MySQL 的 join 功能弱爆了?

{"type":"doc","content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"大家好，我是历小冰，今天我们来学习和吐槽一下 MySQL 的 Join 功能。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"关于MySQL 的 join，大家一定了解过很多它的“轶事趣闻”，比如两表 join 要小表驱动大表，阿里开发者规范禁止三张表以上的 join 操作，MySQL 的 join 功能弱爆了等等。这些规范或者言论亦真亦假，时对时错，需要大家自己对 join 有深入的了解后才能清楚地理解。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"下面，我们就来全面的了解一下 MySQL 的 join 操作。"}]},{"type":"heading","attrs":{"align":null,"level":3},"content":[{"type":"text","text":"正文"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"在日常数据库查询时，我们经常要对多表进行连表操作来一次性获得多个表合并后的数据，这是就要使用到数据库的 join 语法。join 是在数据领域中十分常见的将两个数据集进行合并的操作，如果大家了解的多的话，会发现 MySQL，Oracle，PostgreSQL 和 Spark 都支持该操作。本篇文章的主角是 MySQL，下文没有特别说明的话，就是以 MySQL 的 join 为主语。而 "},{"type":"text","marks":[{"type":"strong"}],"text":"Oracle ，PostgreSQL 和 Spark 则可以算做将其吊打的大boss，其对 join 的算法优化和实现方式都要优于 MySQL。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"MySQL 的 join 有诸多规则，可能稍有不慎，可能一个不好的 join 语句不仅会导致对某一张表的全表查询，还有"},{"type":"text","marks":[{"type":"strong"}],"text":"可能会影响数据库的缓存，导致大部分热点数据都被替换出去，拖累整个数据库性能。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"所以，业界针对 MySQL 的 join 总结了很多规范或者原则，比如说小表驱动大表和禁止三张表以上的 join 操作。下面我们会依次介绍 MySQL join 的算法，和 Oracle 和 Spark 的 join 实现对比，并在其中穿插解答为什么会形成上述的规范或者原则。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"对于 join 操作的实现，大概有 Nested Loop Join (循环嵌套连接)，Hash Join(散列连接) 和 Sort Merge Join(排序归并连接) 三种较为常见的算法"},{"type":"text","text":"，它们各有优缺点和适用条件，接下来我们会依次来介绍。"}]},{"type":"heading","attrs":{"align":null,"level":3},"content":[{"type":"text","text":"MySQL 中的 Nested Loop Join 实现"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"Nested Loop Join 是扫描驱动表，每读出一条记录，就根据 join 的关联字段上的索引去被驱动表中查询对应数据。它适用于被连接的数据子集较小的场景，它也是 MySQL join 的唯一算法实现，关于它的细节我们接下来会详细讲解。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"MySQL 中有两个 Nested Loop Join 算法的变种，分别是 Index Nested-Loop Join 和 Block Nested-Loop Join。"}]},{"type":"heading","attrs":{"align":null,"level":4},"content":[{"type":"text","text":"Index Nested-Loop Join 算法"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"下面，我们先来初始化一下相关的表结构和数据"}]},{"type":"codeblock","attrs":{"lang":null},"content":[{"type":"text","text":"CREATE TABLE `t1` (\n `id` int(11) NOT NULL,\n `a` int(11) DEFAULT NULL,\n `b` int(11) DEFAULT NULL,\n PRIMARY KEY (`id`),\n KEY `a` (`a`)\n) ENGINE=InnoDB;\n\ndelimiter ;;\n# 定义存储过程来初始化t1\ncreate procedure init_data()\nbegin\n declare i int;\n set i=1;\n while(i<=10000)do\n insert into t1 values(i, i, i);\n set i=i+1;\n end while;\nend;;\ndelimiter ;\n# 调用存储过来来初始化t1\ncall init_data();\n# 创建并初始化t2\ncreate table t2 like t1;\ninsert into t2 (select * from t1 where id<=500)"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"有上述命令可知，这两个表都有一个主键索引 id 和一个索引 a，字段 b 上无索引。存储过程 init_data 往表 t1 里插入了 10000 行数据，在表 t2 里插入的是 500 行数据。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"为了避免 MySQL 优化器会自行选择表作为驱动表，影响分析 SQL 语句的执行过程，我们直接使用 straight_join 来让 MySQL 使用固定的连接表顺序进行查询，如下语句中，t1是驱动表，t2是被驱动表。"}]},{"type":"codeblock","attrs":{"lang":null},"content":[{"type":"text","text":"select * from t2 straight_join t1 on (t2.a=t1.a);"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"使用我们之前"},{"type":"link","attrs":{"href":"https://mp.weixin.qq.com/s/88sGSpVYfGBREH-vZkl_jg","title":null},"content":[{"type":"text","text":"文章"}]},{"type":"text","text":"介绍的 explain 命令查看一下该语句的执行计划。