关于MapReduce中自定义Combine类（一）

MRJobConfig

     public static fina COMBINE_CLASS_ATTR

     属性COMBINE_CLASS_ATTR = "mapreduce.job.combine.class"

     ————子接口（F4） JobContent

           方法getCombinerClass

             ————子实现类 JobContextImpl

                 实现getCombinerClass方法：

                 public Class<? extends Reducer<?,?,?,?>> getCombinerClass()

                          throws ClassNotFoundException {

                      return (Class<? extends Reducer<?,?,?,?>>)

                        conf.getClass(COMBINE_CLASS_ATTR, null);

                 }

                 因为JobContextImpl是MRJobConfig子类

                 所以得到了父类MRJobConfig的COMBINE_CLASS_ATTR属性

                 ————子类Job

                     public void setCombinerClass(Class<? extends Reducer> cls

                               ) throws IllegalStateException {

                     ensureState(JobState.DEFINE);

                     conf.setClass(COMBINE_CLASS_ATTR, cls, Reducer.class);

                     }

                因为JobContextImpl是MRJobConfig子类,

                而Job是JobContextImpl的子类

                所以也有COMBINE_CLASS_ATTR属性

                通过setCombinerClass设置了父类MRJobConfig的属性

 

 

MRJobConfig

    ————子接口JobContent

        方法getCombinerClass

        ————子实现类 JobContextImpl

            ————子类 Job

        ————子实现类 TaskAttemptContext

            继承了方法getCombinerClass

 

Task   

   $CombinerRunner(Task的内部类)   

            该内部类有方法create：

            public static <K,V> CombinerRunner<K,V> create(JobConf job,

                               TaskAttemptID taskId,

                               Counters.Counter inputCounter,

                               TaskReporter reporter,

                               org.apache.hadoop.mapreduce.OutputCommitter committer

                              ) throws ClassNotFoundException

            {

                  Class<? extends Reducer<K,V,K,V>> cls =

                    (Class<? extends Reducer<K,V,K,V>>) job.getCombinerClass();

                  if (cls != null) {

                    return new OldCombinerRunner(cls, job, inputCounter, reporter);

                  }

                  // make a task context so we can get the classes

                  org.apache.hadoop.mapreduce.TaskAttemptContext taskContext =

                    new org.apache.hadoop.mapreduce.task.TaskAttemptContextImpl(job, taskId,

                        reporter);

                  Class<? extends org.apache.hadoop.mapreduce.Reducer<K,V,K,V>> newcls =

                    (Class<? extends org.apache.hadoop.mapreduce.Reducer<K,V,K,V>>)

                       taskContext.getCombinerClass();

                  if (newcls != null) {

                    return new NewCombinerRunner<K,V>(newcls, job, taskId, taskContext,

                                                      inputCounter, reporter, committer);

                  }

                  return null;

            }

                  其中这一段应该是旧的API

                  Class<? extends Reducer<K,V,K,V>> cls =

                          (Class<? extends Reducer<K,V,K,V>>) job.getCombinerClass();

                  if (cls != null) {

                          return new OldCombinerRunner(cls, job, inputCounter, reporter);

                  }

                  而这个是新的API

                  org.apache.hadoop.mapreduce.TaskAttemptContext taskContext =

                    new org.apache.hadoop.mapreduce.task.TaskAttemptContextImpl(job, taskId,

                        reporter);

                  Class<? extends org.apache.hadoop.mapreduce.Reducer<K,V,K,V>> newcls =

                    (Class<? extends org.apache.hadoop.mapreduce.Reducer<K,V,K,V>>)

                       taskContext.getCombinerClass();

                  if (newcls != null) {

                    return new NewCombinerRunner<K,V>(newcls, job, taskId, taskContext,

                                                      inputCounter, reporter, committer);

                  }

                  return null;

                  （不知道为什么要写全名，去掉那些包名、向上/下转型和各种泛型的话，看起来就会清晰很多？）

                  而TaskAttemptContext是JobContent的子实现类，所以继承了getCombinerClass方法

                  而且，这里用的是多态，其调用的是子实现类TaskAttemptContextImpl的getCombinerClass方法

                  （TaskAttemptContextImpl继承了JobContextImpl，而JobContextImpl实现了该方法）

                  所以最终get到了属性COMBINE_CLASS_ATTR，即得到了我们通过job.setCombinerClass的xxxC

                    而这个xxxC是给了newcls，而newcls是给了NewCombinerRunner的构造函数的reducerClassc参数

                      NewCombinerRunner(Class reducerClass,

                          JobConf job,

                          org.apache.hadoop.mapreduce.TaskAttemptID taskId,

                          org.apache.hadoop.mapreduce.TaskAttemptContext context,

                          Counters.Counter inputCounter,

                          TaskReporter reporter,

                          org.apache.hadoop.mapreduce.OutputCommitter committer)

                      {

                          super(inputCounter, job, reporter);

                          this.reducerClass = reducerClass;

                          this.taskId = taskId;

                          keyClass = (Class<K>) context.getMapOutputKeyClass();

                          valueClass = (Class<V>) context.getMapOutputValueClass();

                          comparator = (RawComparator<K>) context.getCombinerKeyGroupingComparator();

                          this.committer = committer;

