MySQL 与 InnoDB 的死锁问题

我有一个中央数据库服务器和几个“工作”服务器，它们同时执行这样的查询：

UPDATE job_queue 
SET
  worker = '108.166.81.112',
  attempts = attempts + 1,
  started = '2014-01-14 10:34:03',
  token = '13eb3e6a8c3e1becb34051e08f19fd62'
WHERE completed = '0000-00-00 00:00:00'
  AND (started = '0000-00-00 00:00:00' OR started < '2014-01-14 10:29:03')
  AND attempts < 2
ORDER BY priority DESC, inserted
LIMIT 1

有时我的 job_queue 表会锁定，如果我运行“SHOW ENGINE INNODB STATUS”，我会得到如下信息：

------------------------
LATEST DETECTED DEADLOCK
------------------------
140114 10:34:15
*** (1) TRANSACTION:
TRANSACTION 0 46984514, ACTIVE 0 sec, process no 590, OS thread id 140366633146112 fetching rows
mysql tables in use 1, locked 1
LOCK WAIT 20 lock struct(s), heap size 3024, 545 row lock(s)
MySQL thread id 677401, query id 19385205 10.179.103.110 root init
UPDATE job_queue SET worker='108.166.81.112', attempts=attempts+1, started='2014-01-14 10:34:03', token='13eb3e6a8c3e1becb34051e08f19fd62' WHERE completed='0000-00-00 00:00:00' AND (started='0000-00-00 00:00:00' OR started<'2014-01-14 10:29:03') AND attempts<2 ORDER BY priority DESC, inserted LIMIT 1
*** (1) WAITING FOR THIS LOCK TO BE GRANTED:
RECORD LOCKS space id 0 page no 245767 n bits 128 index `PRIMARY` of table `database`.`job_queue` trx id 0 46984514 lock_mode X waiting
Record lock, heap no 34 PHYSICAL RECORD: n_fields 12; compact format; info bits 0
 0: len 3; hex 800210; asc    ;; 1: len 6; hex 000002cced25; asc      %;; 2: len 7; hex 000003c00f1970; asc       p;; 3: len 30; hex 4f3a31343a2243425343616368654170704a6f62223a363a7b733a31393a; asc O:14:"CBSCacheAppJob":6:{s:19:;...(truncated); 4: len 1; hex 80; asc  ;; 5: len 8; hex 800012513c58bf24; asc    Q<X $;; 6: len 8; hex 800012513c58cc17; asc    Q<X  ;; 7: len 14; hex 31302e3137392e3130332e313333; asc 10.179.103.133;; 8: len 1; hex 81; asc  ;; 9: len 8; hex 800012513c58cc32; asc    Q<X 2;; 10: len 0; hex ; asc ;; 11: len 30; hex 353264393033616162656634346239626536306463346438666432303066; asc 52d903aabef44b9be60dc4d8fd200f;...(truncated);

*** (2) TRANSACTION:
TRANSACTION 0 46984485, ACTIVE 17 sec, process no 590, OS thread id 140366633547520 starting index read, thread declared inside InnoDB 500
mysql tables in use 1, locked 1
4 lock struct(s), heap size 1216, 2 row lock(s), undo log entries 1
MySQL thread id 676723, query id 19385209 10.179.103.133 root init
UPDATE job_queue SET worker='10.179.103.133', attempts=attempts+1, started='2014-01-14 10:34:03', token='efd21d0d34f44badbc30386db4dd252e' WHERE completed='0000-00-00 00:00:00' AND (started='0000-00-00 00:00:00' OR started<'2014-01-14 10:29:03') AND attempts<2 ORDER BY priority DESC, inserted LIMIT 1
*** (2) HOLDS THE LOCK(S):
RECORD LOCKS space id 0 page no 245767 n bits 128 index `PRIMARY` of table `database`.`job_queue` trx id 0 46984485 lock_mode X locks rec but not gap
Record lock, heap no 34 PHYSICAL RECORD: n_fields 12; compact format; info bits 0
 0: len 3; hex 800210; asc    ;; 1: len 6; hex 000002cced25; asc      %;; 2: len 7; hex 000003c00f1970; asc       p;; 3: len 30; hex 4f3a31343a2243425343616368654170704a6f62223a363a7b733a31393a; asc O:14:"CBSCacheAppJob":6:{s:19:;...(truncated); 4: len 1; hex 80; asc  ;; 5: len 8; hex 800012513c58bf24; asc    Q<X $;; 6: len 8; hex 800012513c58cc17; asc    Q<X  ;; 7: len 14; hex 31302e3137392e3130332e313333; asc 10.179.103.133;; 8: len 1; hex 81; asc  ;; 9: len 8; hex 800012513c58cc32; asc    Q<X 2;; 10: len 0; hex ; asc ;; 11: len 30; hex 353264393033616162656634346239626536306463346438666432303066; asc 52d903aabef44b9be60dc4d8fd200f;...(truncated);

