一起学Hive——总结常用的Hive优化技巧

今天总结本人在使用Hive过程中的一些优化技巧，希望给大家带来帮助。Hive优化最体现程序员的技术能力，面试官在面试时最喜欢问的就是Hive的优化技巧。

技巧1.控制reducer数量

下面的内容是我们每次在hive命令行执行SQL时都会打印出来的内容：

In order to change the average load for a reducer (in bytes):
  set hive.exec.reducers.bytes.per.reducer=<number>
In order to limit the maximum number of reducers:
  set hive.exec.reducers.max=<number>
In order to set a constant number of reducers:
  set mapreduce.job.reduces=<number>

很多人都会有个疑问，上面的内容是干什么用的。我们一一来解答，先看

set hive.exec.reducers.bytes.per.reducer=<number>，这个一条Hive命令，用于设置在执行SQL的过程中每个reducer处理的最大字节数量。可以在配置文件中设置，也可以由我们在命令行中直接设置。如果处理的数据量大于number，就会多生成一个reudcer。例如，number = 1024K，处理的数据是1M，就会生成10个reducer。我们来验证下上面的说法是否正确：

1. 执行set hive.exec.reducers.bytes.per.reducer=200000;命令，设置每个reducer处理的最大字节是200000。
2. 执行sql:

select user_id,count(1) as cnt 
  from orders group by user_id limit 20; 

执行上面的sql时会在控制台打印出信息：

  Number of reduce tasks not specified. Estimated from input data size: 159
In order to change the average load for a reducer (in bytes):
  set hive.exec.reducers.bytes.per.reducer=<number>
In order to limit the maximum number of reducers:
  set hive.exec.reducers.max=<number>
In order to set a constant number of reducers:
  set mapreduce.job.reduces=<number>
Starting Job = job_1538917788450_0020, Tracking URL = http://hadoop-master:8088/proxy/application_1538917788450_0020/
Kill Command = /usr/local/src/hadoop-2.6.1/bin/hadoop job  -kill job_1538917788450_0020
Hadoop job information for Stage-1: number of mappers: 1; number of reducers: 159

控制台打印的信息中第一句话：Number of reduce tasks not specified. Estimated from input data size: 159。翻译成中文：没有指定reducer任务数量，根据输入的数据量估计会有159个reducer任务。然后在看最后一句话：number of mappers: 1; number of reducers: 159。确定该SQL最终生成159个reducer。因此如果我们知道数据的大小，只要通过set hive.exec.reducers.bytes.per.reducer命令设置每个reducer处理数据的大小就可以控制reducer的数量。

接着看
set hive.exec.reducers.max=<number> 这也是一条Hive命令，用于设置Hive的最大reducer数量，如果我们设置number为50，表示reducer的最大数量是50。
我们来验证下这个说法是否正确：

1. 执行命令set hive.exec.reducers.max=8;设置reducer的数量为8。
2. 继续执行sql:

select user_id,count(1) as cnt 
  from orders group by user_id limit 20; 

会在控制台打印出如下信息：

Number of reduce tasks not specified. Estimated from input data size: 8
In order to change the average load for a reducer (in bytes):
  set hive.exec.reducers.bytes.per.reducer=<number>
In order to limit the maximum number of reducers:
  set hive.exec.reducers.max=<number>
In order to set a constant number of reducers:
  set mapreduce.job.reduces=<number>
Starting Job = job_1538917788450_0020, Tracking URL = http://hadoop-master:8088/proxy/application_1538917788450_0020/
Kill Command = /usr/local/src/hadoop-2.6.1/bin/hadoop job  -kill job_1538917788450_0020
Hadoop job information for Stage-1: number of mappers: 1; number of reducers: 8

控制台打印的信息中第一句话：Number of reduce tasks not specified. Estimated from input data size: 8。reducer的数据量为8，正好验证了我们的说法。set set hive.exec.reducers.max=8;命令是设置reducer的数量的上界。

最后来看 set mapreduce.job.reduces=<number>命令。这条Hive命令是设置reducer的数据，在执行sql会生成多少个reducer处理数据。使用和上面同样的方法来验证set mapreduce.job.reduces=这条命令。

1. 执行命令set mapreduce.job.reduces=5;设置reducer的数量为8。
2. 继续执行sql:

select user_id,count(1) as cnt 
  from orders group by user_id limit 20; 

