元数据锁(Metadata Lock,以下简称MDL)计划写三篇,这篇主要是介绍MDL的引入背景和基本概念,后两篇会着重介绍MySQL 5.7(包括8.0)及5.6中如何定位MDL问题。
在线上进行DDL操作时,相对于其可能带来的系统负载,其实,我们最担心的还是MDL其可能导致的阻塞问题。
一旦DDL操作因获取不到MDL被阻塞,后续其它针对该表的其它操作都会被阻塞。典型如下,如阻塞稍久的话,我们会看到Threads_running飙升,CPU告警。mysql> show processlist;
+----+-----------------+-----------+-----------+---------+------+---------------------------------+------------------------------------+
|Id|User|Host| db |Command|Time|State|Info|
+----+-----------------+-----------+-----------+---------+------+---------------------------------+------------------------------------+
|4| event_scheduler | localhost | NULL |Daemon|122|Waitingon empty queue | NULL |
|9| root | localhost | NULL |Sleep|57| | NULL |
|12| root | localhost | employees |Query|40|Waitingfortable metadatalock| alter table slowtech.t1 add c1int|
|13| root | localhost | employees |Query|35|Waitingfortable metadatalock|select*fromslowtech.t1 |
|14| root | localhost | employees |Query|30|Waitingfortable metadatalock|select*fromslowtech.t1 |
|15| root | localhost | employees |Query|19|Waitingfortable metadatalock|select*fromslowtech.t1 |
|16| root | localhost | employees |Query|10|Waitingfortable metadatalock|select*fromslowtech.t1 |
|17| root | localhost | employees |Query|0| starting | show processlist |
+----+-----------------+-----------+-----------+---------+------+---------------------------------+------------------------------------+
8rowsinset(0.00sec)
如果发生在线上,无疑会影响到业务。所以,一般建议将DDL操作放到业务低峰期做,其实有两方面的考虑,1. 避免对系统负载产生较大影响。2. 减少DDL被阻塞的概率。
MDL引入的背景
MDL是MySQL 5.5.3引入的,主要用于解决两个问题,
RR事务隔离级别下不可重复读的问题
如下所示,演示环境,MySQL 5.5.0。session1>begin;
QueryOK,0rows affected (0.00sec)
session1>select*fromt1;
+------+------+
| id | name |
+------+------+
|1| a |
|2| b |
+------+------+
2rowsinset(0.00sec)
session2> alter table t1 add c1int;
QueryOK,2rows affected (0.02sec)
Records:2Duplicates:0Warnings:0
session1>select*fromt1;
Emptyset(0.00sec)
session1> commit;
QueryOK,0rows affected (0.00sec)
session1>select*fromt1;
+------+------+------+
| id | name | c1 |
+------+------+------+
|1| a | NULL |
|2| b | NULL |
+------+------+------+
2rowsinset(0.00sec)
可以看到,虽然是RR隔离级别,但在开启事务的情况下,第二次查询却没有结果。
主从复制问题
包括主从数据不一致,主从复制中断等。
如下面的主从数据不一致。session1> create table t1(idint,name varchar(10)) engine=innodb;
QueryOK,0rows affected (0.00sec)
session1>begin;
QueryOK,0rows affected (0.00sec)
session1> insertintot1 values(1,'a');
QueryOK,1row affected (0.00sec)
session2> truncate table t1;
QueryOK,0rows affected (0.46sec)
session1> commit;
QueryOK,0rows affected (0.35sec)
session1>select*fromt1;
Emptyset(0.00sec)
再来看看从库的结果session1>select*fromslowtech.t1;
+------+------+------+
| id | name | c1 |
+------+------+------+
|1| a | NULL |
+------+------+------+
1rowinset(0.00sec)
看看binlog的内容,可以看到,truncate操作记录在前,insert操作记录在后。# at 7140
#180714 19:32:14 server id 1 end_log_pos 7261 Query thread_id=31 exec_time=0 error_code=0
SET TIMESTAMP=1531567934/*!*/;
create table t1(idint,name varchar(10)) engine=innodb
/*!*/;
# at 7261
#180714 19:32:30 server id 1 end_log_pos 7333 Query thread_id=32 exec_time=0 error_code=0
SET TIMESTAMP=1531567950/*!*/;
BEGIN
/*!*/;
# at 7333
#180714 19:32:30 server id 1 end_log_pos 7417 Query thread_id=32 exec_time=0 error_code=0
SET TIMESTAMP=1531567950/*!*/;
