hadoop - combineFileInputFormat 中的 isSplitable 不起作用

Question

我有数千个小文件，我想用 combineFileInputFormat 处理它们。

在combineFileInputFormat中，一个mapper有多个小文件，每个文件不会被分割。

像这样的一个小输入文件的片段，

vers,3
period,2015-01-26-18-12-00,438469546,449329626,complete
config,libdvm.so,chromeview
pkgproc,com.futuredial.digitchat,10021,,0ns:10860078
pkgpss,com.futuredial.digitchat,10021,,0ns:9:6627:6627:6637:5912:5912:5912
pkgsvc-run,com.futuredial.digitchat,10021,.LiveScreenService,1,0n:10860078
pkgsvc-start,com.futuredial.digitchat,10021,.LiveScreenService,1,0n:10860078
pkgproc,com.google.android.youtube,10103,,0ns:10860078
pkgpss,com.google.android.youtube,10103,,0ns:9:12986:13000:13021:11552:11564:11580
pkgsvc-     run,com.google.android.youtube,10103,com.google.android.apps.youtube.app.offline.transfer.OfflineTransferService,1,0n:10860078
pkgsvc-    start,com.google.android.youtube,10103,com.google.android.apps.youtube.app.offline.transfer.OfflineTransferService,1,0n:10860078

我想将整个文件内容传递给映射器。但是，hadoop 将文件分成了一半。

比如上面的文件可能会被拆分成

vers,3
period,2015-01-26-18-12-00,438469546,449329626,complete
config,libdvm.so,chromeview
pkgproc,com.futuredial.digitchat,#the line has been cut

但是我想要处理整个文件的内容。

这是我的代码，引用 Reading file as single record in hadoop

驱动代码

public class CombineSmallfiles {

public static void main(String[] args) throws IOException, ClassNotFoundException, InterruptedException {

    Configuration conf = new Configuration();
    String[] otherArgs = new GenericOptionsParser(conf, args).getRemainingArgs();
    if (otherArgs.length != 2) {
        System.err.println("Usage: conbinesmallfiles <in> <out>");
        System.exit(2);
    }

    conf.setInt("mapred.min.split.size", 1);
    conf.setLong("mapred.max.split.size", 26214400); // 25m
    //conf.setLong("mapred.max.split.size", 134217728); // 128m

    //conf.setInt("mapred.reduce.tasks", 5);

    Job job = new Job(conf, "combine smallfiles");
    job.setJarByClass(CombineSmallfiles.class);
    job.setMapperClass(CombineSmallfileMapper.class);
    //job.setReducerClass(IdentityReducer.class);
    job.setNumReduceTasks(0);

    job.setMapOutputKeyClass(Text.class);
    job.setMapOutputValueClass(Text.class);
    job.setOutputKeyClass(Text.class);
    job.setOutputValueClass(Text.class);
    MultipleOutputs.addNamedOutput(job,"pkgproc",TextOutputFormat.class,Text.class,Text.class);
    MultipleOutputs.addNamedOutput(job,"pkgpss",TextOutputFormat.class,Text.class,Text.class);
    MultipleOutputs.addNamedOutput(job,"pkgsvc",TextOutputFormat.class,Text.class,Text.class);

    job.setInputFormatClass(CombineSmallfileInputFormat.class);
    job.setOutputFormatClass(TextOutputFormat.class);

    FileInputFormat.addInputPath(job, new Path(otherArgs[0]));
    FileOutputFormat.setOutputPath(job, new Path(otherArgs[1]));

    int exitFlag = job.waitForCompletion(true) ? 0 : 1;
    System.exit(exitFlag);

}

}

我的映射器代码

public class CombineSmallfileMapper extends Mapper<NullWritable, Text, Text, Text> {

    private Text file = new Text();
    private MultipleOutputs mos;
    private String period;
    private Long elapsed;

