问题
Recently I encountered a deadlock when deleting records (Note that the isolation level is REPEATABLE READ, MySQL 5.7)
Here is the repro steps
1 Create a new table
CREATE TABLE `t` (
`id` bigint(20) NOT NULL AUTO_INCREMENT,
`name` varchar(32) NOT NULL,
PRIMARY KEY (`id`),
KEY `p_name` (`name`)
) ENGINE=InnoDB CHARSET=utf8;
2 Prepare 3 records
insert into t (name) value ('A'), ('C'), ('D');
3
+====================================+============================================================+
| Session A | Session B |
+====================================+============================================================+
| begin; | |
+------------------------------------+------------------------------------------------------------+
| | begin; |
+------------------------------------+------------------------------------------------------------+
| delete from t where name = 'C'; | |
+------------------------------------+------------------------------------------------------------+
| | delete from t where name = 'C'; --Blocked! |
+------------------------------------+------------------------------------------------------------+
| insert into t (name) values ('B'); | |
+------------------------------------+------------------------------------------------------------+
| | ERROR 1213 (40001): Deadlock found when trying to get lock |
+------------------------------------+------------------------------------------------------------+
The result of show engine innodb status as below shown (LATEST DETECTED DEADLOCK section)
LATEST DETECTED DEADLOCK
------------------------
*** (1) TRANSACTION:
TRANSACTION 3631, ACTIVE 21 sec starting index read
mysql tables in use 1, locked 1
LOCK WAIT 2 lock struct(s), heap size 1136, 1 row lock(s)
MySQL thread id 13, OS thread handle 123145439432704, query id 306 localhost root updating
delete from t where name = 'C'
*** (1) WAITING FOR THIS LOCK TO BE GRANTED:
RECORD LOCKS space id 69 page no 4 n bits 72 index p_name of table `jacky`.`t` trx id 3631 lock_mode X waiting
Record lock, heap no 4 PHYSICAL RECORD: n_fields 2; compact format; info bits 32
0: len 1; hex 43; asc C;;
1: len 8; hex 8000000000000018; asc ;;
*** (2) TRANSACTION:
TRANSACTION 3630, ACTIVE 29 sec inserting
mysql tables in use 1, locked 1
5 lock struct(s), heap size 1136, 4 row lock(s), undo log entries 2
MySQL thread id 14, OS thread handle 123145439711232, query id 307 localhost root update
insert into t (name) values ('B')
*** (2) HOLDS THE LOCK(S):
RECORD LOCKS space id 69 page no 4 n bits 72 index p_name of table `jacky`.`t` trx id 3630 lock_mode X
Record lock, heap no 4 PHYSICAL RECORD: n_fields 2; compact format; info bits 32
0: len 1; hex 43; asc C;;
1: len 8; hex 8000000000000018; asc ;;
*** (2) WAITING FOR THIS LOCK TO BE GRANTED:
RECORD LOCKS space id 69 page no 4 n bits 72 index p_name of table `jacky`.`t` trx id 3630 lock_mode X locks gap before rec insert intention waiting
Record lock, heap no 4 PHYSICAL RECORD: n_fields 2; compact format; info bits 32
0: len 1; hex 43; asc C;;
1: len 8; hex 8000000000000018; asc ;;
As the Innodb status shown, session B is waiting next-key lock C, and session A hold a record lock C and waiting gap lock on C;
As we all know that
DELETE FROM ... WHERE ... sets an exclusive next-key lock on every record the search encounters
A next-key lock is a combination of a record lock on the index record and a gap lock on the gap before the index record.
Q1: I guess if session B firstly got the gap lock (part of next-key), and then waiting for the record lock. Thereby, the latter insert in session A was blocked by session B (due to the gap lock), and eventually result in a dead lock. Right?
Q2: As the C is purged from an index, does the gap lock hold by session B should be ('A', 'D')? If so, why the session A is waiting the insert intension lock on range (, 'C')?
Q3: Why session B has 1 row lock(s), and session A has 4 row lock(s)?
Q4: When change index p_name to a unique index, we still get the deadlock due to gap lock, it's weird. It behaves different from official doc which states only record lock is required.
DELETE FROM ... WHERE ... sets an exclusive next-key lock on every record the search encounters. However, only an index record lock is required for statements that lock rows using a unique index to search for a unique row.
However, it is okay when using primary key id to execute delete (steps as below shown). Is this a bug in MySQL?
1 Prepare data
delete from t;
insert into t (id, name) value (1, 'A'), (3, 'C'), (5, 'D');
2
+-------------------------------------------+--------------------------------------+
| Session A | Session B |
+-------------------------------------------+--------------------------------------+
| begin; | |
| | begin; |
| delete from t where id = 3; | |
| | delete from t where id = 3; Blocked! |
| insert into t (id, name) values (2, 'B'); | |
| | |
| commit; | |
+-------------------------------------------+--------------------------------------+
回答1:
From "WAITING FOR THIS LOCK TO BE GRANTED" part of transaction 3631, we could see:
RECORD LOCKS space id 69 page no 4 n bits 72 index p_name of table `jacky`.`t` trx id 3631 lock_mode X waiting
Record lock, heap no 4 PHYSICAL RECORD: n_fields 2; compact format; info bits 32
- 3631 is waiting a record lock. The corresponding index content is {"name":"C", "id": 24}.
