hudi 数据写入中的生产者-消费者模式

生产者-消费者模式用来协调数据生产和消费速度不一致问题，在hudi中数据写入时非常依赖该设计模式，且中间涉及一些比较好用的工具类可以直接拿来用，例如：ObjectSizeCalculator 来预估对象实例大小、BoundedInMemoryQueue 通过对象大小来调整记录上限的队列。

涉及的主要类

BoundedInMemoryQueue

基于内存的有界队列，特点：

• 预先分配队列内存，通过数据采样来预估队列中实例占用字节大小，并动态调整队列上限。
• 通过信号量 Semaphore 进行并发控制，代表要缓存的最大记录数。
• 支持转换函数，生产者数据通过转换函数进行转换后写入队列。

主要代码流程：

• 队列新增数据：

//  生产者向队列插入一条记录
public void insertRecord(I t) throws Exception {
    // If already closed, throw exception
    if (isWriteDone.get()) {
      throw new IllegalStateException("Queue closed for enqueueing new entries");
    }
    // We need to stop queueing if queue-reader has failed and exited.
    throwExceptionIfFailed();
    //  获取信号量 -1，可以会阻塞
    rateLimiter.acquire();
    // We are retrieving insert value in the record queueing thread to offload computation
    // around schema validation
    // and record creation to it.

    //  对记录进行类型转换
    //  HoodieRecord ---->  HoodieInsertValueGenResult
    final O payload = transformFunction.apply(t);

    // 通过记录采样，调整队列上限
    adjustBufferSizeIfNeeded(payload);

    // 存储到内部的阻塞队列 LinkedBlockingQueue
    queue.put(Option.of(payload));
  }

• 队列记录数上限调整

//  通过采样调整队列上限  
private void adjustBufferSizeIfNeeded(final O payload) throws InterruptedException {
    //  数据大小采样判断，默认64条记录采样一次
    if (this.samplingRecordCounter.incrementAndGet() % RECORD_SAMPLING_RATE != 0) {
      return;
    }
    //  通过 ObjectSizeCalculator.getObjectSize(t); 获取对象大小
    final long recordSizeInBytes = payloadSizeEstimator.sizeEstimate(payload);
    //  基于历史数据计算记录的平均字节大小
    final long newAvgRecordSizeInBytes =
        Math.max(1, (avgRecordSizeInBytes * numSamples + recordSizeInBytes) / (numSamples + 1));

    // 当前允许缓存的数据条数
    final int newRateLimit =
        (int) Math.min(RECORD_CACHING_LIMIT, Math.max(1, this.memoryLimit / newAvgRecordSizeInBytes));

    //  如果要缓存的记录数量有任何变化，那么我们增加或减少信号量，以将速率限制调整为新计算的值。
    if (newRateLimit > currentRateLimit) {
      //  增加信号量
      rateLimiter.release(newRateLimit - currentRateLimit);
    } else if (newRateLimit < currentRateLimit) {
      //  减少信号量
      rateLimiter.acquire(currentRateLimit - newRateLimit);
    }

    currentRateLimit = newRateLimit;

    avgRecordSizeInBytes = newAvgRecordSizeInBytes;

    numSamples++;
  }

• 消费队列数据：

private Option readNextRecord() {
  if (this.isReadDone.get()) {
    return Option.empty();
  }
  // 释放信号量
  rateLimiter.release();

  Option newRecord = Option.empty();
  while (expectMoreRecords()) {
    try {
      throwExceptionIfFailed();
      // 带有超时时间的poll
      newRecord = queue.poll(RECORD_POLL_INTERVAL_SEC, TimeUnit.SECONDS);
      if (newRecord != null) {
        break;
      }
    } catch (InterruptedException e) {
      LOG.error("error reading records from queue", e);
      throw new HoodieException(e);
    }
  }
  // Check one more time here as it is possible producer erred out and closed immediately
  throwExceptionIfFailed();

  if (newRecord != null && newRecord.isPresent()) {
    return newRecord;
  } else {
    // We are done reading all the records from internal iterator.
    this.isReadDone.set(true);
    return Option.empty();
  }
}

