PostgreSQL 不要使用kill -9 杀 Postgresq 用户进程

转载：http://francs3.blog.163.com/blog/static/4057672720109854858308/

Postgresql 8.3.3

今天应用反映数据库很慢，有些SQL执行一天都没执行完,
检查数据库，发现大量会话在更新同一张表 tbl_tmp, 产生大量行锁申请等侍。

--等于等侍状态的SQL (被堵住的SQL)

anpstat=# select datname,procpid,query_start,current_query,waiting,client_addr from pg_stat_activity where waiting='t';
datname | procpid |          query_start          |                                                           current_query                                                           | waiting | client_addr
----------+---------+-------------------------------+-----------------------------------------------------------------------------------------------------------------------------------+---------+-------------
sanpstat |   14044 | 2010-10-07 16:46:40.386904+08 |     update tbl_tmp     set responsed=true ,    feedback =$1,    gmt_modified=now(),    modifier=$2    where seq=$3     | t       | 127.0.0.1
sanpstat |   27839 | 2010-10-07 16:49:18.022499+08 |     update tbl_tmp     set responsed=true ,    feedback =$1,    gmt_modified=now(),    modifier=$2    where seq=$3     | t       | 127.0.0.1
sanpstat |    3539 | 2010-10-07 20:28:13.212075+08 |     update tbl_tmp set   feedback =$1 where         seq=$2                                         | t       | 127.0.0.1
sanpstat |    3894 | 2010-10-07 20:53:43.466517+08 |     update tbl_tmp     set responsed=true ,    feedback =$1,    gmt_modified=now(),    modifier=$2    where seq=$3     | t       | 127.0.0.1
sanpstat |   10130 | 2010-10-08 09:37:51.253871+08 |     update tbl_tmp     set responsed=true ,    feedback =$1,    gmt_modified=now(),    modifier=$2    where seq=$3     | t       | 127.0.0.1
sanpstat |    9083 | 2010-10-08 08:08:59.221976+08 |     update tbl_tmp     set responsed=true ,    feedback =$1,    gmt_modified=now(),    modifier=$2    where seq=$3     | t       | 127.0.0.1
sanpstat |   10038 | 2010-10-08 09:25:55.434459+08 |     update tbl_tmp     set responsed=true ,    feedback =$1,    gmt_modified=now(),    modifier=$2    where seq=$3     | t       | 127.0.0.1
sanpstat |   10241 | 2010-10-08 09:51:13.990492+08 |     update tbl_tmp     set responsed=true ,    feedback =$1,    gmt_modified=now(),    modifier=$2    where seq=$3     | t       | 127.0.0.1                                                                    | t       | 127.0.0.1
sanpstat |   11147 | 2010-10-08 11:11:21.574665+08 |     update tbl_tmp     set responsed=true ,    feedback =$1,    gmt_modified=now(),    modifier=$2    where seq=$3     | t       | 127.0.0.1
sanpstat |   11168 | 2010-10-08 11:17:17.158246+08 |     update tbl_tmp     set responsed=true ,    feedback =$1,    gmt_modified=now(),    modifier=$2    where seq=$3     | t       | 127.0.0.1
sanpstat |   11926 | 2010-10-08 11:54:10.704641+08 |     update tbl_tmp     set responsed=true ,    feedback =$1,    gmt_modified=now(),    modifier=$2    where seq=$3     | t       | 127.0.0.1
sanpstat |   11959 | 2010-10-08 11:56:07.021072+08 |     update tbl_tmp     set responsed=true ,    feedback =$1,    gmt_modified=now(),    modifier=$2    where seq=$3     | t       | 127.0.0.1

