问题？

一个工作任务，统计一个按天分区每天都有百亿条数据条的hive表中account字段的非重用户数（大概两千万）。后来又更改为按id字段分别统计每个id的用户数。
我很轻易的跳出来了count（distinct account）这个句子。然后写上了一行查询，等待了四个小时，然后map反着跑
就知道没这么容易的任务。。
然后想起来Hive SQL 基于的mapreduce是并行计算，百亿的数据可不是平时测试时的mysql里的几百条数据。
这么想来应该是map和reduce的内存不够，

set mapreduce.map.memory.mb=48192; set mapreduce.reduce.memory.mb=48192;

执行语句

select count（distinct account） from…where…

继续mapreduce，三个小时后报错error in shuffle in fetcher#3. shuffle过程又出问题了。
找呀找，reducer只有1？ 那还怎么并行？果断

set mapred.reduce.tasks=1000；

又进行查询，发现reducer 还是1，为什么？

原来因为加入distinct，map阶段不能用combine消重，数据输出为（key，value）形式然后在reduce阶段进行消重。
重点是，Hive在处理COUNT这种“全聚合(full aggregates)”计算时，它会忽略用户指定的Reduce Task数，而强制使用1。

解决办法：

转换为子查询，转化为两个mapreduce任务 先select distinct的字段，然后在count（），这样去重就会分发到不同的reduce块，count依旧是一个reduce但是只需要计数即可。

select count（*） from （select distinct account form tablename where…）t;

这样大概半小时可以得到结果。

后来需求改变为对这个表按account的类型（字段名为id）统计每个类型的account非重复数。
如果按照上述方法，在查询条件添加 where id=..，这样每个查询都需要半小时，效率很低。
优化方法：利用gourp by 按id，account分组，存入一个临时表 只需要对临时表进行统计即可

insert overwrite table temp select id，account，count(1) as num from tablename group by id，account；

这样temp表里的数据直接就是非重数据，并且按id升序排序，按id筛选 count(*)即可。 sum(num)也可统计总数。


华丽的分割线

count（distinct id）的原理


SELECT COUNT( DISTINCT id ) FROM TABLE_NAME WHERE …;

count（distinct id）从执行计划上面来看：只有一个reducer任务（即使你设置reducer任务为100个，实际上也没有用），所有的id都
会聚集到同一个reducer任务进行去重然后在聚合，这非常容易造成数据倾斜.

| 1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38 | STAGE DEPENDENCIES:
Stage-1 is a root stage
Stage-0 depends on stages: Stage-1

STAGE PLANS:
Stage: Stage-1
Map Reduce
Map Operator Tree:
TableScan
alias: emp_ct
Statistics: Num rows: 42 Data size: 171 Basic stats: COMPLETE Column stats: NONE
Select Operator
expressions: dept_num (type: int)
outputColumnNames: _col0
Statistics: Num rows: 42 Data size: 171 Basic stats: COMPLETE Column stats: NONE
Reduce Output Operator
key expressions: _col0 (type: int)
sort order: +
Statistics: Num rows: 42 Data size: 171 Basic stats: COMPLETE Column stats: NONE
Reduce Operator Tree:
Group By Operator
aggregations: count(DISTINCT KEY._col0:0._col0)
mode: complete
outputColumnNames: _col0
Statistics: Num rows: 1 Data size: 8 Basic stats: COMPLETE Column stats: NONE
File Output Operator
compressed: false
Statistics: Num rows: 1 Data size: 8 Basic stats: COMPLETE Column stats: NONE
table:
input format: org.apache.hadoop.mapred.TextInputFormat
output format: org.apache.hadoop.hive.ql.io.HiveIgnoreKeyTextOutputFormat
serde: org.apache.hadoop.hive.serde2.lazy.LazySimpleSerDe

