポイントタスクなどのさまざまな調整タスクを追加することが含まれます。

SeaTunnelTask では、タスクはSourceFlowLifeCycle 、 SinkFlowLifeCycle 、 TransformFlowLifeCycle 、 ShuffleSinkFlowLifeCycle 、 ShuffleSourceFlowLifeCycleに変換されます。

たとえば、 SourceFlowLifeCycleクラスとSinkFlowLifeCycleクラスは次のようになります。

ソースフローライフサイクル

@Override public void init() throws Exception { this.splitSerializer = sourceAction.getSource().getSplitSerializer(); this.reader = sourceAction .getSource() .createReader( new SourceReaderContext( indexID, sourceAction.getSource().getBoundedness(), this, metricsContext, eventListener)); this.enumeratorTaskAddress = getEnumeratorTaskAddress(); } @Override public void open() throws Exception { reader.open(); register(); } public void collect() throws Exception { if (!prepareClose) { if (schemaChanging()) { log.debug("schema is changing, stop reader collect records"); Thread.sleep(200); return; } reader.pollNext(collector); if (collector.isEmptyThisPollNext()) { Thread.sleep(100); } else { collector.resetEmptyThisPollNext(); /** * The current thread obtain a checkpoint lock in the method {@link * SourceReader#pollNext( Collector)}. When trigger the checkpoint or savepoint, * other threads try to obtain the lock in the method {@link * SourceFlowLifeCycle#triggerBarrier(Barrier)}. When high CPU load, checkpoint * process may be blocked as long time. So we need sleep to free the CPU. */ Thread.sleep(0L); } if (collector.captureSchemaChangeBeforeCheckpointSignal()) { if (schemaChangePhase.get() != null) { throw new IllegalStateException( "previous schema changes in progress, schemaChangePhase: " + schemaChangePhase.get()); } schemaChangePhase.set(SchemaChangePhase.createBeforePhase()); runningTask.triggerSchemaChangeBeforeCheckpoint().get(); log.info("triggered schema-change-before checkpoint, stopping collect data"); } else if (collector.captureSchemaChangeAfterCheckpointSignal()) { if (schemaChangePhase.get() != null) { throw new IllegalStateException( "previous schema changes in progress, schemaChangePhase: " + schemaChangePhase.get()); } schemaChangePhase.set(SchemaChangePhase.createAfterPhase()); runningTask.triggerSchemaChangeAfterCheckpoint().get(); log.info("triggered schema-change-after checkpoint, stopping collect data"); } } else { Thread.sleep(100); } }

SourceFlowLifeCycleでは、データの読み取りはcollectメソッドで実行されます。データが読み取られると、そのデータはSeaTunnelSourceCollectorに配置されます。データが受信されると、コレクターはメトリックを更新し、データを下流のコンポーネントに送信します。

 @Override public void collect(T row) { try { if (row instanceof SeaTunnelRow) { String tableId = ((SeaTunnelRow) row).getTableId(); int size; if (rowType instanceof SeaTunnelRowType) { size = ((SeaTunnelRow) row).getBytesSize((SeaTunnelRowType) rowType); } else if (rowType instanceof MultipleRowType) { size = ((SeaTunnelRow) row).getBytesSize(rowTypeMap.get(tableId)); } else { throw new SeaTunnelEngineException( "Unsupported row type: " + rowType.getClass().getName()); } sourceReceivedBytes.inc(size); sourceReceivedBytesPerSeconds.markEvent(size); flowControlGate.audit((SeaTunnelRow) row); if (StringUtils.isNotEmpty(tableId)) { String tableName = getFullName(TablePath.of(tableId)); Counter sourceTableCounter = sourceReceivedCountPerTable.get(tableName); if (Objects.nonNull(sourceTableCounter)) { sourceTableCounter.inc(); } else { Counter counter = metricsContext.counter(SOURCE_RECEIVED_COUNT + "#" + tableName); counter.inc(); sourceReceivedCountPerTable.put(tableName, counter); } } } sendRecordToNext(new Record<>(row)); emptyThisPollNext = false; sourceReceivedCount.inc(); sourceReceivedQPS.markEvent(); } catch (IOException e) { throw new RuntimeException(e); } } public void sendRecordToNext(Record<?> record) throws IOException { synchronized (checkpointLock) { for (OneInputFlowLifeCycle<Record<?>> output : outputs) { output.received(record); } } }

