本文梳理、归纳了在之前工作中常见的一些执行计划。了解ORACLE中有哪些可能的执行计划,以及什么情况下适合哪种执行计划是进行SQL优化的基础。

表访问相关

Full Table Scans

 全表扫描首先会读取表中的所有行,然后过滤掉不满足条件的数据。全表扫描时,数据库会以此读取HWM下的所有格式化了的数据块,此时数据库通常会做multiblock read来提高性能,单次读取的数据块由DB_FILE_MULTIBLOCK_READ_COUNT参数指定。 以下情况会做全表扫描:

  1. 查询列上不存在索引
  2. 在索引列上使用了函数
  3. 执行SELECT COUNT(*)语句,存在索引,但是索引包含空值
  4. 未使用B-TREE索引的前导列。如存在employees(first_name,last_name)的索引,但是查询条件为WHERE last_name='KING'。但是优化器有可能选择index skip scan
  5. 查询选择性很低的时候
  6. 统计信息陈旧
  7. 当表很小的时候,包含的数据块数n小于DB_FILE_MULTIBLOCK_READ_COUNT参数指定的值
  8. 当表具有很高的并行度的时候
  9. 使用了FULL的hints

 下面是一个做全表扫描的列子,其执行计划的关键字为'TABLE ACCESS FULL':

SQL> select owner,table_name from test_env.tb_table_list where owner='AUDSYS';

OWNER                TABLE_NAME
-------------------- ------------------------------
AUDSYS               AUD$UNIFIED

SQL> select * from table(dbms_xplan.display_cursor(null, null, 'basic'));

PLAN_TABLE_OUTPUT
--------------------------------------------------------------------------------------------

EXPLAINED SQL STATEMENT:
------------------------
select owner,table_name from test_env.tb_table_list where owner='AUDSYS'

Plan hash value: 1475094007

-------------------------------------------
| Id  | Operation         | Name          |
-------------------------------------------
|   0 | SELECT STATEMENT  |               |
|   1 |  TABLE ACCESS FULL| TB_TABLE_LIST |

PLAN_TABLE_OUTPUT
--------------------------------------------------------------------------------------------

Table Access By Rowid

 ROWID是数据在数据库中存储位置的内部表示,它是用来定位单个行最快的方式。通常数据库通过索引检索数据行或者指定rowid查询的时候会使用这种访问方式。

SQL> select owner,table_name from test_env.tb_table_list where TABLE_NAME='ACCESS$';

OWNER                TABLE_NAME
-------------------- ------------------------------
SYS                  ACCESS$

SQL> select * from table(dbms_xplan.display_cursor(null, null, 'basic'));

PLAN_TABLE_OUTPUT
-----------------------------------------------------------------------------------------------------

EXPLAINED SQL STATEMENT:
------------------------
select owner,table_name from test_env.tb_table_list where
TABLE_NAME='ACCESS$'

Plan hash value: 3473397811

------------------------------------------------------------------------
| Id  | Operation                           | Name                     |
------------------------------------------------------------------------
|   0 | SELECT STATEMENT                    |                          |

PLAN_TABLE_OUTPUT
-----------------------------------------------------------------------------------------------------

|   1 |  TABLE ACCESS BY INDEX ROWID BATCHED| TB_TABLE_LIST            |
|   2 |   INDEX RANGE SCAN                  | IDX_TB_TABLE_LIST_TBNAME |
------------------------------------------------------------------------

TABLE ACCESS BY INDEX ROWID BATCHED表示数据库通过索引中得到的rowid来检索数据,BATCHED的访问方式,表示数据库会从索引中检索一批ROWID,然后按块顺序访问行,减少访问数据块的次数来提升性能。这是ORACLE 12C的一个新特性。在11g中上面的执行计划表示为TABLE ACCESS BY INDEX ROWID

Sample Table Scans

 这是抽样检索数据的数据访问方式。使用SAMPLE关键字对表中数据进行抽样的时候会使用该执行计划,在执行计划中表现为TABLE ACCESS SAMPLE关键字。它有以下两种形式:

  1. SAMPLE (sample_percent)数据库读取表中指定百分比的行数据。
  2. SAMPLE BLOCK (sample_percent)数据库会读取指定百分比的表数据块。

 sample_percent的百分比在[0.000001,100) 范围内。

B-tree索引相关

Index Unique Scans

 只有通过CREATE UNIQUE INDEX创建唯一索引,并且在查询的时候谓词条件为等于的时候才会以该方式访问数据。唯一约束(unique和primary key),但是创建的非唯一索引是不足以让查询走Index Unique Scan的。在执行计划中的关键字是INDEX UNIQUE SCAN

SQL> CREATE UNIQUE INDEX TEST_ENV.IDX_TB_TABLE_LIST_SID ON TEST_ENV.TB_TABLE_LIST(SID);

索引已创建。

SQL> SELECT SID,OWNER,TABLE_NAME FROM TEST_ENV.TB_TABLE_LIST WHERE SID='58223719640640AB8C603BC1B15D5C51';

SID                                  OWNER                TABLE_NAME
------------------------------------ -------------------- ------------------------------
58223719640640AB8C603BC1B15D5C51     SYS                  ACCESS$

SQL> select * from table(dbms_xplan.display_cursor(null, null, 'basic'));

PLAN_TABLE_OUTPUT
-----------------------------------------------------------------------------------------------------------------

EXPLAINED SQL STATEMENT:
------------------------
SELECT SID,OWNER,TABLE_NAME FROM TEST_ENV.TB_TABLE_LIST WHERE
SID='58223719640640AB8C603BC1B15D5C51'

Plan hash value: 3115192837

-------------------------------------------------------------
| Id  | Operation                   | Name                  |
-------------------------------------------------------------
|   0 | SELECT STATEMENT            |                       |

PLAN_TABLE_OUTPUT
-----------------------------------------------------------------------------------------------------------------

|   1 |  TABLE ACCESS BY INDEX ROWID| TB_TABLE_LIST         |
|   2 |   INDEX UNIQUE SCAN         | IDX_TB_TABLE_LIST_SID |
-------------------------------------------------------------


