PostgreSQL 与 12306 抢火车票的思考

标签

PostgreSQL,门禁广告,数组,范围类型,抢购,排他约束,大盘分析,广告查询,火车票

背景

抢火车票是很有意思的一个课题,对IT人的智商以及IT系统的健壮性,尤其是数据库的功能和性能都是一种挑战.

我记得很多年前写过一篇PostgreSQL varbit和火车票相关的文章,翻出来看看,发现又和BIT有关.

回顾一下我之前写的

《基于 阿里云 RDS PostgreSQL 打造实时用户画像推荐系统》

《门禁广告销售系统需求剖析 与 PostgreSQL数据库实现》

PostgreSQL的bit功能还是很强大的,阿里云RDS PostgreSQL的bitpack也是用户实际应用中的需求提炼的新功能,大伙一起来给阿里云提需求.

打造属于国人的PostgreSQL.

正文

在PostgreSQL 中可以使用varbit存储比特位,下面模拟一个简单的应用场景.

马上春节了,火车票又到了销售旺季,一票难求依旧.

下面就以火车票销售为例来介绍一下PostgreSQL varbit类型的用法.

测试环境 :

PostgreSQL 9.2.1

测试表 :

列车信息表 :

create table train

(id int primary key,--主键

go_date date,-- 发车日期

train_num name,-- 车次

station text[] -- 途径站点

);

车厢或bucket信息表 :

create table train_bucket

(id int primary key,--主键

tid int references train (id),-- 关联列车ID

bno int,-- 车厢或bucket号

sit_level text,-- 席别

sit_cnt int,-- 该车厢或bucket的座位总数

sit_remain int,-- 剩余座位

sit_bit varbit -- 座位BIT位,已售座位用1表示,未售座位用0表示

);

位置信息表 :

create table train_sit

(id serial8 primary key,-- 主键

tid int references train (id),--关联列车ID

tbid int references train_bucket(id),--关联bucket表ID

sit_no int,-- 座位号,来自train_bucket.sit_bit的位置信息.

station_bit varbit -- 途径站点组成的BIT位信息,已售站点用1表示,未售站点用0表示.

);

创建索引 :

create index idx_train_bucket_sit_remain on train_bucket(sit_remain) where sit_remain>0;

create index idx_train_sit_station_bit on train_sit (station_bit) where station_bit<>repeat('1',13)::varbit;

插入测试数据,1趟火车,途径14个站点.

insert into train values (1,'2013-01-20','D645',array['上海南','嘉兴','杭州南','诸暨','义乌','金华','衢州','上饶','鹰潭','新余','宜春','萍乡','株洲','长沙']);

插入测试数据,共计200W个车厢或bucket,每个车厢98个位置.

insert into train_bucket values (generate_series(1,1000000),1,generate_series(1,'一等座',98,repeat('0',98)::varbit);

insert into train_bucket values (generate_series(1000001,2000000),generate_series(1000001,'二等座',98)::varbit);

创建取数组中元素位置的函数 :

create or replace function array_pos (a anyarray,b anyelement) returns int as $$

declare

i int;

begin

for i in 1..array_length(a,1) loop

if b=a[i] then

return i;

end if;

i := i+1;

end loop;

return null;

end;

$$ language plpgsql;

创建购票函数 :

create or replace function buy

(

inout i_train_num name,

inout i_fstation text,

inout i_tstation text,

inout i_go_date date,

out o_slevel text,

out o_bucket_no int,

out o_sit_no int,

out o_order_status boolean

)

returns record as $$

declare

curs1 refcursor;

curs2 refcursor;

v_row int;

v_station text[];

v_train_id int;

v_train_bucket_id int;

v_train_sit_id int;

v_from_station_idx int;

v_to_station_idx int;

v_station_len int;

begin

set enable_seqscan=off;

v_row := 0;

o_order_status := false;

select array_length(station,1),station,id,array_pos(station,i_fstation),i_tstation)

into v_station_len,v_station,v_train_id,v_from_station_idx,v_to_station_idx

from train where train_num=i_train_num and go_date = i_go_date;

if ( found and array_pos(v_station,i_fstation) is not null

and array_pos(v_station,i_tstation) is not null

and array_pos(v_station,i_fstation) < array_pos(v_station,i_tstation)

) then

else

o_order_status := false;

return;

end if;

open curs2 for select tid,tbid,sit_no from train_sit

where (station_bit & bitsetvarbit(repeat('0',v_station_len-1)::varbit,v_from_station_idx-1,v_to_station_idx-v_from_station_idx,1)) = repeat('0',v_station_len-1)::varbit

and station_bit <> repeat('1',v_station_len-1)::varbit

-- and ctid not in (select locked_row from pgrowlocks('train_sit')) -- 耗时约300毫秒,用它来解决热点锁等待不划算.

