2

Flume作为Hadoop生态系统中的一员,可以说是功能最大的数据收集系统,Flume的模型也比较简单,通过agent不断级连,来打通数据源与最终目的地(一般为HDFS)。下图结构说明了Flume中的数据流。
Flume架构

我今天要说的是Channel部分,具体来说是MemoryChannel的分析,其他概念像source、sink大家可以去官方文档查看。

注意:

本文章中的Flume源码为1.6.0版本。

Event

Event是Flume中对数据的抽象,分为两部分:header与body,和http中的header与body很类似。

Flume中按Event为单位操作数据,不同的source、sink在必要时会自动在原始数据与Event之间做转化。

Channel

Channel充当了Source与Sink之间的缓冲区。Channel的引入,使得source与sink之间的藕合度降低,source只管像Channel发数据,sink只需从Channel取数据。
此外,有了Channel,不难得出下面结论:

  • source与sink可以为N对N的关系

  • source发数据的速度可以大于sink取数据的速度(在Channel不满的情况下)

Transaction

Channel采用了Transaction(事务)机制来保证数据的完整性,这里的事务和数据库中的事务概念类似,但并不是完全一致,其语义可以参考下面这个图:
channel_transaction

source端通过commit操作像Channel放置数据,sink端通过commit操作从Channel取数据。

那么事务是如何保证数据的完整性的呢?看下面有两个agent的情况:
图片描述

数据流程:

  1. source 1产生Event,通过“put”、“commit”操作将Event放到Channel 1

  2. sink 1通过“take”操作从Channel 1中取出Event,并把它发送到Source 2

  3. source 2通过“put”、“commit”操作将Event放到Channel 2

  4. source 2sink 1发送成功信号,sink 1“commit”步骤2中的“take”操作(其实就是删除Channel 1中的Event)

说明:

在任何时刻,Event至少在一个Channel中是完整有效的

Memory Channe

Flume中提供的Channel实现主要有三个:

  • Memory Channel,event保存在Java Heap中。如果允许数据小量丢失,推荐使用

  • File Channel,event保存在本地文件中,可靠性高,但吞吐量低于Memory Channel

  • JDBC Channel,event保存在关系数据中,一般不推荐使用

不同的Channel主要在于Event存放的位置不同,今天我着重讲一下比较简单的Memory Channel的源码。

首先看一下MemoryChannel中比较重要的成员变量:

 // lock to guard queue, mainly needed to keep it locked down during resizes
// it should never be held through a blocking operation
private Object queueLock = new Object();

//queue为Memory Channel中存放Event的地方,这里用了LinkedBlockingDeque来实现
@GuardedBy(value = "queueLock")
private LinkedBlockingDeque<Event> queue;

//下面的两个信号量用来做同步操作,queueRemaining表示queue中的剩余空间,queueStored表示queue中的使用空间

// invariant that tracks the amount of space remaining in the queue(with all uncommitted takeLists deducted)
// we maintain the remaining permits = queue.remaining - takeList.size()
// this allows local threads waiting for space in the queue to commit without denying access to the
// shared lock to threads that would make more space on the queue
private Semaphore queueRemaining;
// used to make "reservations" to grab data from the queue.
// by using this we can block for a while to get data without locking all other threads out
// like we would if we tried to use a blocking call on queue
private Semaphore queueStored;

//下面几个变量为配置文件中Memory Channel的配置项
// 一个事务中Event的最大数目
private volatile Integer transCapacity;
// 向queue中添加、移除Event的等待时间
private volatile int keepAlive;
// queue中,所有Event所能占用的最大空间
private volatile int byteCapacity;
private volatile int lastByteCapacity;
// queue中,所有Event的header所能占用的最大空间占byteCapacity的比例
private volatile int byteCapacityBufferPercentage;
// 用于标示byteCapacity中剩余空间的信号量
private Semaphore bytesRemaining;
// 用于记录Memory Channel的一些指标,后面可以通过配置监控来观察Flume的运行情况
private ChannelCounter channelCounter;

