标题比较长，不过只有这样才能把问题描述清楚。
问题详细描述为，在FIRST_ROWS优化模式下，将会话排序和比较方式设置为语义模式，即忽略大小写模式，对字段进行LIKE模糊查询，可能导致错误的结果。
关于大小写不敏感的查询的详细描述，可以参考：http://yangtingkun.itpub.net/post/468/460324
下面直接看问题的现象：
SQL> CREATE TABLE T1 (ID NUMBER PRIMARY KEY, NAME VARCHAR2(30));
表已创建。
SQL> CREATE INDEX IND_T1_NAME ON T1(NAME);
索引已创建。
SQL> INSERT INTO T1 SELECT ROWNUM, CHR(64 + ROWNUM)
2 FROM ALL_OBJECTS WHERE ROWNUM <= 26;
已创建26行。
SQL> COMMIT;
提交完成。
SQL> ALTER SESSION SET NLS_COMP = LINGUISTIC;
会话已更改。
SQL> ALTER SESSION SET NLS_SORT = BINARY_CI;
会话已更改。
SQL> SELECT * FROM T1 WHERE NAME LIKE 'a%';
ID NAME
---------- ------------------------------
1 A
SQL> SELECT /*+ FIRST_ROWS */ * FROM T1 WHERE NAME LIKE 'a%';
未选定行
只要修改上面提到的关键点中的任意一个，就不会产生这个错误的现象：
SQL> SELECT /*+ ALL_ROWS */ * FROM T1 WHERE NAME LIKE 'a%';
ID NAME
---------- ------------------------------
1 A
SQL> SELECT /*+ FIRST_ROWS */ * FROM T1 WHERE NAME = 'a';
ID NAME
---------- ------------------------------
1 A
SQL> ALTER SESSION SET NLS_SORT = BINARY;
会话已更改。
SQL> ALTER SESSION SET NLS_COMP = BINARY;
会话已更改。
SQL> SELECT /*+ FIRST_ROWS */ * FROM T1 WHERE NAME LIKE 'A%';
ID NAME
---------- ------------------------------
1 A
SQL> ALTER SESSION SET NLS_COMP = LINGUISTIC;
会话已更改。
SQL> ALTER SESSION SET NLS_SORT = BINARY_CI;
会话已更改。
SQL> SELECT /*+ FIRST_ROWS */ * FROM T1 WHERE NAME LIKE 'A%';
ID NAME
---------- ------------------------------
1 A
SQL> SELECT /*+ FIRST_ROWS */ * FROM T1 WHERE NAME LIKE 'a';
未选定行
通过上面的几个查询可以看到，问题和FIRST_ROWS，LIKE操作以及基于语义的排序直接相关，下面看看Oracle在异常情况下采用了何种执行计划：
SQL> SET AUTOT ON EXP
SQL> SELECT /*+ FIRST_ROWS */ * FROM T1 WHERE NAME LIKE 'a';
未选定行
执行计划
----------------------------------------------------------
Plan hash value: 3350237141
-------------------------------------------------------------------------------------------
| Id | Operation | Name | Rows | Bytes | Cost (%CPU)| Time |
-------------------------------------------------------------------------------------------
| 0 | SELECT STATEMENT | | 1 | 30 | 35 (0)| 00:00:01 |
|* 1 | VIEW | index$_join$_001 | 1 | 30 | 35 (0)| 00:00:01 |
|* 2 | HASH JOIN | | | | | |
|* 3 | INDEX RANGE SCAN | IND_T1_NAME | 1 | 30 | 3 (34)| 00:00:01 |
| 4 | INDEX FAST FULL SCAN| SYS_C006622 | 1 | 30 | 33 (0)| 00:00:01 |
-------------------------------------------------------------------------------------------
Predicate Information (identified by operation id):
---------------------------------------------------
1 - filter("NAME" LIKE 'a')
2 - access(ROWID=ROWID)
3 - access("NAME" LIKE 'a')
Note
-----
- dynamic sampling used for this statement
由于索引中并不包含语义查询的结果，因此Oracle这里必须访问表才能得到最终的结果，因此这个执行计划是错误的：
SQL> SELECT * FROM T1 WHERE NAME = 'a';
ID NAME
---------- ------------------------------
1 A
执行计划
----------------------------------------------------------
Plan hash value: 3617692013
--------------------------------------------------------------------------
| Id | Operation | Name | Rows | Bytes | Cost (%CPU)| Time |
--------------------------------------------------------------------------
| 0 | SELECT STATEMENT | | 1 | 30 | 3 (0)| 00:00:01 |
|* 1 | TABLE ACCESS FULL| T1 | 1 | 30 | 3 (0)| 00:00:01 |
--------------------------------------------------------------------------
Predicate Information (identified by operation id):
---------------------------------------------------
1 - filter(NLSSORT("NAME",'nls_sort=''BINARY_CI''')=HEXTORAW('6100')
)
Note
-----
- dynamic sampling used for this statement
SQL> SELECT /*+ INDEX(T1) */ * FROM T1 WHERE NAME = 'a';
ID NAME
---------- ------------------------------
1 A
执行计划
----------------------------------------------------------
Plan hash value: 159298173
-------------------------------------------------------------------------------------------
| Id | Operation | Name | Rows | Bytes | Cost (%CPU)| Time |
-------------------------------------------------------------------------------------------
| 0 | SELECT STATEMENT | | 1 | 30 | 827 (1)| 00:00:10 |
|* 1 | TABLE ACCESS BY INDEX ROWID| T1 | 1 | 30 | 827 (1)| 00:00:10 |
| 2 | INDEX FULL SCAN | SYS_C006622 | 26 | | 26 (0)| 00:00:01 |
-------------------------------------------------------------------------------------------
Predicate Information (identified by operation id):