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https://static001.geekbang.org/infoq/01/01dcc43580344b48221015a1a4da8a12.png","alt":"","title":null,"style":null,"href":null,"fromPaste":true,"pastePass":true}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"从上图可以看到，t1 表上的 a 字段是由索引的，join 过程中使用了该索引，因此该 SQL 语句的执行流程如下："}]},{"type":"bulletedlist","content":[{"type":"listitem","content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"从 t2 表中读取一行数据 L1；"}]}]},{"type":"listitem","content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"使用L1 的 a 字段，去 t1 表中作为条件进行查询；"}]}]},{"type":"listitem","content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"取出 t1 中满足条件的行， 跟 L1组成相应的行，成为结果集的一部分；"}]}]},{"type":"listitem","content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"重复执行，直到扫描完 t2 表。"}]}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"这个流程我们就称之为 Index Nested-Loop Join，简称 NLJ，它对应的流程图如下所示。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https://static001.geekbang.org/infoq/1b/1b16faceda497270943ff4efb001802b.png","alt":"Nest-Loop-Join","title":null,"style":null,"href":null,"fromPaste":true,"pastePass":true}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"需要注意的是，在第二步中，根据 a 字段去表t1中查询时，使用了索引，所以每次扫描只会扫描一行(从explain结果得出，根据不同的案例场景而变化)。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"假设驱动表的行数是N，被驱动表的行数是 M。因为在这个 join 语句执行过程中，驱动表是走全表扫描，而被驱动表则使用了索引，并且驱动表中的每一行数据都要去被驱动表中进行索引查询，所以整个 join 过程的近似复杂度是 N"},{"type":"text","marks":[{"type":"italic"}],"text":"2"},{"type":"text","text":"log2M。显然，N 对扫描行数的影响更大，因此这种情况下应该让小表来做驱动表。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"当然，这一切的前提是 join 的关联字段是 a，并且 t1 表的 a 字段上有索引。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"如果没有索引时，再用上图的执行流程时，每次到 t1 去匹配的时候，就要做一次全表扫描。这也导致整个过程的时间复杂度编程了 N * M，这是不可接受的。所以，当没有索引时，MySQL 使用 Block Nested-Loop Join 算法。"}]},{"type":"heading","attrs":{"align":null,"level":4},"content":[{"type":"text","text":"Block Nested-Loop Join"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"Block Nested-Loop Join的算法，简称 BNL，它是 MySQL 在被驱动表上无可用索引时使用的 join 算法，其具体流程如下所示："}]},{"type":"bulletedlist","content":[{"type":"listitem","content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"把表 t2 的数据读取当前线程的 join_buffer 中，在本篇文章的示例 SQL 没有在 t2 上做任何条件过滤，所以就是讲 t2 整张表 放入内存中；"}]}]},{"type":"listitem","content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"扫描表 t1，每取出一行数据，就跟 join_buffer 中的数据进行对比，满足 join 条件的，则放入结果集。"}]}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"比如下面这条 SQL"}]},{"type":"codeblock","attrs":{"lang":null},"content":[{"type":"text","text":"select * from t2 straight_join t1 on (t2.b=t1.b);"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"这条语句的 explain 结果如下所示。可以看出"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https://static001.geekbang.org/infoq/79/796a0e9e13ab9e5a6633f8f5620f8532.png","alt":"","title":null,"style":null,"href":null,"fromPaste":true,"pastePass":true}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"可以看出，这次 join 过程对 t1 和 t2 都做了一次全表扫描，并且将表 t2 中的 500 条数据全部放入内存 join_"},{"type":"text","marks":[{"type":"italic"}],"text":"buffer 中，并且对于表 t1 中的每一行数据，都要去 join_"},{"type":"text","text":"buffer 中遍历一遍，都要做 500 次对比，所以一共要进行 500 * 10000 次内存对比操作，具体流程如下图所示。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https://static001.geekbang.org/infoq/85/85b425b082a129a3ba033fa8fd6509e0.png","alt":"BNL","title":null,"style":null,"href":null,"fromPaste":true,"pastePass":true}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"主要注意的是，"},{"type":"text","marks":[{"type":"strong"}],"text":"第一步中，并不是将表 t2 中的所有数据都放入 join_"},{"type":"text","marks":[{"type":"italic"},{"type":"strong"}],"text":"buffer，而是根据具体的 SQL 语句，而放入不同行的数据和不同的字段"},{"type":"text","marks":[{"type":"italic"}],"text":"。比如下面这条 join 语句则只会将表 t2 中符合 b >= 100 的数据的 b 字段存入 join_"},{"type":"text","text":"buffer。"