会在控制台打印出如下信息：

Number of reduce tasks not specified. Defaulting to jobconf value of: 5
In order to change the average load for a reducer (in bytes):
  set hive.exec.reducers.bytes.per.reducer=<number>
In order to limit the maximum number of reducers:
  set hive.exec.reducers.max=<number>
In order to set a constant number of reducers:
  set mapreduce.job.reduces=<number>
Starting Job = job_1538917788450_0026, Tracking URL = http://hadoop-master:8088/proxy/application_1538917788450_0026/
Kill Command = /usr/local/src/hadoop-2.6.1/bin/hadoop job  -kill job_1538917788450_0026
Hadoop job information for Stage-1: number of mappers: 1; number of reducers: 5

根据Number of reduce tasks not specified. Defaulting to jobconf value of: 5和number of mappers: 1; number of reducers: 5这两句话，可以知道生成5个reducer。

如果我们将数量由5改成15。还是执行select user_id,count(1) as cnt
from orders group by user_id limit 20;SQL，在控制台打印的信息是：

Launching Job 1 out of 1
Number of reduce tasks not specified. Defaulting to jobconf value of: 15
In order to change the average load for a reducer (in bytes):
  set hive.exec.reducers.bytes.per.reducer=<number>
In order to limit the maximum number of reducers:
  set hive.exec.reducers.max=<number>
In order to set a constant number of reducers:
  set mapreduce.job.reduces=<number>
Starting Job = job_1538917788450_0027, Tracking URL = http://hadoop-master:8088/proxy/application_1538917788450_0027/
Kill Command = /usr/local/src/hadoop-2.6.1/bin/hadoop job  -kill job_1538917788450_0027
Hadoop job information for Stage-1: number of mappers: 1; number of reducers: 15

可见reducer的数量已经由5变为15个。

小结，控制hive中reducer的数量由三种方式，分别是：

set hive.exec.reducers.bytes.per.reducer=<number> 
set hive.exec.reducers.max=<number>
set mapreduce.job.reduces=<number>

其中 set mapreduce.job.reduces=<number>的方式优先级最高， set hive.exec.reducers.max=<number>优先级次之， set hive.exec.reducers.bytes.per.reducer=<number> 优先级最低。从hive0.14开始，一个reducer处理文件的大小的默认值是256M。

reducer的数量并不是越多越好，我们知道有多少个reducer就会生成多少个文件，小文件过多在hdfs中就会占用大量的空间，造成资源的浪费。如果reducer数量过小，导致某个reducer处理大量的数据（数据倾斜就会出现这样的现象），没有利用hadoop的分而治之功能，甚至会产生OOM内存溢出的错误。使用多少个reducer处理数据和业务场景相关，不同的业务场景处理的办法不同。

技巧2.使用Map join

sql中涉及到多张表的join，当有一张表的大小小于1G时，使用Map Join可以明显的提高SQL的效率。如果最小的表大于1G，使用Map Join会出现OOM的错误。
用法：

select /*+ MAPJOIN(table_a)*/,a.*,b.* from table_a a join table_b b on a.id = b.id

技巧3.使用distinct + union all代替union

如果遇到要使用union去重的场景，使用distinct + union all比使用union的效果好。
distinct + union all的用法：

select count(distinct *) 
from (
select order_id,user_id,order_type from orders where order_type='0' union all
select order_id,user_id,order_type from orders where order_type='1' union all 
select order_id,user_id,order_type from orders where order_type='1'
)a;

union的用法：

select count(*) 
from(
select order_id,user_id,order_type from orders where order_type='0' union
select order_id,user_id,order_type from orders where order_type='0' union
select order_id,user_id,order_type from orders where order_type='1')t;

技巧4.解决数据倾斜的通用办法

数据倾斜的现象：任务进度长时间维持在99%，只有少量reducer任务完成，未完成任务数据读写量非常大，超过10G。在聚合操作是经常发生。
通用解决方法：set hive.groupby.skewindata=true;
将一个map reduce拆分成两个map reduce。

说说我遇到过的一个场景，需用统计某个一天每个用户的访问量，SQL如下：

select t.user_id,count(*) from user_log t group by t.user_id

执行这条语句之后，发现任务维持在99%达到一个小时。后面自己分析user_log表，发现user_id有很多数据为null。user_id为null的数据会有一个reducer来处理，导致出现数据倾斜的现象。解决方法有两种：
1、通过where条件过滤掉user_id为null的记录。
2、将为null的user_id设置一个随机数值。保证所有数据平均的分配到所有的reducer中处理。