小编给大家分享一下Oracle里的常见执行计划有哪些,相信大部分人都还不怎么了解,因此分享这篇文章给大家参考一下,希望大家阅读完这篇文章后大有收获,下面让我们一起去了解一下吧!
-
与表访问相关的执行计划
Oracle数据库里面与表访问有关的的两种方法:全表扫描和ROWID扫描。反应在执行计划上,与全表扫描对应的执行计划中的关键字是“TABLE ACCESS FULL”,与ROWID扫描对应的执行计划中的关键字是"TABLE ACCESS BY USER ROWID"或“TABLE ACCESS BY INDEX ROWID”.
我们来看一下与表访问的相关的执行计划,先执行如下SQL:
SQL> select empno,ename,rowid from emp where ename='TURNER';
EMPNO ENAME ROWID
———- ———- ——————
7844 TURNER AAAVREAAEAAAACXAAJ
SQL> select * from table(dbms_xplan.display_cursor(null,null,'ALL'));
PLAN_TABLE_OUTPUT
——————————————————————————–
SQL_ID3bjd8ps607cau, child number 0
————————————-
select empno,ename,rowid from emp where ename='TURNER'
Plan hash value: 3956160932
————————————————————————–
| Id | Operation | Name | Rows | Bytes | Cost (%CPU)| Time |
————————————————————————–
| 0 | SELECT STATEMENT | | | | 3 (100)| |
|* 1 | TABLE ACCESS FULL| EMP | 1 | 22 | 3 (0)| 00:00:01 |
从上述显示内容中可以看出,目标sql的执行计划走的是对表EMP的全表扫描,全表扫描在执行计划中对应的关键字就是“TABLE ACCESS FULL”
将上述sql改写成以指定的ROWID的方式执行:
SQL> select empno,ename from emp where rowid='AAAVREAAEAAAACXAAJ';
EMPNO ENAME
———- ———-
7844 TURNER
SQL> select * from table(dbms_xplan.display_cursor(null,null,'ALL'));
PLAN_TABLE_OUTPUT
——————————————————————————————-
SQL_ID8n08pmh26ud05, child number 0
————————————-
select empno,ename from emp where rowid='AAAVREAAEAAAACXAAJ'
Plan hash value: 1116584662
———————————————————————————–
| Id | Operation | Name | Rows | Bytes | Cost (%CPU)| Time |
———————————————————————————–
| 0 | SELECT STATEMENT | | | |1 (100)| |
| 1 | TABLE ACCESS BY USER ROWID| EMP |1 | 22 |1 (0)| 00:00:01 |
从上述显示内容可以看出,现在该sql的执行计划走的对表emp的rowid扫描,其对应的执行计划关键字“TABLE ACCESS BY USER ROWID”。
注意,ROWID扫描所对应的的执行计划的关键字还有可能是“TABLE ACCESS BY INDEX ROWID”,这取决于访问表时的ROWID来源。如果ROWID是来源于用户手工指定(例如上述指定“select empno,ename from emp where rowid='AAAVREAAEAAAACXAAJ'”),则对应的执行计划关键字是“TABLE ACCESS BY USER ROWID”;如果是ROWID是来源于索引,则对应的执行计划关键字是“TABLE ACCESS BY INDEX ROWID”
表EMP的主键是列EMPNO(即列EMPNO上有主键索引),我们将目标sql改写成如下形式后执行:
SQL> select empno,ename from emp where empno=7369;
EMPNO ENAME
———- ———-
7369 SMITH
SQL> select * from table(dbms_xplan.display_cursor(null,null,'ALL'));
PLAN_TABLE_OUTPUT
——————————————————————————————-
SQL_ID6yzqcfbz5xz3c, child number 0
————————————-
select empno,ename from emp where empno=7369
Plan hash value: 2949544139
————————————————————————————–
| Id | Operation | Name | Rows | Bytes | Cost (%CPU)| Time |
————————————————————————————–
| 0 | SELECT STATEMENT | | | | 1 (100)| |
| 1 | TABLE ACCESS BY INDEX ROWID| EMP | 1 | 10 | 1 (0)| 00:00:01 |
PLAN_TABLE_OUTPUT
——————————————————————————————-
|* 2 | INDEX UNIQUE SCAN | PK_EMP | 1 | | 0 (0)| |
————————————————————————————–
从上述显示内容可以看出,此时目标sql的执行计划在访问表emp走的是对表emp的rowid扫描,因为这里的rowid是来源于索引PK_emp,所以其对应的执行计划关键字是“TABLE ACCESS BY INDEX ROWID”.
2.与B树索引相关的执行计划
常见的与B树索引访问相关的方法有:包括索引唯一扫描,索引范围扫描,索引全扫描,索引快速全扫描和索引跳跃式扫描,反应在执行计划上,与索引唯一扫描对应的关键字“INDEX UNIQUE SCAN”,与索引范围扫描对应的关键字是“INDEX RANGE SCAN”,与索引全扫描对应的关键字是“INDEX FULL SCAN”,与索引快速全扫描对应的关键字是“INDEX FAST FULL SCAN”,与索引跳跃式扫描对应的关键字是“INDEX SKIP SCAN”.
下面来看一下与B树索引访问相关的的执行计划实例。创建一个测试表EMPLOYEE:
SQL> create table employee(gender varchar2(1),employee_id number);
Table created.
SQL> insert into employee values('F','99');
1 row created.
SQL> insert into employee values('F','100');
1 row created.
SQL> insert into employee values('M','101');
1 row created.
SQL> insert into employee values('M','102');
1 row created.
SQL> insert into employee values('M','103');
1 row created.
SQL> insert into employee values('M','104');
1 row created.
SQL> insert into employee values('M','105');
1 row created.
SQL> commit;
Commit complete.
SQL> create unique index idx_uni_emp on employee(employee_id);
Index created.
SQL> select * from employee where employee_id=100;
G EMPLOYEE_ID
– ———–
F 100
SQL> select plan_table_output from table(dbms_xplan.display_cursor(null,null,'ALL'));
PLAN_TABLE_OUTPUT
——————————————————————————————-
SQL_IDbum8qv24s6tqp, child number 0
————————————-
select * from employee where employee_id=100
Plan hash value: 1887894887
——————————————————————————————-
| Id | Operation | Name | Rows | Bytes | Cost (%CPU)| Time |
——————————————————————————————-
| 0 | SELECT STATEMENT | | | |1 (100)| |
| 1 | TABLE ACCESS BY INDEX ROWID| EMPLOYEE |1 | 15 |1 (0)| 00:00:01 |
PLAN_TABLE_OUTPUT
——————————————————————————————-
|* 2 | INDEX UNIQUE SCAN | IDX_UNI_EMP |1 | |0 (0)| |
——————————————————————————————-
从上述显示内容可以看出,此sql的执行计划走的是索引IDX_UNI_EMP的索引唯一扫描,索引唯一扫描在执行计划中对应的关键字就是“INDEX UNIQUE SCAN”.
Drop掉上述唯一索引IDX_UNI_EMP
SQL> drop index idx_uni_emp;
Index dropped.
SQL> create index idx_emp_1 on employee(employee_id);
Index created.
SQL> select * from employee where employee_id=100;
G EMPLOYEE_ID
– ———–
F 100
SQL> select plan_table_output from table(dbms_xplan.display_cursor(null,null,'ALL'));
PLAN_TABLE_OUTPUT
——————————————————————————————-
SQL_IDbum8qv24s6tqp, child number 0
————————————-
select * from employee where employee_id=100
Plan hash value: 2428325319
—————————————————————————————–
| Id | Operation | Name| Rows| Bytes | Cost (%CPU)| Time|
—————————————————————————————–
| 0 | SELECT STATEMENT |||| 2 (100)||
| 1 | TABLE ACCESS BY INDEX ROWID| EMPLOYEE| 1 | 15 | 2 (0)| 00:00:01 |
PLAN_TABLE_OUTPUT
——————————————————————————————-
|* 2 | INDEX RANGE SCAN | IDX_EMP_1 | 1 || 1 (0)| 00:00:01 |
—————————————————————————————–
从上述显示内容中可以看出,现在改sql的执行计划走是对索引idx_emp_1的索引范围扫描,索引范围扫描在执行计划中对应的关键字就是“INDEX RANGE SCAN”。
truncate表EMPLOYEE中的数据:
SQL> truncate table employee;
Table truncated.
更新插入10000条记录:
SQL> begin
2 for i in 1..5000 loop
3 insert into employee values('F',i);
4 end loop;
5 commit;
6 end;
7 /
PL/SQL procedure successfully completed.
SQL> begin
2 for i in 5001..10000 loop
3 insert into employee values('M',i);
4 end loop;
5 commit;
6 end;
7 /
PL/SQL procedure successfully completed.
SQL> select gender,count(*) from employee group by gender;
G COUNT(*)
– ———-
M5000
F5000
对表EMPLOYEE收集一下统计信息:
SQL> exec dbms_stats.gather_table_stats(ownname=>'SCOTT',tabname=>'EMPLOYEE',estimate_percent=>100,cascade=>TRUE,no_invalidate=>false,method_opt=>'FOR ALL COLUMNS SIZE 1');
PL/SQL procedure successfully completed.
SQL> set autotrace traceonly
SQL> show user
USER is "SCOTT"
SQL> select employee_id from employee;
10000 rows selected.
Execution Plan
———————————————————-
Plan hash value: 2119105728
——————————————————————————
| Id | Operation | Name | Rows | Bytes | Cost (%CPU)| Time |
——————————————————————————
| 0 | SELECT STATEMENT | | 10000 | 40000 | 7 (0)| 00:00:01 |
| 1 | TABLE ACCESS FULL| EMPLOYEE | 10000 | 40000 | 7 (0)| 00:00:01 |
——————————————————————————
Statistics
———————————————————-
1 recursive calls
0 db block gets
689 consistent gets
0 physical reads
0 redo size
174308 bytes sent via SQL*Net to client
7850 bytes received via SQL*Net from client
668 SQL*Net roundtrips to/from client
0 sorts (memory)
0 sorts (disk)
10000 rows processed
明明上述SQL查询字段employee_id可以通过扫描索引idx_emp_1得到,但oracle依然选择了对employee的全表扫描。
此时就算我们使用Hint强制让oracle扫描索引idx_emp_1,从如下结果可以看到,oracle依然选择了对表employee的全表扫描(即hint失效了)
SQL> select /* index(employee idx_emp_1)*/employee_id from employee;
10000 rows selected.
Execution Plan
———————————————————-
Plan hash value: 2119105728
——————————————————————————
| Id | Operation | Name | Rows | Bytes | Cost (%CPU)| Time |
——————————————————————————
| 0 | SELECT STATEMENT | | 10000 | 40000 | 7 (0)| 00:00:01 |
| 1 | TABLE ACCESS FULL| EMPLOYEE | 10000 | 40000 | 7 (0)| 00:00:01 |
——————————————————————————
Statistics
———————————————————-
1 recursive calls
0 db block gets
689 consistent gets
0 physical reads
0 redo size
174308 bytes sent via SQL*Net to client
7850 bytes received via SQL*Net from client
668 SQL*Net roundtrips to/from client
0 sorts (memory)
0 sorts (disk)
10000 rows processed
会出现上述现象是因为oracle无论如何总会保证目标sql结果的正确性,可能会得到错误结果的执行路径orale是不会考虑的。对于idx_emp_1而言,它是一个单键值的B树索引,索引NULL值不会存储在其中,那么一旦列employee_id中出现null值(虽然这里实际上并没有null值),则扫描索引IDX_EMP_1的结果就会漏掉那些employee_id为NULL的值,这也就意味这个如果orale在执行上述sql选择了扫描索引idx_emp_1,那么执行结果就可能是不准的。这种情况下,oracle当然不会考虑扫描idx_emp_1,即使我们使用Hint。
如果这里我们想让oracle在执行上述sql扫描索引idx_emp_1,则必须将列employee_id的属性改成not null。这就相当于告诉oracle,这里列employee_id上不会有null值,你就放心的扫描idx_emp_1吧。
SQL> select employee_id from employee;
10000 rows selected.
Execution Plan
———————————————————-
Plan hash value: 3918702848
———————————————————————————-
| Id | Operation | Name | Rows | Bytes | Cost (%CPU)| Time |
———————————————————————————-
| 0 | SELECT STATEMENT | | 10000 | 40000 | 7 (0)| 00:00:01 |
| 1 | INDEX FAST FULL SCAN| IDX_EMP_1 | 10000 | 40000 | 7 (0)| 00:00:01 |
———————————————————————————-
从上述显示内容可以看出,现在sql的执行计划走的是idx_emp_1的索引快速全扫描,索引快速全扫描在执行计划中对应的关键字就是“INDEX FAST FULL SCAN”
现在我们加上强制索引IDX_EMP_1的hint,再次执行该sql:
SQL> select /*+ index(employee idx_emp_1) */ employee_id from employee;
10000 rows selected.
Execution Plan
———————————————————-
Plan hash value: 438557521
——————————————————————————
| Id | Operation | Name | Rows | Bytes | Cost (%CPU)| Time |
——————————————————————————
| 0 | SELECT STATEMENT | | 10000 | 40000 | 20 (0)| 00:00:01 |
| 1 | INDEX FULL SCAN | IDX_EMP_1 | 10000 | 40000 | 20 (0)| 00:00:01 |
——————————————————————————
从上述显示内容可以看出,现在SQL的执行计划走的是对索引idx_emp_1的索引全扫描,索引全扫描在执行计划中对应的关键字就是“INDEX FULL SCAN”.
DROP掉单键值B树索引IDX_EMP_1;
SQL> drop index idx_emp_1;
Index dropped.
SQL> create index index_emp_2 on employee(gender,employee_id);
Index created.
SQL> set autot trace
SQL> select * from employee where employee_id=101;
Execution Plan
———————————————————-
Plan hash value: 2052968723
——————————————————————————–
| Id | Operation | Name | Rows | Bytes | Cost (%CPU)| Time |
——————————————————————————–
| 0 | SELECT STATEMENT | | 1 | 6 | 3 (0)| 00:00:01 |
|* 1 | INDEX SKIP SCAN | INDEX_EMP_2 | 1 | 6 | 3 (0)| 00:00:01 |
——————————————————————————–
Predicate Information (identified by operation id):
—————————————————
1 – access("EMPLOYEE_ID"=101)
filter("EMPLOYEE_ID"=101)
Statistics
———————————————————-
0 recursive calls
0 db block gets
6 consistent gets
0 physical reads
0 redo size
600 bytes sent via SQL*Net to client
524 bytes received via SQL*Net from client
2 SQL*Net roundtrips to/from client
0 sorts (memory)
0 sorts (disk)
1 rows processed
从上述显示内容可以看出,现在sql的执行计划走的是对索引IDX_EMP_2的索引跳跃式扫描,索引跳跃式扫描在执行计划中对应的关键字就是“INDEX SKIP SCAN”.
以上是“Oracle里的常见执行计划有哪些”这篇文章的所有内容,感谢各位的阅读!相信大家都有了一定的了解,希望分享的内容对大家有所帮助,如果还想学习更多知识,欢迎关注亿速云行业资讯频道!
原创文章,作者:carmelaweatherly,如若转载,请注明出处:https://blog.ytso.com/205007.html