PostgreSQL 秒杀场景优化

先说结果：
TPS 从 2855 优化到 231376。

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 t1
      Table "public.t1"
 Column |  Type   | Modifiers 
--------+---------+-----------
 id     | integer | not null
 info   | 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 void  
 LANGUAGE plpgsql  
AS $function$ 
declare 
begin 
  perform 1 from t1 where id=i_id for update nowait; 
  update t1 set info=now()::text where id=i_id; 
  exception when others then  
  return; 
end; 
$function$;  

postgres@digoal-> cat test1.sql
\setrandom id 1 1
select f1(:id);

压测advisory lock的方式：
postgres@digoal-> cat test.sql
\setrandom id 1 1
update 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               | 184731
schemaname          | public
relname             | t1
seq_scan            | 0
seq_tup_read        | 0
idx_scan            | 0
idx_tup_fetch       | 0
n_tup_ins           | 0
n_tup_upd           | 0
n_tup_del           | 0
n_tup_hot_upd       | 0
n_live_tup          | 0
n_dead_tup          | 0
n_mod_since_analyze | 0
last_vacuum         | 
last_autovacuum     | 
last_analyze        | 
last_autoanalyze    | 
vacuum_count        | 0
autovacuum_count    | 0
analyze_count       | 0
autoanalyze_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 query
scaling factor: 1
query mode: prepared
number of clients: 20
number of threads: 20
duration: 60 s
number of transactions actually processed: 792029
latency average: 1.505 ms
latency stddev: 4.275 ms
tps = 13196.542846 (including connections establishing)
tps = 13257.270709 (excluding connections establishing)
statement latencies in milliseconds:
        0.002625        \setrandom id 1 1
        1.502420        select f1(:id);

postgres=# select * from pg_stat_all_tables where relname='t1';
-[ RECORD 1 ]-------+-------
relid               | 184731
schemaname          | public
relname             | t1
seq_scan            | 0
seq_tup_read        | 0
idx_scan            | 896963   //  大多数是无用功
idx_tup_fetch       | 896963   //  大多数是无用功
n_tup_ins           | 0
n_tup_upd           | 41775
n_tup_del           | 0
n_tup_hot_upd       | 41400
n_live_tup          | 0
n_dead_tup          | 928
n_mod_since_analyze | 41774
last_vacuum         | 
last_autovacuum     | 
last_analyze        | 
last_autoanalyze    | 
vacuum_count        | 0
autovacuum_count    | 0
analyze_count       | 0
autoanalyze_count   | 0


postgres@digoal-> pgbench -M prepared -n -r -P 1 -f ./test.sql -c 20 -j 20 -T 60
......
transaction type: Custom query
scaling factor: 1
query mode: prepared
number of clients: 20
number of threads: 20
duration: 60 s
number of transactions actually processed: 1392372
latency average: 0.851 ms
latency stddev: 2.475 ms
tps = 23194.831054 (including connections establishing)
tps = 23400.411501 (excluding connections establishing)
statement latencies in milliseconds:
        0.002594        \setrandom id 1 1
        0.848536        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               | 184731
schemaname          | public
relname             | t1
seq_scan            | 0
seq_tup_read        | 0
idx_scan            | 1368933  //  大多数是无用功
idx_tup_fetch       | 1368933   //  大多数是无用功
n_tup_ins           | 0
n_tup_upd           | 54957
n_tup_del           | 0
n_tup_hot_upd       | 54489
n_live_tup          | 0
n_dead_tup          | 1048
n_mod_since_analyze | 54957
last_vacuum         | 
last_autovacuum     | 
last_analyze        | 
last_autoanalyze    | 
vacuum_count        | 0
autovacuum_count    | 0
analyze_count       | 0
autoanalyze_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 void  
 LANGUAGE plpgsql  
AS $function$ 
declare 
  a_lock boolean := false;
begin 
  select pg_try_advisory_xact_lock(i_id) into a_lock;
  if a_lock then
    update t1 set info=now()::text where id=i_id; 
  end if;
  exception when others then  
  return; 
end; 
$function$;  