BoundedInMemoryQueueConsumer

消费者实现类

消费者会将数据下发给 HoodieWriteHandle 进行写入。flink 使用 ExplicitWriteHandler 操作 FlinkAppendHandle、FlinkCreateHandle ， 使用 UpdateHandler 操作 FlinkMergeHandle 完成数据的写入。

/**
 * Consume entries from queue and execute callback function.
 */
public abstract class BoundedInMemoryQueueConsumer {

  /**
   *  BoundedInMemoryExecutor 调用，启动消费者来下发数据
   * API to de-queue entries to memory bounded queue.
   *
   * @param queue In Memory bounded queue
   */
  public O consume(BoundedInMemoryQueue queue) throws Exception {
    Iterator iterator = queue.iterator();

    while (iterator.hasNext()) {
      // 消费记录并下发给 HoodieWriteHandle
      consumeOneRecord(iterator.next());
    }
    // 队列数据消费完毕，后置处理
    // Notifies done
    finish();
    // 返回状态统计信息
    return getResult();
  }
  /**
   * Consumer One record.
   */
  protected abstract void consumeOneRecord(I record);
  /**
   * Notifies implementation that we have exhausted consuming records from queue.
   */
  protected abstract void finish();
  /**
   * Return result of consuming records so far.
   */
  protected abstract O getResult();

}

BoundedInMemoryQueueProducer

生产者接口及实现类

• IteratorBasedQueueProducer 基于迭代器来生产记录，在插入、更新时使用。
• FunctionBasedQueueProducer 基于Function来生产记录，在合并日志log文件和数据parquet文件时使用，以便提供RealTimeView。

public class IteratorBasedQueueProducer implements BoundedInMemoryQueueProducer {
  // input iterator for producing items in the buffer.
  private final Iterator inputIterator;

  public IteratorBasedQueueProducer(Iterator inputIterator) {
    this.inputIterator = inputIterator;
  }

  @Override
  public void produce(BoundedInMemoryQueue queue) throws Exception {
    LOG.info("starting to buffer records");
    // 从传递的迭代器来生产数据
    while (inputIterator.hasNext()) {
      queue.insertRecord(inputIterator.next());
    }
    LOG.info("finished buffering records");
  }
}

public class FunctionBasedQueueProducer implements BoundedInMemoryQueueProducer {

  private static final Logger LOG = LogManager.getLogger(FunctionBasedQueueProducer.class);

  private final Function, Boolean> producerFunction;

  public FunctionBasedQueueProducer(Function, Boolean> producerFunction) {
    this.producerFunction = producerFunction;
  }

  @Override
  public void produce(BoundedInMemoryQueue queue) {
    LOG.info("starting function which will enqueue records");
    producerFunction.apply(queue);
    LOG.info("finished function which will enqueue records");
  }
}

BoundedInMemoryExecutor

通过有界的内存队列协调生产者和生产者的执行器。

public class BoundedInMemoryExecutor {
  // Executor service used for launching writer thread.
  private final ExecutorService executorService;
  // Used for buffering records which is controlled by HoodieWriteConfig#WRITE_BUFFER_LIMIT_BYTES.
  private final BoundedInMemoryQueue queue;
  // Producers
  private final List> producers;
  // Consumer
  private final Option> consumer;

  public BoundedInMemoryExecutor(final long bufferLimitInBytes, BoundedInMemoryQueueProducer producer,
      Option> consumer, final Function transformFunction) {
    this(bufferLimitInBytes, Arrays.asList(producer), consumer, transformFunction, new DefaultSizeEstimator<>());
  }

  public BoundedInMemoryExecutor(final long bufferLimitInBytes, List> producers,
      Option> consumer, final Function transformFunction,
      final SizeEstimator sizeEstimator) {
    this.producers = producers;
    this.consumer = consumer;
    //  线程数量为 producers + consumer
    // Ensure single thread for each producer thread and one for consumer
    this.executorService = Executors.newFixedThreadPool(producers.size() + 1);
    //  基于内存的有界队列
    this.queue = new BoundedInMemoryQueue<>(bufferLimitInBytes, transformFunction, sizeEstimator);
  }

  /**
   * Callback to implement environment specific behavior before executors (producers/consumer) run.
   */
  public void preExecute() {
    // Do Nothing in general context
  }

