Oracle 学习之 数据仓库（二） Dimension 的理解

   在数据仓库中，有事实表、维度表两个概念。

   事实表是数据仓库结构中的中央表，它包含联系事实与维度表的数字度量值和键。事实数据表包含描述业务（例如产品销售）内特定事件的数据。

   维度表是维度属性的集合。是分析问题的一个窗口。是人们观察数据的特定角度，是考虑问题时的一类属性，属性的集合构成一个维。

   如图示

我们以sh用户下的sales表和times表来看，

SALES为事实表

SQL> desc sales
 Name					   Null?    Type
 ----------------------------------------- -------- ----------------------------
 PROD_ID				   NOT NULL NUMBER
 CUST_ID				   NOT NULL NUMBER
 TIME_ID				   NOT NULL DATE
 CHANNEL_ID				   NOT NULL NUMBER
 PROMO_ID				   NOT NULL NUMBER
 QUANTITY_SOLD				   NOT NULL NUMBER(10,2)
 AMOUNT_SOLD				   NOT NULL NUMBER(10,2)

TIMES为维度表

SQL> desc times
 Name					   Null?    Type
 ----------------------------------------- -------- ----------------------------
 TIME_ID				   NOT NULL DATE
 DAY_NAME				   NOT NULL VARCHAR2(9)
 DAY_NUMBER_IN_WEEK			   NOT NULL NUMBER(1)
 DAY_NUMBER_IN_MONTH			   NOT NULL NUMBER(2)
 CALENDAR_WEEK_NUMBER			   NOT NULL NUMBER(2)
 FISCAL_WEEK_NUMBER			   NOT NULL NUMBER(2)
 WEEK_ENDING_DAY			   NOT NULL DATE
 WEEK_ENDING_DAY_ID			   NOT NULL NUMBER
 CALENDAR_MONTH_NUMBER			   NOT NULL NUMBER(2)
 FISCAL_MONTH_NUMBER			   NOT NULL NUMBER(2)
 CALENDAR_MONTH_DESC			   NOT NULL VARCHAR2(8)
 CALENDAR_MONTH_ID			   NOT NULL NUMBER
 FISCAL_MONTH_DESC			   NOT NULL VARCHAR2(8)
 FISCAL_MONTH_ID			   NOT NULL NUMBER
 DAYS_IN_CAL_MONTH			   NOT NULL NUMBER
 DAYS_IN_FIS_MONTH			   NOT NULL NUMBER
 END_OF_CAL_MONTH			   NOT NULL DATE
 END_OF_FIS_MONTH			   NOT NULL DATE
 CALENDAR_MONTH_NAME			   NOT NULL VARCHAR2(9)
 FISCAL_MONTH_NAME			   NOT NULL VARCHAR2(9)
 CALENDAR_QUARTER_DESC			   NOT NULL CHAR(7)
 CALENDAR_QUARTER_ID			   NOT NULL NUMBER
 FISCAL_QUARTER_DESC			   NOT NULL CHAR(7)
 FISCAL_QUARTER_ID			   NOT NULL NUMBER
 DAYS_IN_CAL_QUARTER			   NOT NULL NUMBER
 DAYS_IN_FIS_QUARTER			   NOT NULL NUMBER
 END_OF_CAL_QUARTER			   NOT NULL DATE
 END_OF_FIS_QUARTER			   NOT NULL DATE
 CALENDAR_QUARTER_NUMBER		   NOT NULL NUMBER(1)
 FISCAL_QUARTER_NUMBER			   NOT NULL NUMBER(1)
 CALENDAR_YEAR				   NOT NULL NUMBER(4)
 CALENDAR_YEAR_ID			   NOT NULL NUMBER
 FISCAL_YEAR				   NOT NULL NUMBER(4)
 FISCAL_YEAR_ID 			   NOT NULL NUMBER
 DAYS_IN_CAL_YEAR			   NOT NULL NUMBER
 DAYS_IN_FIS_YEAR			   NOT NULL NUMBER
 END_OF_CAL_YEAR			   NOT NULL DATE
 END_OF_FIS_YEAR			   NOT NULL DATE

