sql优化-派生表与inner-join
2021/5/30 19:21:41
本文主要是介绍sql优化-派生表与inner-join,对大家解决编程问题具有一定的参考价值,需要的程序猿们随着小编来一起学习吧!
首先来说明一下派生表?
外部的表查询的结果集是从子查询中生成的.如下形式:
select ... from (select ....) dt
如上形式中括号中的查询的结果作为外面select语句的查询源,派生表必须指定别名,因此后面的dt必须指定。派生表和临时表差不多,但是在select语句中派生表比临时表要容易,因为派生表不用创建。
一个有关派生表优化的实例。
开发同事发来一个sql优化,涉及到4张表,表中的数据都不是很大,但是查询起来真的很慢。服务器性能又差,查询总是超时。
四张表的表结构如下:
Table: t_info_settingCreate Table: CREATE TABLE `t_info_setting` ( `id` int(11) NOT NULL AUTO_INCREMENT, `parent_key` varchar(32) NOT NULL, `column_name` varchar(32) NOT NULL, `column_key` varchar(32) NOT NULL, `storage_way` tinyint(4) DEFAULT '0', `check_way` tinyint(4) DEFAULT '0', `remark` varchar(500) DEFAULT '', `operator` varchar(128) DEFAULT '', `status` int(11) DEFAULT '1', `update_time` datetime DEFAULT CURRENT_TIMESTAMP ON UPDATE CURRENT_TIMESTAMP, PRIMARY KEY (`id`), KEY `column_key` (`column_key`) ) ENGINE=InnoDB AUTO_INCREMENT=9 DEFAULT CHARSET=utf8
t_info_setting
Table: t_articles_statusCreate Table: CREATE TABLE `t_articles_status` ( `id` int(11) unsigned NOT NULL AUTO_INCREMENT, `linkId` varchar(36) NOT NULL, `column_key` varchar(32) NOT NULL, `status` int(11) DEFAULT '50000', `operator_time` timestamp NULL DEFAULT CURRENT_TIMESTAMP ON UPDATE CURRENT_TIMESTAMP, PRIMARY KEY (`id`), KEY `article_status` (`linkId`,`column_key`) ) ENGINE=InnoDB AUTO_INCREMENT=22232 DEFAULT CHARSET=utf81 row in set (0.00 sec)
t_articles_status
Table: t_article_operationsCreate Table: CREATE TABLE `t_article_operations` ( `id` int(11) NOT NULL AUTO_INCREMENT, `linkId` varchar(36) NOT NULL, `column_key` varchar(32) NOT NULL DEFAULT '', `type` varchar(16) DEFAULT '', `operator` varchar(128) DEFAULT '', `operator_time` timestamp NULL DEFAULT CURRENT_TIMESTAMP ON UPDATE CURRENT_TIMESTAMP, PRIMARY KEY (`id`), KEY `article_operation` (`linkId`,`column_key`), KEY `operator_time` (`operator_time`) ) ENGINE=InnoDB AUTO_INCREMENT=23316 DEFAULT CHARSET=utf81 row in set (0.00 sec)
t_article_operations
Table: t_articlesCreate Table: CREATE TABLE `t_articles` ( `id` int(11) unsigned NOT NULL AUTO_INCREMENT, `linkId` varchar(36) DEFAULT NULL, `source` int(11) DEFAULT '0', `title` varchar(150) NOT NULL, `author` varchar(150) NOT NULL, `tags` varchar(200) DEFAULT NULL, `abstract` varchar(512) DEFAULT NULL, `content` mediumtext, `thumbnail` varchar(256) DEFAULT NULL, `sourceId` varchar(24) DEFAULT '', `accessoryUrl` text, `relatedStock` text, `contentUrl` text, `secuInfo` text, `market` varchar(10) DEFAULT 'hk', `code` varchar(10) DEFAULT '', `updator` varchar(64) DEFAULT '', `createTime` timestamp NULL DEFAULT CURRENT_TIMESTAMP COMMENT '建立时间', `updateTime` timestamp NULL DEFAULT CURRENT_TIMESTAMP ON UPDATE CURRENT_TIMESTAMP COMMENT '更新时间', PRIMARY KEY (`id`), UNIQUE KEY `linkId` (`linkId`) ) ENGINE=InnoDB AUTO_INCREMENT=15282 DEFAULT CHARSET=utf8
t_articles
上面四张表,由上面的自增字段的值可以知道表的数据并不是很大,最大的表也就2万多行,表中的索引情况已经一目了然。开发同学给出的sql语句如下:
( SELECT 'daily' AS category, e.linkId, e.title, e.updateTime FROM ( SELECT DISTINCT b.column_key, b. STATUS, b.linkId FROM t_info_setting a inner JOIN t_articles_status b ON a.column_key = b.column_key inner JOIN t_article_operations c ON b.column_key = c.column_key WHERE a.parent_key = 'daily' AND a. STATUS = 1 AND b. STATUS = 80000 ORDER BY c.operator_time DESC LIMIT 1 ) AS d inner JOIN t_articles e ON d.linkId = e.linkId )UNION ALL ( SELECT 'ipo' AS category, e.linkId, e.title, e.updateTime FROM ( SELECT DISTINCT b.column_key, b. STATUS, b.linkId FROM t_info_setting a inner JOIN t_articles_status b ON a.column_key = b.column_key inner JOIN t_article_operations c ON b.column_key = c.column_key WHERE a.parent_key = 'ipo' AND a. STATUS = 1 AND b. STATUS = 80000 ORDER BY c.operator_time DESC LIMIT 1 ) AS d inner JOIN t_articles e ON d.linkId = e.linkId )UNION ALL ( SELECT 'research' AS category, e.linkId, e.title, e.updateTime FROM ( SELECT DISTINCT b.column_key, b. STATUS, b.linkId FROM t_info_setting a inner JOIN t_articles_status b ON a.column_key = b.column_key inner JOIN t_article_operations c ON b.column_key = c.column_key WHERE a.parent_key = 'research' AND a. STATUS = 1 AND b. STATUS = 80000 ORDER BY c.operator_time DESC LIMIT 1 ) AS d inner JOIN t_articles e ON d.linkId = e.linkId )UNION ALL ( SELECT 'news' AS category, e.linkId, e.title, e.updateTime FROM ( SELECT DISTINCT b.column_key, b. STATUS, b.linkId FROM t_info_setting a inner JOIN t_articles_status b ON a.column_key = b.column_key inner JOIN t_article_operations c ON b.column_key = c.column_key WHERE a.parent_key = 'news' AND a. STATUS = 1 AND b. STATUS = 80000 ORDER BY c.operator_time DESC LIMIT 1 ) AS d inner JOIN t_articles e ON d.linkId = e.linkId )
开发给的sql
原sql很长大概有107行,但是分析这条sql发现了使用了三个union联合查询,然后每条联合的sql语句基本是一模一样的,只是改变了a.parent_key = 'research'这个条件。这说明我们只需要分析其中的一条sql即可。
SELECT 'research' AS category, e.linkId, e.title, e.updateTime FROM ( -- 这里使用了派生表 SELECT DISTINCT --a b.column_key, b. STATUS, b.linkId FROM t_info_setting a inner JOIN t_articles_status b ON a.column_key = b.column_key inner JOIN t_article_operations c ON b.column_key = c.column_key -- c WHERE a.parent_key = 'research' AND a. STATUS = 1 AND b. STATUS = 80000 ORDER BY c.operator_time DESC LIMIT 1 ) AS d inner JOIN t_articles e ON d.linkId = e.linkId -- b
首先:这条sql语句中使用了派生表,分析里面的子查询,最后有一个limit 1也就是只查出一条数据,并且是按照operator_time 进行排序,那么distinct的去重复就是不需要的。再看子查询中查询出了三个字段,但是在b处和e表进行联合查询的时候只使用了linkId 这一个字段,因此子查询中多余的两个字段需要去掉。
在表t_article_operations上有一个符合索引,我们知道mysql在使用复合索引时,采用最左原则,因此在c处的联合查询我们需要加上linkId ,根据上面分析,改写sql如下:
select 'research' as category, e.linkId, e.title, e.updateTimefrom ( select b.linkId -- 去除不必要的列、distinct操作 from t_info_setting a inner join t_articles_status b on a.column_key=b.column_key inner join t_article_operations c on b.linkId=c.linkId and b.column_key=c.column_key -- 关联条件应包含linkId where a.parent_key='research' and a.status=1 and b.status=80000 order by c.operator_time desc limit 1) dinner join t_articles e on d.linkId=e.linkId;
然后查看下改写前后两个sql的执行计划。
改写后的执行计划:
+----+-------------+------------+--------+-------------------+----------------+---------+-----------------------------------------------------------+-------+-------------+| id | select_type | table | type | possible_keys | key | key_len | ref | rows | Extra |+----+-------------+------------+--------+-------------------+----------------+---------+-----------------------------------------------------------+-------+-------------+| 1 | PRIMARY | <derived2> | system | NULL | NULL | NULL | NULL | 1 | NULL || 1 | PRIMARY | e | const | linkId | linkId | 111 | const | 1 | NULL || 2 | DERIVED | c | index | article_operation | operator_time | 5 | NULL | 14711 | NULL || 2 | DERIVED | a | ref | column_key | column_key | 98 | wlb_live_contents.c.column_key | 1 | Using where || 2 | DERIVED | b | ref | article_status | article_status | 208 | wlb_live_contents.c.linkId,wlb_live_contents.c.column_key | 1 | Using where |+----+-------------+------------+--------+-------------------+----------------+---------+-----------------------------------------------------------+-------+-------------+
改写之后的单个sql很快就有了结果,大概0.12秒就有了结束,而原来的sql会超时结束的。
在原sql语句中使用了union,因为我们最后的结果并不要求去重复,只是四个结果集的联合,因此这里我们可以使用union all代替上面的union。
改写后的执行计划DERIVED表示使用了派生表,我们看到在e表与派生表进行inner查询的使用了索引。
分析:
之前看到一种说法是,在数据表和派生表联合进行查询时,不能使用索引,但是上面的的执行计划说明使用了索引(e表和派生表联合查询,e表使用了索引)。【究竟要怎么用还需进一步研究】
改写sql:
上面使用了派生表,其实数据量比较大时,派生表的效率并不是很高的,上面的查询我们试着用4张表的联合查询来改写。
改写之后的sql如下:
SELECT 'research' AS category, e.linkId, e.title, e.updateTimeFROM t_info_setting aINNER JOIN t_articles_status b ON a.column_key = b.column_keyINNER JOIN t_article_operations c ON b.linkId = c.linkIdAND b.column_key = c.column_keyINNER JOIN t_articles e ON c.linkId = e.linkIdWHERE a.parent_key = 'research'AND a. STATUS = 1AND b. STATUS = 80000ORDER BY c.operator_time DESCLIMIT 1
查看执行计划:
+----+-------------+-------+-------+-------------------+----------------+---------+-----------------------------------------------------------+------+-------------+| id | select_type | table | type | possible_keys | key | key_len | ref | rows | Extra |+----+-------------+-------+-------+-------------------+----------------+---------+-----------------------------------------------------------+------+-------------+| 1 | SIMPLE | c | index | article_operation | operator_time | 5 | NULL | 1 | NULL || 1 | SIMPLE | a | ref | column_key | column_key | 98 | wlb_live_contents.c.column_key | 1 | Using where || 1 | SIMPLE | b | ref | article_status | article_status | 208 | wlb_live_contents.c.linkId,wlb_live_contents.c.column_key | 1 | Using where || 1 | SIMPLE | e | ref | linkId | linkId | 111 | wlb_live_contents.c.linkId | 1 | NULL |+----+-------------+-------+-------+-------------------+----------------+---------+-----------------------------------------------------------+------+-------------+
根据执行计划,这个inner join的执行计划是要比上面的使用派生表的执行计划要高一些。
说明:
1:在使用联合查询的时候,可以考虑联合查询的键上创建索引,效率可能会高点。
2:可以考虑在order by的键上创建索引。
3:根据数据可以知道,t_article_operations本质上是一个流水表,记录日志类信息,不应出现在日常查询中。解决此种查询的办法:operator_time保存在t_articles_status中,查询彻底移除t_article_operations,或临时方法:t_article_operations只保留短期数据,历史记录定期迁移至其他表。
这篇关于sql优化-派生表与inner-join的文章就介绍到这儿,希望我们推荐的文章对大家有所帮助,也希望大家多多支持为之网!
- 2024-11-23Springboot应用的多环境打包入门
- 2024-11-23Springboot应用的生产发布入门教程
- 2024-11-23Python编程入门指南
- 2024-11-23Java创业入门:从零开始的编程之旅
- 2024-11-23Java创业入门:新手必读的Java编程与创业指南
- 2024-11-23Java对接阿里云智能语音服务入门详解
- 2024-11-23Java对接阿里云智能语音服务入门教程
- 2024-11-23JAVA对接阿里云智能语音服务入门教程
- 2024-11-23Java副业入门:初学者的简单教程
- 2024-11-23JAVA副业入门:初学者的实战指南