シンクフローライフサイクル

@Override public void received(Record<?> record) { try { if (record.getData() instanceof Barrier) { long startTime = System.currentTimeMillis(); Barrier barrier = (Barrier) record.getData(); if (barrier.prepareClose(this.taskLocation)) { prepareClose = true; } if (barrier.snapshot()) { try { lastCommitInfo = writer.prepareCommit(); } catch (Exception e) { writer.abortPrepare(); throw e; } List<StateT> states = writer.snapshotState(barrier.getId()); if (!writerStateSerializer.isPresent()) { runningTask.addState( barrier, ActionStateKey.of(sinkAction), Collections.emptyList()); } else { runningTask.addState( barrier, ActionStateKey.of(sinkAction), serializeStates(writerStateSerializer.get(), states)); } if (containAggCommitter) { CommitInfoT commitInfoT = null; if (lastCommitInfo.isPresent()) { commitInfoT = lastCommitInfo.get(); } runningTask .getExecutionContext() .sendToMember( new SinkPrepareCommitOperation<CommitInfoT>( barrier, committerTaskLocation, commitInfoSerializer.isPresent() ? commitInfoSerializer .get() .serialize(commitInfoT) : null), committerTaskAddress) .join(); } } else { if (containAggCommitter) { runningTask .getExecutionContext() .sendToMember( new BarrierFlowOperation(barrier, committerTaskLocation), committerTaskAddress) .join(); } } runningTask.ack(barrier); log.debug( "trigger barrier [{}] finished, cost {}ms. taskLocation [{}]", barrier.getId(), System.currentTimeMillis() - startTime, taskLocation); } else if (record.getData() instanceof SchemaChangeEvent) { if (prepareClose) { return; } SchemaChangeEvent event = (SchemaChangeEvent) record.getData(); writer.applySchemaChange(event); } else { if (prepareClose) { return; } writer.write((T) record.getData()); sinkWriteCount.inc(); sinkWriteQPS.markEvent(); if (record.getData() instanceof SeaTunnelRow) { long size = ((SeaTunnelRow) record.getData()).getBytesSize(); sinkWriteBytes.inc(size); sinkWriteBytesPerSeconds.markEvent(size); String tableId = ((SeaTunnelRow) record.getData()).getTableId(); if (StringUtils.isNotBlank(tableId)) { String tableName = getFullName(TablePath.of(tableId)); Counter sinkTableCounter = sinkWriteCountPerTable.get(tableName); if (Objects.nonNull(sinkTableCounter)) { sinkTableCounter.inc(); } else { Counter counter = metricsContext.counter(SINK_WRITE_COUNT + "#" + tableName); counter.inc(); sinkWriteCountPerTable.put(tableName, counter); } } } } } catch (Exception e) { throw new RuntimeException(e); } }

タスク実行

CoordinatorServiceでは、 initメソッドを通じて物理プランが生成され、次にrunメソッドが呼び出されて実際にタスクが開始されます。

 CoordinatorService { jobMaster.init( runningJobInfoIMap.get(jobId).getInitializationTimestamp(), false); ... jobMaster.run(); } JobMaster { public void run() { ... physicalPlan.startJob(); ... } }

JobMasterでは、タスクを開始するときにPhysicalPlanのstartJobメソッドを呼び出します。

 public void startJob() { isRunning = true; log.info("{} state process is start", getJobFullName()); stateProcess(); } private synchronized void stateProcess() { if (!isRunning) { log.warn(String.format("%s state process is stopped", jobFullName)); return; } switch (getJobStatus()) { case CREATED: updateJobState(JobStatus.SCHEDULED); break; case SCHEDULED: getPipelineList() .forEach( subPlan -> { if (PipelineStatus.CREATED.equals( subPlan.getCurrPipelineStatus())) { subPlan.startSubPlanStateProcess(); } }); updateJobState(JobStatus.RUNNING); break; case RUNNING: case DOING_SAVEPOINT: break; case FAILING: case CANCELING: jobMaster.neverNeedRestore(); getPipelineList().forEach(SubPlan::cancelPipeline); break; case FAILED: case CANCELED: case SAVEPOINT_DONE: case FINISHED: stopJobStateProcess(); jobEndFuture.complete(new JobResult(getJobStatus(), errorBySubPlan.get())); return; default: throw new IllegalArgumentException("Unknown Job State: " + getJobStatus()); } }

