Pig、Hive、MapReduce 解决分组 Top K 问题(转)

问题:

有如下数据文件 city.txt (id, city, value)

cat city.txt 
1 wh 500
2 bj 600
3 wh 100
4 sh 400
5 wh 200
6 bj 100
7 sh 200
8 bj 300
9 sh 900
需要按 city 分组聚合,然后从每组数据中取出前两条value最大的记录。

1、这是实际业务中经常会遇到的 group TopK 问题,下面来看看 pig 如何解决:

1 a = load '/data/city.txt'  using PigStorage(' 'as (id:chararray, city:chararray, value:int);
2 b = group by city;
3 c = foreach b {c1=order by value desc; c2=limit c1 2; generate group,c2.value;};
4 d = stream c through `sed 's/[(){}]//g'`;
5 dump d;
结果:
1 (bj,600,300)
2 (sh,900,400)
3 (wh,500,200)
这几行代码其实也实现了mysql中的 group_concat 函数的功能:
1 a = load '/data/city.txt'  using PigStorage(' 'as (id:chararray, city:chararray, value:int);
2 b = group by city;
3 c = foreach b {c1=order by value desc;  generate group,c1.value;};
4 d = stream c through `sed 's/[(){}]//g'`;
5 dump d;
结果:
1 (bj,600,300,100)
2 (sh,900,400,200)
3 (wh,500,200,100)

2、下面我们再来看看hive如何处理group topk的问题:

本质上HSQL和sql有很多相同的地方,但HSQL目前功能还有很多缺失,至少不如原生态的SQL功能强大,

比起PIG也有些差距,如果SQL中这类分组topk的问题如何解决呢?

1 select from city a where
2 2>(select count(1) from city where cname=a.cname and value>a.value)
3 distribute by a.cname sort by a.cname,a.value desc;
http://my.oschina.net/leejun2005/blog/78904

但是这种写法在HQL中直接报语法错误了,下面我们只能用hive udf的思路来解决了:

排序city和value,然后对city计数,最后where过滤掉city列计数器大于k的行即可。

好了,上代码:

(1)定义UDF:

01 package com.example.hive.udf;
02 import org.apache.hadoop.hive.ql.exec.UDF;
03       
04 public final class Rank extends UDF{
05     private int  counter;
06     private String last_key;
07     public int evaluate(final String key){
08       if ( !key.equalsIgnoreCase(this.last_key) ) {
09          this.counter = 0;
10          this.last_key = key;
11       }
12       return this.counter++;
13     }
14 }
(2)注册jar、建表、导数据,查询:
1 add jar Rank.jar;
2 create temporary function rank as 'com.example.hive.udf.Rank';
3 create table city(id int,cname string,value int) row format delimited fields terminated by ' ';
4 LOAD DATA LOCAL INPATH 'city.txt' OVERWRITE INTO TABLE city;
5 select cname, value from (
6     select cname,rank(cname) csum,value from (
7         select id, cname, value from city distribute by cname sort by cname,value desc
8     )a
9 )b where csum < 2;

(3)结果:

 

1 bj  600
2 bj  300
3 sh  900
4 sh  400
5 wh  500
6 wh  200
可以看到,hive相比pig来说,处理起来稍微复杂了点,但随着hive的日渐完善,以后比pig更简洁也说不定。

REF:hive中分组取前N个值的实现

http://baiyunl.iteye.com/blog/1466343

 

3、最后我们来看一下原生态的MR:

 

