从 MySQL 5.5 升级到 5.7 后查询更频繁地遇到死锁

Question

最近，我们使用 AWS DMS 服务将生产数据库迁移到 Amazon RDS，并将版本从 5.5 升级到 5.7。 在那之后，我们经常在重复的键更新查询和更新查询上遇到我们插入的死锁问题。 而在 MySQL 5.5 中它非常小。

例如，假设我们的一个表结构如下。

CREATE TABLE `job_notification` (
  `id` int(11) NOT NULL AUTO_INCREMENT,
  `uid` int(11) NOT NULL,
  `job_id` int(11) NOT NULL,
  `created_time` int(11) NOT NULL,
  `updated_time` int(11) NOT NULL,
  `notify_status` tinyint(3) DEFAULT '0'
  PRIMARY KEY (`id`),
  UNIQUE KEY `uid` (`uid`,`job_id`),
) ENGINE=InnoDB AUTO_INCREMENT=58303732 DEFAULT CHARSET=utf8 COLLATE=utf8_bin

我们的插入查询如下...

    INSERT INTO job_notification (uid, notify_status, updated_time, created_time, job_id) VALUES
('24832194',1,1571900253,1571900253,'734749'),
('24832194',1,1571900254,1571900254,'729161'),
('24832194',1,1571900255,1571900255,'713225'),
('24832194',1,1571900256,1571900256,'701897'),
('24832194',1,1571900257,1571900257,'682155'),
('24832194',1,1571900258,1571900258,'730817'),
('24832194',1,1571900259,1571900259,'717162'),
('24832194',1,1571900260,1571900260,'712884'),
('24832194',1,1571900261,1571900261,'708267'),
('24832194',1,1571900262,1571900262,'701855'),
('24832194',1,1571900263,1571900263,'702129'),
('24832194',1,1571900264,1571900264,'726738'),
('24832194',1,1571900265,1571900265,'725105'),
('24832194',1,1571900266,1571900266,'709306'),
('24832194',1,1571900267,1571900267,'702218'),
('24832194',1,1571900268,1571900268,'700966'),
('24832194',1,1571900269,1571900269,'693848'),
('24832194',1,1571900270,1571900270,'730793'),
('24832194',1,1571900271,1571900271,'729352'),
('24832194',1,1571900272,1571900272,'729043'),
('24832194',1,1571900273,1571900273,'724631'),
('24832194',1,1571900274,1571900274,'718394'),
('24832194',1,1571900275,1571900275,'711702'),
('24832194',1,1571900276,1571900276,'707765'),
('24832194',1,1571900277,1571900277,'692288'),
('24832194',1,1571900278,1571900278,'735549'),
('24832194',1,1571900279,1571900279,'730786'),
('24832194',1,1571900280,1571900280,'706814'),
('24832194',1,1571900281,1571900281,'688999'),
('24832194',1,1571900282,1571900282,'685079'),
('24832194',1,1571900283,1571900283,'686661'),
('24832194',1,1571900284,1571900284,'722110'),
('24832194',1,1571900285,1571900285,'715277'),
('24832194',1,1571900286,1571900286,'701846'),
('24832194',1,1571900287,1571900287,'730105'),
('24832194',1,1571900288,1571900288,'725579')
 ON DUPLICATE KEY UPDATE notify_status=VALUES(notify_status), updated_time=VALUES(updated_time)

我们的更新查询如下...

update job_notification set notify_status = 3 where uid = 51032194 and job_id in (616661, 656221, 386760, 189461, 944509, 591552, 154153, 538703, 971923, 125080, 722110, 715277, 701846, 725579, 686661, 685079)

这些查询在 MySQL 5.5 中运行良好，具有相同的数据包大小和索引，但在迁移后此类查询经常出现死锁......

NB：我们的系统是高级并发系统。 innodb_deadlock_detect被禁用。 innodb_lock_wait_timeout为 50。

innodb_buffer_pool_size为 50465865728

当我们解释查询时，它给出了更好的执行计划。 尽管如此，我们还是经常遇到死锁，因此其他查询也会变慢。

两个查询都使用 pythons Mysqldb package 作为不同的 API 线程（不同的连接）执行，并在 MySQL DB 中启用了自动提交。

解释Output

explain update job_notification SET notify_status = 3 where uid = 51032194 and job_id in (616661, 656221, 386760, 189461, 944509, 591552, 154153, 538703, 971923, 125080, 722110, 715277, 701846, 725579, 686661, 685079);
+----+--------+------------+------------+-------+---------------+------+-----+-------+------+----------+--------+
| id | select_type | table                    | partitions | type  | possible_keys | key  | key_len | ref         | rows | filtered |Extra       |
+----+----------+------------+------------+-------+---------------+------+---------+-------------+------+----------+----------+
|  1 | UPDATE      | job_notification | NULL       | range | uid           | uid  | 8       | const,const |   27 |   100.00 | Using where |
+----+-------------+--------------------------+------------+-------+---------------+------+---------+-------------+--------+-------------+

Answer 1

如果这主要是一个多对多映射表，请去掉 if id并按照http://mysql.rjweb.org/doc.php/index_cookbook_mysql#many_to_many_mapping_table 中的建议的 rest

这将使查询在两个系统上运行得更快。 更快 = 更少的死锁机会。

让我们看看僵局； 可能还有其他事情发生。 使用SHOW ENGINE INNODB STATUS; 死锁发生后立即执行。