案例：改写SQL消除in-clause

现象描述

in-clause/any-clause是常见的SQL语句约束条件，有时in或any后面的clause都是常量，类似于：

select
count(1)
from calc_empfyc_c1_result_tmp_t1
where ls_pid_cusr1 in (‘20120405’, ‘20130405’);

或者

select
count(1)
from calc_empfyc_c1_result_tmp_t1
where ls_pid_cusr1 in any(‘20120405’, ‘20130405’);

但是也有一些如下的特殊用法：

SELECT
*
FROM test1 t1, test2 t2
WHERE t1.a = any(values(t2.a),(t2.b));

其中，a、b为t2中的两列，“t1.a = any(values(t2.ba,(t2.b))”等价于“t1.a = t2.a or t1.a = t2.b”。

因此join-condition实质上是一个不等式，这种不等值的join操作必须走nestloop，对应执行计划如下：

                                                           QUERY PLAN
---------------------------------------------------------------------------------------------------------------------------------
 Nested Loop  (cost=0.00..138614.38 rows=2309100 width=16) (actual time=0.152..19225.483 rows=1000 loops=1)
   Join Filter: (SubPlan 1)
   Rows Removed by Join Filter: 999000
   ->  Seq Scan on test1 t1  (cost=0.00..31.49 rows=2149 width=8) (actual time=0.021..3.309 rows=1000 loops=1)
   ->  Materialize  (cost=0.00..42.23 rows=2149 width=8) (actual time=0.331..1265.810 rows=1000000 loops=1000)
         ->  Seq Scan on test2 t2  (cost=0.00..31.49 rows=2149 width=8) (actual time=0.013..0.268 rows=1000 loops=1)
   SubPlan 1
     ->  Values Scan on "*VALUES*"  (cost=0.00..0.03 rows=2 width=4) (actual time=2890.741..7372.739 rows=1999000 loops=1000000)
 Total runtime: 19227.328 ms
(9 rows)

优化说明

测试发现由于两表结果集过大，导致nestloop耗时过长，超过一小时未返回结果，因此性能优化的关键是消除nestloop，让join走更高效的hashjoin。从语义等价的角度消除any-clause，SQL改写如下：

SELECT
*
FROM (
    SELECT * FROM test1 t1, test2 t2 WHERE t1.a = t2.a
    UNION
    SELECT * FROM test1 t1, test2 t2 WHERE t1.a = t2.b
);

优化后的SQL查询由两个等值join的子查询构成，而每个子查询都可以走更适合此场景的hashjoin。优化后的执行计划如下

                                                           QUERY PLAN
---------------------------------------------------------------------------------------------------------------------------------
 HashAggregate  (cost=1634.99..2096.81 rows=46182 width=16) (actual time=6.369..6.772 rows=1000 loops=1)
   Group By Key: t1.a, t1.b, t2.a, t2.b
   ->  Append  (cost=58.35..1173.17 rows=46182 width=16) (actual time=0.833..3.414 rows=2000 loops=1)
         ->  Hash Join  (cost=58.35..355.67 rows=23091 width=16) (actual time=0.832..1.590 rows=1000 loops=1)
               Hash Cond: (t1.a = t2.a)
               ->  Seq Scan on test1 t1  (cost=0.00..31.49 rows=2149 width=8) (actual time=0.015..0.156 rows=1000 loops=1)
               ->  Hash  (cost=31.49..31.49 rows=2149 width=8) (actual time=0.531..0.531 rows=1000 loops=1)
                      Buckets: 32768  Batches: 1  Memory Usage: 40kB
                     ->  Seq Scan on test2 t2  (cost=0.00..31.49 rows=2149 width=8) (actual time=0.010..0.199 rows=1000 loops=1)
         ->  Hash Join  (cost=58.35..355.67 rows=23091 width=16) (actual time=0.694..1.421 rows=1000 loops=1)
               Hash Cond: (t1.a = t2.b)
               ->  Seq Scan on test1 t1  (cost=0.00..31.49 rows=2149 width=8) (actual time=0.010..0.160 rows=1000 loops=1)
               ->  Hash  (cost=31.49..31.49 rows=2149 width=8) (actual time=0.524..0.524 rows=1000 loops=1)
                      Buckets: 32768  Batches: 1  Memory Usage: 40kB
                     ->  Seq Scan on test2 t2  (cost=0.00..31.49 rows=2149 width=8) (actual time=0.008..0.177 rows=1000 loops=1)
 Total runtime: 7.759 ms
(16 rows)