文章首发在公众号(龙台的技术笔记),之后同步到博客园和个人网站:xiaomage.info
优化项目代码过程中发现一个千万级数据深分页问题,缘由是这样的
库里有一张耗材 MCS_PROD 表,通过同步外部数据中台多维度数据,在系统内部组装为单一耗材产品,最终同步到 ES 搜索引擎
MySQL 同步 ES 流程如下:
在这里问题也就出现了,MySQL 查询分页 OFFSET 越深入,性能越差,初步估计线上 MCS_PROD 表中记录在 1000w 左右
如果按照每页 10 条,OFFSET 值会拖垮查询性能,进而形成一个 "性能深渊"
同步类代码针对此问题有两种优化方式:
文章目录如下:
MySQL VERSION
mysql> select version(); +-----------+ | version() | +-----------+ | 5.7.30 | +-----------+ 1 row in set (0.01 sec)
表结构说明
借鉴公司表结构,字段、长度以及名称均已删减
mysql> DESC MCS_PROD; +-----------------------+--------------+------+-----+---------+----------------+ | Field | Type | Null | Key | Default | Extra | +-----------------------+--------------+------+-----+---------+----------------+ | MCS_PROD_ID | int(11) | NO | PRI | NULL | auto_increment | | MCS_CODE | varchar(100) | YES | | | | | MCS_NAME | varchar(500) | YES | | | | | UPDT_TIME | datetime | NO | MUL | NULL | | +-----------------------+--------------+------+-----+---------+----------------+ 4 rows in set (0.01 sec)
通过测试同学帮忙造了 500w 左右数据量
mysql> SELECT COUNT(*) FROM MCS_PROD; +----------+ | count(*) | +----------+ | 5100000 | +----------+ 1 row in set (1.43 sec)
SQL 语句如下
因为功能需要满足 增量拉取的方式,所以会有数据更新时间的条件查询,以及相关 查询排序(此处有坑)
SELECT MCS_PROD_ID, MCS_CODE, MCS_NAME, UPDT_TIME FROM MCS_PROD WHERE UPDT_TIME >= '1970-01-01 00:00:00.0' ORDER BY UPDT_TIME LIMIT xx, xx
LIMIT 子句可以被用于强制 SELECT 语句返回指定的记录数。LIMIT 接收一个或两个数字参数,参数必须是一个整数常量
如果给定两个参数,第一个参数指定第一个返回记录行的偏移量,第二个参数指定返回记录行的最大数
举个简单的例子,分析下 SQL 查询过程,掌握深分页性能为什么差
mysql> SELECT MCS_PROD_ID,MCS_CODE,MCS_NAME FROM MCS_PROD WHERE (UPDT_TIME >= '1970-01-01 00:00:00.0') ORDER BY UPDT_TIME LIMIT 100000, 1; +-------------+-------------------------+------------------+---------------------+ | MCS_PROD_ID | MCS_CODE | MCS_NAME | UPDT_TIME | +-------------+-------------------------+------------------+---------------------+ | 181789 | XA601709733186213015031 | 尺、桡骨LC-DCP骨板 | 2020-10-19 16:22:19 | +-------------+-------------------------+------------------+---------------------+ 1 row in set (3.66 sec) mysql> EXPLAIN SELECT MCS_PROD_ID,MCS_CODE,MCS_NAME FROM MCS_PROD WHERE (UPDT_TIME >= '1970-01-01 00:00:00.0') ORDER BY UPDT_TIME LIMIT 100000, 1; +----+-------------+----------+------------+-------+---------------+------------+---------+------+---------+----------+-----------------------+ | id | select_type | table | partitions | type | possible_keys | key | key_len | ref | rows | filtered | Extra | +----+-------------+----------+------------+-------+---------------+------------+---------+------+---------+----------+-----------------------+ | 1 | SIMPLE | MCS_PROD | NULL | range | MCS_PROD_1 | MCS_PROD_1 | 5 | NULL | 2296653 | 100.00 | Using index condition | +----+-------------+----------+------------+-------+---------------+------------+---------+------+---------+----------+-----------------------+ 1 row in set, 1 warning (0.01 sec)
简单说明下上面 SQL 执行过程:
MySQL 耗费了 大量随机 I/O 在回表查询聚簇索引的数据上,而这 100000 次随机 I/O 查询数据不会出现在结果集中
如果系统并发量稍微高一点,每次查询扫描超过 100000 行,性能肯定堪忧,另外 LIMIT 分页 OFFSET 越深,性能越差(多次强调)
关于 MySQL 深分页优化常见的大概有以下三种策略:
上面三点都能大大的提升查询效率,核心思想就是让 MySQL 尽可能扫描更少的页面,获取需要访问的记录后再根据关联列回原表查询所需要的列
子查询深分页优化语句如下:
