Hadoop实现Secondary Sort（一）

Hadoop的MapReduce模型支持基于key的排序，即在一次MapReduce之后，结果都是按照key的大小排序的。但是在很多应用情况下，我们需要对映射在一个key下的value集合进行排序，即“secondary sort”。 

    在《hadoop the definate guide》的P227的“secondary sort”章节中，以<year,temperature>为例，在map阶段按照year来分发temperature，在reduce阶段按照同一year对应的temperature大小排序。 

本文以<String, Int>格式为例，先介绍已知类型的二次排序，再介绍泛型<key,value>集合的二次排序。 

设输入为如下的序列对： 

str1 3 

str2 2 

str1 1 

str3 9 

str2 10 

我们期望的输出结果为： 

str1 1,3 

str2 2,10 

str3 9 

1 value集合的二次排序 

（1）先定义一个TextInt类，将String及int对象封装为一个整体。由于TextInt在mapreduce中要作为key进行比较，必须实现WritableComparable接口。如下： 

public class TextInt implements WritableComparable<TextInt>{ private String text ; private int value; public TextInt(){} public TextInt(String text , int value){ this.text = text; this.value = value; } public String getFirst(){ return this.text; } public int getSecond(){ return this.value; } @Override public void readFields(DataInput in) throws IOException { text = in.readUTF(); value = in.readInt(); } @Override public void write(DataOutput out) throws IOException { out.writeUTF(text); out.writeInt(value); } @Override public int compareTo(TextInt that) { return this.text.compareTo(that.text); } } 

（2）我们用KeyValueTextInputFormat的方式来读取输入文件，以tab等分割符切分输入的<key,value>对，key，value都是Text类型。所以在mapper阶段需要把将value还原到int数据，同时封装String及int为TextInt； 

@Override public void map(Text key, Text value, OutputCollector<TextInt,IntWritable> output, Reporter reporter) throws IOException { int intValue = Integer.parseInt(value.toString()); TextInt ti = new TextInt(key.toString(), intValue); output.collect(ti, new IntWritable(intValue)); } 

（3）在reduce阶段，为了方便查看输出数据，我们把同一个string对应的int数据封装在一起，如下： 

@Override public void reduce(TextInt key, Iterator<IntWritable> values, OutputCollector<Text, Text> output, Reporter reporter) throws IOException { StringBuffer combineValue = new StringBuffer(); while( values.hasNext()){ int value = values.next().get(); combineValue.append(value + ","); } output.collect(new Text(key.getFirst()),newText(combineValue.toString())); } 

（4）上面的map及reduce跟普通的mapreduce没什么区别，很容易理解，但真正实现二次排序的是以下两个comparator及一个partitioner。 

1)TextIntComparator：先比较TextInt的String，再比较value； 

public static class TextIntComparator extends WritableComparator{ public TextIntComparator(){ super(TextInt.class, true); //注册comparator } @Override public int compare(WritableComparable a, WritableComparable b) { TextInt o1 = (TextInt)a; TextInt o2 = (TextInt)b; if ( !o1.getFirst().equals(o2.getFirst())){ return o1.getFirst().compareTo(o2.getFirst()); } else{ return o1.getSecond() - o2.getSecond(); } }} 

2)TextComparator: 只比较TextInt中的String。 

public static class TextComparator extends WritableComparator{ public TextComparator(){ super(TextInt.class, true); } @Override public int compare(WritableComparable a, WritableComparable b) { TextInt o1 = (TextInt)a; TextInt o2 = (TextInt)b; return o1.getFirst().compareTo(o2.getFirst()); } } 

3) PartitionByText：根据TextInt中的String来分割TextInt对象： 

public static class PartitionByText implementsPartitioner<TextInt, IntWritable>{ @Override public int getPartition(TextInt key, IntWritable value, intnumPartitions) { return (key.getFirst().hashCode() & Integer.MAX_VALUE) %numPartitions; } @Override public void configure(JobConf job) {} } 

（5）OK，基础工作都完成了，现在看实际的job调用： 

//……define input & output //定义Job JobConf conf = new JobConf(Join.class); conf.setJobName("sort by value"); //add inputpath: FileInputFormat.addInputPath(conf, new Path(input)); conf.setInputFormat(KeyValueTextInputFormat.class); conf.setMapperClass(Mapper.class);conf.setMapOutputKeyClass(TextInt.class); conf.setMapOutputValueClass(IntWritable.class); conf.setOutputKeyComparatorClass(TextIntComparator.class); conf.setOutputValueGroupingComparator(TextComparator.class);conf.setPartitionerClass(PartitionByText.class); conf.setOutputKeyClass(Text.class); conf.setOutputValueClass(Text.class); conf.setReducerClass(Reduce.class); conf.setOutputFormat(TextOutputFormat.class); //output FileOutputFormat.setOutputPath(conf, new Path(output)); JobClient.runJob(conf); 