    @Override
    public void setup(Context context) throws IOException, InterruptedException {
        mos = new MultipleOutputs(context);
    }
    @Override
    protected void map(NullWritable key, Text value, Context context) throws IOException, InterruptedException {
        String file_name = context.getConfiguration().get("map.input.file.name");
        String [] filename_tokens = file_name.split("_");
        String uuid = filename_tokens[0];
        String [] datetime_tokens;
        try{
        datetime_tokens = filename_tokens[1].split("-");
        }catch(ArrayIndexOutOfBoundsException err){
            throw new ArrayIndexOutOfBoundsException(file_name);
        }
        String year,month,day,hour,minute,sec,msec;
        year = datetime_tokens[0];
        month = datetime_tokens[1];
        day = datetime_tokens[2];
        hour = datetime_tokens[3];
        minute = datetime_tokens[4];
        sec = datetime_tokens[5];
        msec = datetime_tokens[6];
        String datetime = year+"-"+month+"-"+"-"+day+" "+hour+":"+minute+":"+sec+"."+msec;
        String content = value.toString();
        String []lines = content.split("\n");
        for(int u = 0;u<lines.length;u++){
            String line = lines[u];
            String []tokens = line.split(",");
            if(tokens[0].equals("period")){
                period = tokens[1];
                try{
                long startTime = Long.valueOf(tokens[2]);
                long endTime = Long.valueOf(tokens[3]);
                elapsed = endTime-startTime;
                }catch(NumberFormatException err){
                    throw new NumberFormatException(line);
                }
            }else if(tokens[0].equals("pkgproc")){
                String proc_info = "";
                try{
                proc_info += period+","+String.valueOf(elapsed)+","+tokens[2]+","+tokens[3];
                }catch(ArrayIndexOutOfBoundsException err){
                    throw new ArrayIndexOutOfBoundsException("pkgproc: "+content+ "line:"+line);
                }
                for(int i = 4;i<tokens.length;i++){
                    String []state_info = tokens[i].split(":");
                    String state = "";
                    state += ","+state_info[0].charAt(0)+","+state_info[0].charAt(1)+","+state_info[0].charAt(2)+","+state_info[1];
                    mos.write("pkgproc",new Text(tokens[1]), new Text(proc_info+state+','+uuid+','+datetime));
                }
            }else if(tokens[0].equals("pkgpss")){
                String proc_info = "";
                proc_info += period+","+String.valueOf(elapsed)+","+tokens[2]+","+tokens[3];
                for(int i = 4;i<tokens.length;i++){
                    String []state_info = tokens[i].split(":");
                    String state = "";
                    state += ","+state_info[0].charAt(0)+","+state_info[0].charAt(1)+","+state_info[0].charAt(2)+","+state_info[1]+","+state_info[2]+","+state_info[3]+","+state_info[4]+","+state_info[5]+","+state_info[6]+","+state_info[7];
                    mos.write("pkgpss",new Text(tokens[1]), new Text(proc_info+state+','+uuid+','+datetime));
                }
            }else if(tokens[0].startsWith("pkgsvc")){
                String []stateName = tokens[0].split("-");
                String proc_info = "";
                //tokens[2] = uid, tokens[3] =  serviceName
                proc_info += stateName[1]+','+period+","+String.valueOf(elapsed)+","+tokens[2]+","+tokens[3];
                String opcount = tokens[4];
                for(int i = 5;i<tokens.length;i++){
                    String []state_info = tokens[i].split(":");
                    String state = "";
                    state += ","+state_info[0].charAt(0)+","+state_info[0].charAt(1)+","+state_info[1];
                    mos.write("pkgsvc",new Text(tokens[1]), new Text(proc_info+state+','+opcount+','+uuid+','+datetime));
                }
            }
        }
    }

}

我的 CombineFileInputFormat，它覆盖了 isSplitable 并返回 false

public class CombineSmallfileInputFormat extends CombineFileInputFormat<NullWritable, Text> {

    @Override
    public RecordReader<NullWritable, Text> createRecordReader(InputSplit split, TaskAttemptContext context) throws IOException {

        return new CombineFileRecordReader<NullWritable,Text>((CombineFileSplit) split,context,WholeFileRecordReader.class);
    }
    @Override
    protected boolean isSplitable(JobContext context,Path file ){
        return false;
    }

}

整个文件记录器

   public class WholeFileRecordReader extends RecordReader<NullWritable, Text> {
    //private static final Logger LOG = Logger.getLogger(WholeFileRecordReader.class);

      /** The path to the file to read. */
      private final Path mFileToRead;
      /** The length of this file. */
      private final long mFileLength;

      /** The Configuration. */
      private final Configuration mConf;

      /** Whether this FileSplit has been processed. */
      private boolean mProcessed;
      /** Single Text to store the file name of the current file. */
    //  private final Text mFileName;
      /** Single Text to store the value of this file (the value) when it is read. */
      private final Text mFileText;

      /**
       * Implementation detail: This constructor is built to be called via
       * reflection from within CombineFileRecordReader.
       *
       * @param fileSplit The CombineFileSplit that this will read from.
       * @param context The context for this task.
       * @param pathToProcess The path index from the CombineFileSplit to process in this record.
       */
      public WholeFileRecordReader(CombineFileSplit fileSplit, TaskAttemptContext context,
          Integer pathToProcess) {
        mProcessed = false;
        mFileToRead = fileSplit.getPath(pathToProcess);
        mFileLength = fileSplit.getLength(pathToProcess);
        mConf = context.getConfiguration();
        context.getConfiguration().set("map.input.file.name", mFileToRead.getName());

        assert 0 == fileSplit.getOffset(pathToProcess);
        //if (LOG.isDebugEnabled()) {
          //LOG.debug("FileToRead is: " + mFileToRead.toString());
          //LOG.debug("Processing path " + pathToProcess + " out of " + fileSplit.getNumPaths());