---------------------------------------------------
1 - filter(NLSSORT("NAME",'nls_sort=''BINARY_CI''')=HEXTORAW('6100') )
Note
-----
- dynamic sampling used for this statement
上面的两个执行计划已经说明了问题的关键，Oracle对于语义的排序无法通过索引获取，必须要访问表或者相应的函数索引，详细描述可以参考文章开头部分给出的链接。
而采用了FIRST_ROWS优化模式后，当操作为LIKE时，Oracle优化器选择了错误的执行计划进行了优化，采用索引的范围扫描代替了表，从而引发了错误：
SQL> SELECT /*+ INDEX_JOIN(T1 IND_T1_NAME SYS_C006622) */ *
2 FROM T1
3 WHERE NAME LIKE 'a';
未选定行
执行计划
----------------------------------------------------------
Plan hash value: 3350237141
-------------------------------------------------------------------------------------------
| Id | Operation | Name | Rows | Bytes | Cost (%CPU)| Time |
-------------------------------------------------------------------------------------------
| 0 | SELECT STATEMENT | | 1 | 30 | 35 (0)| 00:00:01 |
|* 1 | VIEW | index$_join$_001 | 1 | 30 | 35 (0)| 00:00:01 |
|* 2 | HASH JOIN | | | | | |
|* 3 | INDEX RANGE SCAN | IND_T1_NAME | 1 | 30 | 3 (34)| 00:00:01 |
| 4 | INDEX FAST FULL SCAN| SYS_C006622 | 1 | 30 | 33 (0)| 00:00:01 |
-------------------------------------------------------------------------------------------
Predicate Information (identified by operation id):
---------------------------------------------------
1 - filter("NAME" LIKE 'a')
2 - access(ROWID=ROWID)
3 - access("NAME" LIKE 'a')
Note
-----
- dynamic sampling used for this statement
现在没有使用FIRST_ROWS，而采用HINT也达到了相同的效果。而解决这个问题的方法就是通过HINT来避免索引范围扫描的发生。
SQL> ALTER SESSION SET OPTIMIZER_MODE = FIRST_ROWS;
会话已更改。
SQL> SELECT * FROM T1 WHERE NAME LIKE 'a';
未选定行
执行计划
----------------------------------------------------------
Plan hash value: 3350237141
-------------------------------------------------------------------------------------------
| Id | Operation | Name | Rows | Bytes | Cost (%CPU)| Time |
-------------------------------------------------------------------------------------------
| 0 | SELECT STATEMENT | | 1 | 30 | 35 (0)| 00:00:01 |
|* 1 | VIEW | index$_join$_001 | 1 | 30 | 35 (0)| 00:00:01 |
|* 2 | HASH JOIN | | | | | |
|* 3 | INDEX RANGE SCAN | IND_T1_NAME | 1 | 30 | 3 (34)| 00:00:01 |
| 4 | INDEX FAST FULL SCAN| SYS_C006622 | 1 | 30 | 33 (0)| 00:00:01 |
-------------------------------------------------------------------------------------------
Predicate Information (identified by operation id):
---------------------------------------------------
1 - filter("NAME" LIKE 'a')
2 - access(ROWID=ROWID)
3 - access("NAME" LIKE 'a')
Note
-----
- dynamic sampling used for this statement
SQL> SELECT /*+ FULL(T1) */ * FROM T1 WHERE NAME LIKE 'a';
ID NAME
---------- ------------------------------
1 A
执行计划
----------------------------------------------------------
Plan hash value: 3617692013
--------------------------------------------------------------------------
| Id | Operation | Name | Rows | Bytes | Cost (%CPU)| Time |
--------------------------------------------------------------------------
| 0 | SELECT STATEMENT | | 1 | 30 | 3 (0)| 00:00:01 |
|* 1 | TABLE ACCESS FULL| T1 | 1 | 30 | 3 (0)| 00:00:01 |
--------------------------------------------------------------------------
Predicate Information (identified by operation id):
---------------------------------------------------
1 - filter("NAME" LIKE 'a')
Note
-----
- dynamic sampling used for this statement
SQL> SELECT /*+ NO_INDEX(T1) */ * FROM T1 WHERE NAME LIKE 'a';
ID NAME
---------- ------------------------------
1 A
执行计划
----------------------------------------------------------
Plan hash value: 3617692013
--------------------------------------------------------------------------
| Id | Operation | Name | Rows | Bytes | Cost (%CPU)| Time |
--------------------------------------------------------------------------
| 0 | SELECT STATEMENT | | 1 | 30 | 3 (0)| 00:00:01 |
|* 1 | TABLE ACCESS FULL| T1 | 1 | 30 | 3 (0)| 00:00:01 |
--------------------------------------------------------------------------
Predicate Information (identified by operation id):
---------------------------------------------------
1 - filter("NAME" LIKE 'a')
Note
-----
- dynamic sampling used for this statement
查询metalink，Oracle在Doc ID: Note:5252496.8明确说明了这个bug，这个bug会在Oracle10.2.0.4和11.1.0.6中被Fixed。