mysql> SELECT MCS_PROD_ID,MCS_CODE,MCS_NAME FROM MCS_PROD WHERE MCS_PROD_ID >= ( SELECT m1.MCS_PROD_ID FROM MCS_PROD m1 WHERE m1.UPDT_TIME >= '1970-01-01 00:00:00.0' ORDER BY m1.UPDT_TIME LIMIT 3000000, 1) LIMIT 1; +-------------+-------------------------+------------------------+ | MCS_PROD_ID | MCS_CODE | MCS_NAME | +-------------+-------------------------+------------------------+ | 3021401 | XA892010009391491861476 | 金属解剖型接骨板T型接骨板A | +-------------+-------------------------+------------------------+ 1 row in set (0.76 sec) mysql> EXPLAIN SELECT MCS_PROD_ID,MCS_CODE,MCS_NAME FROM MCS_PROD WHERE MCS_PROD_ID >= ( SELECT m1.MCS_PROD_ID FROM MCS_PROD m1 WHERE m1.UPDT_TIME >= '1970-01-01 00:00:00.0' ORDER BY m1.UPDT_TIME LIMIT 3000000, 1) LIMIT 1; +----+-------------+----------+------------+-------+---------------+------------+---------+------+---------+----------+--------------------------+ | id | select_type | table | partitions | type | possible_keys | key | key_len | ref | rows | filtered | Extra | +----+-------------+----------+------------+-------+---------------+------------+---------+------+---------+----------+--------------------------+ | 1 | PRIMARY | MCS_PROD | NULL | range | PRIMARY | PRIMARY | 4 | NULL | 2296653 | 100.00 | Using where | | 2 | SUBQUERY | m1 | NULL | range | MCS_PROD_1 | MCS_PROD_1 | 5 | NULL | 2296653 | 100.00 | Using where; Using index | +----+-------------+----------+------------+-------+---------------+------------+---------+------+---------+----------+--------------------------+ 2 rows in set, 1 warning (0.77 sec)
根据执行计划得知,子查询 table m1 查询是用到了索引。首先在 索引上拿到了聚集索引的主键 ID 省去了回表操作,然后第二查询直接根据第一个查询的 ID 往后再去查 10 个就可以了
"延迟关联" 深分页优化语句如下:
mysql> SELECT MCS_PROD_ID,MCS_CODE,MCS_NAME FROM MCS_PROD INNER JOIN (SELECT m1.MCS_PROD_ID FROM MCS_PROD m1 WHERE m1.UPDT_TIME >= '1970-01-01 00:00:00.0' ORDER BY m1.UPDT_TIME LIMIT 3000000, 1) AS MCS_PROD2 USING(MCS_PROD_ID); +-------------+-------------------------+------------------------+ | MCS_PROD_ID | MCS_CODE | MCS_NAME | +-------------+-------------------------+------------------------+ | 3021401 | XA892010009391491861476 | 金属解剖型接骨板T型接骨板A | +-------------+-------------------------+------------------------+ 1 row in set (0.75 sec) mysql> EXPLAIN SELECT MCS_PROD_ID,MCS_CODE,MCS_NAME FROM MCS_PROD INNER JOIN (SELECT m1.MCS_PROD_ID FROM MCS_PROD m1 WHERE m1.UPDT_TIME >= '1970-01-01 00:00:00.0' ORDER BY m1.UPDT_TIME LIMIT 3000000, 1) AS MCS_PROD2 USING(MCS_PROD_ID); +----+-------------+------------+------------+--------+---------------+------------+---------+-----------------------+---------+----------+--------------------------+ | id | select_type | table | partitions | type | possible_keys | key | key_len | ref | rows | filtered | Extra | +----+-------------+------------+------------+--------+---------------+------------+---------+-----------------------+---------+----------+--------------------------+ | 1 | PRIMARY | <derived2> | NULL | ALL | NULL | NULL | NULL | NULL | 2296653 | 100.00 | NULL | | 1 | PRIMARY | MCS_PROD | NULL | eq_ref | PRIMARY | PRIMARY | 4 | MCS_PROD2.MCS_PROD_ID | 1 | 100.00 | NULL | | 2 | DERIVED | m1 | NULL | range | MCS_PROD_1 | MCS_PROD_1 | 5 | NULL | 2296653 | 100.00 | Using where; Using index | +----+-------------+------------+------------+--------+---------------+------------+---------+-----------------------+---------+----------+--------------------------+ 3 rows in set, 1 warning (0.00 sec)
思路以及性能与子查询优化一致,只不过采用了 JOIN 的形式执行
关于 LIMIT 深分页问题,核心在于 OFFSET 值,它会 导致 MySQL 扫描大量不需要的记录行然后抛弃掉
我们可以先使用书签 记录获取上次取数据的位置,下次就可以直接从该位置开始扫描,这样可以 避免使用 OFFEST
假设需要查询 3000000 行数据后的第 1 条记录,查询可以这么写