          //try {
            //FileSystem fs = FileSystem.get(mConf);
            //assert fs.getFileStatus(mFileToRead).getLen() == mFileLength;
          //} catch (IOException ioe) {
            //// oh well, I was just testing.
          //}
        //}

        //mFileName = new Text();
        mFileText = new Text();
      }

      /** {@inheritDoc} */
      @Override
      public void close() throws IOException {
        mFileText.clear();
      }

      /**
       * Returns the absolute path to the current file.
       *
       * @return The absolute path to the current file.
       * @throws IOException never.
       * @throws InterruptedException never.
       */
      @Override
      public NullWritable getCurrentKey() throws IOException, InterruptedException {
        return NullWritable.get();
      }

      /**
       * <p>Returns the current value.  If the file has been read with a call to NextKeyValue(),
       * this returns the contents of the file as a BytesWritable.  Otherwise, it returns an
       * empty BytesWritable.</p>
       *
       * <p>Throws an IllegalStateException if initialize() is not called first.</p>
       *
       * @return A BytesWritable containing the contents of the file to read.
       * @throws IOException never.
       * @throws InterruptedException never.
       */
      @Override
      public Text getCurrentValue() throws IOException, InterruptedException {
        return mFileText;
      }

      /**
       * Returns whether the file has been processed or not.  Since only one record
       * will be generated for a file, progress will be 0.0 if it has not been processed,
       * and 1.0 if it has.
       *
       * @return 0.0 if the file has not been processed.  1.0 if it has.
       * @throws IOException never.
       * @throws InterruptedException never.
       */
      @Override
      public float getProgress() throws IOException, InterruptedException {
        return (mProcessed) ? (float) 1.0 : (float) 0.0;
      }

      /**
       * All of the internal state is already set on instantiation.  This is a no-op.
       *
       * @param split The InputSplit to read.  Unused.
       * @param context The context for this task.  Unused.
       * @throws IOException never.
       * @throws InterruptedException never.
       */
      @Override
      public void initialize(InputSplit split, TaskAttemptContext context)
          throws IOException, InterruptedException {
        // no-op.
      }

      /**
       * <p>If the file has not already been read, this reads it into memory, so that a call
       * to getCurrentValue() will return the entire contents of this file as Text,
       * and getCurrentKey() will return the qualified path to this file as Text.  Then, returns
       * true.  If it has already been read, then returns false without updating any internal state.</p>
       *
       * @return Whether the file was read or not.
       * @throws IOException if there is an error reading the file.
       * @throws InterruptedException if there is an error.
       */
      @Override
      public boolean nextKeyValue() throws IOException, InterruptedException {
        if (!mProcessed) {
          if (mFileLength > (long) Integer.MAX_VALUE) {
            throw new IOException("File is longer than Integer.MAX_VALUE.");
          }
          byte[] contents = new byte[(int) mFileLength];

          FileSystem fs = mFileToRead.getFileSystem(mConf);
          FSDataInputStream in = null;
          try {
            // Set the contents of this file.
            in = fs.open(mFileToRead);
            IOUtils.readFully(in, contents, 0, contents.length);
            mFileText.set(contents, 0, contents.length);

          } finally {
            IOUtils.closeQuietly(in);
          }
          mProcessed = true;
          return true;
        }
        return false;
      }

}

我希望每个mapper都解析多个小文件，每个小文件不能分割。

但是，上面的代码将剪切(拆分)我的输入文件并引发解析错误(因为我的解析器会将行拆分为标记)。

在我的概念中，combineFileInputFormat 会将多个文件收集到一个分片中，每个分片将提供给一个映射器。因此，一个映射器可以处理多个文件。

在我的代码中，最大输入拆分设置为 25MB，所以我认为问题在于 combineFileInputFormat 将拆分输入拆分的小文件的最后一部分以满足拆分大小限制。

但是，我重写了 isSplitable 并返回 false，但它仍然拆分小文件。

正确的做法是什么？

我不确定是否可以为映射器指定文件数量，而不是指定输入拆分大小？

Answer 1

在构造函数代码中使用 setMaxSplitSize() 方法，它应该可以工作，它理想地告诉分割大小，

public class CFInputFormat extends CombineFileInputFormat<FileLineWritable, Text> {
  public CFInputFormat(){
    super();
    setMaxSplitSize(67108864); // 64 MB, default block size on hadoop
  }
  public RecordReader<FileLineWritable, Text> createRecordReader(InputSplit split, TaskAttemptContext context) throws IOException{
    return new CombineFileRecordReader<FileLineWritable, Text>((CombineFileSplit)split, context, CFRecordReader.class);
  }
  @Override
  protected boolean isSplitable(JobContext context, Path file){
    return false;
  }
}