首页 > 代码库 > 【Flume】【*】深入flume-ng的三大组件——source，channel，sink

【Flume】【*】深入flume-ng的三大组件——source，channel，sink

2024-11-23 02:24:39   202人阅读

概览

flume-ng中最重要的核心三大组件就是source,channel,sink
source负责从源端收集数据，产出event
channel负责暂存event，以备下游取走消费
sink负责消费通道中的event，写到最终的输出端上
技术分享

以上是总体的一个简单结构图，下面我们来深入每一个组件的内部看看：

1、Source

source接口的定义如下：
@InterfaceAudience.Public
@InterfaceStability.Stable
public interface Source extends LifecycleAware, NamedComponent {
  /**
   * Specifies which channel processor will handle this source‘s events.
   *
   * @param channelProcessor
   */
  public void setChannelProcessor(ChannelProcessor channelProcessor);
  /**
   * Returns the channel processor that will handle this source‘s events.
   */
  public ChannelProcessor getChannelProcessor();
}
source生成event并且调用配置的channelprocessor的相关方法，持续的将events存入配置的channel里
channelProcessor中有通道选择器和拦截器链，该过程处在source端收到数据和放入通道直接

该source继承了两个类

1、NamedComponent 

负责给每一个组件取一个唯一标识，就是名字，这个名字来源于我们的配置

2、LifecycleAware

负责组件的启停和状态维护

Source接口的直接实现类是AbstractSource抽象类

该类中就定义了通道处理器
还有一个生命状态周期的枚举类型
public enum LifecycleState {
  IDLE, START, STOP, ERROR;
  public static final LifecycleState[] START_OR_ERROR = new LifecycleState[] {
      START, ERROR };
  public static final LifecycleState[] STOP_OR_ERROR = new LifecycleState[] {
      STOP, ERROR };
}
这里就定义了一个组件会有的4种状态
实现接口的启停组件方法，方法体中只有一个状态的赋值，具体实现，我们来看一个具体的Source——ExecSource
读取配置方面很简单，这里就不说了，看下start方法
public void start() {
    logger.info("Exec source starting with command:{}", command);
    executor = Executors.newSingleThreadExecutor();
    runner = new ExecRunnable(shell, command, getChannelProcessor(), sourceCounter,
        restart, restartThrottle, logStderr, bufferCount, batchTimeout, charset);
    // FIXME: Use a callback-like executor / future to signal us upon failure.
    runnerFuture = executor.submit(runner);
    /*
     * NB: This comes at the end rather than the beginning of the method because
     * it sets our state to running. We want to make sure the executor is alive
     * and well first.
     */
    sourceCounter.start();
    super.start();
    logger.debug("Exec source started");
  }
该方法内部就是启动了一个线程去执行我们配置的终端命令
前面一篇文章也说过从入口分析如何调用到该start方法，以及调用频率：http://blog.csdn.net/simonchi/article/details/42970373


