OutputFormat输出过程的学习

            花了大约1周的时间，终于把MapReduce的5大阶段的源码学习结束掉了，收获不少，就算本人对Hadoop学习的一个里程碑式的纪念吧。今天花了一点点的时间，把MapReduce的最后一个阶段，输出OutputFormat给做了分析，这个过程跟InputFormat刚刚好是对着干的，二者极具对称性。为什么这么说呢，待我一一分析。

            OutputFormat过程的作用就是定义数据key-value的输出格式，给你处理好后的数据，究竟以什么样的形式输出呢，才能让下次别人拿到这个文件的时候能准确的提取出里面的数据。这里，我们撇开这个话题，仅仅我知道的一些定义的数据格式的方法，比如在Redis中会有这样的设计:

[key-length][key][value-length][value][key-length][key][value-length][value]...

或者说不一定非要省空间,直接搞过分隔符

[key]   [value]\n

[key]   [value]\n

[key]   [value]\n

.....

这样逐行读取，再以空格隔开，取出里面的键值对，这么做简单是简单，就是不紧凑，空间浪费得有点多。在MapReduce的OutputFormat的有种格式用的就是这种方式。

        首先必须得了解OutputFormat里面到底有什么东西:

public interface OutputFormat<K, V> {

  /** 
   * Get the {@link RecordWriter} for the given job.
   * 获取输出记录键值记录
   *
   * @param ignored
   * @param job configuration for the job whose output is being written.
   * @param name the unique name for this part of the output.
   * @param progress mechanism for reporting progress while writing to file.
   * @return a {@link RecordWriter} to write the output for the job.
   * @throws IOException
   */
  RecordWriter<K, V> getRecordWriter(FileSystem ignored, JobConf job,
                                     String name, Progressable progress)
  throws IOException;

  /** 
   * Check for validity of the output-specification for the job.
   *  
   * <p>This is to validate the output specification for the job when it is
   * a job is submitted.  Typically checks that it does not already exist,
   * throwing an exception when it already exists, so that output is not
   * overwritten.</p>
   * 作业运行之前进行的检测工作，例如配置的输出目录是否存在等
   *
   * @param ignored
   * @param job job configuration.
   * @throws IOException when output should not be attempted
   */
  void checkOutputSpecs(FileSystem ignored, JobConf job) throws IOException;
}

很简单的2个方法，RecordWriter比较重要，后面的key-value的写入操作都是根据他来完成的。但是他是一个接口，在MapReduce中，我们用的最多的他的子类是FileOutputFormat：

/** A base class for {@link OutputFormat}. */
public abstract class FileOutputFormat<K, V> implements OutputFormat<K, V> {

他是一个抽象类，但是实现了接口中的第二个方法checkOutputSpecs()方法：

public void checkOutputSpecs(FileSystem ignored, JobConf job) 
    throws FileAlreadyExistsException, 
           InvalidJobConfException, IOException {
    // Ensure that the output directory is set and not already there
    Path outDir = getOutputPath(job);
    if (outDir == null && job.getNumReduceTasks() != 0) {
      throw new InvalidJobConfException("Output directory not set in JobConf.");
    }
    if (outDir != null) {
      FileSystem fs = outDir.getFileSystem(job);
      // normalize the output directory
      outDir = fs.makeQualified(outDir);
      setOutputPath(job, outDir);
      
      // get delegation token for the outDir's file system
      TokenCache.obtainTokensForNamenodes(job.getCredentials(), 
                                          new Path[] {outDir}, job);
      
      // check its existence
      if (fs.exists(outDir)) {
    	//如果输出目录以及存在，则抛异常
        throw new FileAlreadyExistsException("Output directory " + outDir + 
                                             " already exists");
      }
    }
  }

就是检查输出目录在不在的操作。在这个类里还出现了一个辅助类：

public static Path getTaskOutputPath(JobConf conf, String name) 
  throws IOException {
    // ${mapred.out.dir}
    Path outputPath = getOutputPath(conf);
    if (outputPath == null) {
      throw new IOException("Undefined job output-path");
    }
    
    //根据OutputCommitter获取输出路径
    OutputCommitter committer = conf.getOutputCommitter();
    Path workPath = outputPath;
    TaskAttemptContext context = new TaskAttemptContext(conf,
                TaskAttemptID.forName(conf.get("mapred.task.id")));
    if (committer instanceof FileOutputCommitter) {
      workPath = ((FileOutputCommitter)committer).getWorkPath(context,
                                                              outputPath);
    }
    