                      }

Task          

  MapTask

        $MapOutputBuffer

            private CombinerRunner<K,V> combinerRunner;

            $SpillThread类($表示内部类)

                combinerRunner = CombinerRunner.create(job, getTaskID(),

                                             combineInputCounter,

                                             reporter, null);

                //此时，我们得到了设置好的合并类                            

                if (combinerRunner == null) {

                      // spill directly

                      DataInputBuffer key = new DataInputBuffer();

                      while (spindex < mend &&

                          kvmeta.get(offsetFor(spindex % maxRec) + PARTITION) == i) {

                        final int kvoff = offsetFor(spindex % maxRec);

                        int keystart = kvmeta.get(kvoff + KEYSTART);

                        int valstart = kvmeta.get(kvoff + VALSTART);

                        key.reset(kvbuffer, keystart, valstart - keystart);

                        getVBytesForOffset(kvoff, value);

                        writer.append(key, value);

                        ++spindex;

                      }

                } else {

                      int spstart = spindex;

                      while (spindex < mend &&

                          kvmeta.get(offsetFor(spindex % maxRec)

                                    + PARTITION) == i) {

                        ++spindex;

                      }

                      // Note: we would like to avoid the combiner if we've fewer

                      // than some threshold of records for a partition

                      if (spstart != spindex) {

                        combineCollector.setWriter(writer);

                        RawKeyValueIterator kvIter =

                          new MRResultIterator(spstart, spindex);

                        combinerRunner.combine(kvIter, combineCollector);

                      }

                }

            

            再查看combine函数

            在Task的内部类NewCombinerRunner下

            public void combine(RawKeyValueIterator iterator,

                                OutputCollector<K,V> collector)

                throws IOException, InterruptedException,ClassNotFoundException

            {

              // make a reducer

              org.apache.hadoop.mapreduce.Reducer<K,V,K,V> reducer =

                (org.apache.hadoop.mapreduce.Reducer<K,V,K,V>)

                  ReflectionUtils.newInstance(reducerClass, job);

              org.apache.hadoop.mapreduce.Reducer.Context

                   reducerContext = createReduceContext(reducer, job, taskId,

                                                        iterator, null, inputCounter,

                                                        new OutputConverter(collector),

                                                        committer,

                                                        reporter, comparator, keyClass,

                                                        valueClass);

              reducer.run(reducerContext);

            }

            上面的reducerClass就是我们传入的xxxC

            最终是通过反射创建了一个xxxC对象，并将其强制向上转型为Reducer实例对象，

            然后调用了向上转型后对象的run方法（当前的xxxC没有run方法，调用的是父类Reduce的run）

            在类Reducer中，run方法如下

            /**

           * 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);

            try {

              while (context.nextKey()) {

                reduce(context.getCurrentKey(), context.getValues(), context);

                // If a back up store is used, reset it

                Iterator<VALUEIN> iter = context.getValues().iterator();

                if(iter instanceof ReduceContext.ValueIterator) {

                  ((ReduceContext.ValueIterator<VALUEIN>)iter).resetBackupStore();       

                }

              }

            } finally {

              cleanup(context);

            }

          }

          有由于多态，此时调用的reduce是子类xxxC中的reduce方法

         （多态态性质：子类复写了该方法，则实际上执行的是子类中的该方法）

          所以说，我们自定义combine用的类的时候，应该继承Reducer类，并且复写reduce方法

          且其输入形式：(以wordcount为例)

       reduce(Text key, Iterable<IntWritable> values, Context context)

       其中key是单词个数，而values是个数列表，也就是value1、value2........

       注意，此时已经是列表，即<键，list<值1、值2、值3.....>>

       (之所以得到这个结论，是因为我当时使用的combine类是WCReduce，

        即Reduce和combine所用的类是一样的，通过对代码的分析，传入值的结构如果是<lkey,value>的话，是不可能做到combine的啊——即所谓的对相同值合并，求计数的累积和，这根本就是两个步骤，对key相同的键值对在map端就进行了一次合并了，合并成了<key,value list>，然后才轮到combine接受直接换个形式的输入，并处理——我们的处理是求和，然后再输出到context，进入reduce端的shuffle过程。

        然后我在reduce中遍历了用syso输出

        结果发现是0，而这实际上是因为经过一次遍历，我的指针指向的位置就不对了啊，

        )

嗯，自己反复使用以下的代码，不断的组合、注释，去测试吧~就会得出这样的结论了

1. /reduce
2.     publicstaticclassWCReduce extends Reducer<Text,IntWritable,Text,IntWritable>{
3.         private final IntWritableValueOut=newIntWritable();
4.         @Override
5.         protectedvoid reduce(Text key,Iterable<IntWritable> values,
6.                 Context context)  throws IOException,InterruptedException{
7.             for(IntWritable value : values){
8.                 System.out.println(value.get()+"--");
9.             }
10.  
11. //            int total = 0 ;
12. //            for (IntWritable value : values) {
13. //                total += value.get();
14. //            }
15. //            ValueOut.set(total);
16. //            context.write(key, ValueOut);
17.         }
18.  
19.     }
20.           
21. job.setCombinerClass(WCReduce.class);

 

 

来自为知笔记(Wiz)

附件列表