mysql> SELECT MCS_PROD_ID,MCS_CODE,MCS_NAME FROM MCS_PROD WHERE MCS_PROD_ID < 3000000 ORDER BY UPDT_TIME LIMIT 1; +-------------+-------------------------+---------------------------------+ | MCS_PROD_ID | MCS_CODE | MCS_NAME | +-------------+-------------------------+---------------------------------+ | 127 | XA683240878449276581799 | 股骨近端-1螺纹孔锁定板(纯钛)YJBL01 | +-------------+-------------------------+---------------------------------+ 1 row in set (0.00 sec) mysql> EXPLAIN SELECT MCS_PROD_ID,MCS_CODE,MCS_NAME FROM MCS_PROD WHERE MCS_PROD_ID < 3000000 ORDER BY UPDT_TIME LIMIT 1; +----+-------------+----------+------------+-------+---------------+------------+---------+------+------+----------+-------------+ | id | select_type | table | partitions | type | possible_keys | key | key_len | ref | rows | filtered | Extra | +----+-------------+----------+------------+-------+---------------+------------+---------+------+------+----------+-------------+ | 1 | SIMPLE | MCS_PROD | NULL | index | PRIMARY | MCS_PROD_1 | 5 | NULL | 2 | 50.00 | Using where | +----+-------------+----------+------------+-------+---------------+------------+---------+------+------+----------+-------------+ 1 row in set, 1 warning (0.00 sec)
好处是很明显的,查询速度超级快,性能都会稳定在毫秒级,从性能上考虑碾压其它方式
不过这种方式局限性也比较大,需要一种类似连续自增的字段,以及业务所能包容的连续概念,视情况而定
上图是阿里云 OSS Bucket 桶内文件列表,大胆猜测是不是可以采用书签记录的形式完成
以下言论可能会打破你对 order by 所有 美好 YY
先说结论吧,当 LIMIT OFFSET 过深时,会使 ORDER BY 普通索引失效(联合、唯一这些索引没有测试)
mysql> EXPLAIN SELECT MCS_PROD_ID,MCS_CODE,MCS_NAME,UPDT_TIME FROM MCS_PROD WHERE (UPDT_TIME >= '1970-01-01 00:00:00.0') ORDER BY UPDT_TIME LIMIT 100000, 1; +----+-------------+----------+------------+-------+---------------+------------+---------+------+---------+----------+-----------------------+ | id | select_type | table | partitions | type | possible_keys | key | key_len | ref | rows | filtered | Extra | +----+-------------+----------+------------+-------+---------------+------------+---------+------+---------+----------+-----------------------+ | 1 | SIMPLE | MCS_PROD | NULL | range | MCS_PROD_1 | MCS_PROD_1 | 5 | NULL | 2296653 | 100.00 | Using index condition | +----+-------------+----------+------------+-------+---------------+------------+---------+------+---------+----------+-----------------------+ 1 row in set, 1 warning (0.00 sec)
先来说一下这个 ORDER BY 执行过程:
按照 UPDT_TIME 排序可能在内存中完成,也可能需要使用外部排序,取决于排序所需的内存和参数 SORT_BUFFER_SIZE
SORT_BUFFER_SIZE 是 MySQL 为排序开辟的内存。如果排序数据量小于 SORT_BUFFER_SIZE,排序会在内存中完成。如果数据量过大,内存放不下,则会利用磁盘临时文件排序
针对 SORT_BUFFER_SIZE 这个参数在网上查询到有用资料比较少,大家如果测试过程中存在问题,可以加微信一起沟通
OFFSET 100000 时,通过 key Extra 得知,没有使用磁盘临时文件排序,这个时候把 OFFSET 调整到 500000
一首凉凉送给写这个 SQL 的同学,发现了 Using filesort
mysql> EXPLAIN SELECT MCS_PROD_ID,MCS_CODE,MCS_NAME,UPDT_TIME FROM MCS_PROD WHERE (UPDT_TIME >= '1970-01-01 00:00:00.0') ORDER BY UPDT_TIME LIMIT 500000, 1; +----+-------------+----------+------------+------+---------------+------+---------+------+---------+----------+-----------------------------+ | id | select_type | table | partitions | type | possible_keys | key | key_len | ref | rows | filtered | Extra | +----+-------------+----------+------------+------+---------------+------+---------+------+---------+----------+-----------------------------+ | 1 | SIMPLE | MCS_PROD | NULL | ALL | MCS_PROD_1 | NULL | NULL | NULL | 4593306 | 50.00 | Using where; Using filesort | +----+-------------+----------+------------+------+---------------+------+---------+------+---------+----------+-----------------------------+ 1 row in set, 1 warning (0.00 sec)
Using filesort 表示在索引之外,需要额外进行外部的排序动作,性能必将受到严重影响
所以我们应该 结合相对应的业务逻辑避免常规 LIMIT OFFSET,采用 # 深分页优化 章节进行修改对应业务
最后有一点需要声明下,MySQL 本身并不适合单表大数据量业务
因为 MySQL 应用在企业级项目时,针对库表查询并非简单的条件,可能会有更复杂的联合查询,亦或者是大数据量时存在频繁新增或更新操作,维护索引或者数据 ACID 特性上必然存在性能牺牲
如果设计初期能够预料到库表的数据增长,理应构思合理的重构优化方式,比如 ES 配合查询、分库分表、TiDB 等解决方式
参考资料: