`
落地窗
  • 浏览: 429331 次
  • 性别: Icon_minigender_1
  • 来自: 北京
社区版块
存档分类
最新评论

Oracle 物化视图 Dimension

阅读更多
要理解oracle中的dimension,首先要搞清楚dimension和dimension table之间的区别。dimension table是table,和关系数据库中的其他table一样,存放数据,需要实际的存储空间。而dimension则只是一个逻辑结构,定义了 dimension table中的一个列或一组列于其他列之间的一个层次关系,dimension只保存定义,可以将其理解为一种特定的constraint。所以,dimension不是一种必须存在的结构,但是,创建dimension对于数据仓库中一些复杂的查询重写有着相当重要的意义。而查询重写,则是数据仓库性能优化的一个不二法门。

数据仓库中由于数据量巨大,一些聚合计算等 操作往往通过物化视图预先计算存储。但是,不可能对所有维度的所有可能的聚合操作都建立物化视图,一则空间不允许,二则刷新时间也不允许。那么,在对某些聚合操作的sql进行查询重写时,就希望能利用已经存在的物化视图,尽管他们的聚合操作条件不完全一致。而dimension定义的各个level之间的 层次关系,对于一些上卷(rolling up)和下钻(drilling down)操作的查询重写的判断是相当重要的,而dimension中定义的attributes对于使用不同的列来做分组的查询重写起作用。

一个典型的dimension定义如下:
CREATE DIMENSION products_dim
LEVEL product IS (products.prod_id)
LEVEL subcategory IS (products.prod_subcategory)
LEVEL category IS (products.prod_category)
HIERARCHY prod_rollup (
product CHILD OF
subcategory CHILD OF
category
)
ATTRIBUTE product_info LEVEL product DETERMINES
(products.prod_name, products.prod_desc,
prod_weight_class, prod_unit_of_measure,
prod_pack_size, prod_status, prod_list_price, prod_min_price)
ATTRIBUTE subcategory DETERMINES
(prod_subcategory, prod_subcategory_desc)
ATTRIBUTE category DETERMINES
(prod_category, prod_category_desc);

dimension 中三个重要的属性:level,hierarchy,attribute。其中level定义了一个或一组列为一个整体,而hierarchy则定义了各 个level之间的层次关系,父level和子level之间是一种1:N的关系,而且,在dimension中可以指定多个hierarchy层次关 系。attribute则定义了level和其他列的一个1:1的关系,但这种1:1的关系不一定是可逆的,比如上面的列子,根据 product_info,也就是prod_id,可以确定prod_name,但不一定要求prod_name就能确定prod_id。

而 且,各个level之间的列不一定要来自同一个table,对于雪花模型,dimension table可能被规范化为许多的小表,则dimension中的level可能是来自不同表中的列。这是需要在dimension中指定join key来指出各个表之间的关联列。例如:

CREATE DIMENSION customers_dim
LEVEL customer IS (customers.cust_id)
LEVEL city IS (customers.cust_city)
LEVEL state IS (customers.cust_state_province)
LEVEL country IS (countries.country_id)
LEVEL subregion IS (countries.country_subregion)
LEVEL region IS (countries.country_region)
HIERARCHY geog_rollup (
customer CHILD OF
city CHILD OF
state CHILD OF
country CHILD OF
subregion CHILD OF
region
JOIN KEY (customers.country_id) REFERENCES country);

如果不指定skip when null子句,每个level中都不允许出现null值。

通过dbms_dimension.describe_dimension可以查看dimension的定义。
通 过dbms_dimension.validate_dimension可以检查dimension是否定义正确,在执行之前需要执行 ultdim.sql创建一个dimension_exceptions表,如果定义有误,则会在dimension_exceptions中查到相应的记录。在9i里,validate_dimension在dbms_olap包中。



####################################


在数据仓库环境中,我们通常利用物化视图强大的查询重写功能来提升统计查询的性能,但是物化视图的查询重写功能有时候无法智能地判断查询中一些相关联的条件,以至于影响性能。比如我们有一张销售表sales,用于存储订单的详细信息,包含交易日期、顾客编号和销售量。我们创建一张物化视图,按月存储累计销量信息,假如这时候我们要查询按季度或者按年度统计销量信息,Oracle是否能够智能地转换查询重写呢?我们知道交易日期中的日期意味着月,月意味着所处的季度,季度意味着年度,但是Oracle却是无法智能地判断这其中的关系,因此无法利用物化视图查询重写来返回我们季度或年度的销量信息,而是直接查询基表,导致性能产生问题。