PhysicalPlanでは、タスクを開始するとタスクのステータスがSCHEDULEDに更新され、その後SubPlanの start メソッドが呼び出され続けます。

 public void startSubPlanStateProcess() { isRunning = true; log.info("{} state process is start", getPipelineFullName()); stateProcess(); } private synchronized void stateProcess() { if (!isRunning) { log.warn(String.format("%s state process not start", pipelineFullName)); return; } PipelineStatus state = getCurrPipelineStatus(); switch (state) { case CREATED: updatePipelineState(PipelineStatus.SCHEDULED); break; case SCHEDULED: try { ResourceUtils.applyResourceForPipeline(jobMaster.getResourceManager(), this); log.debug( "slotProfiles: {}, PipelineLocation: {}", slotProfiles, this.getPipelineLocation()); updatePipelineState(PipelineStatus.DEPLOYING); } catch (Exception e) { makePipelineFailing(e); } break; case DEPLOYING: coordinatorVertexList.forEach( task -> { if (task.getExecutionState().equals(ExecutionState.CREATED)) { task.startPhysicalVertex(); task.makeTaskGroupDeploy(); } }); physicalVertexList.forEach( task -> { if (task.getExecutionState().equals(ExecutionState.CREATED)) { task.startPhysicalVertex(); task.makeTaskGroupDeploy(); } }); updatePipelineState(PipelineStatus.RUNNING); break; case RUNNING: break; case FAILING: case CANCELING: coordinatorVertexList.forEach( task -> { task.startPhysicalVertex(); task.cancel(); }); physicalVertexList.forEach( task -> { task.startPhysicalVertex(); task.cancel(); }); break; case FAILED: case CANCELED: if (checkNeedRestore(state) && prepareRestorePipeline()) { jobMaster.releasePipelineResource(this); restorePipeline(); return; } subPlanDone(state); stopSubPlanStateProcess(); pipelineFuture.complete( new PipelineExecutionState(pipelineId, state, errorByPhysicalVertex.get())); return; case FINISHED: subPlanDone(state); stopSubPlanStateProcess(); pipelineFuture.complete( new PipelineExecutionState( pipelineId, getPipelineState(), errorByPhysicalVertex.get())); return; default: throw new IllegalArgumentException("Unknown Pipeline State: " + getPipelineState()); } }

SubPlanでは、すべてのタスクにリソースが適用されます。リソースの適用はResourceManagerを通じて行われます。リソースの適用中は、ユーザー定義のタグに基づいてノードが選択され、特定のノードでタスクが実行され、リソースの分離が実現されます。

 public static void applyResourceForPipeline( @NonNull ResourceManager resourceManager, @NonNull SubPlan subPlan) { Map<TaskGroupLocation, CompletableFuture<SlotProfile>> futures = new HashMap<>(); Map<TaskGroupLocation, SlotProfile> slotProfiles = new HashMap<>(); // TODO If there is no enough resources for tasks, we need add some wait profile subPlan.getCoordinatorVertexList() .forEach( coordinator -> futures.put( coordinator.getTaskGroupLocation(), applyResourceForTask( resourceManager, coordinator, subPlan.getTags()))); subPlan.getPhysicalVertexList() .forEach( task -> futures.put( task.getTaskGroupLocation(), applyResourceForTask( resourceManager, task, subPlan.getTags()))); futures.forEach( (key, value) -> { try { slotProfiles.put(key, value == null ? null : value.join()); } catch (CompletionException e) { // do nothing } }); // set it first, avoid can't get it when get resource not enough exception and need release // applied resource subPlan.getJobMaster().setOwnedSlotProfiles(subPlan.getPipelineLocation(), slotProfiles); if (futures.size() != slotProfiles.size()) { throw new NoEnoughResourceException(); } } public static CompletableFuture<SlotProfile> applyResourceForTask( ResourceManager resourceManager, PhysicalVertex task, Map<String, String> tags) { // TODO custom resource size return resourceManager.applyResource( task.getTaskGroupLocation().getJobId(), new ResourceProfile(), tags); } public CompletableFuture<List<SlotProfile>> applyResources( long jobId, List<ResourceProfile> resourceProfile, Map<String, String> tagFilter) throws NoEnoughResourceException { waitingWorkerRegister(); ConcurrentMap<Address, WorkerProfile> matchedWorker = filterWorkerByTag(tagFilter); if (matchedWorker.isEmpty()) { log.error("No matched worker with tag filter {}.", tagFilter); throw new NoEnoughResourceException(); } return new ResourceRequestHandler(jobId, resourceProfile, matchedWorker, this) .request(tagFilter); }

