基于代价的查询改写 · k8s原理与实践

[TOC] # 基于代价的查询改写 OceanBase 数据库目前只支持一种基于代价的查询改写——或展开（OR-EXPANSION）。数据库中很多高级的改写规则（例如 complex view merge 和窗口函数改写）都需要基于代价进行改写，OceanBase 数据库后续版本会支持这些复杂的改写规则。 ## 或展开（OR-EXPANSION） OR-EXPANSION 是将一个查询改写成若干个用 UNION 组成的子查询，可以为每个子查询提供更优的优化空间，但是也会导致多个子查询的执行，所以这个改写需要基于代价去判断。 OR-EXPANSION 的改写主要有如下三个作用: * 允许每个分支使用不同的索引来加速查询。如下例所示，Q1 会被改写成 Q2 的形式，其中 Q2 中的谓词`LNNVL(t1.a = 1)`保证了这两个子查询不会生成重复的结果。如果不进行改写，Q1 一般来说会选择主表作为访问路径，对于 Q2 来说，如果 t1 上存在索引（a）和索引（b），那么该改写可能会让 Q2 中的每一个子查询选择索引作为访问路径。 ~~~ Q1: obclient>SELECT * FROM t1 WHERE t1.a = 1 OR t1.b = 1; Q2: obclient>SELECT * FROM t1 WHERE t1.a = 1 UNION ALL SELECT * FROM t1.b = 1 AND LNNVL(t1.a = 1); ~~~ 完整示例如下： ~~~ obclient>CREATE TABLE t1(a INT, b INT, c INT, d INT, e INT, INDEX IDX_a(a), INDEX IDX_b(b)); Query OK, 0 rows affected (0.17 sec) /*如果不进行 OR-EXPANSION 的改写，该查询只能使用主表访问路径*/ obclient> EXPLAIN SELECT/*+NO_REWRITE()*/ * FROM t1 WHERE t1.a = 1 OR t1.b = 1; +--------------------------------------------------------------+ | Query Plan | +--------------------------------------------------------------+ | =================================== |ID|OPERATOR |NAME|EST. ROWS|COST| ----------------------------------- |0 |TABLE SCAN|t1 |4 |649 | =================================== Outputs & filters: ------------------------------------- 0 - output([t1.a], [t1.b], [t1.c], [t1.d], [t1.e]), filter([t1.a = 1 OR t1.b = 1]), access([t1.a], [t1.b], [t1.c], [t1.d], [t1.e]), partitions(p0) /*改写之后，每个子查询能使用不同的索引访问路径*/ obclient>EXPLAIN SELECT * FROM t1 WHERE t1.a = 1 OR t1.b = 1; +------------------------------------------------------------------------+ | Query Plan | +------------------------------------------------------------------------+ | ========================================= |ID|OPERATOR |NAME |EST. ROWS|COST| ----------------------------------------- |0 |UNION ALL | |3 |190 | |1 | TABLE SCAN|t1(idx_a)|2 |94 | |2 | TABLE SCAN|t1(idx_b)|1 |95 | ========================================= Outputs & filters: ------------------------------------- 0 - output([UNION(t1.a, t1.a)], [UNION(t1.b, t1.b)], [UNION(t1.c, t1.c)], [UNION(t1.d, t1.d)], [UNION(t1.e, t1.e)]), filter(nil) 1 - output([t1.a], [t1.b], [t1.c], [t1.d], [t1.e]), filter(nil), access([t1.a], [t1.b], [t1.c], [t1.d], [t1.e]), partitions(p0) 2 - output([t1.a], [t1.b], [t1.c], [t1.d], [t1.e]), filter([lnnvl(t1.a = 1)]), access([t1.a], [t1.b], [t1.c], [t1.d], [t1.e]), partitions(p02 ~~~ * 允许每个分支使用不同的连接算法来加速查询，避免使用笛卡尔联接。如下例所示，Q1 会被改写成 Q2 的形式。对于 Q1 来说，它的联接方式只能是 NESTED LOOP JOIN (笛卡尔乘积)，但是被改写之后，每个子查询都可以选择 NESTED LOOP JOIN、HASH JOIN 或者 MERGE JOIN，这样会有更多的优化空间。 ~~~ Q1: obclient>SELECT * FROM t1, t2 WHERE t1.a = t2.a OR t1.b = t2.b; Q2: obclient>SELECT * FROM t1, t2 WHERE t1.a = t2.a UNION ALL SELECT * FROM t1, t2 WHERE t1.b = t2.b AND LNNVL(t1.a = t2.a); ~~~ 完整示例如下： ~~~ obclient> CREATE TABLE t1(a INT, b INT); Query OK, 0 rows affected (0.17 sec) obclient> CREATE TABLE t2(a INT, b INT); Query OK, 0 rows affected (0.13 sec) /*如果不进行改写，只能使用 NESTED LOOP JOIN*/ obclient> EXPLAIN SELECT/*+NO_REWRITE()*/ * FROM t1, t2 WHERE t1.a = t2.a OR t1.b = t2.b; +--------------------------------------------------------------------------+ | Query Plan | +--------------------------------------------------------------------------+ | =========================================== |ID|OPERATOR |NAME|EST. ROWS|COST | ------------------------------------------- |0 |NESTED-LOOP JOIN| |3957 |585457| |1 | TABLE SCAN |t1 |1000 |499 | |2 | TABLE SCAN |t2 |4 |583 | =========================================== Outputs & filters: ------------------------------------- 0 - output([t1.a], [t1.b], [t2.a], [t2.b]), filter(nil), conds(nil), nl_params_([t1.a], [t1.b]) 1 - output([t1.a], [t1.b]), filter(nil), access([t1.a], [t1.b]), partitions(p0) 2 - output([t2.a], [t2.b]), filter([? = t2.a OR ? = t2.b]), access([t2.a], [t2.b]), partitions(p0) /*被改写之后，每个子查询都使用了 