PostgreSQL 性能优化之 - 大锁与long sql/xact的蝴蝶效应(泛洪)
背景
在使用数据库时，某些特定的因素结合在一起，就可能引发蝴蝶效应。
导致数据库性能的急剧下降。
本文要讲的是和以下几个因素有关的：
因素1
PG的锁排队机制，即使没有获得锁，只要在锁队列中就会造成锁竞争。
session A lock1 get
session B lock2 wait lock1 release
session C lock3 可能和lock1, lock2 冲突
session D lock4 可能和lock1,2,3 冲突
因素2
SQL响应变慢后，服务端程序通常会增加到数据库的连接来处理拥塞的请求
因素3
数据库的性能会随着连接数增加到一个值（通常是核数的3倍）后，性能开始下降
因素4
对象锁在事务结束时释放
例如在事务中查询表时，表的共享锁需要等到事务结束时释放。
以上4个因素加在一起，就可能引发一次应用级别的故障。
模拟
开启lock跟踪：
log_lock_waits = on
deadlock_timeout = 1s
创建测试表
postgres=# create table test(id int primary key, info text, crt_time timestamp);
CREATE TABLE
postgres=# insert into test select generate_series(1,10000000),md5(random()::text),clock_timestamp();
INSERT 0 10000000
测试脚本
vi test1.sql
\setrandom id 1 10000000
update test set info=info where id=:id;
1. pgbench1 模拟数据更新A。使用10个链接（假设正常只需要10个）
正常的性能
pgbench -M prepared -n -r -P 1 -f ./test1.sql -c 10 -j 10 -T 10
progress: 2.0 s, 65994.3 tps, lat 0.149 ms stddev 0.038
progress: 3.0 s, 67706.5 tps, lat 0.145 ms stddev 0.051
progress: 4.0 s, 72865.0 tps, lat 0.135 ms stddev 0.048
progress: 5.0 s, 77664.2 tps, lat 0.126 ms stddev 0.032
progress: 6.0 s, 77138.9 tps, lat 0.127 ms stddev 0.037
progress: 7.0 s, 75941.3 tps, lat 0.129 ms stddev 0.061
progress: 8.0 s, 77328.8 tps, lat 0.127 ms stddev 0.036
开启长时间更新请求
pgbench -M prepared -n -r -P 1 -f ./test1.sql -c 10 -j 10 -T 10000
2. 模拟一个查询长事务，查询表A
也可能是后台的whole vacuum prevent wrapper
postgres=# begin;
BEGIN
postgres=# select * from test limit 1;
id | info | crt_time
----+----------------------------------+----------------------------
1 | e86e219d51c39d16f78d77cf697395ca | 2016-03-16 16:07:49.814487
(1 row)
暂不结束事务, 持有test表的shared lock.