*** (2) WAITING FOR THIS LOCK TO BE GRANTED:
RECORD LOCKS space id 0 page no 57 n bits 120 index `PRIMARY` of table `database`.`job_queue` trx id 0 46984485 lock_mode X waiting
Record lock, heap no 2 PHYSICAL RECORD: n_fields 12; compact format; info bits 0
 0: len 3; hex 800001; asc    ;; 1: len 6; hex 000002ccdab1; asc       ;; 2: len 7; hex 000003c0352b3f; asc     5+?;; 3: len 30; hex 4f3a31323a224175746f50696c6f744a6f62223a363a7b733a31383a2200; asc O:12:"AutoPilotJob":6:{s:18:" ;...(truncated); 4: len 1; hex 82; asc  ;; 5: len 8; hex 800012513c58af57; asc    Q<X W;; 6: len 8; hex 800012513c58bf22; asc    Q<X ";; 7: len 14; hex 3130382e3136362e38312e313132; asc 108.166.81.112;; 8: len 1; hex 81; asc  ;; 9: len 8; hex 800012513c58bf23; asc    Q<X #;; 10: len 0; hex ; asc ;; 11: len 30; hex 616331376430346339326163613366323330646164323239363764336266; asc ac17d04c92aca3f230dad22967d3bf;...(truncated);

*** WE ROLL BACK TRANSACTION (1)
------------
TRANSACTIONS
------------
Trx id counter 0 46989905
Purge done for trx's n:o < 0 46986227 undo n:o < 0 0
History list length 24
LIST OF TRANSACTIONS FOR EACH SESSION:
---TRANSACTION 0 0, not started, process no 590, OS thread id 140366628529920
MySQL thread id 703864, query id 20047015 localhost root
SHOW ENGINE INNODB STATUS
---TRANSACTION 0 46989894, not started, process no 590, OS thread id 140366636758784
MySQL thread id 702822, query id 20046897 10.179.1.63 root
---TRANSACTION 0 46986223, ACTIVE 39782 sec, process no 590, OS thread id 140366626322176
25 lock struct(s), heap size 3024, 710 row lock(s), undo log entries 9
MySQL thread id 677706, query id 19994505 10.179.103.114 root
Trx read view will not see trx with id >= 0 46986224, sees < 0 46986224

任何对表的进一步写入都会超时，直到我重新启动 MySQL 服务器（或手动终止死锁作业）：

PHP Fatal error:  Lock wait timeout exceeded; try restarting transaction(Query: "UPDATE job_queue SET worker='108.166.81.250', attempts=attempts+1, started='2014-01-14 21:27:45', token='369eae55a7f0eacad3b678a3410de8e4' WHERE completed='0000-00-00 00:00:00' AND (started='0000-00-00 00:00:00' OR started<'2014-01-14 21:22:45') AND attempts<2 ORDER BY priority DESC, inserted LIMIT 1") in /utilities/Database.php on line 53

谁能向我解释为什么这个查询会导致死锁？我的印象是 InnoDB 表上的所有查询都是原子发生的。有任何想法吗？

这导致了僵局，因为UPDATE查询正在锁定表中的所有行，并且根据所使用的索引（或缺乏索引），两个不同的会话可能会以稍微不同的顺序锁定它们。请记住UPDATE, DELETE, and SELECT ... FOR UPDATE将锁定它们遇到的所有行，无论这些行是否匹配所有行WHERE条件与否。因此，在使用它们时，您应该通过使用索引（最好是主键）并避免模糊或广泛的选择条件，努力确保它们遇到尽可能少的行。

我对工作队列的建议几乎是通用的：尽可能少地锁定，尽可能少地锁定，并且始终按照确定的顺序。所以，一般来说：

1. 使用非锁定读取（常规SELECT) 通过查找您的工人知道如何做且目前无人认领的工作来找到要做的工作 (lease_owner IS NULL AND lease_expiry IS NULL——或类似的）。
2. 选择一项工作（如果你敢的话，也可以选择几个，但其中一项要简单得多，并且通常可以实现完全可接受的性能）。
3. Update your work item (to claim it, but for any case it needs updating as well):
   1. 开启交易。
   2. 锁定您选择的工作项目SELECT ... FOR UPDATE-- 如果不再无人认领，则中止并选择另一个。
   3. 使用您的工作人员 ID 和租约的到期时间更新您选择的工作项目。
   4. 立即提交您的交易。
4. 开始处理您租用的工作项目。
5. 在其他一些过程中，另一个轮询器会查找废弃的工作并取消声明它（通过上面相同的更新过程）。

通过这种设计，您可以轻松获得非常高的吞吐量（每秒数千个作业），并且基本上没有争用，也没有排序问题。选择不太可能与其他轮询器发生冲突的工作的优化既简单又有效（例如，选择作业 ID 或类似的模数，以避免作业匮乏）。关键是要记住，工作选择上的冲突是okay——只要中止并重试，一切都会进展得很快。

工作队列项目/作业的所有锁定写入只能在单行并通过主键only.