已选择 15 行。

Index Range Scans

 通过该索引字段查询结果可能返回多个值的时候,例如>, <, and以及对非唯一索引的=,就会以该方式访问数据,表现在执行计划上就是INDEX RANGE SCAN关键字。

SQL> select owner,table_name from test_env.tb_table_list where TABLE_NAME='ACCESS$';

OWNER                TABLE_NAME
-------------------- ------------------------------
SYS                  ACCESS$

SQL> select * from table(dbms_xplan.display_cursor(null, null, 'basic'));

PLAN_TABLE_OUTPUT
-----------------------------------------------------------------------------------------------------

EXPLAINED SQL STATEMENT:
------------------------
select owner,table_name from test_env.tb_table_list where
TABLE_NAME='ACCESS$'

Plan hash value: 3473397811

------------------------------------------------------------------------
| Id  | Operation                           | Name                     |
------------------------------------------------------------------------
|   0 | SELECT STATEMENT                    |                          |

PLAN_TABLE_OUTPUT
-----------------------------------------------------------------------------------------------------

|   1 |  TABLE ACCESS BY INDEX ROWID BATCHED| TB_TABLE_LIST            |
|   2 |   INDEX RANGE SCAN                  | IDX_TB_TABLE_LIST_TBNAME |
------------------------------------------------------------------------

Index Full Scans

 索引全扫描会有序的读取整个索引,由于索引是有序的,所以它会消除额外的排序操作。 以下情况会优化器会考虑执行索引全扫描:

  • 谓词使用了索引中的列。该列不必是索引的前导列
  • 没有谓词条件,但是查询的列为该索引的列,并且至少有一个列不为空
  • 查询包含ORDER BY 语句并且排序字段为非空且有索引
SQL> select table_name from test_env.tb_table_list order by table_name;
......

TABLE_NAME
------------------------------
XSTREAM$_SERVER_CONNECTION
XSTREAM$_SUBSET_RULES
XSTREAM$_SYSGEN_OBJS
_default_auditing_options_

已选择 2138 行。
SQL> select * from table(dbms_xplan.display_cursor(null, null, 'basic'));

PLAN_TABLE_OUTPUT
--------------------------------------------------------------------------------------

EXPLAINED SQL STATEMENT:
------------------------
select table_name from test_env.tb_table_list order by table_name

Plan hash value: 2901892796

-----------------------------------------------------
| Id  | Operation        | Name                     |
-----------------------------------------------------
|   0 | SELECT STATEMENT |                          |
|   1 |  INDEX FULL SCAN | IDX_TB_TABLE_LIST_TBNAME |

PLAN_TABLE_OUTPUT
--------------------------------------------------------------------------------------

Index Fast Full Scans

 当查询的列仅仅包含索引列,且不需要排序的时候,优化器会考虑该访问方式,它会以磁盘存储位置来读取,不会保证数据的有序性。

SQL> select table_name from test_env.tb_table_list;

TABLE_NAME
------------------------------
XSTREAM$_SERVER_CONNECTION
XSTREAM$_SUBSET_RULES
XSTREAM$_SYSGEN_OBJS
_default_auditing_options_

已选择 2138 行。

SQL> select * from table(dbms_xplan.display_cursor(null, null, 'basic'));

PLAN_TABLE_OUTPUT
---------------------------------------------------------------------------------

EXPLAINED SQL STATEMENT:
------------------------
select table_name from test_env.tb_table_list

Plan hash value: 3670592075

---------------------------------------------------------
| Id  | Operation            | Name                     |
---------------------------------------------------------
|   0 | SELECT STATEMENT     |                          |
|   1 |  INDEX FAST FULL SCAN| IDX_TB_TABLE_LIST_TBNAME |

PLAN_TABLE_OUTPUT
---------------------------------------------------------------------------------

Index Skip Scans

 在复合索引中未使用前导列的情况,并且前导列只存在很少的distinct value。其在执行计划中表现为INDEX SKIP SCAN

SQL> create index test_env.idx_tb_table_list_mul on test_env.tb_table_list(owner,table_name);

索引已创建。
SQL> select /*+ index(t idx_tb_table_list_mul) */sid,owner,table_name,status from test_env.tb_table_list t where table_name='ACCESS$';

SID                                  OWNER                TABLE_NAME                     STATUS
------------------------------------ -------------------- ------------------------------ --------
58223719640640AB8C603BC1B15D5C51     SYS                  ACCESS$                        VALID

SQL> select * from table(dbms_xplan.display_cursor(null, null, 'basic'));

PLAN_TABLE_OUTPUT
------------------------------------------------------------------------------------------------------------------------

EXPLAINED SQL STATEMENT:
------------------------
select /*+ index(t idx_tb_table_list_mul) */sid,owner,table_name,status
from test_env.tb_table_list t where table_name='ACCESS$'

Plan hash value: 124701251

---------------------------------------------------------------------
| Id  | Operation                           | Name                  |
---------------------------------------------------------------------
|   0 | SELECT STATEMENT                    |                       |

PLAN_TABLE_OUTPUT
------------------------------------------------------------------------------------------------------------------------

|   1 |  TABLE ACCESS BY INDEX ROWID BATCHED| TB_TABLE_LIST         |
|   2 |   INDEX SKIP SCAN                   | IDX_TB_TABLE_LIST_MUL |
---------------------------------------------------------------------

Index Join Scans

 对同一个表,多个索引的hash join。当查询结果都是索引中的列,并且存在于不同索引中时会考虑使用该数据访问方式,但是如果访问表的效率更高则会根据rowid去访问表。 Index join的成本是非常的,绝大多数情况下是不会出现这种情况,通过索引去探测表通常成本更低。

SQL> select owner,table_name from test_env.tb_table_list where ini_trans>=4;

OWNER           TABLE_NAME
--------------- ------------------------------
MDSYS           SDO_DIST_METADATA_TABLE
SYS             AW$AWCREATE
SYS             AW$AWCREATE10G
SYS             AW$AWMD
SYS             AW$AWREPORT
SYS             AW$AWXML
SYS             AW$EXPRESS
SYS             AW_OBJ$
SYS             AW_PROP$
SYS             OBJECT_USAGE
SYS             PS$

OWNER           TABLE_NAME
--------------- ------------------------------
SYS             STREAMS$_APPLY_PROGRESS

已选择 12 行。

SQL> SELECT * FROM TABLE(DBMS_XPLAN.DISPLAY_CURSOR(NULL, NULL, 'basic'));

PLAN_TABLE_OUTPUT
-----------------------------------------------------------------------------------------------

EXPLAINED SQL STATEMENT:
------------------------
select owner,table_name from test_env.tb_table_list where ini_trans>=4

Plan hash value: 832528447

---------------------------------------------------------------------
| Id  | Operation                     | Name                        |
---------------------------------------------------------------------
|   0 | SELECT STATEMENT              |                             |
|   1 |  VIEW                         | index$_join$_001            |

PLAN_TABLE_OUTPUT
-----------------------------------------------------------------------------------------------

|   2 |   HASH JOIN                   |                             |
|   3 |    BITMAP CONVERSION TO ROWIDS|                             |
|   4 |     BITMAP INDEX RANGE SCAN   | BIDX_TB_TABLE_LIST_INITRANS |
|   5 |    INDEX FAST FULL SCAN       | IDX_TB_TABLE_LIST_MUL       |
---------------------------------------------------------------------


已选择 17 行。

Bitmap索引相关

 传统的B-tree索引,一条索引对应一条数据。Bitmap的一个索引的键值通常对应一批rowid。 Bitmap索引适用于拥有较低distinct基数并且不经常修改的数据。Bitmap索引不适用于经常进行DML操作的列,因为一个index key指向很多数据行,当对一个索引的列进行修改时,会锁住整个索引条目以及对应的数据行。 还需要注意的一点就是Bitmap索引列是可以包含空值的。

Bitmap Conversion to Rowid

 只要从位图索引中检索行,就会用到该访问方式,进行数据行与bitmap之间的转换。

SQL> select owner,table_name,status,ini_trans from test_env.tb_table_list where ini_trans>=4;

OWNER           TABLE_NAME                     STATUS      INI_TRANS
--------------- ------------------------------ ---------- ----------
SYS             PS$                            VALID               4
SYS             AW_OBJ$                        VALID               4
SYS             AW_PROP$                       VALID               4
SYS             AW$EXPRESS                     VALID               4
SYS             AW$AWMD                        VALID               4
SYS             AW$AWCREATE                    VALID               4
SYS             AW$AWCREATE10G                 VALID               4
SYS             AW$AWXML                       VALID               4
SYS             AW$AWREPORT                    VALID               4
SYS             OBJECT_USAGE                   VALID              30
SYS             STREAMS$_APPLY_PROGRESS        VALID             120

OWNER           TABLE_NAME                     STATUS      INI_TRANS
--------------- ------------------------------ ---------- ----------
MDSYS           SDO_DIST_METADATA_TABLE        VALID             255

已选择 12 行。

SQL> SELECT * FROM TABLE(DBMS_XPLAN.DISPLAY_CURSOR(NULL, NULL, 'basic'));

PLAN_TABLE_OUTPUT
-----------------------------------------------------------------------------------------------

EXPLAINED SQL STATEMENT:
------------------------
select owner,table_name,status,ini_trans from test_env.tb_table_list
where ini_trans>=4

Plan hash value: 3195249317

---------------------------------------------------------------------------
| Id  | Operation                           | Name                        |
---------------------------------------------------------------------------
|   0 | SELECT STATEMENT                    |                             |

PLAN_TABLE_OUTPUT
-----------------------------------------------------------------------------------------------

|   1 |  TABLE ACCESS BY INDEX ROWID BATCHED| TB_TABLE_LIST               |
|   2 |   BITMAP CONVERSION TO ROWIDS       |                             |
|   3 |    BITMAP INDEX RANGE SCAN          | BIDX_TB_TABLE_LIST_INITRANS |
---------------------------------------------------------------------------


已选择 16 行。

Bitmap Index Single Value

 使用bitmap索引的单个键值来查询数据的时候会走该访问方式,在执行计划中表现为BITMAP INDEX SINGLE VALUE关键字。

SQL> SELECT OWNER,TABLE_NAME,TABLESPACE_NAME,STATUS FROM TEST_ENV.TB_TABLE_LIST WHERE TABLESPACE_NAME='SYSTEM';

....
OWNER           TABLE_NAME                     TABLESPACE_NAME      STATUS
--------------- ------------------------------ -------------------- ----------
LBACSYS         OLS$USER_LEVELS                SYSTEM               VALID
LBACSYS         OLS$USER_COMPARTMENTS          SYSTEM               VALID
LBACSYS         OLS$USER_GROUPS                SYSTEM               VALID
LBACSYS         OLS$PROFILES                   SYSTEM               VALID
LBACSYS         OLS$DIP_DEBUG                  SYSTEM               VALID
LBACSYS         OLS$DIP_EVENTS                 SYSTEM               VALID
LBACSYS         OLS$AUDIT                      SYSTEM               VALID
LBACSYS         OLS$AUDIT_ACTIONS              SYSTEM               VALID
SYS             DBMS_SQLPATCH_STATE            SYSTEM               VALID
SYS             DBMS_SQLPATCH_FILES            SYSTEM               VALID
SYS             AQ_SRVNTFN_TABLE_1             SYSTEM               VALID

OWNER           TABLE_NAME                     TABLESPACE_NAME      STATUS
--------------- ------------------------------ -------------------- ----------
SYS             REGISTRY$SQLPATCH_RU_INFO      SYSTEM               VALID

已选择 912 行。
SQL> SELECT * FROM TABLE(DBMS_XPLAN.DISPLAY_CURSOR(NULL, NULL, 'basic'));

PLAN_TABLE_OUTPUT
-------------------------------------------------------------------------------------------

EXPLAINED SQL STATEMENT:
------------------------
SELECT OWNER,TABLE_NAME,TABLESPACE_NAME,STATUS FROM
TEST_ENV.TB_TABLE_LIST WHERE TABLESPACE_NAME='SYSTEM'

Plan hash value: 4282437356

-----------------------------------------------------------------------------
| Id  | Operation                           | Name                          |
-----------------------------------------------------------------------------
|   0 | SELECT STATEMENT                    |                               |

PLAN_TABLE_OUTPUT
-------------------------------------------------------------------------------------------

|   1 |  TABLE ACCESS BY INDEX ROWID BATCHED| TB_TABLE_LIST                 |
|   2 |   BITMAP CONVERSION TO ROWIDS       |                               |
|   3 |    BITMAP INDEX SINGLE VALUE        | BIDX_TB_TABLE_LIST_TABLESPACE |
-----------------------------------------------------------------------------


已选择 16 行。

Bitmap Index Range Scans

 当谓词条件在bitmap索引列上是一个范围值的时候,如Bitmap Conversion to Rowid中的例子,它在执行计划中表现为BITMAP INDEX RANGE SCAN关键字。

Bitmap Merge

 此访问路径合并多个Bitmap,并返回单个Bitmap作为结果。在执行计划中表现为BITMAP MERGE关键字

SQL> select owner,table_name,status,ini_trans from test_env.tb_table_list where ini_trans>=4 and tablespace_name='SYSTEM';

OWNER           TABLE_NAME                     STATUS      INI_TRANS
--------------- ------------------------------ ---------- ----------
SYS             OBJECT_USAGE                   VALID              30

SQL> SELECT * FROM TABLE(DBMS_XPLAN.DISPLAY_CURSOR(NULL, NULL, 'basic'));

PLAN_TABLE_OUTPUT
------------------------------------------------------------------------------------------------------------------------

EXPLAINED SQL STATEMENT:
------------------------
select owner,table_name,status,ini_trans from test_env.tb_table_list
where ini_trans>=4 and tablespace_name='SYSTEM'

Plan hash value: 1400915856

-----------------------------------------------------------------------------
| Id  | Operation                           | Name                          |
-----------------------------------------------------------------------------
|   0 | SELECT STATEMENT                    |                               |

PLAN_TABLE_OUTPUT
------------------------------------------------------------------------------------------------------------------------

|   1 |  TABLE ACCESS BY INDEX ROWID BATCHED| TB_TABLE_LIST                 |
|   2 |   BITMAP CONVERSION TO ROWIDS       |                               |
|   3 |    BITMAP AND                       |                               |
|   4 |     BITMAP INDEX SINGLE VALUE       | BIDX_TB_TABLE_LIST_TABLESPACE |
|   5 |     BITMAP MERGE                    |                               |
|   6 |      BITMAP INDEX RANGE SCAN        | BIDX_TB_TABLE_LIST_INITRANS   |
-----------------------------------------------------------------------------


已选择 19 行。

表连接相关

 在进行多表关联查询的时候,优化器需要考虑以下内容:

  • 数据访问方式同单表查询一样,优化器必须选择是全表扫描还是通过索引来检索表中数据
  • 关联方法即如何对表进行关联。可能的关联方式有,hash join,nested loops,sorted merge。
  • 关联类型关联条件决定了关联的类型,如inner join,outer join等
  • 关联的顺序在进行表关联的时候,优化器需要决定先关联哪些表,哪个表作为驱动表。基本思想就是尽快的过滤掉尽可能多的数据。

Nested Loops Joins

 嵌套循环连接,可以理解为一个嵌套的for循环,通过外部表的每一条数据去匹配内部表的数据:

FOR erow IN outer_table LOOP
  FOR drow IN inner_table LOOP
    return match value
  END LOOP
END LOOP

 嵌套循环连接适用于以下情况:

  • 数据集很小的时候
  • 在FIRST_ROW的优化器模式下关联大表
  • 关联条件可以高效的访问内部表

 一般情况下,只有在数据集比较小并且关联条件存在索引的时候会使用该方式。也可以通过加入以下hints的方式来强制走嵌套训话连接。:

  • USE_NL_WITH_INDEX(table index) 由优化器决定那个表作为驱动表,index是可选的,不指定则由优化器决定
  • ORDERED USE_NL(d) 指定内部表
SQL> select /*+ORDERED USE_NL(B) */A.owner,A.table_name,A.column_name from TEST_ENV.TB_COLUMN_LIST A,TEST_ENV.TB_TABLE_LIST B WHERE A.TABLE_NAME=B.TABLE_NAME AND B.TABLE_NAME='ACCESS$';

OWNER           TABLE_NAME                     COLUMN_NAME
--------------- ------------------------------ --------------------
SYS             ACCESS$                        D_OBJ#
SYS             ACCESS$                        ORDER#
SYS             ACCESS$                        COLUMNS
SYS             ACCESS$                        TYPES

SQL> SELECT * FROM TABLE(DBMS_XPLAN.DISPLAY_CURSOR(NULL, NULL, 'basic'));

PLAN_TABLE_OUTPUT
------------------------------------------------------------------------------------------------------------------------

EXPLAINED SQL STATEMENT:
------------------------
select /*+ORDERED USE_NL(B) */A.owner,A.table_name,A.column_name from
TEST_ENV.TB_COLUMN_LIST A,TEST_ENV.TB_TABLE_LIST B WHERE
A.TABLE_NAME=B.TABLE_NAME AND B.TABLE_NAME='ACCESS$'

Plan hash value: 1684280275

----------------------------------------------------------------------
| Id  | Operation                            | Name                  |
----------------------------------------------------------------------

PLAN_TABLE_OUTPUT
------------------------------------------------------------------------------------------------------------------------

|   0 | SELECT STATEMENT                     |                       |
|   1 |  NESTED LOOPS                        |                       |
|   2 |   TABLE ACCESS BY INDEX ROWID BATCHED| TB_COLUMN_LIST        |
|   3 |    INDEX RANGE SCAN                  | TB_COLUMN_LIST_TBNAME |
|   4 |   INDEX RANGE SCAN                   | TB_TABLE_LIST_TBNAME  |
----------------------------------------------------------------------


已选择 18 行。

Hash Joins

 这是大数据集关联的常见方式,也是我们日常见得最多的关联方式。优化器会选择两个数据集中较小的一个表,然后用关联字段来构建hash表,存放在PGA中(如果PGA大小不够存放该hash表,则会将部分数据放入到临时表中),然后,数据库扫描较大的数据集,探测哈希表以查找满足连接条件的数据行。 同意通过USE_HASH的hints来强制数据库走HASH连接。

SQL> select /*+USE_HASH(A B) */A.owner,A.table_name,A.column_name from TEST_ENV.TB_COLUMN_LIST A,TEST_ENV.TB_TABLE_LIST
B WHERE A.TABLE_NAME=B.TABLE_NAME AND A.OWNER=B.OWNER AND B.TABLE_NAME='ACCESS$';

OWNER           TABLE_NAME                     COLUMN_NAME
--------------- ------------------------------ --------------------
SYS             ACCESS$                        D_OBJ#
SYS             ACCESS$                        ORDER#
SYS             ACCESS$                        COLUMNS
SYS             ACCESS$                        TYPES

SQL> SELECT * FROM TABLE(DBMS_XPLAN.DISPLAY_CURSOR(NULL, NULL, 'basic'));

PLAN_TABLE_OUTPUT
------------------------------------------------------------------------------------------------------------------------

EXPLAINED SQL STATEMENT:
------------------------
select /*+USE_HASH(A B) */A.owner,A.table_name,A.column_name from
TEST_ENV.TB_COLUMN_LIST A,TEST_ENV.TB_TABLE_LIST B WHERE
A.TABLE_NAME=B.TABLE_NAME AND A.OWNER=B.OWNER AND B.TABLE_NAME='ACCESS$'

Plan hash value: 209128112

----------------------------------------------------------------------
| Id  | Operation                            | Name                  |
----------------------------------------------------------------------

PLAN_TABLE_OUTPUT
------------------------------------------------------------------------------------------------------------------------

|   0 | SELECT STATEMENT                     |                       |
|   1 |  HASH JOIN                           |                       |
|   2 |   TABLE ACCESS BY INDEX ROWID BATCHED| TB_TABLE_LIST         |
|   3 |    INDEX RANGE SCAN                  | TB_TABLE_LIST_TBNAME  |
|   4 |   TABLE ACCESS BY INDEX ROWID BATCHED| TB_COLUMN_LIST        |
|   5 |    INDEX RANGE SCAN                  | TB_COLUMN_LIST_TBNAME |
----------------------------------------------------------------------


已选择 19 行。

Sorted Merge Joins

 这是Nested Loops的一个变种。数据库会首先执行SORT JOIN操作对两个数据集进行排序,然后遍历外部数据集的每一行,去匹配内部数据集,第二次开始后面每次的匹配位置,取决于前一次迭代的匹配。 以下情况会选择进行Sorted Merge:

  • 连接条件不是等值连接,如>,<>
  • 由于其他操作需要排序,优化器发现使用排序合并成本更低。
SQL> select A.owner,A.table_name,A.column_name from TEST_ENV.TB_COLUMN_LIST A,TEST_ENV.TB_TABLE_LIST B WHERE A.TABLE_NAME=B.TABLE_NAME AND A.OWNER=B.OWNER AND B.TABLE_NAME='ACCESS$' ORDER BY TABLE_NAME;

....

OWNER           TABLE_NAME                     COLUMN_NAME
--------------- ------------------------------ --------------------
SYS             XSTREAM$_SYSGEN_OBJS           SPARE5
SYS             XSTREAM$_SYSGEN_OBJS           SPARE4
SYS             XSTREAM$_SYSGEN_OBJS           SPARE3
SYS             XSTREAM$_SYSGEN_OBJS           SPARE2
SYS             XSTREAM$_SYSGEN_OBJS           SPARE6
SYS             XSTREAM$_SYSGEN_OBJS           OBJECT_TYPE
SYS             XSTREAM$_SYSGEN_OBJS           OBJECT_NAME
SYS             XSTREAM$_SYSGEN_OBJS           OBJECT_OWNER
SYS             XSTREAM$_SYSGEN_OBJS           SERVER_NAME
SYS             XSTREAM$_SYSGEN_OBJS           SPARE1
SYS             _default_auditing_options_     A

已选择 23104 行。

SQL> SELECT * FROM TABLE(DBMS_XPLAN.DISPLAY_CURSOR(NULL, NULL, 'basic'));

PLAN_TABLE_OUTPUT
------------------------------------------------------------------------------------------------------------

EXPLAINED SQL STATEMENT:
------------------------
select A.owner,A.table_name,A.column_name from TEST_ENV.TB_COLUMN_LIST
A,TEST_ENV.TB_TABLE_LIST B WHERE A.TABLE_NAME=B.TABLE_NAME AND
A.OWNER=B.OWNER ORDER BY TABLE_NAME

Plan hash value: 3270676555

---------------------------------------------------
| Id  | Operation              | Name             |
---------------------------------------------------

PLAN_TABLE_OUTPUT
------------------------------------------------------------------------------------------------------------

|   0 | SELECT STATEMENT       |                  |
|   1 |  MERGE JOIN            |                  |
|   2 |   SORT JOIN            |                  |
|   3 |    INDEX FAST FULL SCAN| PK_TB_TABLE_LIST |
|   4 |   SORT JOIN            |                  |
|   5 |    TABLE ACCESS FULL   | TB_COLUMN_LIST   |
---------------------------------------------------


已选择 19 行。

查询转换相关

 我们发送给ORACLE的目标SQL与最终执行的SQL有可能是不同的,oracle有可能执行一系列查询转换的操作(如视图合并,子查询展开等)将SQL改写成语义等价的其他形式,是否执行查询转换取决于ORACLE对转换后SQL执行效率的评估。

OR扩展

 OR扩展会将OR语句改写为UNION-ALL的形式,这样各个分支各自走索引、分区修剪、表连接等互不干扰。在之前版本中使用的CONCATENATION来执行OR扩展,12.2开始采用了UNION-ALL的方式。

SQL> select A.owner,A.table_name,B.STATUS,A.column_name from TEST_ENV.TB_COLUMN_LIST A,TEST_ENV.TB_TABLE_LIST B
2 WHERE A.TABLE_NAME=B.TABLE_NAME AND A.OWNER=B.OWNER AND (B.TABLE_NAME='ACCESS$' OR B.OWNER='XDB');

......

OWNER           TABLE_NAME                     STATUS     COLUMN_NAME
--------------- ------------------------------ ---------- --------------------
XDB             XDB_INDEX_DDL_CACHE            VALID      CONSTR_OWNER
SYS             ACCESS$                        VALID      TYPES
SYS             ACCESS$                        VALID      D_OBJ#
SYS             ACCESS$                        VALID      ORDER#
SYS             ACCESS$                        VALID      COLUMNS

已选择 170 行。

SQL> SELECT * FROM TABLE(DBMS_XPLAN.DISPLAY_CURSOR(NULL, NULL, 'BASIC'));

PLAN_TABLE_OUTPUT
------------------------------------------------------------------------------------------------------------------------

EXPLAINED SQL STATEMENT:
------------------------
select A.owner,A.table_name,B.STATUS,A.column_name from
TEST_ENV.TB_COLUMN_LIST A,TEST_ENV.TB_TABLE_LIST B WHERE
A.TABLE_NAME=B.TABLE_NAME AND A.OWNER=B.OWNER AND
(B.TABLE_NAME='ACCESS$' OR B.OWNER='XDB')

Plan hash value: 1155047104

-------------------------------------------------------------------------
| Id  | Operation                               | Name                  |

PLAN_TABLE_OUTPUT
------------------------------------------------------------------------------------------------------------------------

-------------------------------------------------------------------------
|   0 | SELECT STATEMENT                        |                       |
|   1 |  NESTED LOOPS                           |                       |
|   2 |   NESTED LOOPS                          |                       |
|   3 |    VIEW                                 | VW_JF_SET$ADB40ABC    |
|   4 |     UNION-ALL                           |                       |
|   5 |      TABLE ACCESS BY INDEX ROWID BATCHED| TB_TABLE_LIST         |
|   6 |       INDEX RANGE SCAN                  | PK_TB_TABLE_LIST      |
|   7 |      TABLE ACCESS BY INDEX ROWID BATCHED| TB_TABLE_LIST         |
|   8 |       INDEX RANGE SCAN                  | TB_TABLE_LIST_TBNAME  |
|   9 |    INDEX RANGE SCAN                     | TB_COLUMN_LIST_TBNAME |

PLAN_TABLE_OUTPUT
------------------------------------------------------------------------------------------------------------------------

|  10 |   TABLE ACCESS BY INDEX ROWID           | TB_COLUMN_LIST        |
-------------------------------------------------------------------------


已选择 25 行。

视图合并

 将视图合并到包含它的查询块中。进行试图合并后优化器可以考虑更多的连接顺序、访问方式和其他的查询转换,以此来获取更好的性能。视图合并又分为简单视图合并和复杂视图合并。1.简单视图合并 对于简单试图,视图合并总是会带来更好的执行计划,所以数据库总是会进行合并而不会考虑成本。 存在以下情况时不会进行简单视图合并:

  • 视图中包含以下结构时:GROUP BY,DISTINCT,OUTER JOIN, MODEL, CONNECT BY, 集合操作, 聚合
  • 视图出现在半连接或者反连接右侧时
  • 视图出现在SELECT中的子查询
  • 外部查询块包含PL/SQL函数

 下面是一个执行了视图合并的例子:

SQL> select A.owner,A.table_name,V_B.STATUS,A.column_name from TEST_ENV.TB_COLUMN_LIST A,
    (SELECT OWNER,TABLE_NAME,STATUS FROM TEST_ENV.TB_TABLE_LIST) V_B
    WHERE A.TABLE_NAME=V_B.TABLE_NAME AND A.OWNER=V_B.OWNER;

SQL> SELECT * FROM TABLE(DBMS_XPLAN.DISPLAY_CURSOR(NULL,NULL,'BASIC'))
PLAN_TABLE_OUTPUT
-------------------------------------------------------------------------------------------

EXPLAINED SQL STATEMENT:
------------------------
select A.owner,A.table_name,V_B.STATUS,A.column_name from
TEST_ENV.TB_COLUMN_LIST A, (SELECT OWNER,TABLE_NAME,STATUS FROM
TEST_ENV.TB_TABLE_LIST) V_B WHERE A.TABLE_NAME=V_B.TABLE_NAME AND
A.OWNER=V_B.OWNER

Plan hash value: 4162551876

---------------------------------------------
| Id  | Operation          | Name           |
---------------------------------------------
|   0 | SELECT STATEMENT   |                |
|   1 |  HASH JOIN         |                |
|   2 |   TABLE ACCESS FULL| TB_TABLE_LIST  |
|   3 |   TABLE ACCESS FULL| TB_COLUMN_LIST |
---------------------------------------------


已选择 18 行。

然后加个no_merge的hint取消视图合并,对比下执行计划。

SQL>  select A.owner,A.table_name,V_B.STATUS,A.column_name from TEST_ENV.TB_COLUMN_LIST A,
      (SELECT /*+no_merge*/OWNER,TABLE_NAME,STATUS FROM TEST_ENV.TB_TABLE_LIST) V_B
      WHERE A.TABLE_NAME=V_B.TABLE_NAME AND A.OWNER=V_B.OWNER;

SQL> select * from table(dbms_xplan.display_cursor(null, null, 'basic'));

PLAN_TABLE_OUTPUT
-------------------------------------------------------------------------------------

EXPLAINED SQL STATEMENT:
------------------------
select A.owner,A.table_name,V_B.STATUS,A.column_name from
TEST_ENV.TB_COLUMN_LIST A, (SELECT /*+no_merge*/OWNER,TABLE_NAME,STATUS
FROM TEST_ENV.TB_TABLE_LIST) V_B WHERE A.TABLE_NAME=V_B.TABLE_NAME AND
A.OWNER=V_B.OWNER

Plan hash value: 3859876297

----------------------------------------------
| Id  | Operation           | Name           |