transaction type: Custom query
scaling factor: 1
query mode: prepared
number of clients: 20
number of threads: 20
duration: 60 s
number of transactions actually processed: 1217195
latency average: 0.973 ms
latency stddev: 3.563 ms
tps = 20283.314001 (including connections establishing)
tps = 20490.143363 (excluding connections establishing)
statement latencies in milliseconds:
        0.002703        \setrandom id 1 1
        0.970209        select f(:id);

postgres=# select * from pg_stat_all_tables where relname='t1';
-[ RECORD 1 ]-------+-------
relid               | 184731
schemaname          | public
relname             | t1
seq_scan            | 0
seq_tup_read        | 0
idx_scan            | 75927
idx_tup_fetch       | 75927
n_tup_ins           | 0
n_tup_upd           | 75927
n_tup_del           | 0
n_tup_hot_upd       | 75902
n_live_tup          | 0
n_dead_tup          | 962
n_mod_since_analyze | 75927
last_vacuum         | 
last_autovacuum     | 
last_analyze        | 
last_autoanalyze    | 
vacuum_count        | 0
autovacuum_count    | 0
analyze_count       | 0
autoanalyze_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 query
scaling factor: 1
query mode: prepared
number of clients: 128
number of threads: 128
duration: 100 s
number of transactions actually processed: 285673
latency average: 44.806 ms
latency stddev: 45.751 ms
tps = 2855.547375 (including connections establishing)
tps = 2855.856976 (excluding connections establishing)
statement latencies in milliseconds:
        0.002509        \setrandom id 1 1
        44.803299       update t1 set info=now()::text where id=:id;

使用for update nowait的并发处理能力：
transaction type: Custom query
scaling factor: 1
query mode: prepared
number of clients: 128
number of threads: 128
duration: 100 s
number of transactions actually processed: 6663253
latency average: 1.919 ms
latency stddev: 2.804 ms
tps = 66623.169445 (including connections establishing)
tps = 66630.307999 (excluding connections establishing)
statement latencies in milliseconds:
        0.001934        \setrandom id 1 1
        1.917297        select f1(:id);

使用advisory lock后的并发处理能力：
transaction type: Custom query
scaling factor: 1
query mode: prepared
number of clients: 80
number of threads: 80
duration: 60 s
number of transactions actually processed: 13883387
latency average: 0.344 ms
latency stddev: 0.535 ms
tps = 231197.323122 (including connections establishing)
tps = 231376.427515 (excluding connections establishing)
statement latencies in milliseconds:
        0.344042        select f(1);
此时的perf top
   PerfTop:   23883 irqs/sec  kernel:32.2%  exact:  0.0% [1000Hz cycles],  (all, 32 CPUs)
--------------------------------------------------------------------------------     

 samples  pcnt function         DSO
 _______ _____ ____________________________ ______________________________________________

10645.00  3.5% GetSnapshotData  /u02/digoal/soft_bak/pgsql9.5/bin/postgres    
 7963.00  2.6% AllocSetAlloc    /u02/digoal/soft_bak/pgsql9.5/bin/postgres    
 4720.00  1.6% _int_malloc      /lib64/libc-2.12.so   
 4270.00  1.4% __schedule       [kernel.kallsyms]     
 4234.00  1.4% fmgr_info_cxt_security       /u02/digoal/soft_bak/pgsql9.5/bin/postgres    
 4217.00  1.4% LWLockAcquire    /u02/digoal/soft_bak/pgsql9.5/bin/postgres    
 3958.00  1.3% hash_search_with_hash_value  /u02/digoal/soft_bak/pgsql9.5/bin/postgres    
 3656.00  1.2% __GI_vfprintf    /lib64/libc-2.12.so   
 3572.00  1.2% update_blocked_averages      [kernel.kallsyms]     
 3338.00  1.1% PostgresMain     /u02/digoal/soft_bak/pgsql9.5/bin/postgres    
 3267.00  1.1% __switch_to      [kernel.kallsyms]     
 3095.00  1.0% __strlen_sse42   /lib64/libc-2.12.so   
 2996.00  1.0% memcpy           /lib64/libc-2.12.so   
 2930.00  1.0% _int_free        /lib64/libc-2.12.so   
 2568.00  0.8% LWLockRelease    /u02/digoal/soft_bak/pgsql9.5/bin/postgres    
 2446.00  0.8% SearchCatCache   /u02/digoal/soft_bak/pgsql9.5/bin/postgres    
 2178.00  0.7% ExecInitExpr     /u02/digoal/soft_bak/pgsql9.5/bin/postgres    
 2053.00  0.7% hash_any         /u02/digoal/soft_bak/pgsql9.5/bin/postgres    
 2035.00  0.7% __GI___libc_malloc           /lib64/libc-2.12.so   
 2009.00  0.7% _raw_spin_lock_irqsave       [kernel.kallsyms]     
 1804.00  0.6% exec_stmt        /u02/digoal/soft_bak/pgsql9.5/lib/plpgsql.so  
 1764.00  0.6% __memset_sse2    /lib64/libc-2.12.so   
 1717.00  0.6% pqParseInput3    /u02/digoal/soft_bak/pgsql9.5/lib/libpq.so.5.8
 1696.00  0.6% do_select        [kernel.kallsyms]     
 1686.00  0.6% __strcpy_ssse3   /lib64/libc-2.12.so   
 1685.00  0.6% update_curr      [kernel.kallsyms]     
 1619.00  0.5% enqueue_entity   [kernel.kallsyms]     
 1607.00  0.5% pfree/u02/digoal/soft_bak/pgsql9.5/bin/postgres    
 1598.00  0.5% doCustom         /u02/digoal/soft_bak/pgsql9.5/bin/pgbench     
 1594.00  0.5% idle_cpu         [kernel.kallsyms]     
 1589.00  0.5% update_cfs_rq_blocked_load   [kernel.kallsyms]     
 1554.00  0.5% lapic_next_deadline          [kernel.kallsyms]     
 1512.00  0.5% update_cfs_shares[kernel.kallsyms]     
 1491.00  0.5% MemoryContextCreate          /u02/digoal/soft_bak/pgsql9.5/bin/postgres    
 1482.00  0.5% _raw_spin_lock   [kernel.kallsyms]     
 1423.00  0.5% palloc           /u02/digoal/soft_bak/pgsql9.5/bin/postgres    
 1419.00  0.5% __GI___sigsetjmp /lib64/libc-2.12.so   
 1412.00  0.5% __cfree          /lib64/libc-2.12.so   
 1399.00  0.5% unix_stream_recvmsg          [kernel.kallsyms]     
 1393.00  0.5% __fget_light     [kernel.kallsyms]     
 1359.00  0.4% ResourceOwnerReleaseInternal /u02/digoal/soft_bak/pgsql9.5/bin/postgres    
 1351.00  0.4% AllocSetFree     /u02/digoal/soft_bak/pgsql9.5/bin/postgres    
 1277.00  0.4% unix_stream_sendmsg          [kernel.kallsyms]     
 1246.00  0.4% __memcmp_sse4_1  /lib64/libc-2.12.so   
 1240.00  0.4% plpgsql_exec_function        /u02/digoal/soft_bak/pgsql9.5/lib/plpgsql.so  
 1225.00  0.4% expression_tree_walker       /u02/digoal/soft_bak/pgsql9.5/bin/postgres    
 1160.00  0.4% exec_stmt_block  /u02/digoal/soft_bak/pgsql9.5/lib/plpgsql.so
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。

最后提一下9.5的新特性, select ,,, for update ,,, skip locked.
http://blog.163.com/digoal@126/blog/static/163877040201551552017215/
如果能做到UPDATE语法里面，就完美了，直接跳过无法获得锁的行。并发能力瞬间提升，也不用advisory了。

[参考]
1. http://www.postgresql.org/docs/9.5/static/functions-admin.html#FUNCTIONS-ADVISORY-LOCKS