
Postgres2015全国用户大会将于11月20至21日在北京丽亭华苑酒店召开。本次大会嘉宾阵容强大,国内顶级PostgreSQL数据库专家将悉数到场,并特邀欧洲、俄罗斯、日本、美国等国家和地区的数据库方面专家助阵:
- Postgres-XC项目的发起人铃木市一(SUZUKI Koichi)
- Postgres-XL的项目发起人Mason Sharp
- pgpool的作者石井达夫(Tatsuo Ishii)
- PG-Strom的作者海外浩平(Kaigai Kohei)
- Greenplum研发总监姚延栋
- 周正中(德哥), PostgreSQL中国用户会创始人之一
- 汪洋,平安科技数据库技术部经理
- ……
|
![]() |
秒杀场景的典型瓶颈在于对同一条记录的多次更新请求,然后只有一个或者少量请求是成功的,其他请求是以失败或更新不到告终。
例如,Iphone的1元秒杀,如果我只放出1台Iphone,我们把它看成一条记录,秒杀开始后,谁先抢到(更新这条记录的锁),谁就算秒杀成功。
例如:
使用一个标记位来表示这条记录是否已经被更新,或者记录更新的次数(几台Iphone)。
update tbl set xxx=xxx,upd_cnt=upd_cnt+1 where id=pk and upd_cnt+1<=5; -- 假设可以秒杀5台
这种方法的弊端:
获得锁的用户在处理这条记录时,可能成功,也可能失败,或者可能需要很长时间,(例如数据库响应慢)在它结束事务前,其他会话只能等着。
等待是非常不科学的,因为对于没有获得锁的用户,等待是在浪费时间。
所以一般的优化处理方法是先使用for update nowait的方式来避免等待,即如果无法即可获得锁,那么就不等待。
例如:
begin;select 1 from tbl where id=pk for update nowait; -- 如果用户无法即刻获得锁,则返回错误。从而这个事务回滚。update tbl set xxx=xxx,upd_cnt=upd_cnt+1 where id=pk and upd_cnt+1<=5;end;
这种方法可以减少用户的等待时间,因为无法即刻获得锁后就直接返回了。
但是这种方法也存在一定的弊端,对于一个商品,如果可以秒杀多台的话,我们用1条记录来存储多台,降低了秒杀的并发性。
因为我们用的是行锁。
解决这个问题办法很多,最终就是要提高并发性,例如:
1. 分段秒杀,把商品数量打散,拆成多个段,从而提高并发处理能力。
总体来说,
优化的思路是减少锁等待时间,避免串行,尽量并行。
优化到这里就结束了吗?显然没有,以上方法任意数据库都可以做到,如果就这样结束怎么体现PostgreSQL的特性呢?
PostgreSQL还提供了一个锁类型,advisory锁,这种锁比行锁更加轻量,支持会话级别和事务级别。(但是需要注意ID是全局的,否则会相互干扰,也就是说,所有参与秒杀或者需要用到advisory lock的ID需要在单个库内保持全局唯一)
例子:
update tbl set xxx=xxx,upd_cnt=upd_cnt+1 where id=pk and upd_cnt+1<=5 and pg_try_advisory_xact_lock(:id);
最后必须要对比一下for update nowait和advisory lock的性能。
下面是在一台本地虚拟机上的测试。
新建一张秒杀表
postgres=# \d t1Table "public.t1"Column | Type | Modifiers--------+---------+-----------id | integer | not nullinfo | text |Indexes:"t1_pkey" PRIMARY KEY, btree (id)
只有一条记录,不断的被更新
postgres=# select * from t1;id | info----+-------------------------------1 | 2015-09-14 09:47:04.703904+08(1 row)
压测for update nowait的方式:
CREATE OR REPLACE FUNCTION public.f1(i_id integer)RETURNS voidLANGUAGE plpgsqlAS $function$declarebeginperform 1 from t1 where id=i_id for update nowait;update t1 set info=now()::text where id=i_id;exception when others thenreturn;end;$function$;
postgres@digoal-> cat test1.sql\setrandom id 1 1select f1(:id);
压测advisory lock的方式:
postgres@digoal-> cat test.sql\setrandom id 1 1update t1 set info=now()::text where id=:id and pg_try_advisory_xact_lock(:id);
清除压测统计数据:
postgres=# select pg_stat_reset();pg_stat_reset---------------(1 row)postgres=# select * from pg_stat_all_tables where relname='t1';-[ RECORD 1 ]-------+-------relid | 184731schemaname | publicrelname | t1seq_scan | 0seq_tup_read | 0idx_scan | 0idx_tup_fetch | 0n_tup_ins | 0n_tup_upd | 0n_tup_del | 0n_tup_hot_upd | 0n_live_tup | 0n_dead_tup | 0n_mod_since_analyze | 0last_vacuum |last_autovacuum |last_analyze |last_autoanalyze |vacuum_count | 0autovacuum_count | 0analyze_count | 0autoanalyze_count | 0
压测结果:
postgres@digoal-> pgbench -M prepared -n -r -P 1 -f ./test1.sql -c 20 -j 20 -T 60......transaction type: Custom queryscaling factor: 1query mode: preparednumber of clients: 20number of threads: 20duration: 60 snumber of transactions actually processed: 792029latency average: 1.505 mslatency stddev: 4.275 mstps = 13196.542846 (including connections establishing)tps = 13257.270709 (excluding connections establishing)statement latencies in milliseconds:0.002625 \setrandom id 1 11.502420 select f1(:id);
postgres=# select * from pg_stat_all_tables where relname='t1';-[ RECORD 1 ]-------+-------relid | 184731schemaname | publicrelname | t1seq_scan | 0seq_tup_read | 0idx_scan | 896963 // 大多数是无用功idx_tup_fetch | 896963 // 大多数是无用功n_tup_ins | 0n_tup_upd | 41775n_tup_del | 0n_tup_hot_upd | 41400n_live_tup | 0n_dead_tup | 928n_mod_since_analyze | 41774last_vacuum |last_autovacuum |last_analyze |last_autoanalyze |vacuum_count | 0autovacuum_count | 0analyze_count | 0autoanalyze_count | 0
