MapReduce的InputFormat过程的学习

          昨天经过几个小时的学习，把MapReduce的第一个阶段的过程学习了一下，也就是最最开始的时候从文件中的Data到key-value的映射，也就是InputFormat的过程。虽说过程不是很难，但是也存在很多细节的。也很少会有人对此做比较细腻的研究，学习。今天，就让我来为大家剖析一下这段代码的原理。我还为此花了一点时间做了几张结构图，便于大家理解。在这里先声明一下，我研究的MapReduce主要研究的是旧版的API，也就是mapred包下的。

          InputFormat最最原始的形式就是一个接口。后面出现的各种Format都是他的衍生类。结构如下，只包含最重要的2个方法:

public interface InputFormat<K, V> {

  /** 
   * Logically split the set of input files for the job.  
   * 
   * <p>Each {@link InputSplit} is then assigned to an individual {@link Mapper}
   * for processing.</p>
   *
   * <p><i>Note</i>: The split is a <i>logical</i> split of the inputs and the
   * input files are not physically split into chunks. For e.g. a split could
   * be <i><input-file-path, start, offset></i> tuple.
   * 
   * @param job job configuration.
   * @param numSplits the desired number of splits, a hint.
   * @return an array of {@link InputSplit}s for the job.
   */
  InputSplit[] getSplits(JobConf job, int numSplits) throws IOException;

  /** 
   * Get the {@link RecordReader} for the given {@link InputSplit}.
   *
   * <p>It is the responsibility of the <code>RecordReader</code> to respect
   * record boundaries while processing the logical split to present a 
   * record-oriented view to the individual task.</p>
   * 
   * @param split the {@link InputSplit}
   * @param job the job that this split belongs to
   * @return a {@link RecordReader}
   */
  RecordReader<K, V> getRecordReader(InputSplit split,
                                     JobConf job, 
                                     Reporter reporter) throws IOException;
}

所以后面讲解，我也只是会围绕这2个方法进行分析。当然我们用的最多的是从文件中获得输入数据，也就是FileInputFormat这个类。继承关系如下:

public abstract class FileInputFormat<K, V> implements InputFormat<K, V>

我们看里面的1个主要方法:

public InputSplit[] getSplits(JobConf job, int numSplits)

返回的类型是一个InputSpilt对象，这是一个抽象的输入Spilt分片概念。结构如下:

public interface InputSplit extends Writable {

  /**
   * Get the total number of bytes in the data of the <code>InputSplit</code>.
   * 
   * @return the number of bytes in the input split.
   * @throws IOException
   */
  long getLength() throws IOException;
  
  /**
   * Get the list of hostnames where the input split is located.
   * 
   * @return list of hostnames where data of the <code>InputSplit</code> is
   *         located as an array of <code>String</code>s.
   * @throws IOException
   */
  String[] getLocations() throws IOException;
}

提供了与数据相关的2个方法。后面这个返回的值会被用来传递给RecordReader里面去的。在想理解getSplits方法之前还有一个类需要理解，FileStatus，里面包装了一系列的文件基本信息方法:

public class FileStatus implements Writable, Comparable {

  private Path path;
  private long length;
  private boolean isdir;
  private short block_replication;
  private long blocksize;
  private long modification_time;
  private long access_time;
  private FsPermission permission;
  private String owner;
  private String group;

.....

看到这里你估计会有点晕了，下面是我做的一张小小类图关系:

可以看到，FileSpilt为了兼容新老版本，继承了新的抽象类InputSpilt，同时附上旧的接口形式的InputSpilt。下面我们看看里面的getspilt核心过程:

/** Splits files returned by {@link #listStatus(JobConf)} when
   * they're too big.*/ 
  @SuppressWarnings("deprecation")
  public InputSplit[] getSplits(JobConf job, int numSplits)
    throws IOException {
	//获取所有的状态文件
    FileStatus[] files = listStatus(job);
    
    // Save the number of input files in the job-conf
    //在job-cof中保存文件的数量
    job.setLong(NUM_INPUT_FILES, files.length);
    long totalSize = 0;                           
    // compute total size,计算文件总的大小
    for (FileStatus file: files) {                // check we have valid files
      if (file.isDir()) {
    	  //如果是目录不是纯文件的直接抛异常
        throw new IOException("Not a file: "+ file.getPath());
      }
      totalSize += file.getLen();
    }

    //用户期待的划分大小，总大小除以spilt划分数目
    long goalSize = totalSize / (numSplits == 0 ? 1 : numSplits);
    //获取系统的划分最小值
    long minSize = Math.max(job.getLong("mapred.min.split.size", 1),
                            minSplitSize);

    // generate splits
    //创建numSplits个FileSpilt文件划分量
    ArrayList<FileSplit> splits = new ArrayList<FileSplit>(numSplits);
    NetworkTopology clusterMap = new NetworkTopology();
    for (FileStatus file: files) {
      Path path = file.getPath();
      FileSystem fs = path.getFileSystem(job);
      long length = file.getLen();
      //获取此文件的block的位置列表
      BlockLocation[] blkLocations = fs.getFileBlockLocations(file, 0, length);
      //如果文件系统可划分
      if ((length != 0) && isSplitable(fs, path)) {
    	//计算此文件的总的block块的大小
        long blockSize = file.getBlockSize();
        //根据期待大小，最小大小，得出最终的split分片大小
        long splitSize = computeSplitSize(goalSize, minSize, blockSize);

        long bytesRemaining = length;
        //如果剩余待划分字节倍数为划分大小超过1.1的划分比例，则进行拆分
        while (((double) bytesRemaining)/splitSize > SPLIT_SLOP) {
          //获取提供数据的splitHost位置
          String[] splitHosts = getSplitHosts(blkLocations, 
              length-bytesRemaining, splitSize, clusterMap);
          //添加FileSplit
          splits.add(new FileSplit(path, length-bytesRemaining, splitSize, 
              splitHosts));
          //数量减少splitSize大小
          bytesRemaining -= splitSize;
        }
        
        if (bytesRemaining != 0) {
          //添加刚刚剩下的没划分完的部分，此时bytesRemaining已经小于splitSize的1.1倍了
          splits.add(new FileSplit(path, length-bytesRemaining, bytesRemaining, 
                     blkLocations[blkLocations.length-1].getHosts()));
        }
      } else if (length != 0) {
    	//不划分，直接添加Spilt
        String[] splitHosts = getSplitHosts(blkLocations,0,length,clusterMap);
        splits.add(new FileSplit(path, 0, length, splitHosts));
      } else { 
        //Create empty hosts array for zero length files
        splits.add(new FileSplit(path, 0, length, new String[0]));
      }
    }
    
    //最后返回FileSplit数组
    LOG.debug("Total # of splits: " + splits.size());
    return splits.toArray(new FileSplit[splits.size()]);
  }

里面有个computerSpiltSize方法很特殊，考虑了很多情况，总之最小值不能小于系统设定的最小值。要与期待值，块大小，系统允许最小值:

protected long computeSplitSize(long goalSize, long minSize,
                                       long blockSize) {
    return Math.max(minSize, Math.min(goalSize, blockSize));
  }

上述过程的相应流程图如下:

3种情况3中年执行流程。

      处理完getSpilt方法然后，也就是说已经把数据从文件中转划到InputSpilt中了，接下来就是给RecordRead去取出里面的一条条的记录了。当然这在FileInputFormat是抽象方法，必须由子类实现的，我在这里挑出了2个典型的子类SequenceFileInputFormat，和TextInputFormat。他们的实现RecordRead方法如下：

public RecordReader<K, V> getRecordReader(InputSplit split,
                                      JobConf job, Reporter reporter)
    throws IOException {

    reporter.setStatus(split.toString());

    return new SequenceFileRecordReader<K, V>(job, (FileSplit) split);
  }

public RecordReader<LongWritable, Text> getRecordReader(
                                          InputSplit genericSplit, JobConf job,
                                          Reporter reporter)
    throws IOException {
    
    reporter.setStatus(genericSplit.toString());
    return new LineRecordReader(job, (FileSplit) genericSplit);
  }

可以看到里面的区别就在于LineRecordReader和SequenceFileRecordReader的不同了，这也就表明2种方式对应于数据的读取方式可能会不一样，继续往里深入看:

/** An {@link RecordReader} for {@link SequenceFile}s. */
public class SequenceFileRecordReader<K, V> implements RecordReader<K, V> {
  
  private SequenceFile.Reader in;
  private long start;
  private long end;
  private boolean more = true;
  protected Configuration conf;

  public SequenceFileRecordReader(Configuration conf, FileSplit split)
    throws IOException {
    Path path = split.getPath();
    FileSystem fs = path.getFileSystem(conf);
    //从文件系统中读取数据输入流
    this.in = new SequenceFile.Reader(fs, path, conf);
    this.end = split.getStart() + split.getLength();
    this.conf = conf;

    if (split.getStart() > in.getPosition())
      in.sync(split.getStart());                  // sync to start

    this.start = in.getPosition();
    more = start < end;
  }

  ......
  
  /**
   * 获取下一个键值对
   */
  public synchronized boolean next(K key, V value) throws IOException {
	//判断还有无下一条记录
    if (!more) return false;
    long pos = in.getPosition();
    boolean remaining = (in.next(key) != null);
    if (remaining) {
      getCurrentValue(value);
    }
    if (pos >= end && in.syncSeen()) {
      more = false;
    } else {
      more = remaining;
    }
    return more;
  }

我们可以看到SequenceFileRecordReader是从输入流in中一个键值，一个键值的读取，另外一个的实现方式如下:

/**
 * Treats keys as offset in file and value as line. 
 */
public class LineRecordReader implements RecordReader<LongWritable, Text> {
  private static final Log LOG
    = LogFactory.getLog(LineRecordReader.class.getName());

  private CompressionCodecFactory compressionCodecs = null;
  private long start;
  private long pos;
  private long end;
  private LineReader in;
  int maxLineLength;

  ....
  
  /** Read a line. */
  public synchronized boolean next(LongWritable key, Text value)
    throws IOException {

    while (pos < end) {
      //设置key 
      key.set(pos);

      //根据位置一行一行读取，设置value
      int newSize = in.readLine(value, maxLineLength,
                                Math.max((int)Math.min(Integer.MAX_VALUE, end-pos),
                                         maxLineLength));
      if (newSize == 0) {
        return false;
      }
      pos += newSize;
      if (newSize < maxLineLength) {
        return true;
      }

      // line too long. try again
      LOG.info("Skipped line of size " + newSize + " at pos " + (pos - newSize));
    }

    return false;
  }

实现的方式为通过读的位置，从输入流中逐行读取key-value。通过这2种方法，就能得到新的key-value，就会用于后面的map操作。

InputFormat的整个流程其实我忽略了很多细节。大体流程如上述所说。