Metadata Lock原理3

 

http://blog.itpub.net/26515977/viewspace-1208250/   腾讯工程师

随着5.5.3引入MDL，更多的Query被“Waiting for table metadata lock”给\'炕\'了
SHOW PROCESSLIST的输出也有之前的"Locked"变得粒度更加细的\'Waiting for table metadata lock\'
引入MDL，当需要访问、修改表结构时，都需要对元数据上锁(读/写)
MDL在Server层，保护表数据结构，而非数据本身。Server比之前版本繁忙多了

MDL意味着DDL，一旦DDL被阻塞，那么面向该表的所有Query都会被挂起，包括Select，不过5.6作了改进，5.5可通过参数控制

假如没有MDL

在5.1版本，MDL的生命随语句的结束而释放，并非在事务提交后结束，意味着事务不是真正的隔离
因为同一条Query在2次执行期间倘若被插入了DDL，那么得到的结果将会不同，下面是个例子

会话1：

mysql> select version();
+------------+
| version()  |
+------------+
| 5.1.72-log |
+------------+
1 row in set (0.00 sec)

mysql> select @@tx_isolation;
+-----------------+
| @@tx_isolation  |
+-----------------+
| REPEATABLE-READ |
+-----------------+
1 row in set (0.00 sec)

mysql> begin;
Query OK, 0 rows affected (0.00 sec)

mysql> select * from t where id=1;
+----+--------+
| id | name   |
+----+--------+
|  1 | python |
+----+--------+
1 row in set (0.04 sec)

会话2：

mysql> alter table t add column comment varchar(200) default \'I use Python\';
Query OK, 3 rows affected (0.02 sec)
Records: 3  Duplicates: 0  Warnings: 0

会话1：

mysql> select * from t where id=1;
Empty set (0.00 sec)

mysql> rollback;
Query OK, 0 rows affected (0.00 sec)

mysql> begin;
Query OK, 0 rows affected (0.00 sec)

mysql> select * from t where id=1;
+----+--------+--------------+
| id | name   | comment      |
+----+--------+--------------+
|  1 | python | I use Python |
+----+--------+--------------+
1 row in set (0.00 sec)

与上面的不同，在5.5 MDL拉长了生命长度，与事务同生共死，只要事务还在，MDL就在，由于事务持有MDL锁，任何DDL在事务期间都休息染指，下面是个例子

会话1：

mysql> select version();
+------------+
| version()  |
+------------+
| 5.5.16-log |
+------------+
1 row in set (0.01 sec)

mysql> begin;
Query OK, 0 rows affected (0.00 sec)

mysql> select * from t order by id;
+----+------+
| id | name |
+----+------+
|  1 | a    |
|  2 | e    |
|  3 | c    |
+----+------+
3 rows in set (0.00 sec)

会话2：

mysql> alter table t add column cc char(10) default \'c lang\'; <<===Hangs

会话3：

mysql> show processlist;
+----+------+-----------+------+---------+------+---------------------------------+-------------------------------------------------------+
| Id | User | Host      | db   | Command | Time | State                           | Info                                                  |
+----+------+-----------+------+---------+------+---------------------------------+-------------------------------------------------------+
|  2 | root | localhost | db1  | Sleep   |  191 |                                 | NULL                                                  |
|  3 | root | localhost | db1  | Query   |  125 | Waiting for table metadata lock | alter table t add column cc char(10) default \'c lang\' |
|  4 | root | localhost | NULL | Query   |    0 | NULL                            | show processlist                                      |
+----+------+-----------+------+---------+------+---------------------------------+-------------------------------------------------------+

回滚了会话1的事务后，我们看下alter主要被谁炕了

mysql> show profiles;
+----------+---------------+-------------------------------------------------------+
| Query_ID | Duration      | Query                                                 |
+----------+---------------+-------------------------------------------------------+
|        1 | 1263.64100500 | alter table t add column dd char(10) default \' Elang\' |
+----------+---------------+-------------------------------------------------------+
1 row in set (0.00 sec)

mysql> show profile for query 1;
+------------------------------+------------+
| Status                       | Duration   |
+------------------------------+------------+
| starting                     |   0.000124 |
| checking permissions         |   0.000015 |
| checking permissions         |   0.000010 |
| init                         |   0.000023 |
| Opening tables               |   0.000063 |
| System lock                  |   0.000068 |
| setup                        |   0.000082 |
| creating table               |   0.034159 |
| After create                 |   0.000185 |
| copy to tmp table            |   0.000309 |
| rename result table          | 999.999999 |
| end                          |   0.004457 |
| Waiting for query cache lock |   0.000024 |
| end                          |   0.000029 |
| query end                    |   0.000009 |
| closing tables               |   0.000030 |
| freeing items                |   0.000518 |
| cleaning up                  |   0.000015 |
+------------------------------+------------+
18 rows in set (0.00 sec)

