MySQL从5.5版本开始,新增元数据锁,也就是metadata lock,简称MDL锁。
MDL锁的主要目是保护表元数据并行请求过程中不被修改,例如会话1在执行SELECT查询数据,会话2要新增一列,这时第二个会话会被阻塞,以免造成会话1前后看到的表结构不一致。
当然了,MDL后来被扩展了,并不只是针对表(TABLE)对象,也包含库(SCHEMA)、表空间(TABLESPACE)、存储程序(过程、函数、触发器、定时器)等对象,也受到MDL的保护。此外,MDL锁也会阻塞全局 COMMIT 行为,比如加了FTWRL(FLUSH TABLE WITH READ LOCK)后,就会持有一个全局的 COMMIT 共享锁,此时会阻止其他会话的事务提交。
我们从 performance_schema.metadata_lock 就可以看到MDL的详细情况,从MySQL 5.7版本开始,还可以从 sys.schema_table_lock_waits 查看MDL阻塞等待的情况。要特别注意的是,MDL锁等待超时阈值由选项 lock_wait_timeout 控制,该选项默认值是 31536000秒,也就是 一年、一年、一年(重要的话重复三遍),建议调低,比如改成5-10分钟,建议最长不超过1小时(想想,这种MDL等待超过1小时还不报警的话,DBA也该下岗了吧)。
另外,想要在PFS(performance_schema)和 sys schema中能看到MDL详情的话,需要先启用相关的设置:(横屏观看)
[root@yejr.me]> use performance_schema; [root@yejr.me]> UPDATE setup_consumers SET ENABLED = 'YES' WHERE NAME ='global_instrumentation'; [root@yejr.me]> UPDATE setup_instruments SET ENABLED = 'YES' WHERE NAME =‘wait/lock/metadata/sql/mdl';
然后就能查看当前是否有MDL锁了:(横屏观看)
# session1执行一个未结束的只读事务
[root@yejr.me]> begin; select * from test.t1 limit 1;# session2 查看MDL详情
[root@yejr.me]> SELECT * FROM metadata_locks\G
1. row **
OBJECT_TYPE: TABLE
OBJECT_SCHEMA: test
OBJECT_NAME: t1
COLUMN_NAME: NULL
OBJECT_INSTANCE_BEGIN: 140334142005184
LOCK_TYPE: SHARED_READ -- 该MDL请求类型为共享读
LOCK_DURATION: TRANSACTION -- 该MDL请求处于事务级别
LOCK_STATUS: GRANTED -- 该MDL请求已获得
SOURCE: sql_parse.cc:5929
OWNER_THREAD_ID: 1892 -- 线程ID
OWNER_EVENT_ID: 13
已有各路大神对MDL做了详细解析,所以本文不再赘述,我更喜欢写点接地气的。所以我们共同分析一个非常经典的MDL等待案例,借此分析遇到这种问题时,该怎么分析和判断,以及如何应对处理。(横屏观看)
session1 | session2 | session3 |
begin; select from t1 limit 1; |
||
alter table t1 add cx int; # 此时该请求被阻塞 |
||
select from t1 limit 1; # 此时该请求被阻塞 |
这时执行 SHOW PROCESSLIST
的话,能看到下面这样的现场(删除了部分无用输出列)(横屏观看)
+------+------+---------------------------------+--------------------------+
| Id | Time | State | Info |
+------+------+---------------------------------+--------------------------+
| 1853 | 0 | starting | show processlist |
| 1854 | 134 | Waiting for table metadata lock | alter table t1 add cx int|
| 1855 | 83 | Waiting for table metadata lock | select * from t1 limit 1 |
+------+------+---------------------------------+--------------------------+
如果只看现场,我们是没办法分析出到底哪个线程导致的MDL阻塞,这正是MySQL 5.7之前版本的痛苦之处,遇到MDL等待,只能靠DBA的经验、感觉去分析,还不一定完全可靠。
但是,5.7版本之后,我们就可以用PFS和sys schema进行跟踪分析了:(横屏观看)
# 1、查看当前的MDL请求状态(删除部分信息)
[root@yejr.me]> select * from performance_schema.metadata_locks;
+-------------+-----------+---------------------+---------------+-------------+------+
| OBJECT_TYPE | OBJECT_NAM| LOCK_TYPE | LOCK_DURATION | LOCK_STATUS | TID |
+-------------+-----------+---------------------+---------------+-------------+------+
| TABLE | t1 | SHARED_READ | TRANSACTION | GRANTED | 1892 |
| GLOBAL | NULL | INTENTION_EXCLUSIVE | STATEMENT | GRANTED | 1893 |
| SCHEMA | NULL | INTENTION_EXCLUSIVE | TRANSACTION | GRANTED | 1893 |
| TABLE | t1 | SHARED_UPGRADABLE | TRANSACTION | GRANTED | 1893 |
| BACKUP LOCK | NULL | INTENTION_EXCLUSIVE | TRANSACTION | GRANTED | 1893 |
| TABLESPACE | test/t1 | INTENTION_EXCLUSIVE | TRANSACTION | GRANTED | 1893 |
| TABLE | #sql-e953_| EXCLUSIVE | STATEMENT | GRANTED | 1893 |
| TABLE | t1 | EXCLUSIVE | TRANSACTION | PENDING | 1893 |
| TABLE | t1 | SHARED_READ | TRANSACTION | PENDING | 1894 |
+-------------+-----------+---------------------+---------------+-------------+------+