如果我们创建一个物化视图

create materialized view sales_month_sum 
enable query rewrite as 
  SELECT t.calendar_month_id,
         prod_id,
         channel_id,
         promo_id,
         SUM (quantity_sold) quantity_sold,
         SUM (amount_sold) amount_sold
    FROM sales s, times t
   WHERE s.time_id = t.time_id
GROUP BY prod_id,
         channel_id,
         promo_id,
         t.calendar_month_id;

如果我们做如下按月的分组查询

SQL> alter session set query_rewrite_enabled=true;
SQL> alter session set query_rewrite_integrity=trusted;
SQL> set autotrace traceonly
SQL> set line 200
SQL>   SELECT t.calendar_month_id,
         prod_id,
         channel_id,
         promo_id,
         SUM (quantity_sold) quantity_sold,
         SUM (amount_sold) amount_sold
    FROM sales s, times t
   WHERE s.time_id = t.time_id
GROUP BY prod_id,
         channel_id,
         promo_id,
         t.calendar_month_id; 

9068 rows selected.


Execution Plan
----------------------------------------------------------
Plan hash value: 3287305789

------------------------------------------------------------------------------------------------
| Id  | Operation		     | Name	       | Rows  | Bytes | Cost (%CPU)| Time     |
------------------------------------------------------------------------------------------------
|   0 | SELECT STATEMENT	     |		       |  9068 |   690K|    13	 (0)| 00:00:01 |
|   1 |  MAT_VIEW REWRITE ACCESS FULL| SALES_MONTH_SUM |  9068 |   690K|    13	 (0)| 00:00:01 |
------------------------------------------------------------------------------------------------

可见查询使用的是物化视图，但是如果我需要按年、季度对数据做分组查询呢？

  SELECT t.calendar_quarter_id,prod_id,
         channel_id,
         promo_id,
         SUM (quantity_sold) quantity_sold,
         SUM (amount_sold) amount_sold
    FROM sales s, times t
   WHERE s.time_id = t.time_id
GROUP BY prod_id,
         channel_id,
         promo_id,
         t.calendar_quarter_id;

这个查看肯定是不能使用物化视图的，执行计划如下

Execution Plan
----------------------------------------------------------
Plan hash value: 3221963832

---------------------------------------------------------------------------------------------------------
| Id  | Operation		      | Name	| Rows	| Bytes | Cost (%CPU)| Time	| Pstart| Pstop |
---------------------------------------------------------------------------------------------------------
|   0 | SELECT STATEMENT	      | 	|  2037 | 79443 |   569   (6)| 00:00:07 |	|	|
|   1 |  HASH GROUP BY		      | 	|  2037 | 79443 |   569   (6)| 00:00:07 |	|	|
|*  2 |   HASH JOIN		      | 	|   918K|    34M|   546   (2)| 00:00:07 |	|	|
|   3 |    PART JOIN FILTER CREATE    | :BF0000 |  1826 | 21912 |    18   (0)| 00:00:01 |	|	|
|   4 |     TABLE ACCESS FULL	      | TIMES	|  1826 | 21912 |    18   (0)| 00:00:01 |	|	|
|   5 |    PARTITION RANGE JOIN-FILTER| 	|   918K|    23M|   525   (2)| 00:00:07 |:BF0000|:BF0000|
|   6 |     TABLE ACCESS FULL	      | SALES	|   918K|    23M|   525   (2)| 00:00:07 |:BF0000|:BF0000|
---------------------------------------------------------------------------------------------------------

Oracle为了是查询重写更加的智能，引入了Dimension的概念。Dimension我们称之为维，它是基于维度表的，用来描述维度表的维度之间的层级关系。

CREATE DIMENSION SH.TIMES_DIM
  LEVEL DAY                            IS 
    (SH.TIMES.TIME_ID)
  LEVEL MONTH                          IS 
    (SH.TIMES.CALENDAR_MONTH_ID)
  LEVEL QUARTER                        IS 
    (SH.TIMES.CALENDAR_QUARTER_ID)
  LEVEL YEAR                           IS 
    (SH.TIMES.CALENDAR_YEAR_ID)
  HIERARCHY CAL_ROLLUP
    (DAY                               CHILD OF
     MONTH                             CHILD OF
     QUARTER                           CHILD OF
     YEAR);

LEVEL定义等级，基于维度表，HIERARCHY关键字定义层级关系。由层级关系，我们知道quarter是由month组成的。

我们再次查询

SQL> SELECT t.calendar_quarter_id,prod_id,
         channel_id,
         promo_id,
         SUM (quantity_sold) quantity_sold,
         SUM (amount_sold) amount_sold
    FROM sales s, times t
   WHERE s.time_id = t.time_id
GROUP BY prod_id,
         channel_id, 
         promo_id,
         t.calendar_quarter_id;

3375 rows selected.