PLAN_TABLE_OUTPUT
-------------------------------------------------------------------------------------

----------------------------------------------
|   0 | SELECT STATEMENT    |                |
|   1 |  HASH JOIN          |                |
|   2 |   VIEW              |                |
|   3 |    TABLE ACCESS FULL| TB_TABLE_LIST  |
|   4 |   TABLE ACCESS FULL | TB_COLUMN_LIST |
----------------------------------------------


已选择 19 行。

 可以看到第二个执行计划出现了view关键字,表示会把V_B这个视图进行单独的处理。

2.复杂视图合并 复杂视图合并会合并包含GROUP BY和DISTINCT操作的视图,优化器会把GROUP BY和DISTINCT操作延迟到连接之后。是否进行复杂视图合并,取决于优化器对合并后性能的评估。抛开成本问题,以下情况不会进行复杂视图合并:

  • 外部查询表没有rowid或者unique column
  • 视图出现在CONNECT BY查询中
  • 视图包含 GROUPING SETS, ROLLUP, 或PIVOT
  • 视图或者外部查询包含MODEL
SQL> (SELECT OWNER,TABLE_NAME,COUNT(*) AS CNT_COLS FROM TEST_ENV.TB_COLUMN_LIST GROUP BY OWNER,TABLE_NAME) V_A,
  2  TEST_ENV.TB_TABLE_LIST B
  3  WHERE V_A.TABLE_NAME=B.TABLE_NAME AND V_A.OWNER=B.OWNER AND B.OWNER='XDB';
SQL> select * from table(dbms_xplan.display_cursor(null, null, 'basic'));

PLAN_TABLE_OUTPUT
----------------------------------------------------------------------------------------

EXPLAINED SQL STATEMENT:
------------------------
SELECT B.OWNER,B.TABLE_NAME,B.TABLESPACE_NAME,B.STATUS,V_A.CNT_COLS
FROM (SELECT A.OWNER,A.TABLE_NAME,COUNT(*) AS CNT_COLS FROM
TEST_ENV.TB_COLUMN_LIST A GROUP BY A.OWNER,A.TABLE_NAME) V_A,
TEST_ENV.TB_TABLE_LIST B WHERE V_A.TABLE_NAME=B.TABLE_NAME AND
V_A.OWNER=B.OWNER AND B.OWNER='XDB'

Plan hash value: 3165065995

------------------------------------------------------------------------

PLAN_TABLE_OUTPUT
----------------------------------------------------------------------------------------

| Id  | Operation                              | Name                  |
------------------------------------------------------------------------
|   0 | SELECT STATEMENT                       |                       |
|   1 |  HASH GROUP BY                         |                       |
|   2 |   NESTED LOOPS                         |                       |
|   3 |    NESTED LOOPS                        |                       |
|   4 |     TABLE ACCESS BY INDEX ROWID BATCHED| TB_TABLE_LIST         |
|   5 |      INDEX RANGE SCAN                  | PK_TB_TABLE_LIST      |
|   6 |     INDEX RANGE SCAN                   | TB_COLUMN_LIST_TBNAME |
|   7 |    TABLE ACCESS BY INDEX ROWID         | TB_COLUMN_LIST        |
------------------------------------------------------------------------

PLAN_TABLE_OUTPUT
----------------------------------------------------------------------------------------

已选择 23 行。

 可以看到执行计划中已经没有VIEW关键字了,并且GROUP BY操作延迟到了嵌套循环连接之后。

谓词推入

 优化器处理带视图的目标SQL的另一种优化手段。数据库可以将推入的谓词来访问索引或者作为过滤条件,进而走基于索引的嵌套循环连接。 还是上面的列子,现在加一个no_merger的hint,避免做视图合并。VIEW PUSHED PREDICATE表明此时做的是谓词推入。

SQL> SELECT B.OWNER,B.TABLE_NAME,B.TABLESPACE_NAME,B.STATUS,V_A.CNT_COLS FROM
(SELECT /*+no_merge*/A.OWNER,A.TABLE_NAME,COUNT(*) AS CNT_COLS FROM TEST_ENV.TB_COLUMN_LIST A GROUP BY A.OWNER,A.TABLE_NAME) V_A,
TEST_ENV.TB_TABLE_LIST B
WHERE V_A.TABLE_NAME=B.TABLE_NAME AND V_A.OWNER=B.OWNER
AND B.OWNER='XDB';

SQL> SELECT * FROM TABLE(DBMS_XPLAN.DISPLAY_CURSOR(NULL, NULL, 'BASIC'));

PLAN_TABLE_OUTPUT
---------------------------------------------------------------------------------------

EXPLAINED SQL STATEMENT:
------------------------
SELECT B.OWNER,B.TABLE_NAME,B.TABLESPACE_NAME,B.STATUS,V_A.CNT_COLS
FROM (SELECT /*+no_merge*/A.OWNER,A.TABLE_NAME,COUNT(*) AS CNT_COLS
FROM TEST_ENV.TB_COLUMN_LIST A GROUP BY A.OWNER,A.TABLE_NAME) V_A,
TEST_ENV.TB_TABLE_LIST B WHERE V_A.TABLE_NAME=B.TABLE_NAME AND
V_A.OWNER=B.OWNER AND B.OWNER='XDB'

Plan hash value: 466935718

------------------------------------------------------------------------

PLAN_TABLE_OUTPUT
---------------------------------------------------------------------------------------

| Id  | Operation                              | Name                  |
------------------------------------------------------------------------
|   0 | SELECT STATEMENT                       |                       |
|   1 |  NESTED LOOPS                          |                       |
|   2 |   TABLE ACCESS BY INDEX ROWID BATCHED  | TB_TABLE_LIST         |
|   3 |    INDEX RANGE SCAN                    | PK_TB_TABLE_LIST      |
|   4 |   VIEW PUSHED PREDICATE                |                       |
|   5 |    SORT GROUP BY                       |                       |
|   6 |     TABLE ACCESS BY INDEX ROWID BATCHED| TB_COLUMN_LIST        |
|   7 |      INDEX RANGE SCAN                  | TB_COLUMN_LIST_TBNAME |
------------------------------------------------------------------------