然后重点说下MemoryChannel里面的MemoryTransaction,它是Transaction类的子类,从其文档来看,一个Transaction的使用模式都是类似的:

 Channel ch = ...
 Transaction tx = ch.getTransaction();
 try {
   tx.begin();
   ...
   // ch.put(event) or ch.take()
   ...
   tx.commit();
 } catch (ChannelException ex) {
   tx.rollback();
   ...
 } finally {
   tx.close();
 }

可以看到一个Transaction主要有、puttakecommitrollback这四个方法,我们在实现其子类时,主要也是实现着四个方法。

Flume官方为了方便开发者实现自己的Transaction,定义了BasicTransactionSemantics,这时开发者只需要继承这个辅助类,并且实现其相应的、doPutdoTakedoCommitdoRollback方法即可,MemoryChannel就是继承了这个辅助类。

private class MemoryTransaction extends BasicTransactionSemantics {
    //和MemoryChannel一样,内部使用LinkedBlockingDeque来保存没有commit的Event
    private LinkedBlockingDeque<Event> takeList;
    private LinkedBlockingDeque<Event> putList;
    private final ChannelCounter channelCounter;
    //下面两个变量用来表示put的Event的大小、take的Event的大小
    private int putByteCounter = 0;
    private int takeByteCounter = 0;

    public MemoryTransaction(int transCapacity, ChannelCounter counter) {
      //用transCapacity来初始化put、take的队列
      putList = new LinkedBlockingDeque<Event>(transCapacity);
      takeList = new LinkedBlockingDeque<Event>(transCapacity);

      channelCounter = counter;
    }

    @Override
    protected void doPut(Event event) throws InterruptedException {
      //doPut操作,先判断putList中是否还有剩余空间,有则把Event插入到该队列中,同时更新putByteCounter
      //没有剩余空间的话,直接报ChannelException
      channelCounter.incrementEventPutAttemptCount();
      int eventByteSize = (int)Math.ceil(estimateEventSize(event)/byteCapacitySlotSize);

      if (!putList.offer(event)) {
        throw new ChannelException(
          "Put queue for MemoryTransaction of capacity " +
            putList.size() + " full, consider committing more frequently, " +
            "increasing capacity or increasing thread count");
      }
      putByteCounter += eventByteSize;
    }

    @Override
    protected Event doTake() throws InterruptedException {
      //doTake操作,首先判断takeList中是否还有剩余空间
      channelCounter.incrementEventTakeAttemptCount();
      if(takeList.remainingCapacity() == 0) {
        throw new ChannelException("Take list for MemoryTransaction, capacity " +
            takeList.size() + " full, consider committing more frequently, " +
            "increasing capacity, or increasing thread count");
      }
      //然后判断,该MemoryChannel中的queue中是否还有空间,这里通过信号量来判断
      if(!queueStored.tryAcquire(keepAlive, TimeUnit.SECONDS)) {
        return null;
      }
      Event event;
      //从MemoryChannel中的queue中取出一个event
      synchronized(queueLock) {
        event = queue.poll();
      }
      Preconditions.checkNotNull(event, "Queue.poll returned NULL despite semaphore " +
          "signalling existence of entry");
      //放到takeList中,然后更新takeByteCounter变量
      takeList.put(event);

      int eventByteSize = (int)Math.ceil(estimateEventSize(event)/byteCapacitySlotSize);
      takeByteCounter += eventByteSize;

