源起

线上执行ALTER时,通过show processlist查看到出现Waiting for table metadata lock ,导致后面的查询都无法执行。

5217122 | create_table_04 | 172.100.207.148:31291 | finance | Query   | 1829 | Waiting for table metadata lock  | ALTER TABLE `Pay` MODIFY COLUMN `pay`  smallint(6) NULL DEFAULT 0 COMMENT '付??'
5217155 | bx_live_dml     | 172.100.210.4:34730   | finance | Prepare | 1714 | Waiting for table metadata lock  | select * from Pay where (pay =4 or pay =8) and projectId ='CSZY0'

由于当时直接把ALTER给kill了,所以基本没有现场,所以下面模拟什么情况MySQL产生这种事故。

MySQL元数据锁

MySQL DBA对于Waiting for table metadata lock肯定不会陌生,一般都是进行alter操作时被堵住了,导致了我们在show processlist 时,看到线程的状态是在等metadata lock。

为了在并发环境下维护表元数据的数据一致性,在表上有活动事务(显式或隐式)的时候,不可以对元数据进行写入操作。因此从MySQL5.5版本开始引入了MDL锁(metadata lock),来保护表的元数据信息,用于解决或者保证DDL操作与DML操作之间的一致性。对于引入MDL,其主要解决了2个问题,一个是事务隔离问题,比如在可重复隔离级别下,会话A在2次查询期间,会话B对表结构做了修改,两次查询结果就会不一致,无法满足可重复读的要求;另外一个是数据复制的问题,比如会话A执行了多条更新语句期间,另外一个会话B做了表结构变更并且先提交,就会导致slave在重做时,先重做alter,再重做update时就会出现复制错误的现象。

所以在对表进行上述操作时,如果表上有活动事务(未提交或回滚),请求写入的会话会等待在Metadata lock wait 。例如下面的这种情形:

若没有MDL锁的保护,则事务2可以直接执行DDL操作,并且导致事务1出错,5.1版本即是如此。5.5版本加入MDL锁就在于保护这种情况的发生,由于事务1开启了查询,那么获得了MDL锁,锁的模式为SHARED_READ,事务2要执行DDL,则需获得EXCLUSIVE锁,两者互斥,所以事务2需要等待。

注:支持事务的InnoDB引擎表和不支持事务的MyISAM引擎表,都会出现Metadata Lock Wait等待现象。一旦出现Metadata Lock Wait等待现象,后续所有对该表的访问都会阻塞在该等待上,导致连接堆积,业务受影响。

引起Metadata lock wait出现的场景:Waiting for table metadata lock

场景一:当前有执行DML操作时执行ALTRE操作。

# SESSION A
mysql> insert into sbtest2 select * from sbtest1;

# SESSION B
mysql> alter table sbtest2 add test1 int;   //等待SESSION A执行完;

# SESSION C
mysql> show processlist;
+-----+------+-----------+--------+---------+------+---------------------------------+-------------------------------------------+
| Id  | User | Host      | db     | Command | Time | State                           | Info                                      |
+-----+------+-----------+--------+---------+------+---------------------------------+-------------------------------------------+
| 267 | root | localhost | sbtest | Query   |    7 | Sending data                    | insert into sbtest2 select * from sbtest1 |
| 271 | root | localhost | sbtest | Query   |    3 | Waiting for table metadata lock | alter table sbtest2 add test1 int         |
| 272 | root | localhost | NULL   | Query   |    0 | starting                        | show processlist                          |
+-----+------+-----------+--------+---------+------+---------------------------------+-------------------------------------------+
3 rows in set (0.00 sec)

