MapReduce 实现全排序的方式

默认情况下，mr只对key排序。我们所说的全排序，即对key的全排序。

1、使用一个reducer

这个是最容易想到的思路，优点是实现简单，缺点也很明显，一个reduce有可能比较慢。

2、重写Partioner类。

通过重写Partition类，把key在一个范围内的发往一个固定的reduce，这样在一个reduce内key是全排序的，在reduce之间按照序号也是排好序的。比如key代表的是一个年龄。我们可以把数据输出到10个reduer。1-10岁之间发往第0个reduce,11-20发往第2个reduce，以此类推。

缺点是这种划分可能不均匀。

3,、使用TotalOrderPartition

我们知道Mapreduce框架在feed数据给reducer之前会对map output key排序，这种排序机制保证了每一个reducer局部有序，hadoop 默认的partitioner是HashPartitioner，它依赖于output key的hashcode，使得相同key会去相同reducer，但是不保证全局有序，如果想要获得全局排序结果（比如获取top N, bottom N），就需要用到TotalOrderPartitioner了，它保证了相同key去相同reducer的同时也保证了全局有序。

 public class HashPartitioner extends Partitioner {
   /** Use {@link Object#hashCode()} to partition. */
   public int getPartition(K key, V value,
                           int numReduceTasks) {
     return (key.hashCode() & Integer.MAX_VALUE) % numReduceTasks;
   }
 }

 /**
  * Partitioner effecting a total order by reading split points from
  * an externally generated source.
  */
 @InterfaceAudience.Public
 @InterfaceStability.Stable
 public class TotalOrderPartitionerextends WritableComparable,V>
     extends Partitioner implements Configurable {
   // by construction, we know if our keytype
   @SuppressWarnings("unchecked") // is memcmp-able and uses the trie
   public int getPartition(K key, V value, int numPartitions) {
     return partitions.findPartition(key);
   }
 }

TotalOrderPartitioner依赖于一个partition file来distribute keys，partition file是一个实现计算好的sequence file，如果我们设置的reducer number是N，那么这个文件包含（N-1）个key分割点，并且是基于key comparator排好序的。TotalOrderPartitioner会检查每一个key属于哪一个reducer的范围内，然后决定分发给哪一个reducer。

InputSampler类的writePartitionFile方法会对input files取样并创建partition file。有三种取样方法：

1. RandomSampler  随机取样

2. IntervalSampler  从s个split里面按照一定间隔取样，通常适用于有序数据

3. SplitSampler  从s个split中选取前n条记录取样

paritition file可以通过TotalOrderPartitioner.setPartitionFile(conf, partitionFile)来设置，在TotalOrderPartitioner instance创建的时候会调用setConf函数，这时会读入partition file中key值，如果key是BinaryComparable(可以认为是字符串类型)的话会构建trie，时间复杂度是O(n), n是树的深度。如果是非BinaryComparable类型就构建BinarySearchNode，用二分查找，时间复杂度O(log(n))，n是reduce数

 boolean natOrder =
   conf.getBoolean(NATURAL_ORDER, true);
 if (natOrder && BinaryComparable.class.isAssignableFrom(keyClass)) {
   partitions = buildTrie((BinaryComparable[])splitPoints, 0,
       splitPoints.length, new byte[0],
       // Now that blocks of identical splitless trie nodes are
       // represented reentrantly, and we develop a leaf for any trie
       // node with only one split point, the only reason for a depth
       // limit is to refute stack overflow or bloat in the pathological
       // case where the split points are long and mostly look like bytes
       // iii...iixii...iii   .  Therefore, we make the default depth
       // limit large but not huge.
       conf.getInt(MAX_TRIE_DEPTH, 200));
 } else {
   partitions = new BinarySearchNode(splitPoints, comparator);
 }

示例程序

 import org.apache.hadoop.conf.Configuration;
 import org.apache.hadoop.fs.Path;
 import org.apache.hadoop.io.Text;
 import org.apache.hadoop.mapreduce.Job;
 import org.apache.hadoop.mapreduce.lib.input.FileInputFormat;
 import org.apache.hadoop.mapreduce.lib.input.KeyValueTextInputFormat;
 import org.apache.hadoop.mapreduce.lib.output.FileOutputFormat;
 import org.apache.hadoop.mapreduce.lib.partition.InputSampler;
 import org.apache.hadoop.mapreduce.lib.partition.InputSampler.RandomSampler;
 import org.apache.hadoop.mapreduce.lib.partition.TotalOrderPartitioner;

 public class TotalSortMR {

     public static int runTotalSortJob(String[] args) throws Exception {
         Path inputPath = new Path(args[0]);
         Path outputPath = new Path(args[1]);
         Path partitionFile = new Path(args[2]);
         int reduceNumber = Integer.parseInt(args[3]);

         // RandomSampler第一个参数表示key会被选中的概率，第二个参数是一个选取samples数，第三个参数是最大读取input splits数
         RandomSampler sampler = new InputSampler.RandomSampler(0.1, 10000, 10);

         Configuration conf = new Configuration();
         // 设置partition file全路径到conf
         TotalOrderPartitioner.setPartitionFile(conf, partitionFile);

         Job job = new Job(conf);
         job.setJobName("Total-Sort");
         job.setJarByClass(TotalSortMR.class);
         job.setInputFormatClass(KeyValueTextInputFormat.class);
         job.setMapOutputKeyClass(Text.class);
         job.setMapOutputValueClass(Text.class);
         job.setNumReduceTasks(reduceNumber);

         // partitioner class设置成TotalOrderPartitioner
         job.setPartitionerClass(TotalOrderPartitioner.class);

         FileInputFormat.setInputPaths(job, inputPath);
         FileOutputFormat.setOutputPath(job, outputPath);
         outputPath.getFileSystem(conf).delete(outputPath, true);

         // 写partition file到mapreduce.totalorderpartitioner.path
         InputSampler.writePartitionFile(job, sampler);

         return job.waitForCompletion(true)? 0 : 1;

     }

     public static void main(String[] args) throws Exception{
         System.exit(runTotalSortJob(args));
     }
 }

上面的例子是采用InputSampler来创建partition file，其实还可以使用mapreduce来创建，可以自定义一个inputformat来取样，将output key输出到一个reducer

ps:hive 0.12实现了parallel ORDER BY(https://issues.apache.org/jira/browse/HIVE-1402)，也是基于TotalOrderPartitioner，非常靠谱的new feature啊