limit 1

for update nowait; -- 也可不加nowait,加了的话如果获取锁失败将返回55P03异常,需要程序重新提交

loop

fetch curs2 into v_train_id,v_train_bucket_id,o_sit_no;

if found then

update train_sit set station_bit=bitsetvarbit(station_bit,1)

where current of curs2;

GET DIAGNOSTICS v_row = ROW_COUNT;

if (v_row = 1) then

select sit_level,bno into o_slevel,o_bucket_no from train_bucket where id=v_train_bucket_id;

close curs2;

o_order_status := true;

return;

end if;

else

close curs2;

exit;

end if;

end loop;

v_row := 0;

open curs1 for select id,tid,strpos(sit_bit::text,'0'),sit_level,bno from train_bucket

where sit_remain>0

-- and ctid not in (select locked_row from pgrowlocks('train_bucket')) -- 耗时约300毫秒,需要程序重新提交.

loop

fetch curs1 into v_train_bucket_id,o_sit_no,o_slevel,o_bucket_no;

if found then

update train_bucket set sit_bit = set_bit(sit_bit,'0')-1,sit_remain = sit_remain-1

where current of curs1;

GET DIAGNOSTICS v_row = ROW_COUNT;

if (v_row = 1) then

close curs1;

exit;

end if;

else

close curs1;

exit;

end if;

end loop;

if v_row = 1 then

insert into train_sit(tid,sit_no,station_bit)

values (

v_train_id,

v_train_bucket_id,

o_sit_no,

bitsetvarbit(repeat('0',1)

);

o_order_status := true;

return;

else

o_order_status := false;

return;

end if;

exception

when others then

o_order_status := false;

return;

end;

$$ language plpgsql;

测试 :

digoal=# select * from buy('D645','2013-01-20');

i_train_num | i_fstation | i_tstation | i_go_date | o_slevel | o_bucket_no | o_sit_no | o_order_status

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

D645 | 杭州南 | 宜春 | 2013-01-20 | 一等座 | 35356 | 9 | t

(1 row)

压力测试

vi test.sql

select * from buy('D645','上海南','长沙','2013-01-20');

不加nowait测试结果 :

ocz@db-172-16-3-150-> pgbench -M prepared -f ./test.sql -n -r -c 16 -j 8 -T 1200 -U postgres digoal

transaction type: Custom query

scaling factor: 1

query mode: prepared

number of clients: 16

number of threads: 8

duration: 1200 s

number of transactions actually processed: 2197407

tps = 1831.143708 (including connections establishing)

tps = 1831.169308 (excluding connections establishing)

statement latencies in milliseconds:

8.734424 select * from buy('D645','2013-01-20');

加nowait测试结果 :

ocz@db-172-16-3-150-> pgbench -M prepared -f ./test.sql -n -r -c 16 -j 16 -T 12 -U postgres digoal

transaction type: Custom query

scaling factor: 1

query mode: prepared

number of clients: 16

number of threads: 16

duration: 12 s

number of transactions actually processed: 93632

tps = 7800.056248 (including connections establishing)

tps = 7818.803904 (excluding connections establishing)

statement latencies in milliseconds:

2.042862 select * from buy('D645','2013-01-20');

小结

1. 需要解决更新热点,降低等待,提高并行处理几率.

本例的热点在 : update train_bucket set sit_bit = set_bit(sit_bit,sit_remain = sit_remain-1

where current of curs1;

以及

update train_sit set station_bit=bitsetvarbit(station_bit,1)

where current of curs2;

对应的游标 :

open curs2 for select tid,用它来解决热点锁等待不划算.

limit 1

for update;

以及

open curs1 for select id,用它来解决热点锁等待不划算.

limit 1

for update;

解决的关键在这里.

如果不能解决热点的问题,那就提高处理速度,精简字段数量和长度,精简索引. 提高更新速度.

2. 减少数据扫描的量.

partial index,避免满座车厢的扫描,以及全程占位位子的扫描.

3. 先查bucket 是否空闲,再查sit是否空闲.

4. 还需要考虑优先级的问题 :

例如有111000和111100两个位子,如果请求要最后两个站的票,应该优先匹配111100,这样更不容易浪费.如下 :

111000 | 000011 = 111011

111100 | 000011 = 222221

参考

1. setbitvarbit

http://blog.163.com/digoal@126/blog/static/163877040201302192427651/