一、测试环境
系统：Centos 6.2
数据库：mysql Ver 14.14 Distrib 5.5.18, for Linux (x86_64) using readline 5.1
percona工具：percona-toolkit-2.2.12
测试数据库大小：tx_ljxz_71--16G、t_log_item--3G
二、在线修改表结构的过程
mysql在线修改表结构
1 按照原始表（original_table）的表结构和DDL语句，新建一个不可见的临时表（tmp_table）
2 在原表上加write lock，阻塞所有更新操作（insert、delete、update等）
3 执行insert into tmp_table select * from original_table
4 rename original_table和tmp_table，最后drop original_table
5 释放 write lock。
我们可以看见在InnoDB执行DDL的时候，原表是只能读不能写的。
pt-online-schema-change在线修改表结构
1 创建一个和你要执行 alter 操作的表一样的空表结构。
2 执行表结构修改，然后从原表中的数据到copy到 表结构修改后的表，
3 在原表上创建触发器将 copy 数据的过程中，在原表的更新操作 更新到新表.
注意:如果表中已经定义了触发器这个工具就不能工作了。
4 copy 完成以后，用rename table 新表代替原表，默认删除原表。
三、安装配置percona-toolkit
1.安装
官网下载即可
tar zxvf percona-toolkit-2.2.12.tar.gzperl Makefile.PL
make
make test
make install
2.测试安装成功
pt-online-schema-change -uroot --alter="modify buff_id int(20)" --dry-run D=tx_ljxz_71,t=t_world_buff
报错：
Cannot connect to MySQL: install_driver(mysql) failed: Attempt to reload DBD/mysql.pm aborted.Compilation failed in require at (eval 21) line 3. at /usr/local/bin/pt-online-schema-change line 2261
后在虚拟机上的centos 6.5上安装测试，没发现同样的问题，经检测，是定制系统的问题
/usr/local/lib64/perl5/Bundle/DBD/mysql.pm/usr/local/lib64/perl5/DBD/mysql.pm
要求perl-DBD-MySQL的安装目录在/usr/lib64/perl5/，将原来的目录移走，重新安装即可
mv /usr/local/lib64/perl5/ /data/backup/yum install perl-DBD-MySQL
重测一次，返回
Operation, tries, wait: copy_rows, 10, 0.25 create_triggers, 10, 1 drop_triggers, 10, 1 swap_tables, 10, 1 update_foreign_keys, 10, 1Starting a dry run. `tx_ljxz_71`.`t_world_buff` will not be altered. Specify --execute instead of --dry-run to alter the table.Creating new table...Created new table tx_ljxz_71._t_world_buff_new OK.Altering new table...Altered `tx_ljxz_71`.`_t_world_buff_new` OK.Not creating triggers because this is a dry run.Not copying rows because this is a dry run.Not swapping tables because this is a dry run.Not dropping old table because this is a dry run.Not dropping triggers because this is a dry run.2015-01-28T17:00:22 Dropping new table...2015-01-28T17:00:22 Dropped new table OK.Dry run complete. `tx_ljxz_71`.`t_world_buff` was not altered.