3. 模拟一个DDL请求A
postgres=# alter table test add column c1 int;
等待test shared lock锁释放
马上会堵塞正常的业务请求，tps降到0
progress: 53.0 s, 0.0 tps, lat -nan ms stddev -nan
progress: 54.0 s, 0.0 tps, lat -nan ms stddev -nan
progress: 55.0 s, 0.0 tps, lat -nan ms stddev -nan
progress: 56.0 s, 0.0 tps, lat -nan ms stddev -nan
progress: 57.0 s, 0.0 tps, lat -nan ms stddev -nan
progress: 58.0 s, 0.0 tps, lat -nan ms stddev -nan
progress: 59.0 s, 0.0 tps, lat -nan ms stddev -nan
4. 这个时候，业务并不知道数据库堵塞了，会增加更多的连接来处理用户的请求。甚至可能把连接塞满。
pgbench2 模拟拥塞更新A，新建500链接
pgbench -M prepared -n -r -P 1 -f ./test1.sql -c 500 -j 500 -T 10000
新增的连接会全部处于这样的状态:
digoal 25434 22068 0 16:21 ? 00:00:00 postgres: postgres postgres [local] PARSE waiting
digoal 25437 22068 0 16:21 ? 00:00:00 postgres: postgres postgres [local] PARSE waiting
5. 结束长事务或结束DDL请求后，锁释放。
锁释放，大量并发的连接开始处理拥塞的请求
此时性能下降了一半
pgbench2
progress: 10.3 s, 270.5 tps, lat 1396.862 ms stddev 3498.526
progress: 11.0 s, 34443.5 tps, lat 64.132 ms stddev 709.718
progress: 12.0 s, 34986.1 tps, lat 14.229 ms stddev 18.469
progress: 13.0 s, 36645.0 tps, lat 13.661 ms stddev 17.686
progress: 14.0 s, 34570.1 tps, lat 14.463 ms stddev 18.716
progress: 15.0 s, 36435.8 tps, lat 13.752 ms stddev 17.621
progress: 16.0 s, 35513.3 tps, lat 14.052 ms stddev 18.087
progress: 17.0 s, 35560.0 tps, lat 14.013 ms stddev 18.159
pgbench1
progress: 59.0 s, 688.7 tps, lat 340.857 ms stddev 2734.371
progress: 60.0 s, 733.0 tps, lat 13.659 ms stddev 18.501
progress: 61.0 s, 816.0 tps, lat 12.237 ms stddev 16.941
progress: 62.0 s, 811.0 tps, lat 12.328 ms stddev 16.715
progress: 63.0 s, 809.9 tps, lat 12.370 ms stddev 17.370
progress: 64.0 s, 750.1 tps, lat 13.338 ms stddev 17.745
将后建立的500个连接释放后，恢复正常的性能
progress: 66.0 s, 1937.8 tps, lat 5.044 ms stddev 12.975
progress: 67.0 s, 64995.8 tps, lat 0.157 ms stddev 0.757
progress: 68.0 s, 73996.3 tps, lat 0.133 ms stddev 0.042
progress: 69.0 s, 78099.4 tps, lat 0.125 ms stddev 0.038
日志
可以追踪到锁等待的源头，但是不能追踪到大锁。
2016-03-16 16:25:57.531 CST,"postgres","postgres",48877,"[local]",56e91894.beed,3,"ALTER TABLE waiting",2016-03-16 16:25:56 CST,13/28,580426398,LOG,00000,"process 48877 still waiting for AccessExclusiveLock on relation 61245 of database 13241 after 1000.048 ms","Process holding the lock: 48557. Wait queue: 48877, 46333, 46331, 46338, 46334, 46339, 46335, 46340, 46337, 46328, 46336.",,,,,"alter table test add column c1 int;",,"ProcSleep, proc.c:1323","psql"
2016-03-16 16:25:57.531 CST,"postgres","postgres",46333,"[local]",56e91871.b4fd,3,"BIND waiting",2016-03-16 16:25:21 CST,4/263058,0,LOG,00000,"process 46333 still waiting for RowExclusiveLock on relation 61245 of database 13241 after 1000.036 ms","Process holding the lock: 48557. Wait queue: 48877, 46333, 46331, 46338, 46334, 46339, 46335, 46340, 46337, 46328, 46336.",,,,,"update test set info=info where id=$1;",,"ProcSleep, proc.c:1323","pgbench"
2016-03-16 16:26:10.191 CST,"postgres","postgres",49812,"[local]",56e918a1.c294,3,"PARSE waiting",2016-03-16 16:26:09 CST,14/29,0,LOG,00000,"process 49812 still waiting for RowExclusiveLock on relation 61245 of database 13241 after 1000.207 ms","Process holding the lock: 48557. Wait queue: 48877, 46333, 此处省略500+ PIDs, 50816, 50817.",,,,,"update test set info=info where id=$1;",8,"ProcSleep, proc.c:1323","pgbench"
