大数据开发之Hive案例篇10-大表笛卡尔积优化

一. 问题描述

需求描述:
表概述:

dt                  时间分区
data_source  数据来源类别
start_date      时间
data_count    当前时间的数量

需要实现的需求

求每个data_source 下start_date 当前累积的data_count

SQL代码:

select dt,
          data_souce，
          start_date,
          data_count,
          sum(data_count) over(partition by data_source order by start_date) as data_cum_count
  from table_name

运行日志:
从日志可以看到，数据倾斜了，redcue一直卡在99%不动，过一段时间就被断开了。

2023-05-30 12:05:40,318 Stage-1 map = 100%,  reduce = 75%, Cumulative CPU 2693.11 sec
2023-05-30 12:05:41,349 Stage-1 map = 100%,  reduce = 76%, Cumulative CPU 2716.81 sec
2023-05-30 12:05:43,411 Stage-1 map = 100%,  reduce = 77%, Cumulative CPU 2774.08 sec
2023-05-30 12:05:45,478 Stage-1 map = 100%,  reduce = 78%, Cumulative CPU 2795.55 sec
2023-05-30 12:05:46,509 Stage-1 map = 100%,  reduce = 79%, Cumulative CPU 2851.83 sec
2023-05-30 12:05:47,547 Stage-1 map = 100%,  reduce = 80%, Cumulative CPU 2880.86 sec
2023-05-30 12:05:51,678 Stage-1 map = 100%,  reduce = 81%, Cumulative CPU 2935.67 sec
2023-05-30 12:05:52,710 Stage-1 map = 100%,  reduce = 84%, Cumulative CPU 3031.14 sec
2023-05-30 12:05:54,772 Stage-1 map = 100%,  reduce = 85%, Cumulative CPU 3086.83 sec
2023-05-30 12:05:56,833 Stage-1 map = 100%,  reduce = 86%, Cumulative CPU 3101.59 sec
2023-05-30 12:06:00,956 Stage-1 map = 100%,  reduce = 87%, Cumulative CPU 3213.04 sec
2023-05-30 12:06:07,173 Stage-1 map = 100%,  reduce = 89%, Cumulative CPU 3332.53 sec
2023-05-30 12:06:08,209 Stage-1 map = 100%,  reduce = 90%, Cumulative CPU 3348.58 sec
2023-05-30 12:06:09,241 Stage-1 map = 100%,  reduce = 93%, Cumulative CPU 3399.05 sec
2023-05-30 12:06:10,272 Stage-1 map = 100%,  reduce = 94%, Cumulative CPU 3456.29 sec
2023-05-30 12:06:12,334 Stage-1 map = 100%,  reduce = 95%, Cumulative CPU 3503.32 sec
2023-05-30 12:06:14,406 Stage-1 map = 100%,  reduce = 96%, Cumulative CPU 3550.1 sec
2023-05-30 12:06:15,433 Stage-1 map = 100%,  reduce = 97%, Cumulative CPU 3576.75 sec
2023-05-30 12:06:19,561 Stage-1 map = 100%,  reduce = 98%, Cumulative CPU 3674.46 sec
2023-05-30 12:06:29,878 Stage-1 map = 100%,  reduce = 99%, Cumulative CPU 3860.69 sec
2023-05-30 12:07:30,726 Stage-1 map = 100%,  reduce = 99%, Cumulative CPU 4349.64 sec
2023-05-30 12:08:31,498 Stage-1 map = 100%,  reduce = 99%, Cumulative CPU 4622.97 sec
2023-05-30 12:09:32,161 Stage-1 map = 100%,  reduce = 99%, Cumulative CPU 4857.09 sec
2023-05-30 12:10:32,788 Stage-1 map = 100%,  reduce = 99%, Cumulative CPU 5046.44 sec
2023-05-30 12:11:33,443 Stage-1 map = 100%,  reduce = 99%, Cumulative CPU 5196.55 sec
2023-05-30 12:12:34,216 Stage-1 map = 100%,  reduce = 99%, Cumulative CPU 5325.04 sec
2023-05-30 12:13:34,952 Stage-1 map = 100%,  reduce = 99%, Cumulative CPU 5454.34 sec
2023-05-30 12:14:35,677 Stage-1 map = 100%,  reduce = 99%, Cumulative CPU 5584.3 sec
2023-05-30 12:15:36,383 Stage-1 map = 100%,  reduce = 99%, Cumulative CPU 5722.47 sec
2023-05-30 12:16:37,011 Stage-1 map = 100%,  reduce = 99%, Cumulative CPU 5796.86 sec
2023-05-30 12:17:37,641 Stage-1 map = 100%,  reduce = 99%, Cumulative CPU 5864.27 sec
2023-05-30 12:18:38,284 Stage-1 map = 100%,  reduce = 99%, Cumulative CPU 5929.96 sec
2023-05-30 12:19:38,916 Stage-1 map = 100%,  reduce = 99%, Cumulative CPU 5999.27 sec
2023-05-30 12:20:39,508 Stage-1 map = 100%,  reduce = 99%, Cumulative CPU 6066.16 sec
2023-05-30 12:21:40,153 Stage-1 map = 100%,  reduce = 99%, Cumulative CPU 6133.75 sec
2023-05-30 12:22:40,776 Stage-1 map = 100%,  reduce = 99%, Cumulative CPU 6202.56 sec
2023-05-30 12:23:41,326 Stage-1 map = 100%,  reduce = 99%, Cumulative CPU 6271.21 sec
2023-05-30 12:24:41,947 Stage-1 map = 100%,  reduce = 99%, Cumulative CPU 6338.7 sec
2023-05-30 12:25:42,696 Stage-1 map = 100%,  reduce = 99%, Cumulative CPU 6406.98 sec
2023-05-30 12:26:43,307 Stage-1 map = 100%,  reduce = 99%, Cumulative CPU 6474.84 sec
2023-05-30 12:27:43,873 Stage-1 map = 100%,  reduce = 99%, Cumulative CPU 6543.65 sec
2023-05-30 12:28:44,449 Stage-1 map = 100%,  reduce = 99%, Cumulative CPU 6610.24 sec
2023-05-30 12:29:45,003 Stage-1 map = 100%,  reduce = 99%, Cumulative CPU 6679.73 sec
2023-05-30 12:30:45,623 Stage-1 map = 100%,  reduce = 99%, Cumulative CPU 6746.93 sec
2023-05-30 12:31:46,118 Stage-1 map = 100%,  reduce = 99%, Cumulative CPU 6822.78 sec
2023-05-30 12:32:46,658 Stage-1 map = 100%,  reduce = 99%, Cumulative CPU 6890.72 sec
2023-05-30 12:33:47,212 Stage-1 map = 100%,  reduce = 99%, Cumulative CPU 6959.17 sec

web页面日志:
从web页面可以看到，reduce被kill的原因是Container被ApplicationMaster给kill掉了

Speculation: attempt_1680276634497_67940_r_000001_1 succeeded first! [2023-05-30 10:56:47.400]Container killed by the ApplicationMaster. [2023-05-30 10:56:47.422]Container killed on request. Exit code is 143 [2023-05-30 10:56:47.442]Container exited with a non-zero exit code 143.

过一段时间整个Job都被kill掉了

二.解决方案

2.1 数据倾斜

因为reduce卡在了99%，所以首先想到的是数据倾斜，后面了解了下，data_source字段确实存在数据倾斜

调整参数:
然后没什么用

-- 加大reduce个数
set mapred.reduce.tasks = 100;
set hive.auto.convert.join = true;
-- 超过一万行就认为是倾斜
set hive.skewjoin.key=100000;

set mapreduce.map.memory.mb=16384;
set mapreduce.reduce.memory.mb=2048;
set yarn.nodemanager.vmem-pmem-ratio=4.1;
set mapreduce.reduce.memory.mb=5120;
set mapred.map.child.java.opts=-Xmx13106M;
set mapreduce.map.java.opts=-Xmx13106M;
set mapreduce.reduce.java.opts=-Xmx13106M;
set mapreduce.task.io.sort.mb=512;
set mapreduce.job.reduce.slowstart.completedmaps=0.8;

调整代码:
将数据倾斜严重的数据，单独拿出来执行
然后也没什么作用

select dt,
          data_souce，
          start_date,
          data_count,
          sum(data_count) over(partition by data_source order by start_date) as data_cum_count
  from table_name
where data_source in (数据倾斜)；

select dt,
          data_souce，
          start_date,
          data_count,
          sum(data_count) over(partition by data_source order by start_date) as data_cum_count
  from table_name
where data_source not in (数据倾斜)；

2.2 SQL改写1:由分析函数改为常规写法

不确定是不是Hive分析函数的问题，然后我将原始的SQL改为了表连接和临时表的方法来解决

代码:

select t1.dt,t1.data_source,t1.start_date,
          sum(data_count)  data_cum_count
  from table_name t1
 left join table_name t2
  on t1.data_souce = t2.data_souce
where t1.start_date >= t2.start_date
group by t1.dt,t1.data_source,t1.start_date;

运行结果:
运行结果中，某一个job也是卡在reduce 99%，但是卡了20分钟左右，就执行成功了
最终SQL在30分钟左右执行完成

同样的逻辑，表连接的方式居然就可以了，而分析函数却不行，估计一个是写内存，一个是写磁盘把。

然而:
然后这个是测试表，只有一个月的数据，补历史数据要补几年的，那么这个SQL肯定只会更慢。

2.3 分析数据分布

最大的一个data_source居然有9w多个，产生的笛卡尔积得有81亿之多，虽然集群有20个节点，资源还不错，执行也要半个小时以上。

不敢想象如果是一年甚至数年的，那这个笛卡尔积只会更大。

所以只能改SQL了

2.4 SQL改写2:重写

我们需要求每一个start_date的累积数量，那么此时我们可以先求每天的，然后求每天累积的，再求当天每一个start_date累积的，加上前一日的累积的，就是最终我们需要的数据。

SQL代码:

with tmp1 as (
select t1.data_source,t1.dt,sum(t1.data_count) as sum_v_dt
    from table_name t1
  group by t1.data_source,t1.dt
 ),
tmp2 as (
select data_source,
       dt,
       sum(sum_v_dt) over( partition by data_source order by dt) as sum_v_cum_dt
  from tmp1
)
select t2.data_source,
       t2.dt,
       t2.start_date,
       nvl(sum(t2.data_count) over(partition by t2.data_source,t2.dt order by t2.start_date),0) + nvl(tmp2.sum_v_cum_dt,0) as sum_v_cum_dt_sdate
  from table_name t2
  join tmp2
 on t2.data_source = tmp2.data_source
and t2.dt = tmp2.dt +1;

运行记录:
最终的运行时间在5分钟左右
就算数据量提升数倍，因为 join的条件由一个 data_source 变为了两个 data_souce 、dt，大大减少了笛卡尔积的数据量，整个代码的计算量也减少了许多。

参考:

1. https://zhuanlan.zhihu.com/p/398374859
2. https://blog.csdn.net/wisgood/article/details/77063606
3. https://www.jianshu.com/p/fe0c5c7f62ed
4. https://www.jianshu.com/p/9fb56b668ea0
5. https://www.jianshu.com/p/d13f2c0db335