  /**
   *   启动全部生产者
   * Start all Producers.
   */
  public ExecutorCompletionService startProducers() {
    //  Latch 控制何时以及哪个生产者线程将关闭队列
    // Latch to control when and which producer thread will close the queue
    final CountDownLatch latch = new CountDownLatch(producers.size());
    final ExecutorCompletionService completionService =
        new ExecutorCompletionService(executorService);

    producers.stream().map(producer -> {
      return completionService.submit(() -> {
        try {
          preExecute();
          // 生产者下发数据到queue
          producer.produce(queue);
        } catch (Throwable e) {
          LOG.error("error producing records", e);
          queue.markAsFailed(e);
          throw e;
        } finally {
          synchronized (latch) {
            latch.countDown();
            if (latch.getCount() == 0) {
              // 将生产标记为已完成，以便消费者能够退出
              // Mark production as done so that consumer will be able to exit
              queue.close();
            }
          }
        }
        return true;
      });
    }).collect(Collectors.toList());
    return completionService;
  }

  /**
   *  启动唯一消费者
   * Start only consumer.
   */
  private Future startConsumer() {
    return consumer.map(consumer -> {
      return executorService.submit(() -> {
        LOG.info("starting consumer thread");
        preExecute();
        try {
          E result = consumer.consume(queue);
          LOG.info("Queue Consumption is done; notifying producer threads");
          return result;
        } catch (Exception e) {
          LOG.error("error consuming records", e);
          queue.markAsFailed(e);
          throw e;
        }
      });
    }).orElse(CompletableFuture.completedFuture(null));
  }

  /**
   *  指定入口
   * Main API to run both production and consumption.
   */
  public E execute() {
    try {
      // 1. 启动生产者，生产数据
      startProducers();
      // 2. 消费数据
      Future future = startConsumer();
      // Wait for consumer to be done
      // 阻塞拿结果
      return future.get();
    } catch (InterruptedException ie) {
      shutdownNow();
      Thread.currentThread().interrupt();
      throw new HoodieException(ie);
    } catch (Exception e) {
      throw new HoodieException(e);
    }
  }

  public boolean isRemaining() {
    return queue.iterator().hasNext();
  }

  public void shutdownNow() {
    executorService.shutdownNow();
  }

  public BoundedInMemoryQueue getQueue() {
    return queue;
  }
}

FlinkLazyInsertIterable

flink 使用的迭代器

通过迭代器来构建 BoundedInMemoryExecutor 将要插入的数据进行下发，computeNext() 为调用入口。
构造函数指定了下发数据集合迭代器，及HoodieWriteHandle的工厂类。

public class FlinkLazyInsertIterable extends HoodieLazyInsertIterable {

  public FlinkLazyInsertIterable(Iterator> recordItr,
                                 boolean areRecordsSorted,
                                 HoodieWriteConfig config,
                                 String instantTime,
                                 HoodieTable hoodieTable,
                                 String idPrefix,
                                 TaskContextSupplier taskContextSupplier,
                                 ExplicitWriteHandleFactory writeHandleFactory) {
    super(recordItr, areRecordsSorted, config, instantTime, hoodieTable, idPrefix, taskContextSupplier, writeHandleFactory);
  }

  /**
   *   执行入口
   * @return
   */
  @Override
  protected List computeNext() {
    // Executor service used for launching writer thread.
    BoundedInMemoryExecutor, HoodieInsertValueGenResult, List> bufferedIteratorExecutor =
        null;
    try {
      final Schema schema = new Schema.Parser().parse(hoodieConfig.getSchema());
      //  创建 BoundedInMemoryExecutor
      bufferedIteratorExecutor = new BoundedInMemoryExecutor<>(hoodieConfig.getWriteBufferLimitBytes(),
          new IteratorBasedQueueProducer<>(inputItr),  // 生产者，对象创建时传递了要下发数据的集合的 Iterator
          Option.of(getExplicitInsertHandler()),      // 消费者，ExplicitWriteHandleFactory 用来创建 HoodieWriteHandle
          getTransformFunction(schema, hoodieConfig)); // 转换函数