- The index name is p_name in table t.
- lock's mode is "lock_mode X"
From "WAITING FOR THIS LOCK TO BE GRANTED" part of transaction 3630, we could see:
*** (2) HOLDS THE LOCK(S):
RECORD LOCKS space id 69 page no 4 n bits 72 index p_name of table `jacky`.`t` trx id 3630 lock_mode X
Record lock, heap no 4 PHYSICAL RECORD: n_fields 2; compact format; info bits 32
0: len 1; hex 43; asc C;;
1: len 8; hex 8000000000000018; asc ;;
*** (2) WAITING FOR THIS LOCK TO BE GRANTED:
RECORD LOCKS space id 69 page no 4 n bits 72 index p_name of table `jacky`.`t` trx id 3630 lock_mode X locks gap before rec insert intention waiting
Record lock, heap no 4 PHYSICAL RECORD: n_fields 2; compact format; info bits 32
0: len 1; hex 43; asc C;;
1: len 8; hex 8000000000000018; asc ;;
- 3630 is waiting a record lock. The corresponding index content is {"name":"C", "id": 24}. waiting lock's mode is "lock_mode X locks gap"
- 3630 is holding a record lock. The corresponding index content is {"name":"C", "id": 24}. Holding lock's mode is "lock_mode X locks"
- The index name is p_name in table t.
- This deadlock is caused by executing "insert into t (name) values ('B')"
According to your reproduce step, session A will send a delete from t where name = 'C'; first, this will lock:
- ('A', 'C'] and ('C', 'D'): next-key lock 'C' and gap lock before 'D';
DELETE FROM ... WHERE ... sets an exclusive next-key lock on every record the search encounters. However, only an index record lock is required for statements that lock rows using a unique index to search for a unique row.
- add a record lock for 'C' corresponding primary index id. here id value should be "26".
Then session B will start and delete from t where name = 'C'; will be executed again. However. For session B, because session A hadn't been committed, 'C' had been locked by session A. However, if execute a delete sql, session B will try to add lock in the following sequence:
- gap lock before 'C': Success, because innodb could add multi gap lock in the same position.
- record lock 'C': Blocked, because session A have holded that lock. session B have to wait it released by session A.
- gap lock before 'D':
At last, session A send insert into t (name) values ('B');. For table t, there are 2 indexes, which are id and name. id is a auto increasement primary integer key, and for name, this sql will try to add a insert intention lock. However, there have been a gap lock which is holded by session B, therefor session A have to wait session B for releasing that gap lock. Now we could see how this dead lock occur. Innodb will choose a session to rollback base on cost. Here session B will be rolled back.
For Q1, the anwser is yes.
For Q2, actually, the deleted record wouldn't be purged from an index before its session commit.
For Q3, row lock number is equal to trx_rows_locked, and in mysql website, its:
TRX_ROWS_LOCKED
The approximate number or rows locked by this transaction. The value might include delete-marked rows that are physically present but not visible to the transaction.
From this article, we could know:
For non-clustered unique index filtering, due to the need to return tables, the number of filtered rows is locked as the unique index plus the number of returned rows.
For non-clustered non-unique index filtering, the gap lock is involved, so more records are locked.
So, trx_rows_locked (gap lock + next-key lock + return table) is 3 after delete in session A. final trx_rows_locked value should be 3 + 1 (insert key lock) after trying to insert.
The following are for the new update questions: I didn't notice delete primary key and unique secondary key before.
After some investigating, I found:
- When deleting a
primary key, which have been deleted and not commited yet, the new delete operation will only requirerecord lockinstead of next-key lock. - When deleting a
secondary unique key, which have been deleted and not commited yet, the new delete operation will requirenext-key lock.
You could use set GLOBAL innodb_status_output_locks=ON; show engine innodb status to see the detail lock status for running trasactions.
回答2:
For Q4, I finally found a comment in MySQL 5.7 source code that explain why using next-key locks, just for reference.
In a search where at most one record in the index may match, we can use a LOCK_REC_NOT_GAP type record lock when locking a non-delete-marked matching record.
Note that in a unique secondary index there may be different delete-marked versions of a record where only the primary key values differ: thus in a secondary index we must use next-key locks when locking delete-marked records
Note above that a UNIQUE secondary index can contain many rows with the same key value if one of the columns is the SQL null. A clustered index under MySQL can never contain null columns because we demand that all the columns in primary key are non-null.
来源：https://stackoverflow.com/questions/62846907/mysql-a-deadlock-occurred-when-deleting-the-same-row-twice-in-the-unique-seco