使用 SELECT/LIMIT 1/FOR UPDATE 时避免死锁的最佳方法是什么？

Question

我有以下两个导致死锁的事务：

交易 1：

SELECT *
FROM bots b
WHERE b.status = :available
AND (b.last_checked_at IS NULL OR b.last_checked_at < :fiveMinsAgo)
ORDER BY b.last_checked_at ASC
LIMIT 1
FOR UPDATE

*AND* (I don't think that this statement is part of the issue)

UPDATE bots b
SET b.last_checked_at = :now
WHERE b.id = :botId

交易 2：

SELECT b.*, bu.*
FROM bots b
LEFT JOIN bot_users bu ON bu.id = b.assigned_to_bot_user_id
WHERE b.currency = :currency
AND (b.status = :available OR (b.status = :reserved AND bu.last_action_at < :threeMinsAgo))
AND ((b.slot_count - b.items_count) >= :amount)
ORDER BY b.keys_count ASC, b.id ASC
LIMIT 1
FOR UPDATE

*OR*

SELECT b.*, bu.*
FROM bots b
LEFT JOIN bot_users bu ON bu.assigned_bot_id = b.id
WHERE b.currency = :currency
AND (b.status = :available OR (b.status = :reserved AND bu.last_action_at < :threeMinsAgo))
AND (b.keys_count >= :amount)
ORDER BY b.keys_count DESC, b.id ASC
LIMIT 1
FOR UPDATE

*AND*

UPDATE bots SET status = :reserved WHERE id = :id

说明：

事务 #1 选择 1 个可用且在过去 5 分钟内未检查的机器人，检查它，然后更新 last_checked_at 字段。

事务 #2（案例 1）正在选择 1 个处理某种货币的机器人，其状态为可用或已保留且“已过期”（3 分钟内没有用户操作），并且有足够的可用插槽。

交易 #2（案例 2）正在选择 1 个处理某种货币的机器人，其状态为可用或已保留且“已过期”（3 分钟内没有用户操作），并且有足够的密钥来满足请求的金额。

注意事项：

在 b.keys_count、b.currency 和 b.status 上有索引。

死锁示例：

------------------------
LATEST DETECTED DEADLOCK
------------------------
2014-07-04 20:33:11 7fc46f270700
*** (1) TRANSACTION:
TRANSACTION 1657246652, ACTIVE 0 sec fetching rows
mysql tables in use 1, locked 1
LOCK WAIT 4 lock struct(s), heap size 1184, 36 row lock(s)
MySQL thread id 30422870, OS thread handle 0x7fc46f437700, query id 1217244940 10.0.88.8 mydb Creating sort index
*** (1) WAITING FOR THIS LOCK TO BE GRANTED:
RECORD LOCKS space id 4752665 page no 3 n bits 176 index `PRIMARY` of table `mydb`.`bots` trx id 1657246652 lock_mode X locks rec but not gap waiting
*** (2) TRANSACTION:
TRANSACTION 1657246644, ACTIVE 0 sec updating or deleting
mysql tables in use 1, locked 1
10 lock struct(s), heap size 2936, 37 row lock(s), undo log entries 1
MySQL thread id 30422999, OS thread handle 0x7fc46f270700, query id 1217244943 10.0.88.8 mydb updating
UPDATE bots SET status = 1 WHERE id = 19
*** (2) HOLDS THE LOCK(S):
RECORD LOCKS space id 4752665 page no 3 n bits 176 index `PRIMARY` of table `mydb`.`bots` trx id 1657246644 lock_mode X locks rec but not gap
*** (2) WAITING FOR THIS LOCK TO BE GRANTED:
RECORD LOCKS space id 4752665 page no 6 n bits 168 index `status_idx` of table `mydb`.`bots` trx id 1657246644 lock_mode X locks rec but not gap waiting
*** WE ROLL BACK TRANSACTION (1)

查询似乎锁定了大部分行，而不仅仅是一个。

防止这些死锁的最佳方法是什么？最好不要在事务#1 中使用 FOR UPDATE，然后尝试使用基于 ID 的选择获取锁，该选择还包括其他条件？例如：

SELECT *
FROM bots b
WHERE b.status = :available
AND (b.last_checked_at IS NULL OR b.last_checked_at < :fiveMinsAgo)
ORDER BY b.last_checked_at ASC
LIMIT 1

SELECT *
FROM bots b
WHERE b.id = :idSelectedAbove
AND b.status = :available
AND (b.last_checked_at IS NULL OR b.last_checked_at < :fiveMinsAgo)
FOR UPDATE

这显然会在一定百分比的时间内失败，具体取决于各种因素。

在使用 ORDER BY, FOR UPDATE 时有什么方法可以只锁定一行吗？

Answer 1

考虑使用GET_LOCK() 模拟记录锁并废弃FOR UPDATES；

选择特定于您要锁定的行的名称。例如'bot_id_1'.

调用SELECT GET_LOCK('bot_id_1',30)锁定名称'bot_id_1'..如果名称可用则返回1并设置锁，如果30秒后锁不可用则返回0（第二个参数是超时时间） .

完成后使用SELECT RELEASE_LOCK('bot_id_1')。

注意；您必须在每个事务中使用相同的名称，并且在事务中再次调用GET_LOCK() 将释放先前设置的锁。

GET_LOCK() docs