> Linux集群 > Hadoop >

hadoop-reduce分析

Map的结果,会通过partition分发到Reducer上,Reducer做完Reduce操作后,通过OutputFormat,进行输出

Map的结果,会通过partition分发到Reducer上,Reducer做完Reduce操作后,通过OutputFormat,进行输出。

 

?
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
* Licensed to the Apache Software Foundation (ASF) under one
 
package org.apache.hadoop.mapreduce;
 
import java.io.IOException;
 
 * Reduces a set of intermediate values which share a key to a smaller set of
public class Reducer<KEYIN,VALUEIN,KEYOUT,VALUEOUT> {
 
  public class Context
    extends ReduceContext<KEYIN,VALUEIN,KEYOUT,VALUEOUT> {
    public Context(Configuration conf, TaskAttemptID taskid,
                   RawKeyValueIterator input,
                   Counter inputKeyCounter,
                   Counter inputValueCounter,
                   RecordWriter<KEYOUT,VALUEOUT> output,
                   OutputCommitter committer,
                   StatusReporter reporter,
                   RawComparator<KEYIN> comparator,
                   Class<KEYIN> keyClass,
                   Class<VALUEIN> valueClass
                   ) throws IOException, InterruptedException {
      super(conf, taskid, input, inputKeyCounter, inputValueCounter,
            output, committer, reporter,
            comparator, keyClass, valueClass);
    }
  }
 
  /**
   * Called once at the start of the task.
   */
  protected void setup(Context context
                       ) throws IOException, InterruptedException {
    // NOTHING
  }
 
  /**
   * This method is called once for each key. Most applications will define
   * their reduce class by overriding this method. The default implementation
   * is an identity function.
   */
  @SuppressWarnings("unchecked")
  protected void reduce(KEYIN key, Iterable<VALUEIN> values, Context context
                        ) throws IOException, InterruptedException {
    for(VALUEIN value: values) {
      context.write((KEYOUT) key, (VALUEOUT) value);
    }
  }
 
  /**
   * Called once at the end of the task.
   */
  protected void cleanup(Context context
                         ) throws IOException, InterruptedException {
    // NOTHING
  }
 
  /**
   * Advanced application writers can use the
   * {@link #run(org.apache.hadoop.mapreduce.Reducer.Context)} method to
   * control how the reduce task works.
   */
  public void run(Context context) throws IOException, InterruptedException {
    setup(context);
    while (context.nextKey()) {
      reduce(context.getCurrentKey(), context.getValues(), context);
    }
    cleanup(context);
  }
}

 

 

Mapper的结果,可能送到可能的Combiner做合并,Combiner在系统中并没有自己的基类,而是用Reducer作为Combiner的基类,他们对外的功能是一样的,只是使用的位置和使用时的上下文不太一样而已。

Mapper最终处理的结果对<key, value>,是需要送到Reducer去合并的,合并的时候,有相同key的键/值对会送到同一个Reducer那,哪个key到哪个Reducer的分配过程,是由Partitioner规定的,它只有一个方法,输入是Map的结果对<key, value>和Reducer的数目,输出则是分配的Reducer(整数编号)。系统缺省的Partitioner是HashPartitioner,它以key的Hash值对Reducer的数目取模,得到对应的Reducer。

 

 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
* Licensed to the Apache Software Foundation (ASF) under one
 
package org.apache.hadoop.mapreduce;
 
 * Partitions the key space.
public abstract class Partitioner<KEY, VALUE> {
  
  /**
   * Get the partition number for a given key (hence record) given the total
   * number of partitions i.e. number of reduce-tasks for the job.
   *  
   * <p>Typically a hash function on a all or a subset of the key.</p>
   *
   * @param key the key to be partioned.
   * @param value the entry value.
   * @param numPartitions the total number of partitions.
   * @return the partition number for the <code>key</code>.
   */
  public abstract int getPartition(KEY key, VALUE value, int numPartitions);
  
}
 
 * Licensed to the Apache Software Foundation (ASF) under one
 
package org.apache.hadoop.mapreduce.lib.partition;
 
import org.apache.hadoop.mapreduce.Partitioner;
 
/** Partition keys by their {@link Object#hashCode()}. */
public class HashPartitioner<K, V> extends Partitioner<K, V> {
 
  /** Use {@link Object#hashCode()} to partition. */
  public int getPartition(K key, V value,
                          int numReduceTasks) {
    return (key.hashCode() & Integer.MAX_VALUE) % numReduceTasks;
  }
 
}

 

Reducer是所有用户定制Reducer类的基类,和Mapper类似,它也有setup,reduce,cleanup和run方法,其中setup和cleanup含义和Mapper相同,reduce是真正合并Mapper结果的地方,它的输入是key和这个key对应的所有value的一个迭代器,同时还包括Reducer的上下文。系统中定义了两个非常简单的Reducer,IntSumReducer和LongSumReducer,分别用于对整形/长整型的value求和。

 

 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
* Licensed to the Apache Software Foundation (ASF) under one
 
package org.apache.hadoop.mapreduce.lib.reduce;
 
import java.io.IOException;
 
public class IntSumReducer<Key> extends Reducer<Key,IntWritable,
                                                Key,IntWritable> {
  private IntWritable result = new IntWritable();
 
  public void reduce(Key key, Iterable<IntWritable> values,
                     Context context) throws IOException, InterruptedException {
    int sum = 0;
    for (IntWritable val : values) {
      sum += val.get();
    }
    result.set(sum);
    context.write(key, result);
  }
 
}

 

Reduce的结果,通过Reducer.Context的方法collect输出到文件中,和输入类似,Hadoop引入了OutputFormat。OutputFormat依赖两个辅助接口:RecordWriter和OutputCommitter,来处理输出。RecordWriter提供了write方法,用于输出<key, value>和close方法,用于关闭对应的输出。OutputCommitter提供了一系列方法,用户通过实现这些方法,可以定制OutputFormat生存期某些阶段需要的特殊操作。我们在TaskInputOutputContext中讨论过这些方法(明显,TaskInputOutputContext是OutputFormat和Reducer间的桥梁)。

OutputFormat和RecordWriter分别对应着InputFormat和RecordReader,系统提供了空输出NullOutputFormat(什么结果都不输出,NullOutputFormat.RecordWriter只是示例,系统中没有定义),LazyOutputFormat(没在类图中出现,不分析),FilterOutputFormat(不分析)和基于文件FileOutputFormat的SequenceFileOutputFormat和TextOutputFormat输出。

基于文件的输出FileOutputFormat利用了一些配置项配合工作,包括mapred.output.compress:是否压缩;mapred.output.compression.codec:压缩方法;mapred.output.dir:输出路径;mapred.work.output.dir:输出工作路径。FileOutputFormat还依赖于FileOutputCommitter,通过FileOutputCommitter提供一些和Job,Task相关的临时文件管理功能。如FileOutputCommitter的setupJob,会在输出路径下创建一个名为_temporary的临时目录,cleanupJob则会删除这个目录。

SequenceFileOutputFormat输出和TextOutputFormat输出分别对应输入的SequenceFileInputFormat和TextInputFormat

 


(责任编辑:IT)