优化器执行计划选择

1 数据准备

CREATE TABLE TEST1 AS SELECT D.* ,trunc(dbms_random.value(0,100)) AS RAN_VAL FROM DBA_OBJECTS D;

2 性能改善

2.1 写法一

SELECT DISTINCT T1.OWNER

  FROM TEST1 T1

 WHERE RAN_VAL

         < (SELECT MAX(T2.RAN_VAL)

              FROM TEST1 T2

             WHERE T1.OWNER = T2.OWNER);

优化器还是很智能的，走了Hash+Filter筛选，避开了Filter连接。

2.2 写法二

SELECT DISTINCT T1.OWNER

FROM TEST1 T1，

     (SELECT T2.OWNER, MAX(T2.RAN_VAL) as MAX_VAL

      FROM TEST1 T2

      GROUP BY T2.OWNER) T3

WHERE T1.OWNER=T3.OWNER

AND T1.RAN_VAL<T3.MAX_VAL;

因为owner列的数据倾斜很大，两表用这种方式连接，结果集类似笛卡儿积。具体参见我的另一篇博客，链接如下：

https://blog.youkuaiyun.com/songjian1104/article/details/91349960

 

2.3 写法三

SELECT DISTINCT T1.OWNER

FROM TEST1 T1，

     (SELECT T2.OWNER, MAX(T2.RAN_VAL) as MAX_VAL

      FROM TEST1 T2

         WHERE ROWNUM>0

      GROUP BY T2.OWNER) T3

WHERE T1.OWNER=T3.OWNER

AND T1.RAN_VAL<T3.MAX_VAL;

我们故意在T3的子查询里面加了个WHERE ROWNUM>0，为了不让T3内联视图展开，避开写法二的执行计划。

这个执行计划与写法一的执行计划类似。

2.4 写法四

SELECT DISTINCT TT.OWNER

FROM (SELECT T1.OWNER, T1.RAN_VAL

             ,MAX(T1.RAN_VAL) OVER(PARTITION BY T1.OWNER) AS MAX_VAL_PEROWNER

      FROM TEST1 T1) TT

WHERE TT.RAN_VAL<TT.MAX_VAL_PEROWNER;

相比之下，这种分析函数改写性能更好，因为减少了扫描次数，同时也避开了自连接。