2、channel

对于通道来说，最重要的就是event的维护，flume的核心就是要中转这些event，所以event一定不能出事

Channel接口定义如下：

@InterfaceAudience.Public
@InterfaceStability.Stable
public interface Channel extends LifecycleAware, NamedComponent {
  /**
   * <p>Puts the given event into the channel.</p>
   * <p><strong>Note</strong>: This method must be invoked within an active
   * {@link Transaction} boundary. Failure to do so can lead to unpredictable
   * results.</p>
   * @param event the event to transport.
   * @throws ChannelException in case this operation fails.
   * @see org.apache.flume.Transaction#begin()
   */
  public void put(Event event) throws ChannelException;
  /**
   * <p>Returns the next event from the channel if available. If the channel
   * does not have any events available, this method must return {@code null}.
   * </p>
   * <p><strong>Note</strong>: This method must be invoked within an active
   * {@link Transaction} boundary. Failure to do so can lead to unpredictable
   * results.</p>
   * @return the next available event or {@code null} if no events are
   * available.
   * @throws ChannelException in case this operation fails.
   * @see org.apache.flume.Transaction#begin()
   */
  public Event take() throws ChannelException;
  /**
   * @return the transaction instance associated with this channel.
   */
  public Transaction getTransaction();
}
通道中的event全部都在事务的管理之中
先来看看这个事务的定义
public interface Transaction {
public enum TransactionState {Started, Committed, RolledBack, Closed };
  /**
   * <p>Starts a transaction boundary for the current channel operation. If a
   * transaction is already in progress, this method will join that transaction
   * using reference counting.</p>
   * <p><strong>Note</strong>: For every invocation of this method there must
   * be a corresponding invocation of {@linkplain #close()} method. Failure
   * to ensure this can lead to dangling transactions and unpredictable results.
   * </p>
   */
  public void begin();
  /**
   * Indicates that the transaction can be successfully committed. It is
   * required that a transaction be in progress when this method is invoked.
   */
  public void commit();
  /**
   * Indicates that the transaction can must be aborted. It is
   * required that a transaction be in progress when this method is invoked.
   */
  public void rollback();
  /**
   * <p>Ends a transaction boundary for the current channel operation. If a
   * transaction is already in progress, this method will join that transaction
   * using reference counting. The transaction is completed only if there
   * are no more references left for this transaction.</p>
   * <p><strong>Note</strong>: For every invocation of this method there must
   * be a corresponding invocation of {@linkplain #begin()} method. Failure
   * to ensure this can lead to dangling transactions and unpredictable results.
   * </p>
   */
  public void close();
}
和我们想想中的一样，也就是一些标准的事务方法的定义，和一个事务状态的枚举类型的定义

基本事务语义抽象类是对它的实现BasicTransactionSemantics

该类定义了两个属性
state状态和initialThreadId，id是唯一的，用来标识事务
构造方法中会赋值为NEW状态，并获取当前事务的一个ID值

重点来看下如下几个方法的具体实现：
 protected void doBegin() throws InterruptedException {}
  protected abstract void doPut(Event event) throws InterruptedException;
  protected abstract Event doTake() throws InterruptedException;
  protected abstract void doCommit() throws InterruptedException;
  protected abstract void doRollback() throws InterruptedException;
  protected void doClose() {}

1、doBegin

没什么好说的，就是检查状态是否NEW，ID是否匹配，没有问题后，将状态修改为OPEN，表示事务打开

2、doPut

检查ID是否匹配，状态是否打开，event是否为空，为空当然这个put就没意义了
关键看具体是怎么put的？
在FileChannel中有个内部静态类
static class FileBackedTransaction extends BasicTransactionSemantics
 private final LinkedBlockingDeque<FlumeEventPointer> takeList;
    private final LinkedBlockingDeque<FlumeEventPointer> putList;
这分别定义了两个双向队列，用于拿和放
protected void doPut(Event event) throws InterruptedException {
      channelCounter.incrementEventPutAttemptCount();
      if(putList.remainingCapacity() == 0) {
        throw new ChannelException("Put queue for FileBackedTransaction " +
            "of capacity " + putList.size() + " full, consider " +
            "committing more frequently, increasing capacity or " +
            "increasing thread count. " + channelNameDescriptor);
      }
      // this does not need to be in the critical section as it does not
      // modify the structure of the log or queue.
      if(!queueRemaining.tryAcquire(keepAlive, TimeUnit.SECONDS)) {
        throw new ChannelFullException("The channel has reached it‘s capacity. "
            + "This might be the result of a sink on the channel having too "
            + "low of batch size, a downstream system running slower than "
            + "normal, or that the channel capacity is just too low. "
            + channelNameDescriptor);
      }
      boolean success = false;
      log.lockShared();
      try {
        FlumeEventPointer ptr = log.put(transactionID, event);
        Preconditions.checkState(putList.offer(ptr), "putList offer failed "
          + channelNameDescriptor);
        queue.addWithoutCommit(ptr, transactionID);
        success = true;
      } catch (IOException e) {
        throw new ChannelException("Put failed due to IO error "
                + channelNameDescriptor, e);
      } finally {
        log.unlockShared();
        if(!success) {
          // release slot obtained in the case
          // the put fails for any reason
          queueRemaining.release();
        }
      }
    }
第一行，跟监控的度量信息有关，表示即将放入通道的event数量+1，监控度量请参考：http://blog.csdn.net/simonchi/article/details/43270461

