粉丝：什么情况下，hive 只会产生一个reduce任务，而没有maptask

今天下午，在微信群里看到粉丝聊天，提到了一个某公司的面试题：

什么情况下，hive 只会产生一个reduce任务，而没有maptask

这个问题是不是很神奇？

我们常规使用的mapreducer任务执行过程大致如下图：

appmaster通过某种策略计算数据源可以做多少分片（getSplits方法），对应的生成固定数量的maptask，假如存在shuffle的话，就根据默认或者指定的reducer数，将数据分散给特定数量的reducer。不存在shuffle的话，reducer可以为零的。

正常逻辑：

mapTask的职责就是负责读数据，做ETL，也可以利用combiner局部聚合；ReduceTask输入严重依赖于mapper输出，所以‘一直’的逻辑仅有reducer无法执行的。

没有maptask，仅有一个reducerTask的hive任务。有点违背我们的使用常识了哦。

其实，正常使用的情况下，hive的sql模式执行引擎还是主要依赖于hadoop的mapreduce计算框架。

但是仅仅依赖于sql，然后依赖于hadoop的mapreduce，显然不能满足hive的野心的。所以hive想提供一个单独的可以使用java编写的hive自己的map/reduce的框架。

你在hive搜索reducer可以直接找到hive自己的reducer。

hql常规解析后依赖的mapreduce主要还是ExecReducer/ExecMapper--通过实现hadoop的mapper和reducer来实现具体执行逻辑。

hive自己的mapreducer有一个实现案例，就是GenericMR，实现源码如下：

public final class GenericMR {
  public void map(final InputStream in, final OutputStream out,
      final Mapper mapper) throws Exception {
    map(new InputStreamReader(in), new OutputStreamWriter(out), mapper);
  }


  public void map(final Reader in, final Writer out, final Mapper mapper) throws Exception {
    handle(in, out, new RecordProcessor() {
      @Override
      public void processNext(RecordReader reader, Output output) throws Exception {
        mapper.map(reader.next(), output);
      }
    });
  }


  public void reduce(final InputStream in, final OutputStream out,
      final Reducer reducer) throws Exception {
    reduce(new InputStreamReader(in), new OutputStreamWriter(out), reducer);
  }


  public void reduce(final Reader in, final Writer out, final Reducer reducer) throws Exception {
    handle(in, out, new RecordProcessor() {
      @Override
      public void processNext(RecordReader reader, Output output) throws Exception {
        reducer.reduce(reader.peek()[0], new KeyRecordIterator(
            reader.peek()[0], reader), output);
      }
    });
  }


  private void handle(final Reader in, final Writer out,
      final RecordProcessor processor) throws Exception {
    final RecordReader reader = new RecordReader(in);
    final OutputStreamOutput output = new OutputStreamOutput(out);


    try {
      while (reader.hasNext()) {
        processor.processNext(reader, output);
      }
    } finally {
      try {
        output.close();
      } finally {
        reader.close();
      }
    }
  }


  private static interface RecordProcessor {
    void processNext(final RecordReader reader, final Output output) throws Exception;
  }


  private static final class KeyRecordIterator implements Iterator {
    private final String key;
    private final RecordReader reader;


    private KeyRecordIterator(final String key, final RecordReader reader) {
      this.key = key;
      this.reader = reader;
    }


    @Override
    public boolean hasNext() {
      return (reader.hasNext() && key.equals(reader.peek()[0]));
    }


    @Override
    public String[] next() {
      if (!hasNext()) {
        throw new NoSuchElementException();
      }


      return reader.next();
    }


    @Override
    public void remove() {
      throw new UnsupportedOperationException();
    }
  }


  private static final class RecordReader {
    private final BufferedReader reader;
    private String[] next;


    private RecordReader(final InputStream in) {
      this(new InputStreamReader(in));
    }


    private RecordReader(final Reader in) {
      reader = new BufferedReader(in);
      next = readNext();
    }


    private String[] next() {
      final String[] ret = next;


      next = readNext();


      return ret;
    }


    private String[] readNext() {
      try {
        final String line = reader.readLine();
        return (line == null ? null : line.split("\t"));
      } catch (final Exception e) {
        throw new RuntimeException(e);
      }
    }


    private boolean hasNext() {
      return next != null;
    }


    private String[] peek() {
      return next;
    }


    private void close() throws Exception {
      reader.close();
      
    }
  }


  private static final class OutputStreamOutput implements Output {
    private final PrintWriter out;


    private OutputStreamOutput(final OutputStream out) {
      this(new OutputStreamWriter(out));
    }


    private OutputStreamOutput(final Writer out) {
      this.out = new PrintWriter(out);
    }


    public void close() throws Exception {
      out.close();
    }


    @Override
    public void collect(String[] record) throws Exception {
      out.println(_join(record, "\t"));
    }


    private static String _join(final String[] record, final String separator) {
      if (record == null || record.length == 0) {
        return "";
      }
      final StringBuilder sb = new StringBuilder();
      for (int i = 0; i < record.length; i++) {
        if (i > 0) {
          sb.append(separator);
        }
        sb.append(record[i]);
      }
      return sb.toString();
    }
  }
}

重点：

常规的mapreducer的reducer输入依赖于mapper的输出的，所以无法单独执行。但是GenericMR的实现reducer里可以直接支持InputStream/Reader，所以就可以直接生成java的指定输入流或者reader即可。

new GenericMR().reduce(new StringReader("a\tb\tc"), new StringWriter(),
          new Reducer() {
        public void reduce(String key, Iterator records,
            Output output) throws Exception {
          while (true) {
            records.next();
          }
        }
});

这个问题确实很另类，浪尖估计是被面试者简历写了精通hive源码，才被大佬由此一问。要不正常使用者，不会注意到这个框架。

不过这个也给大家提个醒，关注框架使用的同时，也要关注框架的历史及发展。

总结一下就是：hive野心很大，不想仅仅限于hql，想提供一个单独的可以用java编写的hive自己的map/reduce计算框架。

欢迎关注浪尖，留言更多比较有趣的问题哦！