}]},{"type":"codeblock","attrs":{"lang":null},"content":[{"type":"text","text":"select t2.b,t1.b from t2 straight_join t1 on (t2.b=t1.b) where t2.b >= 100;"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"join_"},{"type":"text","marks":[{"type":"italic"}],"text":"buffer 并不是无限大的，由 join_"},{"type":"text","text":"buffer_size 控制，默认值为 256K。当要存入的数据过大时，就只有分段存储了，整个执行过程就变成了："}]},{"type":"bulletedlist","content":[{"type":"listitem","content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"扫描表 t2，将符合条件的数据行存入 join_buffer，因为其大小有限，存到100行时满了，则执行第二步；"}]}]},{"type":"listitem","content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"扫描表 t1，每取出一行数据，就跟 join_buffer 中的数据进行对比，满足 join 条件的，则放入结果集；"}]}]},{"type":"listitem","content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"清空 join_buffer；"}]}]},{"type":"listitem","content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"再次执行第一步，直到全部数据被扫描完，由于 t2 表中有 500行数据，所以一共重复了 5次"}]}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"这个流程体现了该算法名称中 Block 的由来，分块去执行 join 操作。因为表 t2 的数据被分成了 5 次存入 join_buffer，导致表 t1 要被全表扫描 5次。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https://static001.geekbang.org/infoq/75/756403ba107d22a393d5ebb7c6b9c2fc.png","alt":null,"title":null,"style":null,"href":null,"fromPaste":true,"pastePass":true}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"如上所示，和表数据可以全部存入 join_buffer 相比，内存判断的次数没有变化，都是两张表行数的乘积，也就是 10000 * 500，但是被驱动表会被多次扫描，每多存入一次，被驱动表就要扫描一遍，影响了最终的执行效率。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"基于上述两种算法，我们可以得出下面的结论，这也是网上大多数对 MySQL join 语句的规范。"}]},{"type":"bulletedlist","content":[{"type":"listitem","content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"被驱动表上有索引，也就是可以使用Index Nested-Loop Join 算法时，可以使用 join 操作。"}]}]},{"type":"listitem","content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"无论是Index Nested-Loop Join 算法或者 Block Nested-Loop Join 都要使用小表做驱动表。"}]}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"因为上述两个 join 算法的时间复杂度"},{"type":"text","marks":[{"type":"strong"}],"text":"至少"},{"type":"text","text":"也和涉及表的行数成一阶关系，并且要花费大量的内存空间，所以阿里开发者规范所说的严格禁止三张表以上的 join 操作也是可以理解的了。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"但是上述这两个算法只是 join 的算法之一，还有"},{"type":"text","marks":[{"type":"strong"}],"text":"更加高效的 join 算法，比如 Hash Join 和 Sorted Merged join。可惜这两个算法 MySQL 的主流版本中目前都不提供，而 Oracle ，PostgreSQL 和 Spark 则都支持，这也是网上吐槽 MySQL 弱爆了的原因"},{"type":"text","text":"(MySQL 8.0 版本支持了 Hash join，但是8.0目前还不是主流版本)。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"其实阿里开发者规范也是在从 Oracle 迁移到 MySQL 时，因为 MySQL 的 join 操作性能太差而定下的禁止三张表以上的 join 操作规定的 。"}]},{"type":"heading","attrs":{"align":null,"level":3},"content":[{"type":"text","text":"Hash Join 算法"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"Hash Join 是扫描驱动表，利用 join 的关联字段在内存中建立散列表，然后扫描被驱动表，每读出一行数据，并从散列表中找到与之对应数据。它是大数据集连接操时的常用方式，适用于驱动表的数据量较小，可以放入内存的场景，它对于"},{"type":"text","marks":[{"type":"strong"}],"text":"没有索引的大表"},{"type":"text","text":"和并行查询的场景下能够提供最好的性能。可惜它只适用于等值连接的场景，比如 on a.id = where b.a_id。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"还是上述两张表 join 的语句，其执行过程如下"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https://static001.geekbang.org/infoq/17/1731717846ae8e281865379c410e3b41.png","alt":"Hash-Join","title":null,"style":null,"href":null,"fromPaste":true,"pastePass":true}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"bulletedlist","content":[{"type":"listitem","content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"将驱动表 t2 中符合条件的数据取出，对其每行的 join 字段值进行 hash 操作，然后存入内存中的散列表中；"}]}]},{"type":"listitem","content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"遍历被驱动表 t1，每取出一行符合条件的数据，也对其 join 字段值进行 hash 操作，拿结果到内存的散列表中查找匹配，如果找到，则成为结果集的一部分。"