利用可能なノードがすべて取得されると、ノードがシャッフルされ、必要なリソースよりも大きいリソースを持つノードがランダムに選択されます。次に、そのノードに接続され、 RequestSlotOperationが送信されます。

 public Optional<WorkerProfile> preCheckWorkerResource(ResourceProfile r) { // Shuffle the order to ensure random selection of workers List<WorkerProfile> workerProfiles = Arrays.asList(registerWorker.values().toArray(new WorkerProfile[0])); Collections.shuffle(workerProfiles); // Check if there are still unassigned slots Optional<WorkerProfile> workerProfile = workerProfiles.stream() .filter( worker -> Arrays.stream(worker.getUnassignedSlots()) .anyMatch( slot -> slot.getResourceProfile() .enoughThan(r))) .findAny(); if (!workerProfile.isPresent()) { // Check if there are still unassigned resources workerProfile = workerProfiles.stream() .filter(WorkerProfile::isDynamicSlot) .filter(worker -> worker.getUnassignedResource().enoughThan(r)) .findAny(); } return workerProfile; } private CompletableFuture<SlotAndWorkerProfile> singleResourceRequestToMember( int i, ResourceProfile r, WorkerProfile workerProfile) { CompletableFuture<SlotAndWorkerProfile> future = resourceManager.sendToMember( new RequestSlotOperation(jobId, r), workerProfile.getAddress()); return future.whenComplete( withTryCatch( LOGGER, (slotAndWorkerProfile, error) -> { if (error != null) { throw new RuntimeException(error); } else { resourceManager.heartbeat(slotAndWorkerProfile.getWorkerProfile()); addSlotToCacheMap(i, slotAndWorkerProfile.getSlotProfile()); } })); }

ノードのSlotService requestSlotリクエストを受信すると、自身の情報を更新し、それをマスターノードに返します。リソースリクエストが期待される結果を満たさない場合、タスクの失敗を示すNoEnoughResourceExceptionがスローされます。リソースの割り当てが成功すると、 task.makeTaskGroupDeploy()でタスクのデプロイメントが開始され、タスクがworkerノードに送信されて実行されます。

 TaskDeployState deployState = deploy(jobMaster.getOwnedSlotProfiles(taskGroupLocation)); public TaskDeployState deploy(@NonNull SlotProfile slotProfile) { try { if (slotProfile.getWorker().equals(nodeEngine.getThisAddress())) { return deployOnLocal(slotProfile); } else { return deployOnRemote(slotProfile); } } catch (Throwable th) { return TaskDeployState.failed(th); } } private TaskDeployState deployOnRemote(@Non Null SlotProfile slotProfile) { return deployInternal( taskGroupImmutableInformation -> { try { return (TaskDeployState) NodeEngineUtil.sendOperationToMemberNode( nodeEngine, new DeployTaskOperation( slotProfile, nodeEngine .getSerializationService() .toData( taskGroupImmutableInformation)), slotProfile.getWorker()) .get(); } catch (Exception e) { if (getExecutionState().isEndState()) { log.warn(ExceptionUtils.getMessage(e)); log.warn( String.format( "%s deploy error, but the state is already in end state %s, skip this error", getTaskFullName(), currExecutionState)); return TaskDeployState.success(); } else { return TaskDeployState.failed(e); } } }); }