这时候Dimension就派上用场了。Dimension用于说明列之间的父子对应关系,以使优化器能够自动转换不同列的关系,利用物化视图的查询功能来提升查询统计性能。下面我们首先创建一张销售交易表sales,包含交易日期、顾客编号和销售量这几个列,用于保存销售订单信息,整个表有42万多条记录;创建另一张表time_hierarchy用于存储交易日期中时间的关系,包含交易日期及其对应的月、季度及年度等信息,然后我们将体验Dimension的强大功能。



Roby@XUE> create table sales

2 (trans_date date, cust_id int, sales_amount number );



Table created.



Roby@XUE> insert /*+ APPEND */ into sales

2 select trunc(sysdate,'year')+mod(rownum,366) TRANS_DATE,

3 mod(rownum,100) CUST_ID,

4 abs(dbms_random.random)/100 SALES_AMOUNT

5 from all_objects

6 /



5926 rows created.



Roby@XUE> commit;



Commit complete.



Roby@XUE> begin

2 for i in 1 .. 6

3 loop

4 insert /*+ APPEND */ into sales

5 select trans_date, cust_id, abs(dbms_random.random)/100 SALES_AMOUNT

6 from sales;

7 commit;

8 end loop;

9 end;

10 /



PL/SQL procedure successfully completed.



Roby@XUE> select count(*) from sales;



COUNT(*)

----------

426672



创建索引组织表time_hierarchy,里面生成了交易日期中日期DAY、月MMYYYY、季度QTY_YYYY、年度YYYY的关系。



Roby@XUE> create table time_hierarchy

2 (day primary key, mmyyyy, mon_yyyy, qtr_yyyy, yyyy)

3 organization index

4 as

5 select distinct

6 trans_date DAY,

7 cast (to_char(trans_date,'mmyyyy') as number) MMYYYY,

8 to_char(trans_date,'mon-yyyy') MON_YYYY,

9 'Q' || ceil( to_char(trans_date,'mm')/3) || ' FY'

10 || to_char(trans_date,'yyyy') QTR_YYYY,

11 cast( to_char( trans_date, 'yyyy' ) as number ) YYYY

12 from sales

13 /



Table created.



接下我们创建一张物化视图mv_sales,用于存储每个客户对应每个月的销量统计信息。



Roby@XUE> create materialized view mv_sales

2 build immediate

3 refresh on demand

4 enable query rewrite

5 as

6 select sales.cust_id, sum(sales.sales_amount) sales_amount,

7 time_hierarchy.mmyyyy

8 from sales, time_hierarchy

9 where sales.trans_date = time_hierarchy.day

10 group by sales.cust_id, time_hierarchy.mmyyyy

11 /



Materialized view created.



我们对基表进行分析,以使优化器能够物化视图的查询重写功能:



Roby@XUE> analyze table sales compute statistics;



Table analyzed.



Roby@XUE> analyze table time_hierarchy compute statistics;



Table analyzed.



设置会话的查询重写功能:



Roby@XUE> alter session set query_rewrite_enabled=true;



Session altered.



Roby@XUE> alter session set query_rewrite_integrity=trusted;



Session altered.



接下来我们按月统计总的销量:



Roby@XUE> select time_hierarchy.mmyyyy, sum(sales_amount)

2 from sales, time_hierarchy

3 where sales.trans_date = time_hierarchy.day

4 group by time_hierarchy.mmyyyy

5 /



MMYYYY SUM(SALES_AMOUNT)

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

12006 4.0574E+11

12007 1.2297E+10

22006 3.6875E+11

32006 3.9507E+11

42006 3.7621E+11

52006 3.8549E+11

62006 3.6641E+11

72006 3.8110E+11

82006 3.8502E+11

92006 3.7278E+11

102006 3.7983E+11

112006 3.7210E+11

122006 3.8364E+11



13 rows selected.



Execution Plan

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

0 SELECT STATEMENT Optimizer=CHOOSE (Cost=4 Card=327 Bytes=8502)

1 0 SORT (GROUP BY) (Cost=4 Card=327 Bytes=8502)

2 1 TABLE ACCESS (FULL) OF 'MV_SALES' (Cost=2 Card=327 Bytes=8502)

Statistics

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

17 recursive calls

0 db block gets

25 consistent gets

4 physical reads



我们可以看到查询使用了查询重写的功能,直接查询物化视图中的查询方案,而不是查询其表,逻辑IO只有25个,性能相当良好。



假如这时候我们要按季度来查询统计销量信息,结果又会是怎样呢?