1、检查队列的剩余空间

2、keepAlive秒时间内获取一个共享信号量，说明put的过程是互斥的

      如果该时间内没有成功获取该信号量，那么event放入失败

3、FlumeEventPointer是用来做检查点机制的，因为这是文件通道，会用日志记录的

      1、将event和事务id绑定到Pointer上
      2、将pointer放入队列尾部
      3、通道中的事件队列FlumeEventQueue添加一个未提交的事件，绑定了事务ID

4、释放共享信号量


3、doTake

 protected Event doTake() throws InterruptedException {
      channelCounter.incrementEventTakeAttemptCount();
      if(takeList.remainingCapacity() == 0) {
        throw new ChannelException("Take list for FileBackedTransaction, capacity " +
            takeList.size() + " full, consider committing more frequently, " +
            "increasing capacity, or increasing thread count. "
               + channelNameDescriptor);
      }
      log.lockShared();
      /*
       * 1. Take an event which is in the queue.
       * 2. If getting that event does not throw NoopRecordException,
       * then return it.
       * 3. Else try to retrieve the next event from the queue
       * 4. Repeat 2 and 3 until queue is empty or an event is returned.
       */
      try {
        while (true) {
          FlumeEventPointer ptr = queue.removeHead(transactionID);
          if (ptr == null) {
            return null;
          } else {
            try {
              // first add to takeList so that if write to disk
              // fails rollback actually does it‘s work
              Preconditions.checkState(takeList.offer(ptr),
                "takeList offer failed "
                  + channelNameDescriptor);
              log.take(transactionID, ptr); // write take to disk
              Event event = log.get(ptr);
              return event;
            } catch (IOException e) {
              throw new ChannelException("Take failed due to IO error "
                + channelNameDescriptor, e);
            } catch (NoopRecordException e) {
              LOG.warn("Corrupt record replaced by File Channel Integrity " +
                "tool found. Will retrieve next event", e);
              takeList.remove(ptr);
            } catch (CorruptEventException ex) {
              if (fsyncPerTransaction) {
                throw new ChannelException(ex);
              }
              LOG.warn("Corrupt record found. Event will be " +
                "skipped, and next event will be read.", ex);
              takeList.remove(ptr);
            }
          }
        }
      } finally {
        log.unlockShared();
      }
    }

1、剩余容量检查

2、检查点机制，日志记录操作，从头部取event

3、从takeList中删除该event


4、doCommit

protected void doCommit() throws InterruptedException {
      int puts = putList.size();
      int takes = takeList.size();
      if(puts > 0) {
        Preconditions.checkState(takes == 0, "nonzero puts and takes "
                + channelNameDescriptor);
        log.lockShared();
        try {
          log.commitPut(transactionID);
          channelCounter.addToEventPutSuccessCount(puts);
          synchronized (queue) {
            while(!putList.isEmpty()) {
              if(!queue.addTail(putList.removeFirst())) {
                StringBuilder msg = new StringBuilder();
                msg.append("Queue add failed, this shouldn‘t be able to ");
                msg.append("happen. A portion of the transaction has been ");
                msg.append("added to the queue but the remaining portion ");
                msg.append("cannot be added. Those messages will be consumed ");
                msg.append("despite this transaction failing. Please report.");
                msg.append(channelNameDescriptor);
                LOG.error(msg.toString());
                Preconditions.checkState(false, msg.toString());
              }
            }
            queue.completeTransaction(transactionID);
          }
        } catch (IOException e) {
          throw new ChannelException("Commit failed due to IO error "
                  + channelNameDescriptor, e);
        } finally {
          log.unlockShared();
        }
      } else if (takes > 0) {
        log.lockShared();
        try {
          log.commitTake(transactionID);
          queue.completeTransaction(transactionID);
          channelCounter.addToEventTakeSuccessCount(takes);
        } catch (IOException e) {
          throw new ChannelException("Commit failed due to IO error "
              + channelNameDescriptor, e);
        } finally {
          log.unlockShared();
        }
        queueRemaining.release(takes);
      }
      putList.clear();
      takeList.clear();
      channelCounter.setChannelSize(queue.getSize());
    }