PLAN_TABLE_OUTPUT
---------------------------------------------------------------------------------------



已选择 23 行。

子查询展开

 在子查询展开中优化器会把嵌套的子查询转换成一个关联查询语句。只有当JOIN语句返回的数据行与原始SQL返回数据行相同,并且子查询中不包含聚合函数(如AVG)时,优化器才会执行此转换。

SQL> SELECT OWNER,TABLE_NAME,COLUMN_NAME FROM TEST_ENV.TB_TABLE_LIST WHERE
  2  (OWNER,TABLE_NAME) IN (SELECT OWNER, TABLE_NAME FROM TEST_ENV.TB_TABLE_LIST);

SQL> SELECT * FROM TABLE(DBMS_XPLAN.DISPLAY_CURSOR(NULL, NULL, 'BASIC'));

PLAN_TABLE_OUTPUT
--------------------------------------------------------------------------------------------

EXPLAINED SQL STATEMENT:
------------------------
SELECT OWNER,TABLE_NAME,COLUMN_NAME FROM TEST_ENV.TB_COLUMN_LIST WHERE
(OWNER,TABLE_NAME) IN (SELECT OWNER, TABLE_NAME FROM
TEST_ENV.TB_TABLE_LIST)

Plan hash value: 2699376560

--------------------------------------------------
| Id  | Operation             | Name             |
--------------------------------------------------

PLAN_TABLE_OUTPUT
--------------------------------------------------------------------------------------------

|   0 | SELECT STATEMENT      |                  |
|   1 |  HASH JOIN            |                  |
|   2 |   INDEX FAST FULL SCAN| PK_TB_TABLE_LIST |
|   3 |   TABLE ACCESS FULL   | TB_COLUMN_LIST   |
--------------------------------------------------


已选择 17 行。

 从上面的执行计划可以看出,执行子查询展开后,表现为两个表进行HASH JOIN。我们可以使用no_unnest的hint来让优化器不对该SQL执行子查询展开。此时的执行计划如下:

SQL> SELECT * FROM TABLE(DBMS_XPLAN.DISPLAY_CURSOR(NULL, NULL, 'BASIC'));

PLAN_TABLE_OUTPUT
--------------------------------------------------------------------------------

EXPLAINED SQL STATEMENT:
------------------------
select owner,table_name,column_name from test_env.tb_column_list where
(owner,table_name) in(select /*+no_unnest*/ owner,table_name from
test_env.tb_table_list)

Plan hash value: 2714665686

-----------------------------------------------
| Id  | Operation          | Name             |
-----------------------------------------------

PLAN_TABLE_OUTPUT
--------------------------------------------------------------------------------

|   0 | SELECT STATEMENT   |                  |
|   1 |  FILTER            |                  |
|   2 |   TABLE ACCESS FULL| TB_COLUMN_LIST   |
|   3 |   INDEX UNIQUE SCAN| PK_TB_TABLE_LIST |
-----------------------------------------------


已选择 17 行。

 不进行子查询展开后的执行计划走的是FILTER的类型的过滤了。

表扩展

 基于索引可以提高查询的性能,但是维护索引也会产生开销。在一个大的分区表中(特别是采用业务时间做PARTITION KEY的分区表),经常DML操作比较频繁的就几个分区,其他分区大多数时候都做的全表检索的时间比较多。在类似这种场景下我们就可以只在比较活跃的分区上面创建索引,而其他分区则不创建。优化器会自动的帮我们在有索引的分区上面通过索引检索数据,没有索引的分区进行全表扫描,然后再把结果进行UNION-ALL操作。 下面是ORACLE官网的一个例子:

  1. 表sales根据time_id进行范围分区
SELECT *
FROM   sales
WHERE  time_id >= TO_DATE('2000-01-01 00:00:00', 'SYYYY-MM-DD HH24:MI:SS')
AND    prod_id = 38;
  1. 执行该语句得到的执行计划如下:
Plan hash value: 3087065703

--------------------------------------------------------------------------
|Id| Operation                                  | Name      |Pstart|Pstop|
--------------------------------------------------------------------------
| 0| SELECT STATEMENT                           |              |    |    |
| 1|  PARTITION RANGE ITERATOR                  |              | 13 | 28 |
| 2|   TABLE ACCESS BY LOCAL INDEX ROWID BATCHED| SALES        | 13 | 28 |
| 3|    BITMAP CONVERSION TO ROWIDS             |              |    |    |
|*4|     BITMAP INDEX SINGLE VALUE              |SALES_PROD_BIX| 13 | 28 |
--------------------------------------------------------------------------

Predicate Information (identified by operation id):
---------------------------------------------------

   4 - access("PROD_ID"=38)
  1. 然后将sales_q4_2003分区的索引禁用
ALTER INDEX sales_prod_bix MODIFY PARTITION sales_q4_2003 UNUSABLE;
ALTER INDEX sales_time_bix MODIFY PARTITION sales_q4_2003 UNUSABLE;
  1. 此时数据库在有索引的分区采用索引检索数据,对没有索引的分区进行全表扫描。
Plan hash value: 2120767686

---------------------------------------------------------------------------
|Id| Operation                                    | Name     |Pstart|Pstop|
---------------------------------------------------------------------------
| 0| SELECT STATEMENT                             |               |   |   |
| 1|  VIEW                                        | VW_TE_2       |   |   |
| 2|   UNION-ALL                                  |               |   |   |
| 3|    PARTITION RANGE ITERATOR                  |               | 13| 27|
| 4|     TABLE ACCESS BY LOCAL INDEX ROWID BATCHED| SALES         | 13| 27|
| 5|      BITMAP CONVERSION TO ROWIDS             |               |   |   |
|*6|       BITMAP INDEX SINGLE VALUE              | SALES_PROD_BIX| 13| 27|
| 7|    PARTITION RANGE SINGLE                    |               | 28| 28|
|*8|     TABLE ACCESS FULL                        | SALES         | 28| 28|
---------------------------------------------------------------------------

Predicate Information (identified by operation id):
---------------------------------------------------

   6 - access("PROD_ID"=38)
   8 - filter("PROD_ID"=38)

 通过将非活跃分区的索引删除的方式可以减少维护索引的成本,优化器会自动的帮我们进行查询的改写,而不影响我们原本的SQL语句。

04-24 02:30