# 2、查看当前的MDL等待状态
[root@yejr.me]> select * from sys.schema_table_lock_waits\G
1. row **
object_schema: test
object_name: t1
waiting_thread_id: 1893 -- 等待的线程ID
waiting_pid: 1854 -- 等待的连接PID
waiting_account: root@localhost
waiting_lock_type: EXCLUSIVE -- 等待的锁类型
waiting_lock_duration: TRANSACTION
waiting_query: alter table t1 drop cx
waiting_query_secs: 134 -- 锁等待时长
waiting_query_rows_affected: 0
waiting_query_rows_examined: 0
blocking_thread_id: 1892 -- 这里的线程ID值对应上面的TID
blocking_pid: 1853 -- 连接PID,对应下方的trx_mysql_thread_id
blocking_account: root@localhost
blocking_lock_type: SHARED_READ -- 阻塞的锁类型
blocking_lock_duration: TRANSACTION
sql_kill_blocking_query: KILL QUERY 1853
sql_kill_blocking_connection: KILL 1853
2. row **
object_schema: test
object_name: t1
waiting_thread_id: 1894
waiting_pid: 1855
waiting_account: root@localhost
waiting_lock_type: SHARED_READ
waiting_lock_duration: TRANSACTION
waiting_query: select * from t1 limit 1
waiting_query_secs: 83
waiting_query_rows_affected: 0
waiting_query_rows_examined: 0
blocking_thread_id: 1892
blocking_pid: 1853
blocking_account: root@localhost
blocking_lock_type: SHARED_READ
blocking_lock_duration: TRANSACTION
sql_kill_blocking_query: KILL QUERY 1853
sql_kill_blocking_connection: KILL 1853
3. row **
object_schema: test
object_name: t1
waiting_thread_id: 1893
waiting_pid: 1854
waiting_account: root@localhost
waiting_lock_type: EXCLUSIVE
waiting_lock_duration: TRANSACTION
waiting_query: alter table t1 drop cx
waiting_query_secs: 83
waiting_query_rows_affected: 0
waiting_query_rows_examined: 0
blocking_thread_id: 1893
blocking_pid: 1854
blocking_account: root@localhost
blocking_lock_type: SHARED_UPGRADABLE
blocking_lock_duration: TRANSACTION
sql_kill_blocking_query: KILL QUERY 1854
sql_kill_blocking_connection: KILL 1854
4. row **
object_schema: test
object_name: t1
waiting_thread_id: 1894
waiting_pid: 1855
waiting_account: root@localhost
waiting_lock_type: SHARED_READ
waiting_lock_duration: TRANSACTION
waiting_query: select * from t1 limit 1
waiting_query_secs: 83
waiting_query_rows_affected: 0
waiting_query_rows_examined: 0
blocking_thread_id: 1893
blocking_pid: 1854
blocking_account: root@localhost
blocking_lock_type: SHARED_UPGRADABLE
blocking_lock_duration: TRANSACTION
sql_kill_blocking_query: KILL QUERY 1854
sql_kill_blocking_connection: KILL 1854
# 3、顺便查看当前是否有活跃未提交的事务,通常持续时间最久的事务也是引起大面积MDL锁阻塞的根源
[root@yejr.me]> select trx_id,trx_state,time_to_sec(timediff(now(), trx_started))
as trx_active_secs, trx_mysql_thread_id as pid, trx_query
from information_schema.innodb_trx
order by trx_active_secs desc \G
1. row **
trx_id: 281479749621520
trx_state: RUNNING
trx_active_secs: 226
pid: 1853
trx_query: select trx_id,trx_state...
上述测试的环境是:
[root@yejr.me]>
...
Server version: 8.0.16 MySQL Community Server - GPL
...
问题分析
从上面的结果可分析得到以下结论:
- 连接PID=1854的那个会话,被连接PID=1853的会话阻塞了
- 连接PID=1855的那个会话,被连接PID=1854的会话阻塞了
- 连接PID=1855被阻塞的源头也可以追溯到PID=1853那个会话
- 也就是:session1阻塞session2,然后 session2阻塞session3
问题解决方法:
- 想要让session2和3都不被阻塞,只需要让session1上持有的锁释放即可
- 虽然上面提示可以执行
KILL QUERY 1853
,但实际上是不管用的,因为PID=1853中导致MDL等待的SQL已经执行结束,只是事务还没提交而已,因此正确的方法是执行KILL 1853
彻底杀掉该连接,回滚相应的事务,释放MDL锁
最后多说一下,MDL是在MySQL server层的锁,而InnoDB层也有表级别上的IS/IX锁,二者并不是一回事。Enjoy MySQL :)
</div>