# SESSION D
mysql> select * from sbtest2 limit 10;    //等待元数据锁;

# SESSION E
mysql> show processlist;
+-----+------+-----------+--------+---------+------+---------------------------------+-------------------------------------------+
| Id  | User | Host      | db     | Command | Time | State                           | Info                                      |
+-----+------+-----------+--------+---------+------+---------------------------------+-------------------------------------------+
| 267 | root | localhost | sbtest | Query   |   20 | Sending data                    | insert into sbtest2 select * from sbtest1 |
| 271 | root | localhost | sbtest | Query   |   13 | Waiting for table metadata lock | alter table sbtest2 add test1 int         |
| 272 | root | localhost | NULL   | Query   |    0 | starting                        | show processlist                          |
| 308 | root | localhost | sbtest | Query   |    3 | Waiting for table metadata lock | select * from sbtest2 limit 10            |
+-----+------+-----------+--------+---------+------+---------------------------------+-------------------------------------------+
4 rows in set (0.00 sec)

从上述例子可以看出,我们在执行DDL语句的时候得事先看一下,进程中是否已经存在某些DML语句占用了表的元数据锁,这样会导致DDL语句处于锁等待状态。一旦出现Waiting for table metadata lock等待现象,后续所有对该表的访问都会阻塞在该等待上,包括读操作,导致连接堆积,业务受影响。

场景二:当前有对表的长时间查询或使用mysqldump/mysqlpump时,使用alter会被堵住。

# SESSION A
mysql> select *,sleep(10) from sbtest2;

# SESSION B
mysql> alter table sbtest2 add test2 int;   //等待SESSION A执行完;

# SESSION C
mysql> show processlist;
+-----+------+-----------+--------+---------+------+---------------------------------+---------------------------------------+
| Id  | User | Host      | db     | Command | Time | State                           | Info                                  |
+-----+------+-----------+--------+---------+------+---------------------------------+---------------------------------------+
| 267 | root | localhost | sbtest | Query   |   12 | User sleep                      | select *,sleep(10) from sbtest2       |
| 271 | root | localhost | sbtest | Query   |    8 | Waiting for table metadata lock | alter table sbtest2 add test3 int     |
| 272 | root | localhost | NULL   | Query   |    0 | starting                        | show processlist                      |
| 311 | root | localhost | NULL   | Query   |    3 | Waiting for table metadata lock | select * from sbtest.sbtest2 limit 10 |
+-----+------+-----------+--------+---------+------+---------------------------------+---------------------------------------+
4 rows in set (0.00 sec)

场景三:显示或者隐式开启事务后未提交或回滚,比如查询完成后未提交或者回滚,使用alter会被堵住。

# SESSION A
mysql> begin;
mysql> select * from sbtest2;

# SESSION B
mysql> alter table sbtest2 add test2 int;   //等待SESSION A执行完;

# SESSION C
mysql> show processlist;
+-----+------+-----------+--------+---------+------+---------------------------------+-----------------------------------+
| Id  | User | Host      | db     | Command | Time | State                           | Info                              |
+-----+------+-----------+--------+---------+------+---------------------------------+-----------------------------------+
| 267 | root | localhost | sbtest | Sleep   |   36 |                                 | NULL                              |
| 271 | root | localhost | sbtest | Query   |   30 | Waiting for table metadata lock | alter table sbtest2 add test2 int |
| 272 | root | localhost | NULL   | Query   |    0 | starting                        | show processlist                  |
+-----+------+-----------+--------+---------+------+---------------------------------+-----------------------------------+
3 rows in set (0.00 sec)

场景四:表上有失败的查询事务,比如查询不存在的列,语句失败返回,但是事务没有提交,此时alter仍然会被堵住。

# SESSION A
mysql> begin;
mysql> select error from sbtest2; 
ERROR 1054 (42S22): Unknown column 'error' in 'field list'

# SESSION B
mysql> alter table sbtest2 add test3 int;    //等待SESSION A提交或回滚;

# SESSION C
mysql> show processlist;
+-----+------+-----------+--------+---------+------+---------------------------------+-----------------------------------+
| Id  | User | Host      | db     | Command | Time | State                           | Info                              |
+-----+------+-----------+--------+---------+------+---------------------------------+-----------------------------------+
| 267 | root | localhost | sbtest | Sleep   |    7 |                                 | NULL                              |
| 271 | root | localhost | sbtest | Query   |    3 | Waiting for table metadata lock | alter table sbtest2 add test3 int |
| 272 | root | localhost | NULL   | Query   |    0 | starting                        | show processlist                  |
| 311 | root | localhost | NULL   | Sleep   |  413 |                                 | NULL                              |
+-----+------+-----------+--------+---------+------+---------------------------------+-----------------------------------+
4 rows in set (0.00 sec)