显然，在"rename result table"处发生了等待，在我们alter一张大表的时候，真正触礁的地方也就这里，这也呼应下面的案例1

案例

案例 1：
5.5下给大表加主键时会锁住读的问题

加锁策略：

1）上MDL读锁
2）操作数据，最耗时，需copy data，简易流程如下：
       a) 创建临时表A，重定义A为修改后的表结构
       b) 从原表读取数据插入到A表
3）将MDL读锁升级为写锁
       c) 删除原表，将A重命名为原表名
4）释放MDL写锁

5.5必须在打开了old_alter_table后才会采取上述策略，但事实上，无论其是否打开
对主键操作都必须copy data，基于此5.6已经改进，从而把MDL写锁仅圈在重命名的操作
这样阻塞时间非常之短，几乎可以认为不阻塞读，在5.5时做alter建议将其打开，减小持锁时间

案例 2：
mysqldump备份不一致导致备库延迟问题

由于5.1中没有引入MDL，所有在mysqldump备份过程中，并发DDL都会对其产生影响，导致备份集不一致
最终表现是使用此备份集恢复的备库在relay主库binlog会出现slave error，造成在备库上的读与主库数据不一致
因为5.5有了MDL，所以–single-transaction时，事务内操作过的表都会持有MDL，因此不会被DDL破坏
例如，mysqldump已经备份了a，b，c表，因为它们在事务内，事务还没提交，它们的MDL不会释放
因此另外的线程如果做a,b,c中任意一张表的DDL操作，都会出现Waiting for table metadata lock，而还没备份到的表不会持有MDL，因此还可以做DDL

监控

lock_wait_timeout

mysql> show variables like \'lock_wait_timeout\';
+-------------------+----------+
| Variable_name     | Value    |
+-------------------+----------+
| lock_wait_timeout | 31536000 |
+-------------------+----------+
1 row in set (0.00 sec)

官方文档解释：

This variable specifies the timeout in seconds for attempts to acquire metadata locks.
The permissible values range from 1 to 31536000 (1 year). The default is 31536000

5.5默认是1年！太长了，个人认为不是很合理，这个需要我们DBA自己调整，50秒会比较好
另外，这个既然针对MDL的超时设置，那为何不命名为"metadata_lock_wait_timeout"呢？

show open tables：当前打开的表，与flush tables结合，能判断表的活性

show status like \'Open%tables\';是上面的简化输出

诊断

对于MyISAM表：

当发现Query由于“Waiting for table metadata lock”而挂起等待MyISAM表，这是因为该表已经被包含在一个未提交的事务，无论其是否有意
而且常见的锁追踪技巧此时也没有作用，比如InnoDB Lock Monitor, InnoDB Status, mysqladmin debug output, INFORMATION_SCHEMA等
事实上，在select一张表之前先SET AUTOCOMMIT=0，那么之后，直到该事务提交或结束后其它DDL才能够得到该表的控制权

测试：

Connection #1:

create table t1 (id int) engine=myisam;
set @@autocommit=0;
select * from t1;
Connection #2:

alter table t1 rename to t2; <-- Hangs

对于InnoDB表：

倘若在InnoDB表遇见此问题，我们能够用SHOW ENGINE INNODB STATUS，优先查看transactions部分

测试：

create table t3 (id int) engine=innodb;
create table t4 (id int) engine=innodb;
delimiter |
CREATE TRIGGER t3_trigger AFTER INSERT ON t3
  FOR EACH ROW BEGIN
    INSERT INTO t4 SET id = NEW.id;
  END;
|
delimiter ;

Connection #1:

begin;
insert into t3 values (1);

Connection #2:

drop trigger if exists t3_trigger; <-- Hangs

mysql> SHOW ENGINE INNODB STATUS\G;
....
....
....
------------
TRANSACTIONS
------------
Trx id counter BF03
Purge done for trx\'s n:o < BD03 undo n:o < 0
History list length 82
LIST OF TRANSACTIONS FOR EACH SESSION:
---TRANSACTION 0, not started
MySQL thread id 4, OS thread handle 0xa7d3fb90, query id 40 localhost root
show engine innodb status
---TRANSACTION BF02, ACTIVE 38 sec
2 lock struct(s), heap size 320, 0 row lock(s), undo log entries 2
MySQL thread id 2, OS thread handle 0xa7da1b90, query id 37 localhost root

...
...
...