Stage: Stage-0
Fetch Operator
limit: -1
Processor Tree:
ListSink

| | —- | —- |


运行示例：注意设置的reducer任务数量实际上是不生效的。

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37		 hive> set mapreduce.job.reduces=5;
hive>
> select count(distinct dept_num)
> from emp_ct;
Query ID = mart_fro_20200320233947_4f60c190-4967-4da6-bf3e-97db786fbc6c
Total jobs = 1
Launching Job 1 out of 1
Number of reduce tasks determined at compile time: 1
In order to change the average load for a reducer (in bytes):
set hive.exec.reducers.bytes.per.reducer=
In order to limit the maximum number of reducers:
set hive.exec.reducers.max=
In order to set a constant number of reducers:
set mapreduce.job.reduces=
Start submit job !
Start GetSplits
GetSplits finish, it costs : 32 milliseconds
Submit job success : job_1584341089622_358496
Starting Job = job_1584341089622_358496, Tracking URL = http://BJHTYD-Hope-25-11.hadoop.jd.local:50320/proxy/application_1584341089622_358496/
Kill Command = /data0/hadoop/hadoop_2.100.31_2019090518/bin/hadoop job -kill job_1584341089622_358496
Hadoop job(job_1584341089622_358496) information for Stage-1: number of mappers: 2; number of reducers: 1
2020-03-20 23:39:58,215 Stage-1(job_1584341089622_358496) map = 0%, reduce = 0%
2020-03-20 23:40:09,628 Stage-1(job_1584341089622_358496) map = 50%, reduce = 0%, Cumulative CPU 2.74 sec
2020-03-20 23:40:16,849 Stage-1(job_1584341089622_358496) map = 100%, reduce = 0%, Cumulative CPU 7.43 sec
2020-03-20 23:40:29,220 Stage-1(job_1584341089622_358496) map = 100%, reduce = 100%, Cumulative CPU 10.64 sec
MapReduce Total cumulative CPU time: 10 seconds 640 msec
Stage-1 Elapsed : 40533 ms job_1584341089622_358496
Ended Job = job_1584341089622_358496
MapReduce Jobs Launched:
Stage-1: Map: 2 Reduce: 1 Cumulative CPU: 10.64 sec HDFS Read: 0.000 GB HDFS Write: 0.000 GB SUCCESS Elapsed : 40s533ms job_1584341089622_358496

Total MapReduce CPU Time Spent: 10s640ms
Total Map: 2 Total Reduce: 1
Total HDFS Read: 0.000 GB Written: 0.000 GB
OK
3
Time taken: 43.025 seconds, Fetched: 1 row(s)


回到顶部

count（distinct id）的解决方案

该怎么解决这个问题呢？实际上解决方法非常巧妙：
我们利用Hive对嵌套语句的支持，将原来一个MapReduce作业转换为两个作业，在第一阶段选出全部的非重复id，在第二阶段再对
这些已消重的id进行计数。这样在第一阶段我们可以通过增大Reduce的并发数，并发处理Map输出。在第二阶段，由于id已经消重，
因此COUNT()操作在Map阶段不需要输出原id数据，只输出一个合并后的计数即可。这样即使第二阶段Hive强制指定一个Reduce Task，
极少量的Map输出数据也不会使单一的Reduce Task成为瓶颈。改进后的SQL语句如下：

SELECT COUNT() FROM (SELECT DISTINCT id FROM TABLE_NAME WHERE … ) t;

查看一下执行计划：

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61		 STAGE DEPENDENCIES:
Stage-1 is a root stage
Stage-2 depends on stages: Stage-1
Stage-0 depends on stages: Stage-2

STAGE PLANS:
Stage: Stage-1
Map Reduce
Map Operator Tree:
TableScan
alias: emp_ct
Statistics: Num rows: 42 Data size: 171 Basic stats: COMPLETE Column stats: NONE
Select Operator
expressions: dept_num (type: int)
outputColumnNames: dept_num
Statistics: Num rows: 42 Data size: 171 Basic stats: COMPLETE Column stats: NONE
Reduce Output Operator
key expressions: dept_num (type: int)
sort order: +
Map-reduce partition columns: dept_num (type: int)
Statistics: Num rows: 42 Data size: 171 Basic stats: COMPLETE Column stats: NONE
Reduce Operator Tree:
Group By Operator
keys: KEY._col0 (type: int)
mode: complete
outputColumnNames: _col0
Statistics: Num rows: 21 Data size: 85 Basic stats: COMPLETE Column stats: NONE
File Output Operator
compressed: false
table:
input format: org.apache.hadoop.mapred.SequenceFileInputFormat
output format: org.apache.hadoop.hive.ql.io.HiveSequenceFileOutputFormat
serde: org.apache.hadoop.hive.serde2.lazybinary.LazyBinarySerDe

Stage: Stage-2
Map Reduce
Map Operator Tree:
TableScan
Reduce Output Operator
sort order:
Statistics: Num rows: 21 Data size: 85 Basic stats: COMPLETE Column stats: NONE
value expressions: _col0 (type: int)
Reduce Operator Tree:
Group By Operator
aggregations: count(VALUE._col0)
mode: complete
outputColumnNames: _col0
Statistics: Num rows: 1 Data size: 8 Basic stats: COMPLETE Column stats: NONE
File Output Operator
compressed: false
Statistics: Num rows: 1 Data size: 8 Basic stats: COMPLETE Column stats: NONE
table:
input format: org.apache.hadoop.mapred.TextInputFormat
output format: org.apache.hadoop.hive.ql.io.HiveIgnoreKeyTextOutputFormat
serde: org.apache.hadoop.hive.serde2.lazy.LazySimpleSerDe

Stage: Stage-0
Fetch Operator
limit: -1
Processor Tree:
ListSink

具体看一下执行结果：注意看reducer任务的数量，第一个reducer任务是5个，第二个是1个.