truncate table t1
/*!*/;
# at 7417
#180714 19:32:30 server id 1 end_log_pos 7444 Xid = 422
COMMIT/*!*/;
# at 7444
#180714 19:32:34 server id 1 end_log_pos 7516 Query thread_id=31 exec_time=0 error_code=0
SET TIMESTAMP=1531567954/*!*/;
BEGIN
/*!*/;
# at 7516
#180714 19:32:24 server id 1 end_log_pos 7611 Query thread_id=31 exec_time=0 error_code=0
SET TIMESTAMP=1531567944/*!*/;
insertintot1 values(1,'a')
/*!*/;
# at 7611
#180714 19:32:34 server id 1 end_log_pos 7638 Xid = 421
COMMIT/*!*/;
如果会话2执行的是drop table操作,还会导致主从中断。
有意思的是,如果会话2执行的是alter table操作,其依旧会被阻塞,阻塞时间受innodb_lock_wait_timeout参数限制。mysql> show processlist;
+----+------+-----------+----------+---------+------+-------------------+---------------------------+
|Id|User|Host| db |Command|Time|State|Info|
+----+------+-----------+----------+---------+------+-------------------+---------------------------+
|54| root | localhost | NULL |Query|0| NULL | show processlist |
|58| root | localhost | slowtech |Sleep|1062| | NULL |
|60| root | localhost | slowtech |Query|11| copy to tmp table | alter table t1 add c1int|
+----+------+-----------+----------+---------+------+-------------------+---------------------------+
3rowsinset(0.00sec)
MDL的基本概念
首先,看看官方的说法,
To ensure transaction serializability, the server must not permit one session to perform a data definition language (DDL) statement on a table that is used in an uncompleted explicitly or implicitly started transaction in another session.
The server achieves this by acquiring metadata locks on tables used within a transaction and deferring release of those locks until the transaction ends.
A metadata lock on a table prevents changes to the table's structure.
This locking approach has the implication that a table that is being used by a transaction within one session cannot be used in DDL statements by other sessions until the transaction ends.
从上面的描述可以看到,
1. MDL出现的初衷就是为了保护一个处于事务中的表的结构不被修改。
2. 这里提到的事务包括两类,显式事务和AC-NL-RO(auto-commit non-locking read-only)事务。显式事务包括两类:1. 关闭AutoCommit下的操作,2. 以begin或start transaction开始的操作。AC-NL-RO可理解为AutoCommit开启下的select操作。
3. MDL是事务级别的,只有在事务结束后才会释放。在此之前,其实也有类似的保护机制,只不过是语句级别的。
需要注意的是,MDL不仅仅适用于表,同样也适用于其它对象,如下表所示,其中,"等待状态"对应的是"show processlist"中的State。
为了提高数据库的并发度,MDL被细分为了11种类型。MDL_INTENTION_EXCLUSIVE
MDL_SHARED
MDL_SHARED_HIGH_PRIO
MDL_SHARED_READ
MDL_SHARED_WRITE
MDL_SHARED_WRITE_LOW_PRIO
MDL_SHARED_UPGRADABLE
MDL_SHARED_READ_ONLY
MDL_SHARED_NO_WRITE
MDL_SHARED_NO_READ_WRITE
MDL_EXCLUSIVE
常用的有MDL_SHARED_READ,MDL_SHARE D_WRITE及MDL_EXCLUSIVE,其分别用于SELECT操作,DML操作及DDL操作。其它类型的对应操作可参考源码sql/mdl.h。
对于MDL_EXCLUSIVE,官方的解释是,
An exclusive metadata lock.
A connection holding this lock can modify both table's metadata and data.
No other type of metadata lock can be granted while this lock is held.
To be used for CREATE/DROP/RENAME TABLE statements and for execution of certain phases of other DDL statements.
简而言之,MDL_EXCLUSIVE是独占锁,在其持有期间是不允许其它类型的MDL被授予,自然也包括SELECT和DML操作。
这也就是为什么DDL操作被阻塞时,后续其它操作也会被阻塞。
关于MDL的补充
1. MDL的最大等待时间由lock_wait_timeout参数决定,其默认值为31536000(365天)。在使用工具进行DDL操作时,这个值就不太合理。事实上,pt-online-schema-change和gh-ost对其就进行了相应的调整,其中,前者60s,后者3s。
2. 如果一个SQL语法上有效,但执行时报错,如,列名不存在,其同样会获取MDL锁,直到事务结束才释放。
下篇预告,《MySQL 5.7中如何定位"Waiting for table metadata lock"》
欢迎阅读公众号其它文章