官方文档对 TRANSACTIONS解释：

TRANSACTIONS

If this section reports lock waits, your applications might have lock contention.
The output can also help to trace the reasons for transaction deadlocks.

除此之外，我们还可以利用information_schema，以下是几个有用SQL：
① 锁等待

SELECT * FROM INNODB_LOCK_WAITS

② 被阻塞的事务

SELECT *
FROM INNODB_LOCKS
WHERE LOCK_TRX_ID IN (SELECT BLOCKING_TRX_ID FROM INNODB_LOCK_WAITS)

或者

SELECT INNODB_LOCKS.*
FROM INNODB_LOCKS
JOIN INNODB_LOCK_WAITS
  ON (INNODB_LOCKS.LOCK_TRX_ID = INNODB_LOCK_WAITS.BLOCKING_TRX_ID)

③ 指定表上的锁

SELECT * FROM INNODB_LOCKS
WHERE LOCK_TABLE = db_name.table_name

④ 事务与锁

SELECT TRX_ID, TRX_REQUESTED_LOCK_ID, TRX_MYSQL_THREAD_ID, TRX_QUERY
FROM INNODB_TRX
WHERE TRX_STATE = \'LOCK WAIT\'

当然，最好的办法还是综合运用前面2种

与table cache的关系

先明白table cache所解决的问题域：fd(文件描述符)打开/关闭太过频繁导致资源消耗
那么它是如何解决的？tc通过cache所有打开的fd，当有新的连接请求时不需重新打开，结束后也不用关闭

DDL操作终究会被阻塞，即使table cache含有其所需的fd。MySQL认为，tc旧条目必须失效
访问该表的DDL操作必须重新打开fd，下面是个测试

会话1：

mysql> show status like \'Open%tables\';
+---------------+-------+
| Variable_name | Value |
+---------------+-------+
| Open_tables   | 26    |  <==当前打开的表数量
| Opened_tables | 2     |  <==已经打开的表数量
+---------------+-------+
2 rows in set (0.00 sec)

会话2：

mysql> alter table t add column Oxx char(20) default \'ORACLE\';
Query OK, 3 rows affected (0.05 sec)
Records: 3  Duplicates: 0  Warnings: 0

会话1：

mysql> select * from t order by id;
+----+------+--------+--------+---------+---------+-------+--------+--------+--------+--------+
| id | name | cc     | dd     | EE      | ff      | OO    | OE     | OF     | OX     | Oxx    |
+----+------+--------+--------+---------+---------+-------+--------+--------+--------+--------+
|  1 | a    | c lang |  Elang |  Golang |  Golang | MySQL | ORACLE | ORACLE | ORACLE | ORACLE |
|  2 | e    | c lang |  Elang |  Golang |  Golang | MySQL | ORACLE | ORACLE | ORACLE | ORACLE |
|  3 | c    | c lang |  Elang |  Golang |  Golang | MySQL | ORACLE | ORACLE | ORACLE | ORACLE |
+----+------+--------+--------+---------+---------+-------+--------+--------+--------+--------+
3 rows in set (0.00 sec)

mysql> show status like \'Open%tables\';
+---------------+-------+
| Variable_name | Value |
+---------------+-------+
| Open_tables   | 27    |
| Opened_tables | 3     |
+---------------+-------+
2 rows in set (0.00 sec)

会话2：

mysql> alter table t add column Oxf char(20) default \'ORACLE\';
Query OK, 3 rows affected (0.06 sec)
Records: 3  Duplicates: 0  Warnings: 0

会话1：

mysql> show status like \'Open%tables\';
+---------------+-------+
| Variable_name | Value |
+---------------+-------+
| Open_tables   | 26    |
| Opened_tables | 3     |
+---------------+-------+
2 rows in set (0.00 sec)

结论：

当需要对"热表"做DDL，需要特别谨慎，否则，容易造成MDL等待，导致连接耗尽或者拖垮Server

参考资料：

http://dinglin.iteye.com/blog/1884696
http://www.xaprb.com/blog/2012/08/28/debugging-metadata-locking-in-mysql-5-5/
http://blog.chinaunix.net/uid-28212952-id-3400571.html
http://www.chriscalender.com/?p=1189
http://www.mysqlperformanceblog.com/2013/02/01/implications-of-metadata-locking-changes-in-mysql-5-5/

By 迦叶

2013-12-17

Good Luck