通过上面的job执行，在output的文件中，就能看到类似的输出结果了： 

str1 1,3 

str2 2,10 

str3 9 

上面介绍了如何实现，在“知其然”的基础上，我们进一步“知其所以然”。首先，我们需要把相同string的int数据分发到同一个reducer，这就是PartitionByText的作用；其次，用TextComparator控制reduce阶段int数据集合的group，即把相同的string对应的int数据组装在一起；最后，用TextIntComparator实现在reduce阶段的排列方式。 

2 泛型value的二次排序 

在上一个章节中，将已知类型的key和value封装成TextInt对象，然后做二次排序。在很多情况下，我们需要对不同类型的key或value做二次排序，或者value是一个二元组/多元组，实现多级的排序，这时可以使用泛型的key或value。 

定义泛型的key/value组合对象，key和value均要实现WritableComparable接口。 

public static class CombinedObject implements WritableComparable{private static Configuration conf = new Configuration(); private Class<? extends WritableComparable> firstClass; private Class<? extends WritableComparable> secondClass; private WritableComparable first ; private WritableComparable second; public CombinedObject(){ } public CombinedObject(Class<? extends WritableComparable>keyClass, Class<? extends WritableComparable> valueClass){ if (keyClass == null || valueClass == null) { throw new IllegalArgumentException("nullvalueClass"); } this.firstClass = keyClass; this.secondClass = valueClass; first = ReflectionUtils.newInstance(firstClass, conf); second = ReflectionUtils.newInstance(secondClass, conf); } public CombinedObject(WritableComparable f, WritableComparables){ this(f.getClass(), s.getClass()); setFirst(f); setSecond(s); } public void setFirst(WritableComparable key){ this.first =key; } public void setSecond(WritableComparable value){ this.second = value; } public WritableComparable getFirst(){ return this.first; } @Override public void readFields(DataInput in) throws IOException { String firstClassName = in.readUTF(); String secondClassName = in.readUTF(); firstClass = (Class<? extends WritableComparable>)WritableName.getClass(firstClassName, conf); secondClass = (Class<? extends WritableComparable>)WritableName.getClass(secondClassName, conf); first = ReflectionUtils.newInstance(firstClass, conf); second = ReflectionUtils.newInstance(secondClass, conf); first.readFields(in); second.readFields(in); } @Override public void write(DataOutput out) throws IOException { out.writeUTF(firstClass.getName()); out.writeUTF(secondClass.getName()); first.write(out); second.write(out); } @Override public boolean equals(Object o) { if( o instanceof CombinedObject){ CombinedObject that = (CombinedObject)o; return that.first.equals(this.first); } return false; } @Override public int hashCode() { return first.hashCode(); } @Override public int compareTo(Object o) { CombinedObject that = (CombinedObject)o; return that.first.compareTo(this.first); } } 

注意： 

在Writable接口的write(DataOutputout)及readFields(DataInput in)函数中，需要考虑firstClass及secondClass的序列化及反序列化。 

在comparator接口的equal,hashCode,compareTo三个具体实现中，只考虑first的比较即可。 

KeyComparator：只比较first 

public static class KeyComparator extends WritableComparator{ public KeyComparator(){ super(CombinedObject.class, true); } @Override public int compare(WritableComparable a, WritableComparable b) { CombinedObject o1 = (CombinedObject)a; CombinedObject o2 = (CombinedObject)b; return o1.first.compareTo(o2.first); } } 

KeyValueComparator:比较first之后，再比较second。 

public static class KeyValueComparator extends WritableComparator{ public KeyValueComparator(){ super(CombinedObject.class, true); } @Override public int compare(WritableComparable a, WritableComparable b) { CombinedObject o1 = (CombinedObject)a; CombinedObject o2 = (CombinedObject)b; if ( !o1.first.equals(o2.first)){ return o1.first.compareTo(o2.first); } else{ return o2.second.compareTo(o1.second); //descend } } } 

KeyPartitioner：根据first来分发CombinedObject。 

public static class KeyPartitioner implementsPartitioner<CombinedObject, Writable>{ @Override public int getPartition(CombinedObject key, Writable value, intnumPartitions) { return (key.getFirst().hashCode() & Integer.MAX_VALUE) %numPartitions; } @Override public void configure(JobConf job) { } } 

具体的调用方式: 

job.setMapperClass(ContentMapper.class); job.setMapOutputKeyClass(CombinedObject.class); // sort value list job.setOutputKeyComparatorClass(KeyValueComparator.class); job.setOutputValueGroupingComparator(KeyComparator.class); job.setPartitionerClass(KeyPartitioner.class); job.setReducerClass(SameInfoReducer.class); 

在这个章节介绍了对泛型的key和value的二次排序，可以设计更多组合形式的<key,value>的排序，也可以自定义多元组构成value，实现更多灵活的排序方式。

本文转自《hadoop开发者》第二期。