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69		 hive> set mapreduce.job.reduces=5;
hive>
> select count(dept_num)
> from (
> select distinct dept_num
> from emp_ct
> ) t1;
Query ID = mart_fro_20200320234453_68ad3780-c3e5-44bc-94df-58a8f2b01f59
Total jobs = 2
Launching Job 1 out of 2
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=
In order to limit the maximum number of reducers:
set hive.exec.reducers.max=
In order to set a constant number of reducers:
set mapreduce.job.reduces=
Start submit job !
Start GetSplits
GetSplits finish, it costs : 13 milliseconds
Submit job success : job_1584341089622_358684
Starting Job = job_1584341089622_358684, Tracking URL = http://BJHTYD-Hope-25-11.hadoop.jd.local:50320/proxy/application_1584341089622_358684/
Kill Command = /data0/hadoop/hadoop_2.100.31_2019090518/bin/hadoop job -kill job_1584341089622_358684
Hadoop job(job_1584341089622_358684) information for Stage-1: number of mappers: 2; number of reducers: 5
2020-03-20 23:45:02,920 Stage-1(job_1584341089622_358684) map = 0%, reduce = 0%
2020-03-20 23:45:23,533 Stage-1(job_1584341089622_358684) map = 50%, reduce = 0%, Cumulative CPU 3.48 sec
2020-03-20 23:45:25,596 Stage-1(job_1584341089622_358684) map = 100%, reduce = 0%, Cumulative CPU 7.08 sec
2020-03-20 23:45:32,804 Stage-1(job_1584341089622_358684) map = 100%, reduce = 20%, Cumulative CPU 9.43 sec
2020-03-20 23:45:34,861 Stage-1(job_1584341089622_358684) map = 100%, reduce = 40%, Cumulative CPU 12.39 sec
2020-03-20 23:45:36,923 Stage-1(job_1584341089622_358684) map = 100%, reduce = 80%, Cumulative CPU 18.47 sec
2020-03-20 23:45:40,011 Stage-1(job_1584341089622_358684) map = 100%, reduce = 100%, Cumulative CPU 23.23 sec
MapReduce Total cumulative CPU time: 23 seconds 230 msec
Stage-1 Elapsed : 46404 ms job_1584341089622_358684
Ended Job = job_1584341089622_358684
Launching Job 2 out of 2
Number of reduce tasks determined at compile time: 1
In order to change the average load for a reducer (in bytes):
set hive.exec.reducers.bytes.per.reducer=
In order to limit the maximum number of reducers:
set hive.exec.reducers.max=
In order to set a constant number of reducers:
set mapreduce.job.reduces=
Start submit job !
Start GetSplits
GetSplits finish, it costs : 47 milliseconds
Submit job success : job_1584341089622_358729
Starting Job = job_1584341089622_358729, Tracking URL = http://BJHTYD-Hope-25-11.hadoop.jd.local:50320/proxy/application_1584341089622_358729/
Kill Command = /data0/hadoop/hadoop_2.100.31_2019090518/bin/hadoop job -kill job_1584341089622_358729
Hadoop job(job_1584341089622_358729) information for Stage-2: number of mappers: 5; number of reducers: 1
2020-03-20 23:45:48,353 Stage-2(job_1584341089622_358729) map = 0%, reduce = 0%
2020-03-20 23:46:05,846 Stage-2(job_1584341089622_358729) map = 20%, reduce = 0%, Cumulative CPU 2.62 sec
2020-03-20 23:46:06,873 Stage-2(job_1584341089622_358729) map = 60%, reduce = 0%, Cumulative CPU 8.49 sec
2020-03-20 23:46:08,931 Stage-2(job_1584341089622_358729) map = 80%, reduce = 0%, Cumulative CPU 11.53 sec
2020-03-20 23:46:09,960 Stage-2(job_1584341089622_358729) map = 100%, reduce = 0%, Cumulative CPU 15.23 sec
2020-03-20 23:46:35,639 Stage-2(job_1584341089622_358729) map = 100%, reduce = 100%, Cumulative CPU 20.37 sec
MapReduce Total cumulative CPU time: 20 seconds 370 msec
Stage-2 Elapsed : 54552 ms job_1584341089622_358729
Ended Job = job_1584341089622_358729
MapReduce Jobs Launched:
Stage-1: Map: 2 Reduce: 5 Cumulative CPU: 23.23 sec HDFS Read: 0.000 GB HDFS Write: 0.000 GB SUCCESS Elapsed : 46s404ms job_1584341089622_358684

Stage-2: Map: 5 Reduce: 1 Cumulative CPU: 20.37 sec HDFS Read: 0.000 GB HDFS Write: 0.000 GB SUCCESS Elapsed : 54s552ms job_1584341089622_358729

Total MapReduce CPU Time Spent: 43s600ms
Total Map: 7 Total Reduce: 6
Total HDFS Read: 0.000 GB Written: 0.000 GB
OK
3
Time taken: 103.692 seconds, Fetched: 1 row(s)


这个解决方案有点类似于set hive.groupby.skew.indata参数的作用！
实际测试：

1
2
3
4
5
6
7
8
9		 select count(distinct dept_num)
from emp_ct


select count(*)
from (
select distinct dept_num
from emp_ct
)