今天看到一条SQL大量使用了with，开发人员为了逻辑清晰，把一些结果集先用with缓存起来，后面还有很多地方用到这个结果集，原始的SQL需要执行2个多小时。优化方法是将把最先缓存的SQL放到用的地方,优化后12s。
下面来模拟这个场景，不用纠结SQL的意义，把当时的SQL抽象就是这样的。可以看到SQL1中先把语句a中的结果缓存起来，当语句b要用的时候，object_id上的索引是用不到的，它只有在结果集中过滤。简单点说，SQL1相对于SQL2来说，谓词没有推进。所以在SQL的写法上，追求书面上的清晰和性能要做一个平衡。
--制造数据
SQL> drop table test purge;
SQL> create table test as select * from dba_objects;
SQL> create index ind_t_object_id on test(object_id) nologging;
SQL> exec dbms_stats.gather_table_stats(user,'test',cascade => true);
SQL> set autotrace traceonly
--SQL1，优化前没有谓词推进
SQL> with a as(select * from test where object_type='TABLE'),
b as (select count(1) from a where object_id<10),
c as (select count(1) from a where object_id>=10 and object_id<20)
select (select * from b) bc,
(select * from c) cc
from dual;
执行计划
----------------------------------------------------------
Plan hash value: 2659770981
----------------------------------------------------------------------------------------------
| Id | Operation | Name | Rows | Bytes | Cost (%CPU)| Time |
----------------------------------------------------------------------------------------------
| 0 | SELECT STATEMENT | | 1 | | 198 (1)| 00:00:03 |
| 1 | VIEW | | 1 | 13 | 6 (0)| 00:00:01 |
| 2 | SORT AGGREGATE | | 1 | 13 | | |
|* 3 | VIEW | | 1600 | 20800 | 6 (0)| 00:00:01 |
| 4 | TABLE ACCESS FULL | SYS_TEMP_0FD9D66| 1600 | 154K| 6 (0)| 00:00:01 |
| 5 | VIEW | | 1 | 13 | 6 (0)| 00:00:01 |
| 6 | SORT AGGREGATE | | 1 | 13 | | |
|* 7 | VIEW | | 1600 | 20800 | 6 (0)| 00:00:01 |
| 8 | TABLE ACCESS FULL | SYS_TEMP_0FD9D66| 1600 | 154K| 6 (0)| 00:00:01 |
| 9 | TEMP TABLE TRANSFORMATION | | | | | |
| 10 | LOAD AS SELECT | SYS_TEMP_0FD9D66| | | | |
|* 11 | TABLE ACCESS FULL | TEST | 1600 | 154K| 196 (1)| 00:00:03 |
| 12 | FAST DUAL | | 1 | | 2 (0)| 00:00:01 |
----------------------------------------------------------------------------------------------
Predicate Information (identified by operation id):
---------------------------------------------------
3 - filter("OBJECT_ID"<10)
7 - filter("OBJECT_ID">=10 AND "OBJECT_ID"<20)
11 - filter("OBJECT_TYPE"='TABLE')
统计信息
----------------------------------------------------------
394 recursive calls
22 db block gets
589 consistent gets
15 physical reads
600 redo size
381 bytes sent via SQL*Net to client
337 bytes received via SQL*Net from client
2 SQL*Net roundtrips to/from client
0 sorts (memory)
0 sorts (disk)
1 rows processed