Roby@XUE> select time_hierarchy.qtr_yyyy, sum(sales_amount)

2 from sales, time_hierarchy

3 where sales.trans_date = time_hierarchy.day

4 group by time_hierarchy.qtr_yyyy

5 /



QTR_YYYY SUM(SALES_AMOUNT)

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

Q1 FY2006 1.1696E+12

Q1 FY2007 1.2297E+10

Q2 FY2006 1.1281E+12

Q3 FY2006 1.1389E+12

Q4 FY2006 1.1356E+12



Execution Plan

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

0 SELECT STATEMENT Optimizer=CHOOSE (Cost=1681 Card=5 Bytes=145)

1 0 SORT (GROUP BY) (Cost=1681 Card=5 Bytes=145)

2 1 NESTED LOOPS (Cost=35 Card=426672 Bytes=12373488)

3 2 TABLE ACCESS (FULL) OF 'SALES' (Cost=35 Card=426672

4 2 INDEX (UNIQUE SCAN) OF 'SYS_IOT_TOP_7828' (UNIQUE)



Statistics

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

14 recursive calls

0 db block gets

428048 consistent gets

599 physical reads



可以看到查询将直接查询基表产生了将近428048个逻辑IO,性能无法满足需求。

接下我们创建一个Dimension表time_hierarchy_dim,用于提醒优化器time_hierarchy表中的DAY列暗示着MMYYYY,MMYYYY又意味着QTY_YYYY,QTY_YYYY又意味着YYYY。然后我们将重新运行上面那个查询,看执行计划发生了怎样的变更。


Roby@XUE> create dimension time_hierarchy_dim

2 level day is time_hierarchy.day

3 level mmyyyy is time_hierarchy.mmyyyy

4 level qtr_yyyy is time_hierarchy.qtr_yyyy

5 level yyyy is time_hierarchy.yyyy

6 hierarchy time_rollup

7 (

8 day child of

9 mmyyyy child of

10 qtr_yyyy child of

11 yyyy

12 )

13 attribute mmyyyy

14 determines mon_yyyy;



Dimension created.



Roby@XUE> select time_hierarchy.qtr_yyyy, sum(sales_amount)

2 from sales, time_hierarchy

3 where sales.trans_date = time_hierarchy.day

4 group by time_hierarchy.qtr_yyyy

5 /



QTR_YYYY SUM(SALES_AMOUNT)

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

Q1 FY2006 1.1696E+12

Q1 FY2007 1.2297E+10

Q2 FY2006 1.1281E+12

Q3 FY2006 1.1389E+12

Q4 FY2006 1.1356E+12



Execution Plan

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

0 SELECT STATEMENT Optimizer=CHOOSE (Cost=14 Card=5 Bytes=195)

1 0 SORT (GROUP BY) (Cost=14 Card=5 Bytes=195)

2 1 HASH JOIN (Cost=7 Card=1157 Bytes=45123)

3 2 VIEW (Cost=4 Card=46 Bytes=598)

4 3 SORT (UNIQUE) (Cost=4 Card=46 Bytes=598)

5 4 INDEX (FAST FULL SCAN) OF 'SYS_IOT_TOP_7828' (UNIQUE)

6 2 TABLE ACCESS (FULL) OF 'MV_SALES' (Cost=2 Card=327



Statistics

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

193 recursive calls

0 db block gets

49 consistent gets

2 physical reads



可以看到创建Dimension后,Oracle已经能够智能地理解交易日期中月度和季度的转换关系,查询使用到物化视图,逻辑IO由原来的428048个减少到49个,性能有了大幅的提升。

同样我们再来统计一下年度的销量信息:



Roby@XUE> select time_hierarchy.yyyy, sum(sales_amount)

2 from sales, time_hierarchy

3 where sales.trans_date = time_hierarchy.day

4 group by time_hierarchy.yyyy

5 /



YYYY SUM(SALES_AMOUNT)

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

2006 4.5721E+12

2007 1.2297E+10



Execution Plan

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

0 SELECT STATEMENT Optimizer=CHOOSE (Cost=10 Card=2 Bytes=66)

1 0 SORT (GROUP BY) (Cost=10 Card=2 Bytes=66)

2 1 HASH JOIN (Cost=7 Card=478 Bytes=15774)




我们再创建一张customer_hierarchy表,用于存储客户代码、邮政编码和地区的关系,然后我们将按不同邮编或地区来查询各自的月度、季度或者年度销量信息。



Roby@XUE> create table customer_hierarchy

2 ( cust_id primary key, zip_code, region )

3 organization index

4 as

5 select cust_id,

6 mod( rownum, 6 ) || to_char(mod( rownum, 1000 ), 'fm0000') zip_code,

7 mod( rownum, 6 ) region

8 from ( select distinct cust_id from sales)

9 /



Table created.



