导语
最近总是听到用 AI 来优化 SQL 的言论，今天心血来潮试了一下，把表结构、统计信息、SQL 语句、执行计划都告诉 AI，AI 给出了一大堆的建议，它会从索引，语句改写，参数调整各个方面给出优化策略，看似面面俱到，但是如果不懂优化理论，随便使用其给出的优化建议，可能的结果就是，一顿操作猛如虎，一看战绩零杠五。所以本人还是老老实实的总结优化技巧吧，这些案例技巧或许某天会成为 AI 的营养餐。
SQL 案例
SQL 案例语句：(实际业务场景太复杂，截取片段来说明本文主题)
SELECT ta.*, tb.* FROM (SELECT * FROM (SELECT a.contactid, a.subs_number, a.log_time, ROW_NUMBER() OVER(PARTITION BY a.contactid, a.subs_number ORDER BY a.log_time DESC) rn, a.log_id FROM a WHERE a.contactid IS NOT NULL AND a.log_time >= '2025-05-30 00:00:00' AND a.log_time <= '2025-06-02') cc WHERE rn = 1) ta LEFT JOIN (SELECT b.*, ROW_NUMBER() OVER(PARTITION BY b.basesn ORDER BY b.create_time DESC) rn FROM b WHERE b.create_time IS NOT NULL) tb ON ta.contactid = tb.basesn AND tb.rn = 1
复制代码
下面支撑该案例 SQL 的测试表结构，符合案例 SQL 特点的测试数据。
CREATE TABLE a(log_id bigint,CONTACTID INT,subs_number INT,log_time datetime,PRIMARY KEY (log_id),KEY idx_logtime(log_time)); CREATE TABLE b(id bigint PRIMARY KEY,basesn INT,create_time datetime,KEY idx_basesn(basesn)); delimiter // CREATE OR REPLACE PROCEDURE P1() IS BEGIN FOR I IN 1 .. 10000 LOOP INSERT INTO a(log_id,contactid,subs_number,log_time) VALUES(i,TRUNC(rand()*8000),TRUNC(rand()*9000),SYSDATE-rand()*90); END LOOP; FOR I IN 1 .. 1000000 LOOP INSERT INTO b(id,basesn,create_time) VALUES(i,TRUNC(rand()*800000),SYSDATE-rand()*90); END LOOP; END; // delimiter ;
复制代码
两表的统计信息如下：
greatsql> SHOW index FROM a;+-------+------------+-------------+--------------+-------------+-----------+-------------+----------+--------+------+------------+---------+---------------+---------+------------+| Table | Non_unique | Key_name | Seq_in_index | Column_name | Collation | Cardinality | Sub_part | Packed | Null | Index_type | Comment | Index_comment | Visible | Expression |+-------+------------+-------------+--------------+-------------+-----------+-------------+----------+--------+------+------------+---------+---------------+---------+------------+| a | 0 | PRIMARY | 1 | log_id | A | 10000 | NULL | NULL | | BTREE | | | YES | NULL || a | 1 | idx_logtime | 1 | log_time | A | 9990 | NULL | NULL | YES | BTREE | | | YES | NULL |+-------+------------+-------------+--------------+-------------+-----------+-------------+----------+--------+------+------------+---------+---------------+---------+------------+2 rows in set (0.00 sec)
greatsql> SHOW index FROM b;+-------+------------+------------+--------------+-------------+-----------+-------------+----------+--------+------+------------+---------+---------------+---------+------------+| Table | Non_unique | Key_name | Seq_in_index | Column_name | Collation | Cardinality | Sub_part | Packed | Null | Index_type | Comment | Index_comment | Visible | Expression |+-------+------------+------------+--------------+-------------+-----------+-------------+----------+--------+------+------------+---------+---------------+---------+------------+| b | 0 | PRIMARY | 1 | id | A | 916864 | NULL | NULL | | BTREE | | | YES | NULL || b | 1 | idx_basesn | 1 | basesn | A | 515268 | NULL | NULL | YES | BTREE | | | YES | NULL |+-------+------------+------------+--------------+-------------+-----------+-------------+----------+--------+------+------------+---------+---------------+---------+------------+2 rows in set (0.00 sec)