HASH JOIN*/ obclient> EXPLAIN SELECT * FROM t1, t2 WHERE t1.a = t2.a OR t1.b = t2.b; +--------------------------------------------------------------------------+ | Query Plan | +--------------------------------------------------------------------------+ |ID|OPERATOR |NAME|EST. ROWS|COST| ------------------------------------- |0 |UNION ALL | |2970 |9105| |1 | HASH JOIN | |1980 |3997| |2 | TABLE SCAN|t1 |1000 |499 | |3 | TABLE SCAN|t2 |1000 |499 | |4 | HASH JOIN | |990 |3659| |5 | TABLE SCAN|t1 |1000 |499 | |6 | TABLE SCAN|t2 |1000 |499 | ===================================== Outputs & filters: ------------------------------------- 0 - output([UNION(t1.a, t1.a)], [UNION(t1.b, t1.b)], [UNION(t2.a, t2.a)], [UNION(t2.b, t2.b)]), filter(nil) 1 - output([t1.a], [t1.b], [t2.a], [t2.b]), filter(nil), equal_conds([t1.a = t2.a]), other_conds(nil) 2 - output([t1.a], [t1.b]), filter(nil), access([t1.a], [t1.b]), partitions(p0) 3 - output([t2.a], [t2.b]), filter(nil), access([t2.a], [t2.b]), partitions(p0) 4 - output([t1.a], [t1.b], [t2.a], [t2.b]), filter(nil), equal_conds([t1.b = t2.b]), other_conds([lnnvl(t1.a = t2.a)]) 5 - output([t1.a], [t1.b]), filter(nil), access([t1.a], [t1.b]), partitions(p0) 6 - output([t2.a], [t2.b]), filter(nil), access([t2.a], [t2.b]), partitions(p0) ~~~ * 允许每个分支分别消除排序，更加快速的获取 TOP-K 结果。如下例所示，Q1 会被改写成 Q2。对于 Q1 来说，执行方式是只能把满足条件的行数找出来，然后进行排序，最终取 TOP-10 结果。对于 Q2 来说，如果存在索引（a,b）, 那么 Q2 中的两个子查询都可以使用索引把排序消除，每个子查询取 TOP-10 结果，然后最终对这 20 行数据排序一下取出最终的 TOP-10 行。 ~~~ Q1: obclient>SELECT * FROM t1 WHERE t1.a = 1 OR t1.a = 2 ORDER BY b LIMIT 10; Q2: obclient>SELECT * FROM (SELECT * FROM t1 WHERE t1.a = 1 ORDER BY b LIMIT 10 UNION ALL SELECT * FROM t1 WHERE t1.a = 2 ORDER BY b LIMIT 10) AS TEMP ORDER BY temp.b LIMIT 10; ~~~ 完整示例如下： ~~~ obclient> CREATE TABLE t1(a INT, b INT, INDEX IDX_a(a, b)); Query OK, 0 rows affected (0.20 sec) /*不改写的话，需要排序最终获取 TOP-K 结果*/ obclient> EXPLAIN SELECT/*+NO_REWRITE()*/ * FROM t1 WHERE t1.a = 1 OR t1.a = 2 ORDER BY b LIMIT 10; +-------------------------------------------------------------------------+ | Query Plan | +-------------------------------------------------------------------------+ | ========================================== |ID|OPERATOR |NAME |EST. ROWS|COST| ------------------------------------------ |0 |LIMIT | |4 |77 | |1 | TOP-N SORT | |4 |76 | |2 | TABLE SCAN|t1(idx_a)|4 |73 | ========================================== Outputs & filters: ------------------------------------- 0 - output([t1.a], [t1.b]), filter(nil), limit(10), offset(nil) 1 - output([t1.a], [t1.b]), filter(nil), sort_keys([t1.b, ASC]), topn(10) 2 - output([t1.a], [t1.b]), filter(nil), access([t1.a], [t1.b]), partitions(p0) /* 进行改写的话，排序算子可以被消除，最终获取 TOP-K 结果*/ obclient>EXPLAIN SELECT * FROM t1 WHERE t1.a = 1 OR t1.a = 2 ORDER BY b LIMIT 10; +-------------------------------------------------------------------------+ | Query Plan | +-------------------------------------------------------------------------+ | =========================================== |ID|OPERATOR |NAME |EST. ROWS|COST| ------------------------------------------- |0 |LIMIT | |3 |76 | |1 | TOP-N SORT | |3 |76 | |2 | UNION ALL | |3 |74 | |3 | TABLE SCAN|t1(idx_a)|2 |37 | |4 | TABLE SCAN|t1(idx_a)|1 |37 | =========================================== Outputs & filters: ------------------------------------- 0 - output([UNION(t1.a, t1.a)], [UNION(t1.b, t1.b)]), filter(nil), limit(10), offset(nil) 1 - output([UNION(t1.a, t1.a)], [UNION(t1.b, t1.b)]), filter(nil), sort_keys([UNION(t1.b, t1.b), ASC]), topn(10) 2 - output([UNION(t1.a, t1.a)], [UNION(t1.b, t1.b)]), filter(nil) 3 - output([t1.a], [t1.b]), filter(nil), access([t1.a], [t1.b]), partitions(p0), limit(10), offset(nil) 4 - output([t1.a], [t1.b]), filter([lnnvl(t1.a = 1)]), access([t1.a], [t1.b]), partitions(p0), limit(10), offset(nil) ~~~