Execution Plan
----------------------------------------------------------
Plan hash value: 3397140165

--------------------------------------------------------------------------------------------------
| Id  | Operation		       | Name		 | Rows  | Bytes | Cost (%CPU)| Time	 |
--------------------------------------------------------------------------------------------------
|   0 | SELECT STATEMENT	       |		 |    20 |  1720 |    36  (14)| 00:00:01 |
|   1 |  HASH GROUP BY		       |		 |    20 |  1720 |    36  (14)| 00:00:01 |
|*  2 |   HASH JOIN		       |		 |   128K|    10M|    33   (7)| 00:00:01 |
|   3 |    VIEW 		       |		 |   849 |  6792 |    19   (6)| 00:00:01 |
|   4 |     HASH UNIQUE 	       |		 |   849 |  6792 |    19   (6)| 00:00:01 |
|   5 |      TABLE ACCESS FULL	       | TIMES		 |  1826 | 14608 |    18   (0)| 00:00:01 |
|   6 |    MAT_VIEW REWRITE ACCESS FULL| SALES_MONTH_SUM |  9068 |   690K|    13   (0)| 00:00:01 |
--------------------------------------------------------------------------------------------------

这次是使用物化视图与times表做关联，性能更高了。

我们对比如下两个查询

SQL>   SELECT t.calendar_quarter_id,
         prod_id,
         channel_id,
         promo_id,
         SUM (quantity_sold) quantity_sold,
         SUM (amount_sold) amount_sold
    FROM sales s, times t
   WHERE s.time_id = t.time_id AND t.calendar_quarter_id = 1769
GROUP BY prod_id,
         channel_id,
         promo_id,
         t.calendar_quarter_id; 

168 rows selected.


Execution Plan
----------------------------------------------------------
Plan hash value: 3397140165

--------------------------------------------------------------------------------------------------
| Id  | Operation		       | Name		 | Rows  | Bytes | Cost (%CPU)| Time	 |
--------------------------------------------------------------------------------------------------
|   0 | SELECT STATEMENT	       |		 |     1 |    86 |    33   (7)| 00:00:01 |
|   1 |  HASH GROUP BY		       |		 |     1 |    86 |    33   (7)| 00:00:01 |
|*  2 |   HASH JOIN		       |		 |  6423 |   539K|    32   (4)| 00:00:01 |
|   3 |    VIEW 		       |		 |    34 |   272 |    19   (6)| 00:00:01 |
|   4 |     HASH UNIQUE 	       |		 |    34 |   272 |    19   (6)| 00:00:01 |
|*  5 |      TABLE ACCESS FULL	       | TIMES		 |    90 |   720 |    18   (0)| 00:00:01 |
|   6 |    MAT_VIEW REWRITE ACCESS FULL| SALES_MONTH_SUM |  9068 |   690K|    13   (0)| 00:00:01 |
--------------------------------------------------------------------------------------------------

使用了物化视图

SQL>SELECT t.calendar_quarter_id,
         prod_id,
         channel_id,
         promo_id,
         SUM (quantity_sold) quantity_sold,
         SUM (amount_sold) amount_sold
    FROM sales s, times t
   WHERE s.time_id = t.time_id AND t.calendar_quarter_desc = '1998-01'
GROUP BY prod_id,
         channel_id,
         promo_id,
         t.calendar_quarter_id;

168 rows selected.