复制代码
语句分析
SQL 有两个派生表 ta，tb，这两表做left join 左外连接。派生表 ta，作为左外连接的左表，内层表 a 有 log_time 过滤条件，该列有单列索引，查询两三天的数据数据量命中几百行，查询一个月左右的数据量命中几千到 1 万左右。派生表 tb，作为左外连接的右表，内层表 b 全表百万级别的数据量，条件create_time is not null过滤性不好。两个派生表都使用了窗口函数ROW_NUMBER()
执行计划分析
语句实际执行计划如下：
EXPLAIN: -> Nested loop left join (cost=22497.56 rows=0) (actual time=6181.328..6182.085 rows=331 loops=1) -> Filter: (cc.rn = 1) (cost=1.21..35.91 rows=30) (actual time=2.712..2.794 rows=331 loops=1) -> Table scan on cc (cost=2.50..2.50 rows=0) (actual time=2.704..2.752 rows=331 loops=1) -> Materialize (cost=0.00..0.00 rows=0) (actual time=2.698..2.698 rows=331 loops=1) -> Window aggregate: row_number() OVER (PARTITION BY a.CONTACTID,a.subs_number ORDER BY a.log_time desc ) (actual time=2.189..2.323 rows=331 loops=1) -> Sort: a.CONTACTID, a.subs_number, a.log_time DESC (cost=149.21 rows=331) (actual time=2.168..2.190 rows=331 loops=1) -> Filter: (a.CONTACTID is not null) (cost=149.21 rows=331) (actual time=0.156..1.847 rows=331 loops=1) -> Index range scan on a using idx_logtime over ('2025-05-30 00:00:00' <= log_time <= '2025-06-02 00:00:00'), with index condition: ((a.log_time >= TIMESTAMP'2025-05-30 00:00:00') and (a.log_time <= TIMESTAMP'2025-06-02 00:00:00')) (cost=149.21 rows=331) (actual time=0.147..1.806 rows=331 loops=1) -> Index lookup on tb using <auto_key0> (basesn=cc.contactid, rn=1) (cost=0.25..766.47 rows=3025) (actual time=18.668..18.668 rows=1 loops=331) -> Materialize (cost=0.00..0.00 rows=0) (actual time=6178.570..6178.570 rows=1000000 loops=1) -> Window aggregate: row_number() OVER (PARTITION BY b.basesn ORDER BY b.create_time desc ) (actual time=2153.616..3469.381 rows=1000000 loops=1) -> Sort: b.basesn, b.create_time DESC (cost=100382.85 rows=998296) (actual time=2153.598..2733.042 rows=1000000 loops=1) -> Filter: (b.create_time is not null) (cost=100382.85 rows=998296) (actual time=0.075..900.074 rows=1000000 loops=1) -> Table scan on b (cost=100382.85 rows=998296) (actual time=0.074..316.051 rows=1000000 loops=1)
1 row in set (6.22 sec)
复制代码
两表 ta，tb 使用Nested loop方式进行连接，ta 表作为外层驱动表，结果集 rows 为 331。 tb 表作为内层循环表，循环扫描 331 次，这些都消耗不多。
此 SQL 耗时多的步骤在对 tb 的内层表 b 进行排序(Sort)，做窗口函数聚合计算(Window aggregate)，再做物化处理(Materialize)这三个步骤了，对一百万的数据做这些处理耗时约 6s，虽然只执行一次，但对 SQL 性能的影响是很大的。现在问题聚焦于能不能减少做这些处理的数据量。
从 b 表的统计信息看，关联字段basesn的选择性不错，本 SQL 最终结果集也只有 331 行，关联字段对 b 表的过滤条件是很好的，当前优化器的行为表现是，因为有窗口函数聚合运算，主查询的关联谓词条件无法推入到 tb 派生表的内部。了解了这一点，想办法改写语句，让关联字段起到过滤作用。
优化方案
这里我想到的解决方案是：对外层查询表的列CONTACTID去重处理，关联到 tb 内层查询中，对满足关联条件的数据做 Sort，Window aggregate，Materialize 这些处理。
为什么增加这一层关联与原语句等价呢，就当作思考题吧，可在评论区评论噢！
语句改写参考如下：
SELECT ta.*, tb.* FROM (SELECT * FROM (SELECT a.contactid, a.subs_number, a.log_time, row_number() OVER(PARTITION BY a.contactid, a.subs_number ORDER BY a.log_time DESC) rn, a.log_id FROM a WHERE a.contactid IS NOT NULL AND a.log_time >= '2025-05-30 00:00:00' AND a.log_time <= '2025-06-02') cc WHERE rn = 1) ta LEFT JOIN (SELECT b.*, row_number() OVER(PARTITION BY b.basesn ORDER BY b.create_time DESC) rn FROM b join (SELECT distinct CONTACTID FROM a WHERE CONTACTID IS NOT NULL AND LOG_TIME >= '2025-05-30 00:00:00' AND LOG_TIME <= '2025-06-02') a1 ON a1.CONTACTID = b.basesn WHERE b.create_time IS NOT NULL) tb ON ta.contactid = tb.basesn AND tb.rn = 1