四、测试在线修改表结构
1.在线修改表结构的时间对比：
mysql在线修改表结构
time mysql -uroot tx_ljxz_71 -e "alter table t_log_item modify role_level int(20)"
| real 18m51.198s | real 18m17.492s | real 19m53.119s |
在innodb_buffer_pool_size预热之后，在线修改表结构的时间为：
| 12m9.507s | 11m23.653s |
11m51.561s |
负载：
top - 10:32:56 up 111 days, 18:40, 2 users, load average: 1.50, 1.42, 0.77Tasks: 223 total, 1 running, 222 sleeping, 0 stopped, 0 zombieCpu(s): 9.2%us, 1.0%sy, 0.0%ni, 85.0%id, 4.8%wa, 0.0%hi, 0.0%si, 0.0%stMem: 16315920k total, 16162676k used, 153244k free, 51036k buffersSwap: 12287992k total, 245416k used, 12042576k free, 6720696k cached
pt-online-schema-change在线修改表结构
time pt-online-schema-change --user=root --critical-load Threads_running=100 --alter="modify role_level int(20)" --execute D=tx_ljxz_71,t=t_log_item
--critical-load值默认的Threads_running为50，不修改会报错：Error copying rows from `tx_ljxz_71`.`t_log_item` to `tx_ljxz_71`.`_t_log_item_new`: Threads_running=51 exceeds its critical threshold 50
| real 18m56.473s | real 19m27.950s | real 18m58.556s |
在innodb_buffer_pool_size预热之后，在线修改表结构的时间为：
| 14m10.936s | 13m41.146s | 13m56.238s |
负载：
top - 10:32:19 up 111 days, 18:53, 2 users, load average: 2.29, 1.75, 0.83Tasks: 230 total, 1 running, 229 sleeping, 0 stopped, 0 zombieCpu(s): 12.2%us, 1.2%sy, 0.0%ni, 65.5%id, 21.1%wa, 0.0%hi, 0.0%si, 0.0%stMem: 12179948k total, 12011356k used, 168592k free, 37348k buffersSwap: 12287992k total, 165424k used, 12122568k free, 2597832k cached
此时，没有执行任何sql语句对数据库进行操作
2.在线修改表结构时对数据库进行增删查改
mysql在线修改表结构
time mysql -uroot tx_ljxz_71 -e "alter table t_log_item modify role_level int(20)"
执行测试脚本，发现除了select语句，其他语句无法继续执行直至表结构修改完毕再进行增删改的操作
在线修改表结构的时间为：
| 11m43.975s | 11m37.980s | 11m22.229s |
此时增删查改的时间为：
| 14m36.781s | 13m40.872s | 13m29.913s |
pt-online-schema-change在线修改表结构
time pt-online-schema-change --user=root --critical-load Threads_running=100 --alter="modify role_level int(20)" --execute D=tx_ljxz_71,t=t_log_item
a.percona-toolkit默认的innodb_lock_wait_timeout为1，retries值为3，需要在mysql的配置文件中修改innodb_lock_wait_timeout值,如果innodb_lock_wait_timeout值不够大（隐患），在执行mysql语句的同时，复制数据表，会报错：
2015-02-03T16:22:48 Error copying rows from `tx_ljxz_71`.`t_log_item` to `tx_ljxz_71`.`_t_log_item_new`: 2015-02-03T16:22:48 DBD::mysql::st execute failed: Lock wait timeout exceeded; try restarting transaction [for Statement "INSERT LOW_PRIORITY IGNORE INTO `tx_ljxz_71`.`_t_log_item_new` (`pid`, `agent_id`, `server_id`, `role_id`, `role_level`, `action`, `item_id`, `amount`, `bag_amount`, `zero_dateline`, `year`, `month`, `day`, `equip_id`, `color`, `fineness`, `start_time`, `end_time`, `bind_type`, `super_unique_id`, `from`) SELECT `pid`, `agent_id`, `server_id`, `role_id`, `role_level`, `action`, `item_id`, `amount`, `bag_amount`, `zero_dateline`, `year`, `month`, `day`, `equip_id`, `color`, `fineness`, `start_time`, `end_time`, `bind_type`, `super_unique_id`, `from` FROM `tx_ljxz_71`.`t_log_item` FORCE INDEX(`PRIMARY`) WHERE ((`pid` >= ?)) AND ((`pid` <= ?)) LOCK IN SHARE MODE /*pt-online-schema-change 20031 copy nibble*/" with ParamValues: 0='13794706058943', 1='13794714479917'] at /usr/local/bin/pt-online-schema-change line 10385.
b.在线修改表结构时，会优先执行sql语句，暂停表结构的修改，暂停的时间如果大于innodb_lock_wait_timeout值会报上面一个错误，暂停的时间小于执行sql语句的，则执行完sql语句后，继续修改表结构,完成修改表结构的时间与在这期间执行sql语句的时间挂钩
Altering `tx_ljxz_71`.`t_log_item`...Creating new table...Created new table tx_ljxz_71._t_log_item_new OK.Altering new table...Altered `tx_ljxz_71`.`_t_log_item_new` OK.2015-02-03T16:25:31 Creating triggers...2015-02-03T16:25:31 Created triggers OK.2015-02-03T16:25:31 Copying approximately 14221411 rows...Copying `tx_ljxz_71`.`t_log_item`: 0% 01:01:13 remainCopying `tx_ljxz_71`.`t_log_item`: 1% 59:04 remainPausing because Threads_running=33.Pausing because Threads_running=30.Copying `tx_ljxz_71`.`t_log_item`: 1% 01:57:15 remainCopying `tx_ljxz_71`.`t_log_item`: 5% 44:33 remain......2015-02-03T16:40:48 Copied rows OK.2015-02-03T16:40:48 Swapping tables...2015-02-03T16:40:50 Swapped original and new tables OK.2015-02-03T16:40:50 Dropping old table...2015-02-03T16:40:59 Dropped old table `tx_ljxz_71`.`_t_log_item_old` OK.2015-02-03T16:40:59 Dropping triggers...2015-02-03T16:40:59 Dropped triggers OK.Successfully altered `tx_ljxz_71`.`t_log_item`.