      return event;
    }

    @Override
    protected void doCommit() throws InterruptedException {
      //该对应一个事务的提交
      //首先判断putList与takeList的相对大小
      int remainingChange = takeList.size() - putList.size();
      //如果takeList小,说明向该MemoryChannel放的数据比取的数据要多,所以需要判断该MemoryChannel是否有空间来放
      if(remainingChange < 0) {
        // 1. 首先通过信号量来判断是否还有剩余空间
        if(!bytesRemaining.tryAcquire(putByteCounter, keepAlive,
          TimeUnit.SECONDS)) {
          throw new ChannelException("Cannot commit transaction. Byte capacity " +
            "allocated to store event body " + byteCapacity * byteCapacitySlotSize +
            "reached. Please increase heap space/byte capacity allocated to " +
            "the channel as the sinks may not be keeping up with the sources");
        }
        // 2. 然后判断,在给定的keepAlive时间内,能否获取到充足的queue空间
        if(!queueRemaining.tryAcquire(-remainingChange, keepAlive, TimeUnit.SECONDS)) {
          bytesRemaining.release(putByteCounter);
          throw new ChannelFullException("Space for commit to queue couldn't be acquired." +
              " Sinks are likely not keeping up with sources, or the buffer size is too tight");
        }
      }
      int puts = putList.size();
      int takes = takeList.size();
      //如果上面的两个判断都过了,那么把putList中的Event放到该MemoryChannel中的queue中。
      synchronized(queueLock) {
        if(puts > 0 ) {
          while(!putList.isEmpty()) {
            if(!queue.offer(putList.removeFirst())) {
              throw new RuntimeException("Queue add failed, this shouldn't be able to happen");
            }
          }
        }
        //清空本次事务中用到的putList与takeList,释放资源
        putList.clear();
        takeList.clear();
      }
      //更新控制queue大小的信号量bytesRemaining,因为把takeList清空了,所以直接把takeByteCounter加到bytesRemaining中。
      bytesRemaining.release(takeByteCounter);
      takeByteCounter = 0;
      putByteCounter = 0;
      //因为把putList中的Event放到了MemoryChannel中的queue,所以把puts加到queueStored中去。
      queueStored.release(puts);
      //如果takeList比putList大,说明该MemoryChannel中queue的数量应该是减少了,所以把(takeList-putList)的差值加到信号量queueRemaining
      if(remainingChange > 0) {
        queueRemaining.release(remainingChange);
      }
      if (puts > 0) {
        channelCounter.addToEventPutSuccessCount(puts);
      }
      if (takes > 0) {
        channelCounter.addToEventTakeSuccessCount(takes);
      }

      channelCounter.setChannelSize(queue.size());
    }

    @Override
    protected void doRollback() {
      //当一个事务失败时,会进行回滚,即调用本方法
      //首先把takeList中的Event放回到MemoryChannel中的queue中。
      int takes = takeList.size();
      synchronized(queueLock) {
        Preconditions.checkState(queue.remainingCapacity() >= takeList.size(), "Not enough space in memory channel " +
            "queue to rollback takes. This should never happen, please report");
        while(!takeList.isEmpty()) {
          queue.addFirst(takeList.removeLast());
        }
        //然后清空putList
        putList.clear();
      }
      //因为清空了putList,所以需要把putList所占用的空间大小添加到bytesRemaining中
      bytesRemaining.release(putByteCounter);
      putByteCounter = 0;
      takeByteCounter = 0;
      //因为把takeList中的Event回退到queue中去了,所以需要把takeList的大小添加到queueStored中
      queueStored.release(takes);
      channelCounter.setChannelSize(queue.size());
    }

  }

MemoryChannel的逻辑相对简单,主要是通过MemoryTransaction中的putListtakeList与MemoryChannel中的queue打交道,这里的queue相当于持久化层,只不过放到了内存中,如果是FileChannel的话,会把这个queue放到本地文件中。下面表示了Event在一个使用了MemoryChannel的agent中数据流向:

source ---> putList ---> queue ---> takeList ---> sink

还需要注意的一点是,这里的事务可以嵌套使用,如下图:
transaction

当有两个agent级连时,sink的事务中包含了一个source的事务,这也应证了前面所说的:

在任何时刻,Event至少在一个Channel中是完整有效的

参考


刘家财
1.7k 声望97 粉丝

Lisp、编译原理爱好者