东方陨

QA常用mysql设计-搜索引擎2020年update

================================================================================

update嵌套查询修改不生效，是因为Navicat版本不支持
UPDATE `A` SET prop_count =15 WHERE id in
(select id from `A` where user_id in (select id from `B` where data_id="xx"
and status=2 and prop_expire_time="9999-12-31 23:59:59"
order by update_time DESC LIMIT 1));

背景：SQL监控接入dpc，日期为云自带的函数，但在本地odps调试时候不可以走云函数，需要自己获取当前时间-1，格式为yyyymmdd 如20191213

mysql中的DATE_FORMAT(NOW(),'%Y-%m-%d')函数

前一天日期

DATE_FORMAT(adddate(now(),-1),'%Y%m%d')

1.格式：
DATE_FORMAT(date,format) 函数用于显示日期或时间数据的不同样式。
1.1参数：date 合法的日期；
format 最终输出的日期/时间；
2.参考：
DATE_FORMAT(NOW(),’%Y-%m-%d’) 格式转换

SELECT DATE_FORMAT(NOW(),'%Y-%m-%d') AS '日期'

输出格式为2019-12-12

如果需要20191212格式的日期，则

mysql> select date_format(current_timestamp, '%Y%m%d')

-> ;

+------------------------------------------+

| date_format(current_timestamp, '%Y%m%d') |

+------------------------------------------+

| 20191224 |

+------------------------------------------+

该死的odps不支持mysql函数

select DATEADD(GETDATE(), -1, 'dd') from database.table_name limit 1;

~~select to_char(DATEADD(GETDATE(), -1, 'dd'),'yyyymmdd') from database.table_name limit 1;~~

但是第二个SQL不行，因为我最后要的是20191212这种的而第一个函数转出来是20191212 15:11:20这种的格式导致结果显示为yy1212

原因：''错误导致应为“” 换成下面的就对了

select to_char(dateadd(GETDATE(),-1,'dd'),"yyyymmdd") from database.table_name limit 1;

第二种SQL为

select replace(split_part(dateadd(from_unixtime(unix_timestamp()),-1,'dd')," ",1),"-","") from database.table_name

思考：mysql一个函数搞定，odps需要多个函数，比较慢和烦人~~不能只用平台自带的配置调用配置ds='${bizdate}'

1技术会废掉 2出去面试找不到工作 3不要用傻瓜式配置，离开平台将没得学习和提高，还是要多思考。本地调试写死日期不明智，因为数据表会保留一周的数据，考虑到可用性和易维护性，还是写成函数吧，不能图省事，要看的长远~

===========================================================================

背景：
业务线埋点业务复杂，正则埋点130条+，基于UT平台经常断开链接以及每次回归成本较大，线上无监控，等着BI T+1发现问题的时候就晚了，对大盘数据造成影响。且QA最了解搜索埋点的不同场景的字段，基于odps做了埋点的监控S1测试方式：
1. 搜索埋点曝光case 21 条，点击case 64 条
其中，曝光case 1条为1-多个卡片，点击case1条为多个卡片（114个case）
// 统计mysql中一个字段的value中K出现的次数，如value值为 K3-22，K8-9，K10-2查看详情-二级页，查找K出现的次数，每个K是一个case

SELECT widget_name, sum(LENGTH(widget_name) - LENGTH( REPLACE(widget_name,'K','')))

from table_name where event_id=2101 and pd_emp_id in ("11","12","13","14","15" ,"16") ORDER BY gmt_modified DESC

1.1 曝光case校验字段：soku_test_ab、engine、item_log、aaid、k、track_info、source_from、search_from