# SESSION D
mysql> select * from information_schema.innodb_trx;
Empty set (0.00 sec)

其实SESSION A中的事务并未开启,但是由于select获取表元数据的语句,语法上是有效的,虽然执行失败了,但是任然不会释放元数据锁,故而导致SESSION B的alter动作被阻塞。

通过SESSION D查看当前打开事务时,你会发现没有,从而找不到原因。所以当出现这种场景时,如何判断是哪个进程导致的呢,我们可以尝试查看表performance_schema. events_statements_current,分析进程状态来进行判断。

mysql> select * from performance_schema. events_statements_current\G
*************************** 1. row ***************************
              THREAD_ID: 293
               EVENT_ID: 32
           END_EVENT_ID: 32
             EVENT_NAME: statement/sql/select
                 SOURCE: socket_connection.cc:101
            TIMER_START: 212721717099954000
              TIMER_END: 212721717213807000
             TIMER_WAIT: 113853000
              LOCK_TIME: 0
               SQL_TEXT: select error from sbtest2
                 DIGEST: 0bbb2d5d1be45e77debea68111264885
            DIGEST_TEXT: SELECT ERROR FROM `sbtest2` 
         CURRENT_SCHEMA: sbtest
            OBJECT_TYPE: NULL
          OBJECT_SCHEMA: NULL
            OBJECT_NAME: NULL
  OBJECT_INSTANCE_BEGIN: NULL
            MYSQL_ERRNO: 1054
      RETURNED_SQLSTATE: 42S22
           MESSAGE_TEXT: Unknown column 'error' in 'field list'
                 ERRORS: 1

然后找到其sid, kill掉该session,也可以kill掉DDL所在的session解决可以解决此问题。

另外,测试时SESSION A要显式开启一个事务,否则查询会隐式回滚结束,无法重现上面的场景。SESSION B执行alter后,没有立即阻塞住,而是立马开始copy to tmp table,这个过程结束后,才进行了MDL锁等待。这怎么解释呢,应该是执行alter操作主要分为创建临时新表->插入老表的数据->临时新表rename to老表三个步骤,在这种情况下,到最后一步才需要MDL锁,所以copy过程中不会阻塞。由于没有查询在进行,而且查询也没有进入innodb层 (失败返回),所以show processlist和information_schema.innodb_trx没有可以参考的信息。

出现以上几种情况时,这个时候如果进行如下操作就会引起MDL:

  • 创建、删除索引。
  • 修改表结构。
  • 表维护操作(optimize table、repair table等)。
  • 删除表。
  • 获取表上表级写锁 (lock table tab_name write)。

使用Profile分析场景三:显示或者隐式开启事务后未提交或回滚,比如查询完成后未提交或者回滚,使用alter会被堵住

# SESSION A
mysql> set profiling=on;
mysql> begin;
mysql> select * from sbtest.sbtest2 limit 1;

# SESSION B
mysql> set profiling=on;
mysql> alter table sbtest.sbtest5 add test2 int;   //等待SESSION A执行完;

# SESSION C
mysql> set profiling=on;
mysql> select * from sbtest.sbtest2 limit 1;

# SESSION D
mysql> set profiling=on;
mysql> show processlist;
+-----+------+-----------+------+---------+------+---------------------------------+------------------------------------------+
| Id  | User | Host      | db   | Command | Time | State                           | Info                                     |
+-----+------+-----------+------+---------+------+---------------------------------+------------------------------------------+
| 325 | root | localhost | NULL | Query   |   25 | Waiting for table metadata lock | alter table sbtest.sbtest2 add test5 int |
| 326 | root | localhost | NULL | Query   |    3 | Waiting for table metadata lock | select * from sbtest.sbtest2 limit 1     |
| 327 | root | localhost | NULL | Query   |    0 | starting                        | show processlist                         |
| 328 | root | localhost | NULL | Sleep   |   50 |                                 | NULL                                     |
+-----+------+-----------+------+---------+------+---------------------------------+------------------------------------------+
4 rows in set (0.00 sec)