01 import java.io.IOException;
02 import java.util.TreeSet;
03  
04 import org.apache.hadoop.conf.Configuration;
05 import org.apache.hadoop.fs.Path;
06 import org.apache.hadoop.io.IntWritable;
07 import org.apache.hadoop.io.LongWritable;
08 import org.apache.hadoop.io.Text;
09 import org.apache.hadoop.mapreduce.Job;
10 import org.apache.hadoop.mapreduce.Mapper;
11 import org.apache.hadoop.mapreduce.Reducer;
12 import org.apache.hadoop.mapreduce.lib.input.FileInputFormat;
13 import org.apache.hadoop.mapreduce.lib.output.FileOutputFormat;
14 import org.apache.hadoop.util.GenericOptionsParser;
15  
16 public class GroupTopK {
17     // 这个 MR 将会取得每组年龄中 id 最大的前 3 个
18     // 测试数据由脚本生成:http://my.oschina.net/leejun2005/blog/76631
19     public static class GroupTopKMapper extends
20             Mapper<LongWritable, Text, IntWritable, LongWritable> {
21         IntWritable outKey = new IntWritable();
22         LongWritable outValue = new LongWritable();
23         String[] valArr = null;
24  
25         public void map(LongWritable key, Text value, Context context)
26                 throws IOException, InterruptedException {
27             valArr = value.toString().split("\t");
28             outKey.set(Integer.parseInt(valArr[2]));// age int
29             outValue.set(Long.parseLong(valArr[0]));// id long
30             context.write(outKey, outValue);
31         }
32     }
33  
34     public static class GroupTopKReducer extends
35             Reducer<IntWritable, LongWritable, IntWritable, LongWritable> {
36  
37         LongWritable outValue = new LongWritable();
38  
39         public void reduce(IntWritable key, Iterable<LongWritable> values,
40                 Context context) throws IOException, InterruptedException {
41             TreeSet<Long> idTreeSet = new TreeSet<Long>();
42             for (LongWritable val : values) {
43                 idTreeSet.add(val.get());
44                 if (idTreeSet.size() > 3) {
45                     idTreeSet.remove(idTreeSet.first());
46                 }
47             }
48             for (Long id : idTreeSet) {
49                 outValue.set(id);
50                 context.write(key, outValue);
51             }
52         }
53     }
54  
55     public static void main(String[] args) throws Exception {
56         Configuration conf = new Configuration();
57         String[] otherArgs = new GenericOptionsParser(conf, args)
58                 .getRemainingArgs();
59  
60         System.out.println(otherArgs.length);
61         System.out.println(otherArgs[0]);
62         System.out.println(otherArgs[1]);
63  
64         if (otherArgs.length != 3) {
65             System.err.println("Usage: GroupTopK <in> <out>");
66             System.exit(2);
67         }
68         Job job = new Job(conf, "GroupTopK");
69         job.setJarByClass(GroupTopK.class);
70         job.setMapperClass(GroupTopKMapper.class);
71         job.setReducerClass(GroupTopKReducer.class);
72         job.setNumReduceTasks(1);
73         job.setOutputKeyClass(IntWritable.class);
74         job.setOutputValueClass(LongWritable.class);
75         FileInputFormat.addInputPath(job, new Path(otherArgs[1]));
76         FileOutputFormat.setOutputPath(job, new Path(otherArgs[2]));
77         System.exit(job.waitForCompletion(true) ? 0 1);
78     }
79 }

hadoop jar GroupTopK.jar GroupTopK /tmp/decli/record_new.txt /tmp/1

结果:

 

hadoop fs -cat /tmp/1/part-r-00000
0       12869695
0       12869971
0       12869976
1       12869813
1       12869870
1       12869951

......

数据验证:

awk '$3==0{print $1}' record_new.txt|sort -nr|head -3
12869976
12869971
12869695

可以看到结果没有问题。 

注:测试数据由以下脚本生成:

http://my.oschina.net/leejun2005/blog/76631

 

PS:

如果说hive类似sql的话,那pig就类似plsql存储过程了:程序编写更自由,逻辑能处理的更强大了。

pig中还能直接通过反射调用java的静态类中的方法,这块内容请参考之前的相关pig博文。

附几个HIVE UDAF链接,有兴趣的同学自己看下:

Hive UDAF和UDTF实现group by后获取top值 http://blog.csdn.net/liuzhoulong/article/details/7789183
hive中自定义函数(UDAF)实现多行字符串拼接为一行 http://blog.sina.com.cn/s/blog_6ff05a2c0100tjw4.html
编写Hive UDAF http://www.fuzhijie.me/?p=118
Hive UDAF开发 http://richiehu.blog.51cto.com/2093113/386113

posted @ 2014-08-22 21:15  stubborn412  阅读(1032)  评论(0编辑  收藏  举报