具体规则如下，示例：
track_info 下的 soku_test_ab 正则规则 [a-z]{1}
track_info 下的 engine 正则规则 \S*
track_info 下的 item_log 正则规则 \S*
track_info 下的 k 正则规则 \S*
track_info 下的 aaid 正则规则 [0-9a-z]{32}
track_info 下的 source_from 正则规则 (home|discover|vip)
track_info 下的 search_from 正则规则 ^[1-9]\d?|1[01]\d|1[01]\d|^10trackinfo不为空spm正则规则a2h0c.8166622.PhoneXXPictureTab\d∗.channeltab\d∗(;|,)∗scm正则规则20140669.search.rgroupset.filter§∗1.2点击case校验字段：sokutestab、engine、itemlog、aaid、k、trackinfo、spm、scm结果页：trackinfo下的sokutestab正则规则[a−z]1trackinfo下的engine正则规则§∗trackinfo下的itemlog正则规则§∗trackinfo下的sourcefrom正则规则(home|discover|vip)trackinfo下的searchfrom正则规则[1−9]\d?trackinfo不为空spm正则规则a2h0c.8166622.PhoneXXPictureTab\d∗.channeltab\d∗(;|,)∗scm正则规则20140669.search.rgroupset.filter§∗1.2点击case校验字段：sokutestab、engine、itemlog、aaid、k、trackinfo、spm、scm结果页：trackinfo下的sokutestab正则规则[a−z]1trackinfo下的engine正则规则§∗trackinfo下的itemlog正则规则§∗trackinfo下的sourcefrom正则规则(home|discover|vip)trackinfo下的searchfrom正则规则[1−9]\d?|^1[01]\d|10|10
track_info 不为空
spm 正则规则 a2h0c.8166622.PhoneXXProgramSeries_\d*.poster_\d*
scm 正则规则 20140669.search.\S*.\S*

默认页：
track_info 下的 soku_test_ab 正则规则 [a-z]{1}
track_info 下的 aaid 正则规则 [0-9a-zA-Z]{32}
track_info 下的 k 正则规则 \S*
track_info 正则规则 \S*
spm 正则规则 a2h0c.8166619.PhoneXXOperate.clearbutton
scm 正则规则 20140669.search.searcharea.clearbutton

进度：曝光埋点需要增加aaid k，进度：100% 曝光埋点新增soku_test_ab、engine、item_log，完成100%

点击埋点新增soku_test_ab、engine、item_log字段，完成100%
共85条case，128个字段

分工：
曝光埋点
点击埋点
测试手段：
每个迭代版本一灰前覆盖测试所有点击曝光埋点
通过正则平台筛选业务线，勾选case生成测试方案，手动运行，生成测试报告（成功、失败、未运行的case）
测试完毕QA发出测试报告
埋点涉及到的各种平台
埋点日志平台：抓取埋点实时日志
XX：埋点正则case地址
FBI监控平台：关注埋点监控日报
埋点数据监控平台
OneData报警平台：全部点击曝光case已接入，但暂不支持正则报警，待RD完善
埋点接入正则case
埋点监控以及报警
埋点相关数据表调研

埋点所有事件表，曝光是没拆的原始日志，是离线表。实时表在特斯拉，很危险，一天上千亿数据
埋点日志15分钟延迟表
埋点小时表
底层表T+1日志

select alldata.allcnt, faildata.failcnt, round(faildata.failcnt*100.0/alldata.allcnt,4) as failratio

from(

( select count(*) as failcnt,'trackInfo' as question

from (

SELECT a.* ,

get_json_object(a.track_info,'$.show_q') as show_q ,

get_json_object(a.track_info,'$.search_q') as search_q

from

(

select *

FROM xx.xx

WHERE ds=20191216 and site='xx' and (device='android' )

and (original_spm='a2h0c.8166622.home.default' or

original_spm='a2h0c.8166619.xx.default')

and app_version>'8.0.0'

WHERE

b.req_id is null or trim(b.req_id)=''

or b.show_q is null or trim(b.show_q)=''

or b.search_q is null or trim(b.search_q)=''

or b.recext is NULL or trim(b.recext)=''

or b.aaid is NULL or trim(b.aaid)=''

or b.alginfo is null or trim(b.alginfo)=''

)faildata

LEFT JOIN (

select count(*) as allcnt, 'trackInfo' as question

FROM xx.xx

WHERE ds=20191210 and site='xx' and (device='android' )

and (original_spm='a2h0c.8166622.home.default' or

original_spm='a2h0c.8166619.xx.default' )

