概述
MapReduce框架对处理结果的输出会根据key值进行默认的排序,这个默认排序可以满足一部分需求,但是也是十分有限的,在我们实际的需求当中,往往有要对reduce输出结果进行二次排序的需求。对于二次排序的实现,网络上已经有很多人分享过了,但是对二次排序的实现原理及整个MapReduce框架的处理流程的分析还是有非常大的出入,而且部分分析是没有经过验证的。本文将通过一个实际的MapReduce二次排序的例子,讲述二次排序的实现和其MapReduce的整个处理流程,并且通过结果和Map、Reduce端的日志来验证描述的处理流程的正确性。
数据样例
60-01-94-59-DD-5F2020-03-30 23:56:51113.40870623.128417
60-01-94-59-DD-5F2020-03-30 23:57:23113.40870623.128417
60-01-94-59-DD-5F2020-03-30 23:57:36113.40870623.128417
60-01-94-59-DD-5F2020-03-30 23:58:01113.40870623.128417
04-ED-33-F8-BB-C32020-03-30 23:56:56113.33595123.126707
04-ED-33-F8-BB-C32020-03-30 23:56:59113.33595123.126707
50-EC-50-1F-7E-E82020-03-30 23:57:11113.33595123.126707
04-ED-33-F8-BB-C32020-03-30 23:57:31113.33595123.126707
50-EC-50-1F-7E-E82020-03-30 23:58:10113.33595123.126707
50-EC-50-1F-7E-E82020-03-30 23:58:28113.33595123.126707
需求:
需求一:
统计每个身份(mac)的运动轨迹.需要有时间(10s时间内被同一个地点设备重复采集时认为一个数据),经纬度,且运动轨迹以时间升序排序
结果样例
04-ED-33-F8-BB-C3 START-->[2020-03-30 23:56:56](23.126707,113.335951)-->[2020-03-30 23:56:59](23.126707,113.335951)-->[2020-03-30 23:57:31](23.126707,113.335951)-->END
50-EC-50-1F-7E-E8 START-->[2020-03-30 23:57:11](23.126707,113.335951)-->[2020-03-30 23:58:10](23.126707,113.335951)-->[2020-03-30 23:58:28](23.126707,113.335951)-->END
60-01-94-59-DD-5F START-->[2020-03-30 23:56:51](23.128417,113.408706)-->[2020-03-30 23:57:23](23.128417,113.408706)-->[2020-03-30 23:57:36](23.128417,113.408706)-->[2020-03-30 23:58:01](23.128417,113.408706)-->END
需求分析
根据需求,有效字段为:采集时间,身份ID,无线AP经度,无线AP纬度
有效字段数据样例如下:
60-01-94-59-DD-5F2020-03-30 23:56:51113.40870623.128417
60-01-94-59-DD-5F2020-03-30 23:57:23113.40870623.128417
60-01-94-59-DD-5F2020-03-30 23:57:36113.40870623.128417
60-01-94-59-DD-5F2020-03-30 23:58:01113.40870623.128417
04-ED-33-F8-BB-C32020-03-30 23:56:56113.33595123.126707
04-ED-33-F8-BB-C32020-03-30 23:56:59113.33595123.126707
50-EC-50-1F-7E-E82020-03-30 23:57:11113.33595123.126707
04-ED-33-F8-BB-C32020-03-30 23:57:31113.33595123.126707
50-EC-50-1F-7E-E82020-03-30 23:58:10113.33595123.126707
50-EC-50-1F-7E-E82020-03-30 23:58:28113.33595123.126707
解决思路
Map端处理
根据需求,我们有一个非常明确的目标就是要对第一列相同的记录,并且对合并后的时间进行排序。我们都知道MapReduce框架不管是默认排序或者是自定义排序都只是对key值进行排序,现在的情况是这些数据不是key值,怎么办?其实我们可以将原始数据的key值和其对应的数据组合成一个新的key值,然后新的key值对应的value还是原始数据中的valu。那么我们就可以将原始数据的map输出变成类似下面的数据结构:
{[60-01-94-59-DD-5F,2020-03-30 23:56:51],2020-03-30 23:56:51,113.408706,23.128417}
{[60-01-94-59-DD-5F,2020-03-30 23:57:23],2020-03-30 23:57:23,113.408706,23.128417}
{[60-01-94-59-DD-5F,2020-03-30 23:57:36],2020-03-30 23:57:36,113.408706,23.128417}
{[60-01-94-59-DD-5F,2020-03-30 23:58:01],2020-03-30 23:58:01,113.408706,23.128417}
{[04-ED-33-F8-BB-C3,2020-03-30 23:56:56],2020-03-30 23:56:56,113.335951,23.126707}
{[04-ED-33-F8-BB-C3,2020-03-30 23:56:59],2020-03-30 23:56:59,113.335951,23.126707}
{[04-ED-33-F8-BB-C3,2020-03-30 23:57:31],2020-03-30 23:57:31,113.335951,23.126707}
{[50-EC-50-1F-7E-E8,2020-03-30 23:57:11],2020-03-30 23:57:11,113.335951,23.126707}
{[50-EC-50-1F-7E-E8,2020-03-30 23:58:10],2020-03-30 23:58:10,113.335951,23.126707}
{[50-EC-50-1F-7E-E8,2020-03-30 23:58:28],2020-03-30 23:58:28,113.335951,23.126707}
那么我们只需要对[]里面的key值进行排序就OK了,然后我们需要自定义一个分区处理器,因为我的目标不是想将新key相同的记录传到一个reduce中,而是想将新key中第一个字段相同的记录放到同一个reduce中进行分组合并,所以我们需要根据新key值的第一个字段来自定义一个分区处理器。通过分区操作后,得到的数据流如下:
Partition1:{[60-01-94-59-DD-5F,2020-03-30 23:56:51],2020-03-30 23:56:51,113.408706,23.128417}、{[60-01-94-59-DD-5F,2020-03-30 23:57:23],2020-03-30 23:57:23,113.408706,23.128417},{[60-01-94-59-DD-5F,2020-03-30 23:57:36],2020-03-30 23:57:36,113.408706,23.128417},{[60-01-94-59-DD-5F,2020-03-30 23:58:01],2020-03-30 23:58:01,113.408706,23.128417}
Partition2:{[04-ED-33-F8-BB-C3,2020-03-30 23:56:56],2020-03-30 23:56:56,113.335951,23.126707},{[04-ED-33-F8-BB-C3,2020-03-30 23:56:59],2020-03-30 23:56:59,113.335951,23.126707},{[04-ED-33-F8-BB-C3,2020-03-30 23:57:31],2020-03-30 23:57:31,113.335951,23.126707}
Partition3:{[50-EC-50-1F-7E-E8,2020-03-30 23:57:11],2020-03-30 23:57:11,113.335951,23.126707},{[50-EC-50-1F-7E-E8,2020-03-30 23:58:10],2020-03-30 23:58:10,113.335951,23.126707},{[50-EC-50-1F-7E-E8,2020-03-30 23:58:28],2020-03-30 23:58:28,113.335951,23.126707}
分区操作完成之后,我调用自己的自定义排序器对新的key值进行排序
{[04-ED-33-F8-BB-C3,2020-03-30 23:56:56],2020-03-30 23:56:56,113.335951,23.126707}
{[04-ED-33-F8-BB-C3,2020-03-30 23:56:59],2020-03-30 23:56:59,113.335951,23.126707}
{[04-ED-33-F8-BB-C3,2020-03-30 23:57:31],2020-03-30 23:57:31,113.335951,23.126707}
{[50-EC-50-1F-7E-E8,2020-03-30 23:57:11],2020-03-30 23:57:11,113.335951,23.126707}
{[50-EC-50-1F-7E-E8,2020-03-30 23:58:10],2020-03-30 23:58:10,113.335951,23.126707}
{[50-EC-50-1F-7E-E8,2020-03-30 23:58:28],2020-03-30 23:58:28,113.335951,23.126707}
{[60-01-94-59-DD-5F,2020-03-30 23:56:51],2020-03-30 23:56:51,113.408706,23.128417}
{[60-01-94-59-DD-5F,2020-03-30 23:57:23],2020-03-30 23:57:23,113.408706,23.128417}
{[60-01-94-59-DD-5F,2020-03-30 23:57:36],2020-03-30 23:57:36,113.408706,23.128417}
{[60-01-94-59-DD-5F,2020-03-30 23:58:01],2020-03-30 23:58:01,113.408706,23.128417}
Reduce端处理
经过Shuffle处理之后,数据传输到Reducer端了。在Reducer端按照组合键的第一个字段进行分组,并且每处理完一次分组之后就会调用一次reduce函数来对这个分组进行处理和输出。最终各个分组的数据结果变成类似下面的数据结构:
04-ED-33-F8-BB-C3 START-->[2020-03-30 23:56:56](23.126707,113.335951)-->[2020-03-30 23:56:59](23.126707,113.335951)-->[2020-03-30 23:57:31](23.126707,113.335951)-->END
50-EC-50-1F-7E-E8 START-->[2020-03-30 23:57:11](23.126707,113.335951)-->[2020-03-30 23:58:10](23.126707,113.335951)-->[2020-03-30 23:58:28](23.126707,113.335951)-->END
60-01-94-59-DD-5F START-->[2020-03-30 23:56:51](23.128417,113.408706)-->[2020-03-30 23:57:23](23.128417,113.408706)-->[2020-03-30 23:57:36](23.128417,113.408706)-->[2020-03-30 23:58:01](23.128417,113.408706)-->END
核心代码
入口类MotionTrajectoryApp
public class MotionTrajectoryApp {
public static void main (String[] args) throws IOException, ClassNotFoundException, InterruptedException {
Configuration conf = new Configuration();
//没有这里可能会报权限错误,因为远程部署用户与当前开发用户不一致
// System.setProperty("HADOOP_USER_NAME","hadoop");
//获取配置 , 连接到hdfs
// conf.set("fs.defaultFS", "hdfs://hadoop01:9000");
// conf.set("hadoop.user.name", "hadoop");
//访问本地
// conf.set("mapreduce.framework.name","local");
// conf.set("fs.defaultFS","file:///");
Job job = Job.getInstance(conf,"Motion-Trajectory");
job.setJarByClass(MotionTrajectoryApp.class);
//指定mapper类和reduce类
job.setMapperClass(MotionTrajectoryMapper.class);
job.setReducerClass(MotionTrajectoryReduce.class);
//指定mappertask输出类型
job.setMapOutputKeyClass(MotionScore.class);
job.setMapOutputValueClass(Text.class);
//指定reducetask输出类型
job.setOutputKeyClass(Text.class);
job.setOutputValueClass(Text.class);
job.setInputFormatClass(TextInputFormat.class);
job.setOutputFormatClass(TextOutputFormat.class);
//设置分区和reduce数量(reduce的数量,和分区的数量对应,因为分区为一个,所以reduce的数量也是一个)
job.setPartitionerClass(MotionTrajectoryPartitioner.class);
job.setNumReduceTasks