1、如果putList不为空，提交的是放入通道的事件数量

2、如果takeList不为空，提交的是从通道拿走的事件数量


5、doRollback

  protected void doRollback() throws InterruptedException {
      int puts = putList.size();
      int takes = takeList.size();
      log.lockShared();
      try {
        if(takes > 0) {
          Preconditions.checkState(puts == 0, "nonzero puts and takes "
              + channelNameDescriptor);
          synchronized (queue) {
            while (!takeList.isEmpty()) {
              Preconditions.checkState(queue.addHead(takeList.removeLast()),
                  "Queue add failed, this shouldn‘t be able to happen "
                      + channelNameDescriptor);
            }
          }
        }
        putList.clear();
        takeList.clear();
        queue.completeTransaction(transactionID);
        channelCounter.setChannelSize(queue.getSize());
        log.rollback(transactionID);
      } catch (IOException e) {
        throw new ChannelException("Commit failed due to IO error "
            + channelNameDescriptor, e);
      } finally {
        log.unlockShared();
        // since rollback is being called, puts will never make it on
        // to the queue and we need to be sure to release the resources
        queueRemaining.release(puts);
      }
    }
在此之前，putList,takeList都没有clear，所以这里可以对着两个双向队列回滚操作
以上是文件通道的实现，如果是内存通道，就没有log的检查点记录了，简单多了，不需要维护状态了。


3、sink

sink的接口定义如下：
@InterfaceAudience.Public
@InterfaceStability.Stable
public interface Sink extends LifecycleAware, NamedComponent {
  /**
   * <p>Sets the channel the sink will consume from</p>
   * @param channel The channel to be polled
   */
  public void setChannel(Channel channel);
  /**
   * @return the channel associated with this sink
   */
  public Channel getChannel();
  /**
   * <p>Requests the sink to attempt to consume data from attached channel</p>
   * <p><strong>Note</strong>: This method should be consuming from the channel
   * within the bounds of a Transaction. On successful delivery, the transaction
   * should be committed, and on failure it should be rolled back.
   * @return READY if 1 or more Events were successfully delivered, BACKOFF if
   * no data could be retrieved from the channel feeding this sink
   * @throws EventDeliveryException In case of any kind of failure to
   * deliver data to the next hop destination.
   */
  public Status process() throws EventDeliveryException;
  public static enum Status {
    READY, BACKOFF
  }
}
sink与一个通道连接，并消费通道中的events，把它们发送到一个配置的目的地
其实和source的原理大部分相同，同样有一个AbstractSink

我们同样看一个具体实现吧，HDFSEventSink
看它的process方法
public Status process() throws EventDeliveryException {
    Channel channel = getChannel();
    Transaction transaction = channel.getTransaction();
    List<BucketWriter> writers = Lists.newArrayList();
    transaction.begin();

     …………………………

     transaction.commit();
      if (txnEventCount < 1) {
        return Status.BACKOFF;
      } else {
        sinkCounter.addToEventDrainSuccessCount(txnEventCount);
        return Status.READY;
      }
    } catch (IOException eIO) {
      transaction.rollback();
      LOG.warn("HDFS IO error", eIO);
      return Status.BACKOFF;
    } catch (Throwable th) {
      transaction.rollback();
      LOG.error("process failed", th);
      if (th instanceof Error) {
        throw (Error) th;
      } else {
        throw new EventDeliveryException(th);
      }
    } finally {
      transaction.close();
    }
  }
这里可以看到，flume-ng在sink端是有事务控制的
事务从 从通道中取event开始，到sink到下一个目的地结束
在这个过程中，任意的失败都会导致事务的回滚，这就保证数据了一致性。

【Flume】【*】深入flume-ng的三大组件——source，channel，sink


联系
我们