)alldata

on alldata.question=faildata.question

);

ROUND() 函数

ROUND 函数用于把数值字段舍入为指定的小数位数。

SQL ROUND() 语法

SELECT ROUND(column_name,decimals) FROM table_name

参数	描述
column_name	必需。要舍入的字段。
decimals	必需。规定要返回的小数位数。

计算一个带条件和不带条件的比例，改为用sum(case when 带条件 then 1 else 0),count(1) 然后再把两个字段作除法

就是

SELECT

SUM(

CASE parent_id

WHEN 0 THEN

'00'

WHEN 1 THEN

'11'

ELSE

'OTHERS'

END ) AS parent_id_new

2019-12-18日改版

select spm, count(distinct case when aaid is null then aaid else null end) over () , count(k) ,
group by spm

sum(expo) over (partition by a.dim) expo_all

或者

a b c d

select
*, a/all as rate
from (

   select count(if(error<>'-1',1,null)) as a
   , count(1) as all
   , count(if(error='a',1,null)) as a1
   -- select sum(if(error<>'-1',1,0)),sum(1),
   from (
       select *,

       case when a=null then 'a'
           when b=null then 'b'
           else '-1'
       end as error
   ) a

)

优化前的SQL：

设计上：

mysql也支持聚合的时候加一些条件，不过一般都是数据分析师才会这么搞，或者BI统计的时候用。日常这么写SQL，要被DBA干死的。因为这样很消耗Mysql的cpu，计算也一般都很慢，在线业务跑这种SQL，那接口几秒钟能返回也是够快了，随便几个并发起来了，库都要被拖挂了。

业务上：SQL查询的是A or B or C or D为空的总和计算/total_log，接入监控报警，一旦报警，无法准确定位是哪个字段出错了，可能是A可能是B可能是C，因为计算的是总和出错率。到时候还需要把SQL粘贴到odps，逐一修改判断哪个字段为空

优化后的SQL：

分别计算A为空B为空C为空D为空的出错率，因为最终的监控只能监控一个字段，所以需要sum(*)输出一个值接入监控系统

校验1个场景下的10个字段，统计，占比，如果哪个字段漏掉了，排查的时候，只需要把监控SQL粘贴到odps去掉sum(*)，清晰地看到ABCD各自的失败率占比。哪个字段为空，准确定位

不需要分多个规则配置，不需要分端，减少冗余无效的复制粘贴以及一堆规则的填写，精简化，报错明显

--odps sql

--********************************************************************--

--author:老子

--create time:2019-12-18 19:29:09

--********************************************************************--

-- select COUNT(spm_nullrate) OVER (),COUNT(scm_nullrate) over() from ( --计算的值占总数的比

-- select sum( case when COALESCE(spm,'')<>'' then 1 else 0 end) as spm_null, sum(1) scm_total,

-- select device, sum( case when COALESCE(spm,'')<>'' then 1 else 0 end) as spm_null,sum(1),

-- sum( case when spm is null or trim(spm)='' then 1 else 0 end) /sum(1) as spm_rate,

-- as可以省略不写如下

-- select

-- sum( case when spm is null or trim(spm)='' then 1 else 0 end) as spm_null, sum(1) log_total,

-- round(sum( case when spm is null or trim(spm)='' then 1 else 0 end) /sum(1) ,4) spm_rate,

--以下SQL的as也可以去掉

-- //通用需求柏拉图必须存在的key的校验

-- public static String[] mustExistKeyOuter = {"spm", "scm"};