タスクの展開

タスクをデプロイすると、タスク情報はリソース割り当て中に取得されたノードに送信されます。

 public TaskDeployState deployTask(@NonNull Data taskImmutableInformation) { TaskGroupImmutableInformation taskImmutableInfo = nodeEngine.getSerializationService().toObject(taskImmutableInformation); return deployTask(taskImmutableInfo); } public TaskDeployState deployTask(@NonNull TaskGroupImmutableInformation taskImmutableInfo) { logger.info( String.format( "received deploying task executionId [%s]", taskImmutableInfo.getExecutionId())); TaskGroup taskGroup = null; try { Set<ConnectorJarIdentifier> connectorJarIdentifiers = taskImmutableInfo.getConnectorJarIdentifiers(); Set<URL> jars = new HashSet<>(); ClassLoader classLoader; if (!CollectionUtils.isEmpty(connectorJarIdentifiers)) { // Prioritize obtaining the jar package file required for the current task execution // from the local, if it does not exist locally, it will be downloaded from the // master node. jars = serverConnectorPackageClient.getConnectorJarFromLocal( connectorJarIdentifiers); } else if (!CollectionUtils.isEmpty(taskImmutableInfo.getJars())) { jars = taskImmutableInfo.getJars(); } classLoader = classLoaderService.getClassLoader( taskImmutableInfo.getJobId(), Lists.newArrayList(jars)); if (jars.isEmpty()) { taskGroup = nodeEngine.getSerializationService().toObject(taskImmutableInfo.getGroup()); } else { taskGroup = CustomClassLoadedObject.deserializeWithCustomClassLoader( nodeEngine.getSerializationService(), classLoader, taskImmutableInfo.getGroup()); } logger.info( String.format( "deploying task %s, executionId [%s]", taskGroup.getTaskGroupLocation(), taskImmutableInfo.getExecutionId())); synchronized (this) { if (executionContexts.containsKey(taskGroup.getTaskGroupLocation())) { throw new RuntimeException( String.format( "TaskGroupLocation: %s already exists", taskGroup.getTaskGroupLocation())); } deployLocalTask(taskGroup, classLoader, jars); return TaskDeployState.success(); } } catch (Throwable t) { logger.severe( String.format( "TaskGroupID : %s deploy error with Exception: %s", taskGroup != null && taskGroup.getTaskGroupLocation() != null ? taskGroup.getTaskGroupLocation().toString() : "taskGroupLocation is null", ExceptionUtils.getMessage(t))); return TaskDeployState.failed(t); } }

ワーカーノードはタスクを受信すると、 TaskExecutionServiceのdeployTaskメソッドを呼び出して、起動時に作成されたスレッドプールにタスクを送信します。

タスクがスレッドプールに送信されると、次のようになります。

 private final class BlockingWorker implements Runnable { private final TaskTracker tracker; private final CountDownLatch startedLatch; private BlockingWorker(TaskTracker tracker, CountDownLatch startedLatch) { this.tracker = tracker; this.startedLatch = startedLatch; } @Override public void run() { TaskExecutionService.TaskGroupExecutionTracker taskGroupExecutionTracker = tracker.taskGroupExecutionTracker; ClassLoader classLoader = executionContexts .get(taskGroupExecutionTracker.taskGroup.getTaskGroupLocation()) .getClassLoader(); ClassLoader oldClassLoader = Thread.currentThread().getContextClassLoader(); Thread.currentThread().setContextClassLoader(classLoader); final Task t = tracker.task; ProgressState result = null; try { startedLatch.countDown(); t.init(); do { result = t.call(); } while (!result.isDone() && isRunning && !taskGroupExecutionTracker.executionCompletedExceptionally()); ... } }

Task.callメソッドが呼び出され、データ同期タスクが実際に実行されます。

クラスローダー

SeaTunnel では、他のコンポーネントクラスとの競合を避けるためにサブクラスを優先するようにデフォルトの ClassLoader が変更されました。

 @Override public synchronized ClassLoader getClassLoader(long jobId, Collection<URL> jars) { log.debug("Get classloader for job {} with jars {}", jobId, jars); if (cacheMode) { // with cache mode, all jobs share the same classloader if the jars are the same jobId = 1L; } if (!classLoaderCache.containsKey(jobId)) { classLoaderCache.put(jobId, new ConcurrentHashMap<>()); classLoaderReferenceCount.put(jobId, new ConcurrentHashMap<>()); } Map<String, ClassLoader> classLoaderMap = classLoaderCache.get(jobId); String key = covertJarsToKey(jars); if (classLoaderMap.containsKey(key)) { classLoaderReferenceCount.get(jobId).get(key).incrementAndGet(); return classLoaderMap.get(key); } else { ClassLoader classLoader = new SeaTunnelChildFirstClassLoader(jars); log.info("Create classloader for job {} with jars {}", jobId, jars); classLoaderMap.put(key, classLoader); classLoaderReferenceCount.get(jobId).put(key, new AtomicInteger(1)); return classLoader; } }

REST API タスク送信

SeaTunnel は、REST API 経由のタスク送信もサポートしています。この機能を有効にするには、 hazelcast.yamlファイルに次の構成を追加します。

 network: rest-api: enabled: true endpoint-groups: CLUSTER_WRITE: enabled: true DATA: enabled: true

