In the scenario below, simple DML is performed on an InnoDB table and the transaction is committed. Occasionally there is a table metadata lock held by a system thread after that.

Note that the table is created with InnoDB persistent statistics turned off.

The behavior apparently changed after this commit in 10.5.1

commit ea37b144094a0c2ebfc6774047fd473c1b2a8658

Author: Marko Mäkelä

Date:   Tue Dec 10 15:42:50 2019 +0200

    MDEV-16678 Prefer MDL to dict_sys.latch for innodb background tasks

The lock can be long enough to cause, for example, ER_LOCK_WAIT_TIMEOUT on non-waiting operations, even though from the user's point of view it is a single-threaded action. Possibly in more complicated concurrent scenarios unexpected deadlocks or at least delays (causing performance issues) can also occur.

The test cases are highly non-deterministic and need to be run with --repeat=N.

The test case with metadata lock info needs a result file to be recorded (with the empty result set upon SELECT from I_S), so that --repeat=N fails when a lock is found. It fails reasonably quickly for me on 10.5 (within ~5 tries). It takes much longer on 10.6, several hundred attempts, even longer on 10.11, although I got it a couple times there, and I've never seen it on 11.4+ so far (although given the probability trend, it doesn't mean much).

The test case with non-waiting drop table doesn't need a result file, but it fails more rarely.

--source include/have_innodb.inc

install soname 'metadata_lock_info';

create table t (a int) engine=InnoDB stats_persistent=0;

start transaction;

insert into t set a = 1;

select * from t;

commit;

select thread_id, lock_mode, lock_type, table_name from information_schema.metadata_lock_info;

drop table t;

uninstall soname 'metadata_lock_info';

 commit;

 select thread_id, lock_mode, lock_type, table_name from information_schema.metadata_lock_info;

 thread_id	lock_mode	lock_type	table_name

+0	MDL_SHARED	Table metadata lock	t

 drop table t;

 uninstall soname 'metadata_lock_info';

Result length mismatch

--source include/have_innodb.inc

create table t (a int) engine=InnoDB stats_persistent=0;

set lock_wait_timeout=0, innodb_lock_wait_timeout=0;

start transaction;

insert into t set a = 1;

select * from t;

commit;

drop table t;

bug.t1                                   [ 20 fail ]

        Test ended at 2024-12-04 14:29:07

CURRENT_TEST: bug.t1

mysqltest: At line 11: query 'drop table t' failed: 1205: Lock wait timeout exceeded; try restarting transaction