      // 开始执行，阻塞获取数据写入的状态信息 WriteStatus
      final List result = bufferedIteratorExecutor.execute();
      assert result != null && !result.isEmpty() && !bufferedIteratorExecutor.isRemaining();
      return result;
    } catch (Exception e) {
      throw new HoodieException(e);
    } finally {
      if (null != bufferedIteratorExecutor) {
        bufferedIteratorExecutor.shutdownNow();
      }
    }
  }

  @SuppressWarnings("rawtypes")
  private ExplicitWriteHandler getExplicitInsertHandler() {
    HoodieWriteHandle handle = ((ExplicitWriteHandleFactory) writeHandleFactory).getWriteHandle();
    //  HoodieAppendHandle
    return new ExplicitWriteHandler(handle);
  }
}

FlinkMergeHelper

用来做merge数据时使用，生产者为历史basefile数据，消费者为 FlinkMergeHandle。FlinkMergeHandle在初始化时已经包含了 logfile的全部数据。

public class FlinkMergeHelper extends AbstractMergeHelper>,
    List, List> {

  private FlinkMergeHelper() {
  }

  private static class MergeHelperHolder {
    private static final FlinkMergeHelper FLINK_MERGE_HELPER = new FlinkMergeHelper();
  }

  public static FlinkMergeHelper newInstance() {
    return FlinkMergeHelper.MergeHelperHolder.FLINK_MERGE_HELPER;
  }

  @Override
  public void runMerge(HoodieTable>, List, List> table,
                       HoodieMergeHandle>, List, List> mergeHandle) throws IOException {
    final GenericDatumWriter gWriter;
    final GenericDatumReader gReader;
    Schema readSchema;
    // Read records from previous version of base file and merge.
    final boolean externalSchemaTransformation = table.getConfig().shouldUseExternalSchemaTransformation();
    HoodieBaseFile baseFile = mergeHandle.baseFileForMerge();

    if (externalSchemaTransformation || baseFile.getBootstrapBaseFile().isPresent()) {
      readSchema = HoodieFileReaderFactory.getFileReader(table.getHadoopConf(), mergeHandle.getOldFilePath()).getSchema();
      gWriter = new GenericDatumWriter<>(readSchema);
      gReader = new GenericDatumReader<>(readSchema, mergeHandle.getWriterSchemaWithMetaFields());
    } else {
      // flink 走这 ==>
      gReader = null;
      gWriter = null;
      readSchema = mergeHandle.getWriterSchemaWithMetaFields();
    }

    BoundedInMemoryExecutor wrapper = null;
    Configuration cfgForHoodieFile = new Configuration(table.getHadoopConf());
    //  构建 base file 读取器
    HoodieFileReader reader = HoodieFileReaderFactory.getFileReader(cfgForHoodieFile, mergeHandle.getOldFilePath());
    try {
      final Iterator readerIterator;
      if (baseFile.getBootstrapBaseFile().isPresent()) {
        readerIterator = getMergingIterator(table, mergeHandle, baseFile, reader, readSchema, externalSchemaTransformation);
      } else {
        // flink ==>   读取 base file （历史parquet）
        readerIterator = reader.getRecordIterator(readSchema);
      }

      ThreadLocal encoderCache = new ThreadLocal<>();
      ThreadLocal decoderCache = new ThreadLocal<>();

      wrapper = new BoundedInMemoryExecutor(table.getConfig().getWriteBufferLimitBytes(),
          new IteratorBasedQueueProducer<>(readerIterator),  // 生产者
          Option.of(new UpdateHandler(mergeHandle)),        //  消费者
          record -> {
              if (!externalSchemaTransformation) {
                // flink ==>  不进行转换
                return record;
              }
             return transformRecordBasedOnNewSchema(gReader, gWriter, encoderCache, decoderCache, (GenericRecord) record);
            }
      );

      wrapper.execute();
    } catch (Exception e) {
      throw new HoodieException(e);
    } finally {
      if (reader != null) {
        // parquet
        reader.close();
      }

      mergeHandle.close();
      if (null != wrapper) {
        wrapper.shutdownNow();
      }
    }
  }
}

后续主要关注 FlinkMergeHelper及FlinkLazyInsertIterable 在flink 调用中的构建。