复制代码
改写后的语句执行计划如下：
EXPLAIN: -> Nested loop left join (cost=111.18 rows=0) (actual time=14.846..15.281 rows=331 loops=1) -> Filter: (cc.rn = 1) (cost=1.21..35.91 rows=30) (actual time=2.668..2.747 rows=331 loops=1) -> Table scan on cc (cost=2.50..2.50 rows=0) (actual time=2.636..2.683 rows=331 loops=1) -> Materialize (cost=0.00..0.00 rows=0) (actual time=2.630..2.630 rows=331 loops=1) -> Window aggregate: row_number() OVER (PARTITION BY a.CONTACTID,a.subs_number ORDER BY a.log_time desc ) (actual time=2.214..2.356 rows=331 loops=1) -> Sort: a.CONTACTID, a.subs_number, a.log_time DESC (cost=149.21 rows=331) (actual time=2.173..2.198 rows=331 loops=1) -> Filter: (a.CONTACTID is not null) (cost=149.21 rows=331) (actual time=0.089..1.784 rows=331 loops=1) -> Index range scan on a using idx_logtime over ('2025-05-30 00:00:00' <= log_time <= '2025-06-02 00:00:00'), with index condition: ((a.log_time >= TIMESTAMP'2025-05-30 00:00:00') and (a.log_time <= TIMESTAMP'2025-06-02 00:00:00')) (cost=149.21 rows=331) (actual time=0.071..1.730 rows=331 loops=1) -> Index lookup on tb using <auto_key0> (basesn=cc.contactid, rn=1) (cost=0.25..2.57 rows=10) (actual time=0.037..0.038 rows=1 loops=331) -> Materialize (cost=0.00..0.00 rows=0) (actual time=12.159..12.159 rows=382 loops=1) -> Window aggregate: row_number() OVER (PARTITION BY b.basesn ORDER BY b.create_time desc ) (actual time=11.614..11.781 rows=382 loops=1) -> Sort: b.basesn, b.create_time DESC (actual time=11.608..11.636 rows=382 loops=1) -> Stream results (cost=237.31 rows=518) (actual time=1.673..11.394 rows=382 loops=1) -> Nested loop inner join (cost=237.31 rows=518) (actual time=1.670..11.247 rows=382 loops=1) -> Filter: (a1.CONTACTID is not null) (cost=214.40..35.91 rows=297) (actual time=1.430..1.545 rows=321 loops=1) -> Table scan on a1 (cost=215.02..221.21 rows=298) (actual time=1.429..1.502 rows=321 loops=1) -> Materialize (cost=215.00..215.00 rows=298) (actual time=1.428..1.428 rows=321 loops=1) -> Table scan on <temporary> (cost=179.02..185.21 rows=298) (actual time=1.303..1.349 rows=321 loops=1) -> Temporary table with deduplication (cost=179.00..179.00 rows=298) (actual time=1.302..1.302 rows=321 loops=1) -> Filter: (a.CONTACTID is not null) (cost=149.21 rows=298) (actual time=0.110..1.143 rows=331 loops=1) -> Index range scan on a using idx_logtime over ('2025-05-30 00:00:00' <= log_time <= '2025-06-02 00:00:00'), with index condition: ((a.log_time >= TIMESTAMP'2025-05-30 00:00:00') and (a.log_time <= TIMESTAMP'2025-06-02 00:00:00')) (cost=149.21 rows=331) (actual time=0.108..1.108 rows=331 loops=1) -> Filter: (b.create_time is not null) (cost=0.48 rows=2) (actual time=0.028..0.030 rows=1 loops=321) -> Index lookup on b using idx_basesn (basesn=a1.CONTACTID) (cost=0.48 rows=2) (actual time=0.027..0.029 rows=1 loops=321)
1 row in set (0.03 sec)
复制代码
可以看出改写后的 SQL 耗时 0.03s，比原来的 6.2s，性能提升了约 200 倍。表面上 SQL 是比原来复杂了一点，但整体执行效率却得到了很大的提升。
总结
SQL 优化的核心思想是减少 I/O 开销，无论什么优化技巧都是围绕这个主题，根据 SQL 具体情况演变出的形形色色的方法而已。万变不离其宗，本案例也是如此。
通过手动改写 SQL，实现谓词下推，减少了内层表需要处理的数据量，从而提升了 SQL 性能。
当然，我们期待 GreatSQL 的优化器能在未来实现这一算法，自动实现谓词下推，不用改动 SQL，即可高效执行 SQL。
无论哪种数据库的优化器，都会或多或少存在一定缺陷，我们优化 DBA 需要做的就是，理解其缺陷，再利用现有资源，帮助其找到好的执行计划，来提升 SQL 性能。
评论