    // ${mapred.out.dir}/_temporary/_${taskid}/${name}
    return new Path(workPath, name);
  }

就是上面OutputCommiter，里面定义了很多和Task,job作业相关的方法。很多时候都会与OutputFormat合作的形式出现。他也有自己的子类实现File OutputCommiter:

public class FileOutputCommitter extends OutputCommitter {

  public static final Log LOG = LogFactory.getLog(
      "org.apache.hadoop.mapred.FileOutputCommitter");
/**
   * Temporary directory name 
   */
  public static final String TEMP_DIR_NAME = "_temporary";
  public static final String SUCCEEDED_FILE_NAME = "_SUCCESS";
  static final String SUCCESSFUL_JOB_OUTPUT_DIR_MARKER =
    "mapreduce.fileoutputcommitter.marksuccessfuljobs";

  public void setupJob(JobContext context) throws IOException {
    JobConf conf = context.getJobConf();
    Path outputPath = FileOutputFormat.getOutputPath(conf);
    if (outputPath != null) {
      Path tmpDir = new Path(outputPath, FileOutputCommitter.TEMP_DIR_NAME);
      FileSystem fileSys = tmpDir.getFileSystem(conf);
      if (!fileSys.mkdirs(tmpDir)) {
        LOG.error("Mkdirs failed to create " + tmpDir.toString());
      }
    }
  }
  ....

在Reduce阶段的后面的写阶段，FileOutputFormat是默认的输出的类型：

//获取输出的key，value
    final RecordWriter<OUTKEY, OUTVALUE> out = new OldTrackingRecordWriter<OUTKEY, OUTVALUE>(
        reduceOutputCounter, job, reporter, finalName);
    
    OutputCollector<OUTKEY,OUTVALUE> collector = 
      new OutputCollector<OUTKEY,OUTVALUE>() {
        public void collect(OUTKEY key, OUTVALUE value)
          throws IOException {
          //将处理后的key,value写入输出流中，最后写入HDFS作为最终结果
          out.write(key, value);
          // indicate that progress update needs to be sent
          reporter.progress();
        }
      };

out就是直接发挥作用的类，但是是哪个Formtat的返回的呢，我们进入OldTrackingRecordWriter继续看：

public OldTrackingRecordWriter(
        org.apache.hadoop.mapred.Counters.Counter outputRecordCounter,
        JobConf job, TaskReporter reporter, String finalName)
        throws IOException {
      this.outputRecordCounter = outputRecordCounter;
      //默认是FileOutputFormat文件输出方式
      this.fileOutputByteCounter = reporter
          .getCounter(FileOutputFormat.Counter.BYTES_WRITTEN);
      Statistics matchedStats = null;
      if (job.getOutputFormat() instanceof FileOutputFormat) {
        matchedStats = getFsStatistics(FileOutputFormat.getOutputPath(job), job);
      }
      fsStats = matchedStats;

      FileSystem fs = FileSystem.get(job);
      long bytesOutPrev = getOutputBytes(fsStats);
      //从配置中获取作业的输出方式
      this.real = job.getOutputFormat().getRecordWriter(fs, job, finalName,
          reporter);
      long bytesOutCurr = getOutputBytes(fsStats);
      fileOutputByteCounter.increment(bytesOutCurr - bytesOutPrev);
    }

果然是我们所想的那样，FileOutputFormat，但是不要忘了它的getRecordWriter()是抽象方法，调用它还必须由它的子类来实现:

 public abstract RecordWriter<K, V> getRecordWriter(FileSystem ignored,
                                               JobConf job, String name,
                                               Progressable progress)
    throws IOException;

在这里我们举出其中在InputFormat举过的例子，TextOutputFormat,SequenceFileOutputFormat，与 TextInputFormat,SequenceFileInputFormat对应。也就说说上面2个子类定义了2种截然不同的输出格式，也就返回了不一样的RecordWriter实现类.在TextOutputFormat中，他定义了一个叫LineRecordWriter的定义：

 public RecordWriter<K, V> getRecordWriter(FileSystem ignored,
                                                  JobConf job,
                                                  String name,
                                                  Progressable progress)
    throws IOException {
	//从配置判断输出是否要压缩
    boolean isCompressed = getCompressOutput(job);
    //配置中获取加在key-value的分隔符
    String keyValueSeparator = job.get("mapred.textoutputformat.separator", 
                                       "\t");
    //根据是否压缩获取相应的LineRecordWriter
    if (!isCompressed) {
      Path file = FileOutputFormat.getTaskOutputPath(job, name);
      FileSystem fs = file.getFileSystem(job);
      FSDataOutputStream fileOut = fs.create(file, progress);
      return new LineRecordWriter<K, V>(fileOut, keyValueSeparator);
    } else {
      Class<? extends CompressionCodec> codecClass =
        getOutputCompressorClass(job, GzipCodec.class);
      // create the named codec
      CompressionCodec codec = ReflectionUtils.newInstance(codecClass, job);
      // build the filename including the extension
      Path file = 
        FileOutputFormat.getTaskOutputPath(job, 
                                           name + codec.getDefaultExtension());
      FileSystem fs = file.getFileSystem(job);
      FSDataOutputStream fileOut = fs.create(file, progress);
      return new LineRecordWriter<K, V>(new DataOutputStream
                                        (codec.createOutputStream(fileOut)),
                                        keyValueSeparator);
    }

他以一个内部类的形式存在于 TextOutputFormat。而在SequenceFileOutputFormat中，他的形式是怎样的呢：

  public RecordWriter<K, V> getRecordWriter(
                                          FileSystem ignored, JobConf job,
                                          String name, Progressable progress)
    throws IOException {
    // get the path of the temporary output file 
    Path file = FileOutputFormat.getTaskOutputPath(job, name);
    
    FileSystem fs = file.getFileSystem(job);
    CompressionCodec codec = null;
    CompressionType compressionType = CompressionType.NONE;
    if (getCompressOutput(job)) {
      // find the kind of compression to do
      compressionType = getOutputCompressionType(job);

      // find the right codec
      Class<? extends CompressionCodec> codecClass = getOutputCompressorClass(job,
	  DefaultCodec.class);
      codec = ReflectionUtils.newInstance(codecClass, job);
    }
    final SequenceFile.Writer out = 
      SequenceFile.createWriter(fs, job, file,
                                job.getOutputKeyClass(),
                                job.getOutputValueClass(),
                                compressionType,
                                codec,
                                progress);

    return new RecordWriter<K, V>() {

        public void write(K key, V value)
          throws IOException {

          out.append(key, value);
        }

        public void close(Reporter reporter) throws IOException { out.close();}
      };
  }

关键的操作都在于SequenceFile.Writer中。有不同的RecordWriter就会有不同的写入数据的方式，这里我们举LineRecordWriter的例子。看看他的写入方法:

//往输出流中写入key-value
    public synchronized void write(K key, V value)
      throws IOException {

      //判断键值对是否为空
      boolean nullKey = key == null || key instanceof NullWritable;
      boolean nullValue = value == null || value instanceof NullWritable;
      
      //如果k-v都为空，则操作失败，不写入直接返回
      if (nullKey && nullValue) {
        return;
      }
      
      //如果key不空，则写入key
      if (!nullKey) {
        writeObject(key);
      }
      
      //如果key,value都不为空，则中间写入k-v分隔符，在这里为\t空格符
      if (!(nullKey || nullValue)) {
        out.write(keyValueSeparator);
      }
      
      //最后写入value
      if (!nullValue) {
        writeObject(value);
      }

在这个方法里，我们就能看出他的存储形式就是我刚刚在上面讲的第二种存储方式。这个方法将会在下面的代码中被执行：

OutputCollector<OUTKEY,OUTVALUE> collector = 
      new OutputCollector<OUTKEY,OUTVALUE>() {
        public void collect(OUTKEY key, OUTVALUE value)
          throws IOException {
          //将处理后的key,value写入输出流中，最后写入HDFS作为最终结果
          out.write(key, value);
          // indicate that progress update needs to be sent
          reporter.progress();
        }
      };

过程可以这么理解:

collector.collect()------->out.write(key, value)------->LineRecordWriter.write(key, value)------->DataOutputStream.write(key, value).

DataOutputStream是内置于LineRecordWriter的作为里面的变量存在的。这样从Reduce末尾阶段到Output的过程也完全打通了。下面可以看看这上面涉及的完整的类目关系。

      下一阶段的学习，可能或偏向于Task，Job阶段的过程分析，更加宏观过程上的一个分析。也可能会分析某个功能块的实现过程，比如Hadoop的IPC过程，据说用了很多JAVA NIO的东西。