--正在执行的 update 语句（堵住了其它SQL）

sanpstat=# select datname,procpid,query_start,current_query,waiting,client_addr from pg_stat_activity where waiting='f'
sanpstat-# and current_query like '%update tbl_tmp %';
datname | procpid |          query_start          |                                                           current_query                                                           | waiting | client_addr
----------+---------+-------------------------------+-----------------------------------------------------------------------------------------------------------------------------------+---------+---------------
sanpstat |    2012 | 2010-10-07 16:52:55.8228+08   |     update tbl_tmp     set responsed=true ,    feedback =$1,    gmt_modified=now(),    modifier=$2    where seq=$3     | f       | 127.0.0.1
sanpstat |   14157 | 2010-10-07 16:41:37.308062+08 |     update tbl_tmp     set responsed=true ,    feedback =$1,    gmt_modified=now(),    modifier=$2    where seq=$3     | f       | 127.0.0.1
sanpstat |   10177 | 2010-10-08 09:43:43.297872+08 |     update tbl_tmp     set responsed=true ,    feedback =$1,    gmt_modified=now(),    modifier=$2    where seq=$3     | f       | 127.0.0.1
sanpstat |    2043 | 2010-10-07 19:25:46.931806+08 |     update tbl_tmp     set responsed=true ,    feedback =$1,    gmt_modified=now(),    modifier=$2    where seq=$3     | f       | 127.0.0.1
sanpstat |    3298 | 2010-10-07 19:37:53.307125+08 |     update tbl_tmp     set responsed=true ,    feedback =$1,    gmt_modified=now(),    modifier=$2    where seq=$3     | f       | 127.0.0.1
sanpstat |    3322 | 2010-10-07 19:40:39.373079+08 |     update tbl_tmp     set responsed=true ,    feedback =$1,    gmt_modified=now(),    modifier=$2    where seq=$3     | f       | 127.0.0.1
sanpstat |    4114 | 2010-10-08 09:25:18.701269+08 |     update tbl_tmp     set responsed=true ,    feedback =$1,    gmt_modified=now(),    modifier=$2    where seq=$3     | f       | 127.0.0.1
sanpstat |    9082 | 2010-10-08 09:18:16.153882+08 |     update tbl_tmp     set show=1 ,    gmt_modified=now(),    modifier=$1    where seq=$2                              | f       | 127.0.0.1
sanpstat |   10042 | 2010-10-08 09:32:55.260732+08 |     update tbl_tmp     set responsed=true ,    feedback =$1,    gmt_modified=now(),    modifier=$2    where seq=$3     | f       | 127.0.0.1
sanpstat |   10278 | 2010-10-08 11:08:55.334562+08 |     update tbl_tmp     set responsed=true ,    feedback =$1,    gmt_modified=now(),    modifier=$2    where seq=$3     | f       | 127.0.0.1
sanpstat |   11810 | 2010-10-08 11:46:24.147652+08 |     update tbl_tmp     set responsed=true ,    feedback =$1,    gmt_modified=now(),    modifier=$2    where seq=$3     | f       | 127.0.0.1
sanpstat |   11964 | 2010-10-08 11:58:32.831916+08 |     update tbl_tmp     set responsed=true ,    feedback =$1,    gmt_modified=now(),    modifier=$2    where seq=$3     |

--用pg_cancel_backend()杀进程

sanpstat=# select pg_cancel_backend(2012);
pg_cancel_backend
-------------------
t
(1 row)

用 pg_cancel_backend()杀进程，虽然显示为 't' ，但进程还在，文档上解释 Cancel a backend's current query,猜想pg_cancel_backend()只能 kill select 语句，而不能kill update语句，后来实验也证实了这一点，有兴趣的朋友可以自己做实验去验证下。

--kill -9 杀进程

kill -9   2012
kill -9 14157
kill -9 10177
kill -9   2043
kill -9   3298
kill -9   3322
kill -9   4114
kill -9   9082
kill -9 10042
kill -9 10278
kill -9 11810
kill -9 11964

这里将正在执行的update tbl_tmp 表的会话kill 掉.

--数据库无法连接

[enterprisedb@sanp-rich-db1 ~]$ psql -h 192.168.3.27 -p 1921 -d postgres -U rmt_db_bak
psql: FATAL: the database system is in recovery mode

备注：进程kill -9 以后，发现数据库无法连接，遭了，还好是个测试库。

--数据库处于恢复模式

-bash-3.2$ ps -ef | grep post
postgres 4627     1 0 Apr12 ?        00:02:24 /home/postgres/bin/postgres -D /mnt/data
postgres 4629 4627 0 Apr12 ?        00:00:04 postgres: logger process
root     14195 14168 0 15:37 pts/7    00:00:00 su - postgres
postgres 14196 14195 0 15:37 pts/7    00:00:00 -bash
postgres 14332 4627 44 15:41 ?        00:00:18 postgres: startup process   recovering 00000001000000DA0000006E
postgres 21495 14196 0 15:41 pts/7    00:00:00 ps -ef
postgres 21496 14196 0 15:41 pts/7    00:00:00 grep post

发现Postgresql 处于 recovery 状态，奇怪，Postgresql 恢复过程中是不允许连接的，在这点上，Oracle似乎要友好些

--结论

postgresql 8.3.3的版本只提供系统函数 pg_cancel_backend(pid int) 杀进程，但是，这个函数只能 kill Select 查询，而updae,delete DML不生效,感兴趣的朋友可以自己做实验验证下；
postgresql 8.4 以后版本提供函数 pg_terminate_backend(pid int), 这个函数功能比前者强大，可以kill 各种DML(SELECT,UPDATE,DELETE,DROP)操作;
对于Postgresql ，最好不要用kill -9 去杀用户进程，因为如果是一个很大的UPDATE,kill后，Postgresql需要很长的时间做Recovery,而在这个RECOVERY过程中，数据库是不可用的，在这点上，Oracle 要友好些，在数据库恢复的过程中数据库依然可用,不太需要因Kill -9 用户进程而造成数据库不可用的情况。