}]}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"可以看出，"},{"type":"text","marks":[{"type":"strong"}],"text":"该算法和 Block Nested-Loop Join 有类似之处，只不过是将无序的 Join Buffer 改为了散列表 hash table，从而让数据匹配不再需要将 join buffer 中的数据全部遍历一遍，而是直接通过 hash，以接近 O(1) 的时间复杂度获得匹配的行"},{"type":"text","text":"，这极大地提高了两张表的 join 速度。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"不过由于 hash 的特性，该算法只能适用于等值连接的场景，其他的连接场景均无法使用该算法。"}]},{"type":"heading","attrs":{"align":null,"level":3},"content":[{"type":"text","text":"Sorted Merge Join 算法"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"Sort Merge Join 则是先根据 join 的关联字段将两张表排序(如果已经排序好了，比如字段上有索引则不需要再排序)，然后在对两张表进行一次归并操作。如果两表已经被排过序，在执行排序合并连接时不需要再排序了，这时Merge Join的性能会优于Hash Join。Merge Join可适于于非等值Join（>，=，<=，但是不包含!=，也即<>）。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"需要注意的是，如果连接的字段已经有索引，也就说已经排好序的话，可以直接进行归并操作，但是如果连接的字段没有索引的话，则它的执行过程如下图所示。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https://static001.geekbang.org/infoq/41/418c22715b31f579506065595ce86001.png","alt":"Sorted_merge_join","title":null,"style":null,"href":null,"fromPaste":true,"pastePass":true}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"bulletedlist","content":[{"type":"listitem","content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"遍历表 t2，将符合条件的数据读取出来，按照连接字段 a 的值进行排序；"}]}]},{"type":"listitem","content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"遍历表 t1，将符合条件的数据读取出来，也按照连接字段 a 的值进行排序；"}]}]},{"type":"listitem","content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"将两个排序好的数据进行归并操作，得出结果集。"}]}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"Sorted Merge Join 算法的主要时间消耗在于对两个表的排序操作，所以如果两个表已经按照连接字段排序过了，该算法甚至比 Hash Join 算法还要快。在一边情况下，该算法是比 Nested Loop Join 算法要快的。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"下面，我们来总结一下上述三种算法的区别和优缺点。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https://static001.geekbang.org/infoq/2c/2c9af716fc68d303e2ea2c9c62a117f8.png","alt":null,"title":null,"style":null,"href":null,"fromPaste":true,"pastePass":true}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"heading","attrs":{"align":null,"level":4},"content":[{"type":"text","text":"对于 Join 操作的理解"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"讲完了 Join 相关的算法，我们这里也聊一聊对于 join 操作的业务理解。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"在业务不复杂的情况下，大多数join并不是无可替代。比如订单记录里一般只有订单用户的 user_id，返回信息时需要取得用户姓名，可能的实现方案有如下几种："}]},{"type":"numberedlist","attrs":{"start":null,"normalizeStart":1},"content":[{"type":"listitem","content":[{"type":"paragraph","attrs":{"indent":0,"number":1,"align":null,"origin":null},"content":[{"type":"text","text":"一次数据库操作，使用 join 操作，订单表和用户表进行 join，连同用户名一起返回；"}]}]},{"type":"listitem","content":[{"type":"paragraph","attrs":{"indent":0,"number":2,"align":null,"origin":null},"content":[{"type":"text","text":"两次数据库操作，分两次查询，第一次获得订单信息和 user"},{"type":"text","marks":[{"type":"italic"}],"text":"id，第二次根据 user"},{"type":"text","text":"id 取姓名，使用代码程序进行信息合并；"}]}]},{"type":"listitem","content":[{"type":"paragraph","attrs":{"indent":0,"number":3,"align":null,"origin":null},"content":[{"type":"text","text":"使用冗余用户名称或者从 ES 等非关系数据库中读取。"}]}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"上述方案都能解决数据聚合的问题，而且基于程序代码来处理，比数据库 join 更容易调试和优化，比如取用户姓名不从数据库中取，而是先从缓存中查找。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"当然， join 操作也不是一无是处，所以技术都有其使用场景，上边这些方案或者规则都是互联网开发团队总结出来的，适用于高并发、轻写重读、分布式、业务逻辑简单的情况，这些场景一般对数据的一致性要求都不高，甚至允许脏读。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"但是，在金融银行或者财务等企业应用场景，join 操作则是不可或缺的，这些应用一般都是低并发、频繁复杂数据写入、CPU密集而非IO密集，主要业务逻辑通过数据库处理甚至包含大量存储过程、对一致性与完整性要求很高的系统。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"link","attrs":{"href":"http://remcarpediem.net/","title":null},"content":[{"type":"text","text":"个人博客，欢迎来玩"}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}}]}