目录
一、背景
二、分析流程
三、验数方法
3.1 数据量级比对
3.2 一致性比对
3.2.1 勾稽验证+md5方法
3.2.2 暴力比对法
3.3 差异数据发现
四、总结
做数据,经常遇到数据验证,很烦很枯燥,即耗时又耗人,但又必须去做。如何去做数据验证,并标准化整个流程,让验数变得轻松。
……
相同表结构数据验证:比如修改表逻辑
相似表结构数据验证:比如修改表字段。
新表数据校验:比如新开发了表,选择一个比对表参考等等
数据验证三步走:
select left_table.pv-right_table.pv as pv_diff,
left_table.user_id_uv - right_table.user_id_uv as user_id_uv_diff,
left_table.order_id_uv - right_table.order_id_uv as order_id_uv_diff,
left_table.city_id_uv - right_table.city_id_uv as city_id_uv_diff
from (
select count(1) as pv,
count(distinct user_id) as user_id_uv,
count(distinct order_id) as order_id_uv,
count(distinct city_id) as city_id_uv
from mart_online.fact_user_order_day
where dt=20190413
)left_table
left outer join (
select count(1) as pv,
count(distinct user_id) as user_id_uv,
count(distinct order_id) as order_id_uv,
count(distinct city_id) as city_id_uv
from mart_test.fact_user_order_day
where dt=20190413
)right_table
on 1=1
左表pv减去右表pv值为:[0],核心字段uv差为:[0] 即两个表数据条数相同
+-------+----------------+------------------+---------------+
|pv_diff|user_id_uv_diff |order_id_uv_diff |city_id_uv_diff|
+-------+----------------+------------------+---------------+
| 0| 0| 0| 0|
+-------+----------------+------------------+---------------+
勾稽是一个小姑娘起的名字,在这里就是看一下左表不为NULL的left_table_num,右表不为NULL的right_table_num,两个表都有的 left_right_equal_num,如果这三个数相等就说明数据是一致的。反之数据肯定不一致,同时可以计算出不一致的条数。
md5:就是计算一行数据的md5值,把它当成key去做比对。尤其是在百亿数据规模的情况下,这种方法也使用。
************ 数据量一致性验证SQL ************* 注意:这里采用 full join
select sum(case when left_table.record_key is not null or left_table.record_key !='' then 1 else 0 end) as left_table_num,
sum(case when right_table.record_key is not null or right_table.record_key !='' then 1 else 0 end) as right_table_num,
sum(case when left_table.record_key = right_table.record_key then 1 else 0 end) as left_right_equal_num
from (
select md5(
concat(
if(user_id is null, '-', cast(user_id as string)),
if(user_name is null, '-', cast(user_name as string)),
if(order_id is null, '-', cast(order_id as string)),
if(city_id is null, '-', cast(city_id as string)),
if(city_name is null, '-', cast(city_name as string)),
if(字段n…… is null, '-', cast(字段n…… as string)),
if(dt is null, '-', cast(dt as string))
)
) as record_key
from mart_online.fact_user_order_day
where dt=20190413
)left_table
full outer join (
select md5(
concat(
if(user_id is null, '-', cast(user_id as string)),
if(user_name is null, '-', cast(user_name as string)),
if(order_id is null, '-', cast(order_id as string)),
if(city_id is null, '-', cast(city_id as string)),
if(city_name is null, '-', cast(city_name as string)),
if(字段n…… is null, '-', cast(字段n…… as string)),
if(dt is null, '-', cast(dt as string))
)
) as record_key
from mart_test.fact_user_order_day
where dt=20190413
)right_table
on left_table.record_key=right_table.record_key
************ 数据量一致性验证报表 *************
[left_table_num]左表中的数据条数,[right_table_num]右表中的条数,[left_right_equal_num]两个表中相等的数据条数。
左表中有[5660]条数据和右表不一致!
+--------------+---------------+--------------------+
|left_table_num|right_table_num|left_right_equal_num|
+--------------+---------------+--------------------+
| 16358699| 16358699| 16353039|
+--------------+---------------+--------------------+
适合具有唯一ID的表,返回空说明验证准确。
select online.*,
test.* from(
select id,
user_id,
user_name,
order_id,
city_id,
city_name
from mart_online.fact_user_order_day
where dt='20190413'
)online
left outer join (
select id,
user_id,
user_name,
order_id,
city_id,
city_name
from mart_test.fact_user_order_day
where dt='20190413'
) test
on test.id=online.id
where test.user_id!=online.user_id
or test.user_name!=online.user_name
or test.order_id!=online.order_id
or test.city_id!= online.city_id
or test.city_name!= online.city_name
发现差异数据的方法很多,这里只讲一个通用的方法:逐条比对法(假定两个表有唯一的ID,如果没有唯一ID,其实md5不一样的数据就不一致),这种方法适合小规模数据,当然我们真是实现的时候是结合一致性验证的情况,直接就能找到差异的数据并打印出来。
select left_table.*,
right_table.*
from (
select *
from mart_online.fact_user_order_day
where dt=20190413
)left_table
full outer join (
select *
from mart_test.fact_user_order_day
where dt=20190413
)right_table
on left_table.id = right_table.id
and left_table.dt = right_table.dt
where COALESCE(left_table.user_id, 0) <> COALESCE(right_table.user_id, 0)
or COALESCE(left_table.user_name, 0) <> COALESCE(right_table.user_name, 0)
or COALESCE(left_table.order_id, 0) <> COALESCE(right_table.order_id, 0)
or COALESCE(left_table.city_id, 0) <> COALESCE(right_table.city_id, 0)
or COALESCE(left_table.city_name, 0) <> COALESCE(right_table.city_name, 0)
or COALESCE(left_table.字段n……, 0) <> COALESCE(right_table.字段n……, 0)
不一致的条数:[5660],case如下表所示:
+-------+----------------+------------------+---------------+---------------+
|id |left_user_id |left_字段n…… |right_user_id |right_字段n…… |
+-------+----------------+------------------+---------------+---------------+
| 0| 1| 哇哈哈| 1| 养乐多|
+-------+----------------+------------------+---------------+---------------+
如上验数SQL,可以通过代码封装,自动生成,就可以做成自动化数据验证的小工具了。真实情况比较复杂,要考虑字段的识别,where条件,两个表是否有唯一ID,没有唯一ID如何处理等等。