在线修改表结构的时间为：
| 15m8.163s | 14m55.007s | 14m58.613s |
此时增删查改的时间为：
| 2m23.536s | 2m10.319s | 2m13.728s |
五、结论
在没有执行sql语句时，mysql在线修改表结构的时间与pt-online-schma-change的基本上相等，无大的差别；pt-online-schma-change方式会占用较多内存，负载也会略高
在mysql预热以后，在线修改表结构的时间，直接修改会比pt-online-schema-change方式略快
在线修改表结构的同时执行mysql语句，mysql直接修改的方式会先修改完表结构再执行sql语句，pt-online-schema-change方式会优先执行sql语句，再复制数据表，复制完毕后再把执行sql语句的结果更新到新表，因此，在时间上，直接修改表结构会比pt-online-schema-change方式，在实际用时上，直接修改会略快
测试脚本：
#!/usr/local/bin/pythonimport threadingfrom os import systemfrom random import randintimport timeimport MySQLdbMYSQL = "/usr/local/bin/mysql -uroot"cxn = MySQLdb.connect(host='localhost', user='root', port=3306, db='tx_ljxz_71')def run1(db): role_level = randint(40, 70) start_time = int(time.time()) system('%s %s -e "select count(*) from t_log_item where role_level=%s"' % (MYSQL, db, role_level)) print '%s select count(*) from t_log_item where role_level=%s use %d sec.' % (db, role_level, int(time.time()) - start_time)def run2(db): month = randint(9, 10) day = randint(20, 30) start_time = int(time.time()) system('%s %s -e "select sum(use_unbind) from t_log_consume_gold where year=2013 and month=%s and day=%s"' % (MYSQL, db, month, day)) print '%s select sum(use_unbind) from t_log_consume_gold where year=2013 and month=%s and day=%s use %d sec.' % (db, month, day, int(time.time()) - start_time)def run3(db): p_id = randint(139249095950, 139349095950) start_time = int(time.time()) try: system('%s %s -e "insert into t_log_item (pid) values(%s)"' % (MYSQL, db, p_id)) except: print "insert into t_log_item failed" else: print '%s insert into t_log_item (pid) values(%s) use %d sec.' % (db, p_id, int(time.time()) - start_time)def run4(db): role_level = randint(1, 15) start_time = int(time.time()) try: system('%s %s -e "delete from t_log_item where role_level=%s;commit"' % (MYSQL, db, role_level)) except: print "delte from t_log_item failed" else: print '%s delete from t_log_item where role_level=%s use %d sec.' % (db, role_level, int(time.time()) - start_time)def run5(db): role_level = randint(16, 20) server_id = randint(1, 100) start_time = int(time.time()) try: system('%s %s -e "update t_log_item set server_id=%s where role_level=%s;commit"' % (MYSQL, db, server_id, role_level)) """cxn.commint() cxn.close()""" except: print "update t_log_item failed" else: print '%s update t_log_item set server_id=%s where role_level=%s use %d sec.' % (db, server_id, role_level, int(time.time()) - start_time)if __name__ == '__main__': threads = [] funcs = [run1, run2, run3, run4, run5] dbs = ['tx_ljxz_71', 'tx_ljxz_71', 'tx_ljxz_71','tx_ljxz_71', 'tx_ljxz_71'] for i in range(20): threads.append(threading.Thread(target=funcs[randint(0,4)], args=(dbs[randint(0,4)],))) for i in range(20): threads[i].start() for i in range(20): threads[i].join()
本文出自 “运维小卒” 博客，转载请与作者联系！