接前文

http://blog.itpub.net/29254281/viewspace-2157111/

这是上文的优化版本

创建试验数据,5天每天一百万随机数据,总共500w数据

create table nums(id int not null primary key);

delimiter $$
create procedure pFastCreateNums(cnt int)
begin
declare s int default 1;
truncate table nums;
insert into nums select s;
while s*2<=cnt do
insert into nums select id+s from nums;
set s=s*2;
end while;
end $$
delimiter ;

call pFastCreateNums(2000000);

drop table if exists t ;
create table t(
query_time date,
ts float,
key(query_time,ts)
);

insert into t select '2018-07-01',round(100000*rand(),2) from nums where id<=1000000;
insert into t select '2018-07-02',round(100000*rand(),2) from nums where id<=1000000;
insert into t select '2018-07-03',round(100000*rand(),2) from nums where id<=1000000;
insert into t select '2018-07-04',round(100000*rand(),2) from nums where id<=1000000;
insert into t select '2018-07-05',round(100000*rand(),2) from nums where id<=1000000;

首先，修正上文的SQL，增加精度,因为在大数据量下,会有显著的误差。

select query_time,v,ts
from (
select t6.query_time,t6.ts,v,seq,
case when @gid=concat(seq,'#',query_time) then @rn:=@rn+1 when @gid:=concat(seq,'#',query_time) then @rn:=1 end s
from (
select query_time,ts,rn,percent,v,v-percent d,seq from (
select t2.query_time,ts,rn,round(rn/total,10) percent from (
select query_time,ts,
case when @gid=query_time then @rn:=@rn+1 when @gid:=query_time then @rn:=1 end rn
from (
select * from t ,(select @gid:='',@rn:=0) vars order by query_time,ts
) t1
) t2 inner join (
select query_time,count(*) total from t group by query_time
) t3 on(t2.query_time=t3.query_time)
) t4 ,
(select 0.71 v,1 seq union all select 0.81,2 union all select 0.91,3) t5
) t6 where d>=0 order by query_time,v,d
) t7 where s=1 order by query_time,seq ;

在ssd环境下,上文的SQL运行时长和结果如下.

148.813 s

前文这个SQL的计算结果是非常精确的
但是计算时间和采样点数量有巨大关系. 假如原始数据是100w，三个百分位数的采样,则数据扩张到300w;4个百分位数的采样,则数据扩张到400w.这是因为使用笛卡尔积扩张了数据的缘故.

优化版本:

select query_time,d,max(ts) ts from (
select t2.query_time,ts,rn,round(rn/total,10) percent,
case
when 0.71>=round(rn/total,10) then 0.71
when 0.81>=round(rn/total,10) then 0.81
when 0.91>=round(rn/total,10) then 0.91
end d
from (
select query_time,ts,
case when @gid=query_time then @rn:=@rn+1 when @gid:=query_time then @rn:=1 end rn
from (
select * from t ,(select @gid:='',@rn:=0) vars order by query_time,ts
) t1
) t2 inner join (
select query_time,count(*) total from t group by query_time
) t3 on(t2.query_time=t3.query_time)
) t6
where d is not null
group by query_time,d

结果:

用时:
33.922 秒

这个版本的优点是增加百分位数采样点,不会增加开销.
缺点是在非常小的样本环境中,会有误差.
不过后续也有手段消除误差.

MySQL百分位数计算(优化版)