-- public static String[] mustExistKeyInner = {"soku_test_ab", "engine", "item_log", "aaid", "k", "source_from", "search_from"};

select device,

sum( case when spm is null or trim(spm)='' then 1 else 0 end) as spm_null, sum(1) as log_total,

round(sum( case when spm is null or trim(spm)='' then 1 else 0 end) /sum(1) ,4) as spm_rate,

sum( case when scm is null or trim(scm)='' then 1 else 0 end) as scm_null,

round(sum( case when scm is null or trim(scm)='' then 1 else 0 end) /sum(1),4) as scm_rate,

sum( case when track_info is null or trim(track_info)='' then 1 else 0 end) as track_info_null,

round(sum( case when track_info is null or trim(track_info)='' then 1 else 0 end) /sum(1),4) as track_info_rate,

sum(case when get_json_object(track_info,'$.soku_test_ab') is null or trim(get_json_object(track_info,'$.soku_test_ab'))='' then 1 else 0 end) as soku_test_ab_null,

round(sum( case when get_json_object(track_info,'$.soku_test_ab') is null or trim(get_json_object(track_info,'$.soku_test_ab'))='' then 1 else 0 end) /sum(1),4) as soku_test_ab_rate,

sum( case when get_json_object(track_info,'$.engine') is null or trim(get_json_object(track_info,'$.engine'))='' then 1 else 0 end) as engine_null,

round(sum( case when get_json_object(track_info,'$.engine') is null or trim(get_json_object(track_info,'$.engine'))='' then 1 else 0 end) /sum(1),4) as engine_rate,

sum( case when get_json_object(track_info,'$.item_log') is null or trim(get_json_object(track_info,'$.item_log'))='' then 1 else 0 end) as item_log_null,

round(sum( case when get_json_object(track_info,'$.item_log') is null or trim(get_json_object(track_info,'$.item_log'))=''then 1 else 0 end) /sum(1),4) item_log_rate,

sum( case when get_json_object(track_info,'$.aaid') is null or trim(get_json_object(track_info,'$.aaid'))='' then 1 else 0 end) as aaid_null,

round(sum( case when get_json_object(track_info,'$.aaid') is null or trim(get_json_object(track_info,'$.aaid'))='' then 1 else 0 end) /sum(1),4) as aaid_rate,

sum( case when get_json_object(track_info,'$.k') is null or trim(get_json_object(track_info,'$.k'))='' then 1 else 0 end) as k_null,

round(sum( case when get_json_object(track_info,'$.k') is null or trim(get_json_object(track_info,'$.k'))='' then 1 else 0 end) /sum(1),4) as k_rate,

sum( case when get_json_object(track_info,'$.source_from') is null or trim(get_json_object(track_info,'$.source_from'))='' then 1 else 0 end) as source_from_null,

round(sum( case when get_json_object(track_info,'$.source_from') is null or trim(get_json_object(track_info,'$.source_from'))='' then 1 else 0 end) /sum(1),4) as source_from_rate,

sum( case when get_json_object(track_info,'$.search_from') is null or trim(get_json_object(track_info,'$.search_from'))='' then 1 else 0 end) as search_from_null,

round(sum( case when get_json_object(track_info,'$.search_from') is null or trim(get_json_object(track_info,'$.search_from'))='' then 1 else 0 end) /sum(1),4) as search_from_rate

from database.table_name（库名.表名）

WHERE ds=20191216 and hh=07 and site='xx' and (device='android' or device='iphone' )

and spm like '%a2h0c.8166619.hotkeyword%' and original_spm like '%a2h0c.8166619.xx%'

and original_spm!='a2h0c.8166619.xxhistory.open'

and original_spm!='a2h0c.8166619.xxhistory.close' GROUP BY device;

-- group by spm,scm,track_info,engine);

运行结果：

由于监控系统规则只能有一个字段报警，故此对所有字段的failrate求和，如需排查问题，则去掉

select device,round(spm_rate+scm_rate+track_info_rate+soku_test_ab_rate+engine_rate+item_log_rate+aaid_rate+k_rate+source_from_rate+search_from_rate ,4)

as total_failrate FROM (

最终接入监控的SQL如下

select device,round(spm_rate+scm_rate+track_info_rate+soku_test_ab_rate+engine_rate+item_log_rate+aaid_rate+k_rate+source_from_rate+search_from_rate ,4)

as total_failrate FROM (

select