この構成を追加すると、Hazelcast ノードは HTTP リクエストを受信できるようになります。

タスクの送信に REST API を使用すると、クライアントは HTTP 要求を送信するノードになり、サーバーは SeaTunnel クラスターになります。

サーバーはリクエストを受信すると、リクエスト URI に基づいて適切なメソッドを呼び出します。

 public void handle(HttpPostCommand httpPostCommand) { String uri = httpPostCommand.getURI(); try { if (uri.startsWith(SUBMIT_JOB_URL)) { handleSubmitJob(httpPostCommand, uri); } else if (uri.startsWith(STOP_JOB_URL)) { handleStopJob(httpPostCommand, uri); } else if (uri.startsWith(ENCRYPT_CONFIG)) { handleEncrypt(httpPostCommand); } else { original.handle(httpPostCommand); } } catch (IllegalArgumentException e) { prepareResponse(SC_400, httpPostCommand, exceptionResponse(e)); } catch (Throwable e) { logger.warning("An error occurred while handling request " + httpPostCommand, e); prepareResponse(SC_500, httpPostCommand, exceptionResponse(e)); } this.textCommandService.sendResponse(httpPostCommand); }

ジョブ送信要求を処理する方法は、パスによって決まります。

 private void handleSubmitJob(HttpPostCommand httpPostCommand, String uri) throws IllegalArgumentException { Map<String, String> requestParams = new HashMap<>(); RestUtil.buildRequestParams(requestParams, uri); Config config = RestUtil.buildConfig(requestHandle(httpPostCommand), false); ReadonlyConfig envOptions = ReadonlyConfig.fromConfig(config.getConfig("env")); String jobName = envOptions.get(EnvCommonOptions.JOB_NAME); JobConfig jobConfig = new JobConfig(); jobConfig.setName( StringUtils.isEmpty(requestParams.get(RestConstant.JOB_NAME)) ? jobName : requestParams.get(RestConstant.JOB_NAME)); boolean startWithSavePoint = Boolean.parseBoolean(requestParams.get(RestConstant.IS_START_WITH_SAVE_POINT)); String jobIdStr = requestParams.get(RestConstant.JOB_ID); Long finalJobId = StringUtils.isNotBlank(jobIdStr) ? Long.parseLong(jobIdStr) : null; SeaTunnelServer seaTunnelServer = getSeaTunnelServer(); RestJobExecutionEnvironment restJobExecutionEnvironment = new RestJobExecutionEnvironment( seaTunnelServer, jobConfig, config, textCommandService.getNode(), startWithSavePoint, finalJobId); JobImmutableInformation jobImmutableInformation = restJobExecutionEnvironment.build(); long jobId = jobImmutableInformation.getJobId(); if (!seaTunnelServer.isMasterNode()) { NodeEngineUtil.sendOperationToMasterNode( getNode().nodeEngine, new SubmitJobOperation( jobId, getNode().nodeEngine.toData(jobImmutableInformation), jobImmutableInformation.isStartWithSavePoint())) .join(); } else { submitJob(seaTunnelServer, jobImmutableInformation, jobConfig); } this.prepareResponse( httpPostCommand, new JsonObject() .add(RestConstant.JOB_ID, String.valueOf(jobId)) .add(RestConstant.JOB_NAME, jobConfig.getName())); }

ここでのロジックはクライアント側と同様です。ローカルモードがないため、ローカルサービスを作成する必要はありません。

クライアント側では、 ClientJobExecutionEnvironmentクラスが論理プランの解析に使用され、同様に、 RestJobExecutionEnvironmentクラスが同じタスクを実行します。

タスクを送信する際、現在のノードがマスターノードでない場合は、マスターノードに情報が送信されます。マスターノードは、コマンドラインクライアントからのコマンドを処理するのと同様に、タスクの送信を処理します。

現在のノードがマスターノードである場合は、 submitJobメソッドを直接呼び出し、後続の処理のためにcoordinatorService.submitJobメソッドを呼び出します。

 private void submitJob( SeaTunnelServer seaTunnelServer, JobImmutableInformation jobImmutableInformation, JobConfig jobConfig) { CoordinatorService coordinatorService = seaTunnelServer.getCoordinatorService(); Data data = textCommandService .getNode() .nodeEngine .getSerializationService() .toData(jobImmutableInformation); PassiveCompletableFuture<Void> voidPassiveCompletableFuture = coordinatorService.submitJob( Long.parseLong(jobConfig.getJobContext().getJobId()), data, jobImmutableInformation.isStartWithSavePoint()); voidPassiveCompletableFuture.join(); }

どちらの送信方法でも、送信側で論理プランを解析し、その情報をマスターノードに送信します。その後、マスターノードは物理プランの解析、割り当て、およびその他の操作を実行します。