Execution Plan
----------------------------------------------------------
Plan hash value: 3221963832

-----------------------------------------------------------------------------------------------------------------
| Id  | Operation		      | Name	| Rows	| Bytes |TempSpc| Cost (%CPU)| Time	| Pstart| Pstop |
-----------------------------------------------------------------------------------------------------------------
|   0 | SELECT STATEMENT	      | 	|  8146 |   373K|	|   848   (2)| 00:00:11 |	|	|
|   1 |  HASH GROUP BY		      | 	|  8146 |   373K|  3632K|   848   (2)| 00:00:11 |	|	|
|*  2 |   HASH JOIN		      | 	| 57459 |  2637K|	|   546   (2)| 00:00:07 |	|	|
|   3 |    PART JOIN FILTER CREATE    | :BF0000 |    91 |  1820 |	|    18   (0)| 00:00:01 |	|	|
|*  4 |     TABLE ACCESS FULL	      | TIMES	|    91 |  1820 |	|    18   (0)| 00:00:01 |	|	|
|   5 |    PARTITION RANGE JOIN-FILTER| 	|   918K|    23M|	|   525   (2)| 00:00:07 |:BF0000|:BF0000|
|   6 |     TABLE ACCESS FULL	      | SALES	|   918K|    23M|	|   525   (2)| 00:00:07 |:BF0000|:BF0000|
-----------------------------------------------------------------------------------------------------------------

没有使用物化视图。

其实条件实质上是一样的，因为t.calendar_quarter_desc = '1998-01' 和t.calendar_quarter_id = 1769 在times表中表示相同的数据。

但是Oracle不知道CALENDAR_QUARTER_DESC与CALENDAR_QUARTER_ID的关系。

我们在创建Dimension时，可以为LEVEL指定属性值。

如下

CREATE DIMENSION SH.TIMES_DIM
  LEVEL DAY                            IS 
    (SH.TIMES.TIME_ID)
  LEVEL MONTH                          IS 
    (SH.TIMES.CALENDAR_MONTH_ID)
  LEVEL QUARTER                        IS 
    (SH.TIMES.CALENDAR_QUARTER_ID)
  LEVEL YEAR                           IS 
    (SH.TIMES.CALENDAR_YEAR_ID)
  HIERARCHY CAL_ROLLUP
    (DAY                               CHILD OF
     MONTH                             CHILD OF
     QUARTER                           CHILD OF
     YEAR)
  ATTRIBUTE QUARTER DETERMINES 
    (SH.TIMES.CALENDAR_QUARTER_DESC,
     SH.TIMES.DAYS_IN_CAL_QUARTER,
     SH.TIMES.END_OF_CAL_QUARTER,
     SH.TIMES.CALENDAR_QUARTER_NUMBER)
  ATTRIBUTE YEAR DETERMINES 
    (SH.TIMES.CALENDAR_YEAR,
     SH.TIMES.DAYS_IN_CAL_YEAR,
     SH.TIMES.END_OF_CAL_YEAR);

我们再次查询

SQL> SELECT t.calendar_quarter_id,
         prod_id,
         channel_id,
         promo_id,
         SUM (quantity_sold) quantity_sold,
         SUM (amount_sold) amount_sold
    FROM sales s, times t
   WHERE s.time_id = t.time_id AND t.calendar_quarter_desc = '1998-01'
GROUP BY prod_id,
         channel_id,
         promo_id,
         t.calendar_quarter_id;  

168 rows selected.


Execution Plan
----------------------------------------------------------
Plan hash value: 3290467316

--------------------------------------------------------------------------------------------------
| Id  | Operation		       | Name		 | Rows  | Bytes | Cost (%CPU)| Time	 |
--------------------------------------------------------------------------------------------------
|   0 | SELECT STATEMENT	       |		 |    20 |  2240 |    33   (7)| 00:00:01 |
|   1 |  HASH GROUP BY		       |		 |    20 |  2240 |    33   (7)| 00:00:01 |
|*  2 |   HASH JOIN		       |		 | 17191 |  1880K|    32   (4)| 00:00:01 |
|   3 |    VIEW 		       | VW_GBF_5	 |    91 |  3094 |    19   (6)| 00:00:01 |
|   4 |     HASH GROUP BY	       |		 |    91 |   728 |    19   (6)| 00:00:01 |
|   5 |      VIEW		       |		 |    91 |   728 |    19   (6)| 00:00:01 |
|   6 |       HASH UNIQUE	       |		 |    91 |  1456 |    19   (6)| 00:00:01 |
|*  7 |        TABLE ACCESS FULL       | TIMES		 |    91 |  1456 |    18   (0)| 00:00:01 |
|   8 |    MAT_VIEW REWRITE ACCESS FULL| SALES_MONTH_SUM |  9068 |   690K|    13   (0)| 00:00:01 |
--------------------------------------------------------------------------------------------------

这次就使用了物化视图。