2016-03-16 16:26:19.367 CST,"postgres","postgres",48877,"[local]",56e91894.beed,4,"ALTER TABLE waiting",2016-03-16 16:25:56 CST,13/28,580426398,LOG,00000,"process 48877 acquired AccessExclusiveLock on relation 61245 of database 13241 after 22836.312 ms",,,,,,"alter table test add column c1 int;",,"ProcSleep, proc.c:1327","psql"
2016-03-16 16:26:19.368 CST,"postgres","postgres",48877,"[local]",56e91894.beed,5,"ALTER TABLE",2016-03-16 16:25:56 CST,13/28,580426398,ERROR,42701,"column ""c1"" of relation ""test"" already exists",,,,,,"alter table test add column c1 int;",,"check_for_column_name_collision, tablecmds.c:5069","psql"
2016-03-16 16:26:19.379 CST,"postgres","postgres",49814,"[local]",56e918a1.c296,4,"PARSE waiting",2016-03-16 16:26:09 CST,15/2,0,LOG,00000,"process 49814 acquired RowExclusiveLock on relation 61245 of database 13241 after 10177.162 ms",,,,,,"update test set info=info where id=$1;",8,"ProcSleep, proc.c:1327","pgbench"
要追踪大锁，
可以使用以下SQL
with t_wait as
(select a.mode,a.locktype,a.database,a.relation,a.page,a.tuple,a.classid,
a.objid,a.objsubid,a.pid,a.virtualtransaction,a.virtualxid,a,
transactionid,b.query,b.xact_start,b.query_start,b.usename,b.datname
from pg_locks a,pg_stat_activity b where a.pid=b.pid and not a.granted),
t_run as
(select a.mode,a.locktype,a.database,a.relation,a.page,a.tuple,
a.classid,a.objid,a.objsubid,a.pid,a.virtualtransaction,a.virtualxid,
a,transactionid,b.query,b.xact_start,b.query_start,
b.usename,b.datname from pg_locks a,pg_stat_activity b where
a.pid=b.pid and a.granted)
select r.locktype,r.mode r_mode,r.usename r_user,r.datname r_db,
r.relation::regclass,r.pid r_pid,
r.page r_page,r.tuple r_tuple,r.xact_start r_xact_start,
r.query_start r_query_start,
now()-r.query_start r_locktime,r.query r_query,w.mode w_mode,
w.pid w_pid,w.page w_page,
w.tuple w_tuple,w.xact_start w_xact_start,w.query_start w_query_start,
now()-w.query_start w_locktime,w.query w_query
from t_wait w,t_run r where
r.locktype is not distinct from w.locktype and
r.database is not distinct from w.database and
r.relation is not distinct from w.relation and
r.page is not distinct from w.page and
r.tuple is not distinct from w.tuple and
r.classid is not distinct from w.classid and
r.objid is not distinct from w.objid and
r.objsubid is not distinct from w.objsubid and
r.transactionid is not distinct from w.transactionid and
r.pid <> w.pid
order by
(( case w.mode
when 'INVALID' then 0
when 'AccessShareLock' then 1
when 'RowShareLock' then 2
when 'RowExclusiveLock' then 3
when 'ShareUpdateExclusiveLock' then 4
when 'ShareLock' then 5
when 'ShareRowExclusiveLock' then 6
when 'ExclusiveLock' then 7
when 'AccessExclusiveLock' then 8
else 0
end ) +
( case r.mode
when 'INVALID' then 0
when 'AccessShareLock' then 1
when 'RowShareLock' then 2
when 'RowExclusiveLock' then 3
when 'ShareUpdateExclusiveLock' then 4
when 'ShareLock' then 5
when 'ShareRowExclusiveLock' then 6
when 'ExclusiveLock' then 7
when 'AccessExclusiveLock' then 8
else 0
end )) desc,r.xact_start;
优化措施或处理措施
1. 养成大锁处理习惯, 配置锁等待超时
2. 应用程序或中间件应该有自动释放空闲连接的功能
3. auto_explain也不会记录锁等待的时间，所以不利于分析原因。只有从日志中才能分析。