然后回滚SESSION A,等待SESSION B和SESSION C执行完,查看profile。

查看SESSION A

# SESSION A
mysql> show profiles;
+----------+-------------+------------------------------------------+
| Query_ID | Duration    | Query                                    |
+----------+-------------+------------------------------------------+
|        1 | 42.81646375 | alter table sbtest.sbtest2 add test5 int |
+----------+-------------+------------------------------------------+
1 row in set, 1 warning (0.00 sec)

mysql> show profile for query 1;
+--------------------------------+----------+
| Status                         | Duration |
+--------------------------------+----------+
| starting                       | 0.000069 |
| checking permissions           | 0.000004 |
| checking permissions           | 0.000004 |
| init                           | 0.000005 |
| Opening tables                 | 0.000128 |
| setup                          | 0.000032 |
| creating table                 | 0.000654 |
| After create                   | 0.000053 |
| Waiting for table metadata loc | 1.000069 |
.....................
| After create                   | 0.000013 |
| Waiting for table metadata loc | 0.871435 |
| After create                   | 0.000042 |
| System lock                    | 0.000013 |
| preparing for alter table      | 0.002475 |
| altering table                 | 9.752928 |
| committing alter table to stor | 0.185624 |
| end                            | 0.000021 |
| query end                      | 0.000010 |
| closing tables                 | 0.000007 |
| freeing items                  | 0.000020 |
| cleaning up                    | 0.000011 |
+--------------------------------+----------+
83 rows in set, 1 warning (0.00 sec)

查看SESSION C

mysql> show profiles;
+----------+-------------+--------------------------------------+
| Query_ID | Duration    | Query                                |
+----------+-------------+--------------------------------------+
|        1 | 10.75216050 | select * from sbtest.sbtest2 limit 1 |
+----------+-------------+--------------------------------------+
1 row in set, 1 warning (0.00 sec)

mysql> show profile for query 1;
+--------------------------------+-----------+
| Status                         | Duration  |
+--------------------------------+-----------+
| starting                       |  0.000080 |
| checking permissions           |  0.000007 |
| Opening tables                 |  0.000012 |
| Waiting for table metadata loc | 10.751829 |
| Opening tables                 |  0.000094 |
| init                           |  0.000019 |
| System lock                    |  0.000010 |
| optimizing                     |  0.000004 |
| statistics                     |  0.000011 |
| preparing                      |  0.000009 |
| executing                      |  0.000003 |
| Sending data                   |  0.000040 |
| end                            |  0.000006 |
| query end                      |  0.000008 |
| closing tables                 |  0.000008 |
| freeing items                  |  0.000014 |
| cleaning up                    |  0.000009 |
+--------------------------------+-----------+
17 rows in set, 1 warning (0.00 sec)

从上述测试可以看出,SESSION C需要打开表时碰到了元数据锁。MySQL不论SESSION A执行的是select还是delete,此时alter table语句无法获取到metadata独占锁,会进行等待;所以会影响SESSION C的读取。

这是最基本的一种情形,这个和MySQL 5.6中的online ddl并不冲突。一般alter table的操作过程中,在after create步骤会获取metadata独占锁,当进行到altering table的过程时(通常是最花时间的步骤),对该表的读写都可以正常进行,这就是online ddl的表现,并不会像之前在整个alter table过程中阻塞写入,当然并不是所有ALTER语句都支持online ddl。

总之,alter table的语句是很危险的(其实他的危险其实是未提交事物或者长事务导致的),在操作之前最好确认对要操作的表没有任何进行中的操作、没有未提交事务、也没有显式事务中的报错语句。如果有alter table的维护任务,在无人监管的时候运行,最好通过lock_wait_timeout设置好超时时间,避免长时间的metedata锁等待。

原文来自微信公众号:运维那点事

via。http://www.yunweipai.com/archives/19818.html