postgres@digoal-> pgbench -M prepared -n -r -P 1 -f ./test.sql -c 20 -j 20 -T 60......transaction type: Custom queryscaling factor: 1query mode: preparednumber of clients: 20number of threads: 20duration: 60 snumber of transactions actually processed: 1392372latency average: 0.851 mslatency stddev: 2.475 mstps = 23194.831054 (including connections establishing)tps = 23400.411501 (excluding connections establishing)statement latencies in milliseconds:0.002594 \setrandom id 1 10.848536 update t1 set info=now()::text where id=:id and pg_try_advisory_xact_lock(:id);
postgres=# select * from pg_stat_all_tables where relname='t1';-[ RECORD 1 ]-------+--------relid | 184731schemaname | publicrelname | t1seq_scan | 0seq_tup_read | 0idx_scan | 1368933 // 大多数是无用功idx_tup_fetch | 1368933 // 大多数是无用功n_tup_ins | 0n_tup_upd | 54957n_tup_del | 0n_tup_hot_upd | 54489n_live_tup | 0n_dead_tup | 1048n_mod_since_analyze | 54957last_vacuum |last_autovacuum |last_analyze |last_autoanalyze |vacuum_count | 0autovacuum_count | 0analyze_count | 0autoanalyze_count | 0
我们注意到,不管用哪种方法,都会浪费掉很多次的无用功扫描。
为了解决无用扫描的问题,可以使用以下函数。(当然,还有更好的方法是对用户透明。)
CREATE OR REPLACE FUNCTION public.f(i_id integer)RETURNS voidLANGUAGE plpgsqlAS $function$declarea_lock boolean := false;beginselect pg_try_advisory_xact_lock(i_id) into a_lock;if a_lock thenupdate t1 set info=now()::text where id=i_id;end if;exception when others thenreturn;end;$function$;
transaction type: Custom queryscaling factor: 1query mode: preparednumber of clients: 20number of threads: 20duration: 60 snumber of transactions actually processed: 1217195latency average: 0.973 mslatency stddev: 3.563 mstps = 20283.314001 (including connections establishing)tps = 20490.143363 (excluding connections establishing)statement latencies in milliseconds:0.002703 \setrandom id 1 10.970209 select f(:id);
postgres=# select * from pg_stat_all_tables where relname='t1';-[ RECORD 1 ]-------+-------relid | 184731schemaname | publicrelname | t1seq_scan | 0seq_tup_read | 0idx_scan | 75927idx_tup_fetch | 75927n_tup_ins | 0n_tup_upd | 75927n_tup_del | 0n_tup_hot_upd | 75902n_live_tup | 0n_dead_tup | 962n_mod_since_analyze | 75927last_vacuum |last_autovacuum |last_analyze |last_autoanalyze |vacuum_count | 0autovacuum_count | 0analyze_count | 0autoanalyze_count | 0
除了吞吐率的提升,我们其实还看到真实的处理数(更新次数)也有提升,所以不仅仅是降低了等待延迟,实际上也提升了处理能力。
最后提供一个物理机上的数据参考,使用128个并发连接,同时对一条记录进行更新:
不做任何优化的并发处理能力:
transaction type: Custom queryscaling factor: 1query mode: preparednumber of clients: 128number of threads: 128duration: 100 snumber of transactions actually processed: 285673latency average: 44.806 mslatency stddev: 45.751 mstps = 2855.547375 (including connections establishing)tps = 2855.856976 (excluding connections establishing)statement latencies in milliseconds:0.002509 \setrandom id 1 144.803299 update t1 set info=now()::text where id=:id;
使用for update nowait的并发处理能力:
transaction type: Custom queryscaling factor: 1query mode: preparednumber of clients: 128number of threads: 128duration: 100 snumber of transactions actually processed: 6663253latency average: 1.919 mslatency stddev: 2.804 mstps = 66623.169445 (including connections establishing)tps = 66630.307999 (excluding connections establishing)statement latencies in milliseconds:0.001934 \setrandom id 1 11.917297 select f1(:id);
使用advisory lock后的并发处理能力:
transaction type: Custom queryscaling factor: 1query mode: preparednumber of clients: 128number of threads: 128duration: 100 snumber of transactions actually processed: 19154754latency average: 0.667 mslatency stddev: 1.054 mstps = 191520.550924 (including connections establishing)tps = 191546.208051 (excluding connections establishing)statement latencies in milliseconds:0.002085 \setrandom id 1 10.664420 select f(:id);
使用advisory lock,性能相比不做任何优化性能提升了约66倍,相比
for update nowait性能提升了约1.8倍。
这种优化可以快速告诉用户是否能秒杀到此类商品,而不需要等待其他用户更新结束后才知道。所以大大降低了RT,提高了吞吐率。
[参考]