问题
I have a stored procedure generating UID's from a "ticket" table, but under load I'm getting lots of deadlocks. I'm calling this procedure many times from multiple concurrent connections whenever my task needs a new UID.
BEGIN
DECLARE a_uid BIGINT(20) UNSIGNED;
START TRANSACTION;
SELECT uid INTO a_uid FROM uid_data FOR UPDATE; # Lock
INSERT INTO uid_data (stub) VALUES ('a') ON DUPLICATE KEY UPDATE uid=uid+1;
SELECT a_uid+1 AS `uid`;
COMMIT;
END
I did consider using:
BEGIN
REPLACE INTO uid_data (stub) VALUES ('a');
SELECT LAST_INSERT_ID();
END
However I wasn't sure if that would be safe with concurrent connections as there's no locking, unlike the first procedure with the SELECT FOR UPDATE
.
Here's the table:
mysql> DESCRIBE uid_data;
+-------+---------------------+------+-----+---------+----------------+
| Field | Type | Null | Key | Default | Extra |
+-------+---------------------+------+-----+---------+----------------+
| uid | bigint(20) unsigned | NO | PRI | NULL | auto_increment |
| stub | char(1) | NO | UNI | NULL | |
+-------+---------------------+------+-----+---------+----------------+
I've setup for read-committed transaction isolation:
mysql> SHOW VARIABLES LIKE 'tx_isolation';
+---------------+-----------------+
| Variable_name | Value |
+---------------+-----------------+
| tx_isolation | READ-COMMITTED |
+---------------+-----------------+
Here's what I'm getting back from SHOW ENGINE INNODB STATUS;
...
... dozens and dozens of the following record locks...
Record lock, heap no 1046 PHYSICAL RECORD: n_fields 2; compact format; info bits 32
0: len 1; hex 61; asc a;;
1: len 8; hex 00000000000335f2; asc 5 ;;
Record lock, heap no 1047 PHYSICAL RECORD: n_fields 2; compact format; info bits 32
0: len 1; hex 61; asc a;;
1: len 8; hex 00000000000335f1; asc 5 ;;
*** (2) WAITING FOR THIS LOCK TO BE GRANTED:
RECORD LOCKS space id 13 page no 4 n bits 1120 index `stub` of table `my_db`.`uid_data` trx id 13AA89 lock_mode X waiting
Record lock, heap no 583 PHYSICAL RECORD: n_fields 2; compact format; info bits 32
0: len 1; hex 61; asc a;;
1: len 8; hex 00000000000334a8; asc 4 ;;
*** WE ROLL BACK TRANSACTION (1)
I'd be grateful if someone could explain what's happening and how they can be avoided.
回答1:
Do this:
CREATE TABLE tickets
(
uid serial
)
Then to get the next uid:
BEGIN
INSERT INTO tickets VALUES (NULL);
SELECT LAST_INSERT_ID();
END
uid serial is equivalent to
uid BIGINT(20) UNSIGNED NOT NULL PRIMARY KEY auto_increment
You shouldn't experience any deadlocks with this approach and can throw as many connections at it as you like.
回答2:
A deadlock occurs in this scenario:
Transaction 1 : requests a lock (SELECT...FOR UPDATE
) and acquires it
Transaction 2 : requests a lock (SELECT...FOR UPDATE
) and must wait
Transaction 1 : tries to insert, hits a duplicate, therefore updates (INSERT...ON DUPLICATE KEY UPDATE
) => deadlock
I am not too sure about the reason, I suspect it has something to do with the ON DUPLICATE KEY UPDATE
. I'm still investigating and will come back if I find out.
[Edit] A deadlock occurs even with:
BEGIN
START TRANSACTION;
SELECT uid FROM uid_data FOR UPDATE;
UPDATE uid_data SET uid = uid +1; -- here, a deadlock would be detected in a blocked, concurrent connection
COMMIT;
END
What about this:
BEGIN
START TRANSACTION;
UPDATE uid_data SET uid = uid +1;
SELECT uid FROM uid_data;
COMMIT;
END
You could drop your stub
colum altogether. The only drawback is that you must initialise your uid_data with one row.
回答3:
You can try using
UPDATE uid_data SET uid = LAST_INSERT_ID(uid+1);
SELECT LAST_INSERT_ID();
on a table like
CREATE TABLE `uid_data` (
`uid` BIGINT(20) UNSIGNED NOT NULL
)
COLLATE='utf8_general_ci'
ENGINE=MyISAM;
This is thread safe and will not lock the table if it is MyISAM (except during the actual update statement).
来源:https://stackoverflow.com/questions/11346256/mysql-deadlocks-with-stored-procedure-generating-uid