Roby@XUE> analyze table customer_hierarchy compute statistics;



Table analyzed.



改写物化视图,查询方案中添加按不同邮编的月度统计销量。



Roby@XUE> drop materialized view mv_sales;



Materialized view dropped.



Roby@XUE> create materialized view mv_sales

2 build immediate

3 refresh on demand

4 enable query rewrite

5 as

6 select customer_hierarchy.zip_code,

7 time_hierarchy.mmyyyy,

8 sum(sales.sales_amount) sales_amount

9 from sales, time_hierarchy, customer_hierarchy

10 where sales.trans_date = time_hierarchy.day

11 and sales.cust_id = customer_hierarchy.cust_id

12 group by customer_hierarchy.zip_code, time_hierarchy.mmyyyy

13 /



Materialized view created.



Roby@XUE> set autotrace traceonly

Roby@XUE> select customer_hierarchy.zip_code,

2 time_hierarchy.mmyyyy,

3 sum(sales.sales_amount) sales_amount

4 from sales, time_hierarchy, customer_hierarchy

5 where sales.trans_date = time_hierarchy.day

6 and sales.cust_id = customer_hierarchy.cust_id

7 group by customer_hierarchy.zip_code, time_hierarchy.mmyyyy

8 /



1216 rows selected.



Execution Plan

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

0 SELECT STATEMENT Optimizer=CHOOSE (Cost=2 Card=409 Bytes=20450)

1 0 TABLE ACCESS (FULL) OF 'MV_SALES' (Cost=2 Card=409 Bytes=20450)



Statistics

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

28 recursive calls

0 db block gets

116 consistent gets

5 physical reads



可以看到如果按不同邮编、不同月度来统计查询的话,优化器将会查询物化视图中的查询方案,性能也是比较可观的。假如我们查不同地区年度的统计销量信息,结果又会是怎样?



Roby@XUE> select customer_hierarchy.region,

2 time_hierarchy.yyyy,

3 sum(sales.sales_amount) sales_amount

4 from sales, time_hierarchy, customer_hierarchy

5 where sales.trans_date = time_hierarchy.day

6 and sales.cust_id = customer_hierarchy.cust_id

7 group by customer_hierarchy.region, time_hierarchy.yyyy

8 /



9 rows selected.



Execution Plan

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

0 SELECT STATEMENT Optimizer=CHOOSE (Cost=1681 Card=9 Bytes=261)

1 0 SORT (GROUP BY) (Cost=1681 Card=9 Bytes=261)

2 1 NESTED LOOPS (Cost=35 Card=426672 Bytes=12373488)

3 2 NESTED LOOPS (Cost=35 Card=426672 Bytes=8106768)

4 3 TABLE ACCESS (FULL) OF 'SALES' (Cost=35 Card=426672

5 3 INDEX (UNIQUE SCAN) OF 'SYS_IOT_TOP_7833' (UNIQUE)

6 2 INDEX (UNIQUE SCAN) OF 'SYS_IOT_TOP_7828' (UNIQUE)



Statistics

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

0 recursive calls

0 db block gets

428047 consistent gets

745 physical reads



可以看到查询性能大有影响。接下我们同样创建dimension sales_dimension,用于说明客户代码和邮编、地区间的关系:



Roby@XUE> drop dimension time_hierarchy_dim

2 /



Dimension dropped.



Roby@XUE> create dimension sales_dimension

2 level cust_id is customer_hierarchy.cust_id

3 level zip_code is customer_hierarchy.zip_code

4 level region is customer_hierarchy.region

5 level day is time_hierarchy.day

6 level mmyyyy is time_hierarchy.mmyyyy

7 level qtr_yyyy is time_hierarchy.qtr_yyyy

8 level yyyy is time_hierarchy.yyyy

9 hierarchy cust_rollup

10 (

11 cust_id child of

12 zip_code child of

13 region

14 )

15 hierarchy time_rollup

16 (

17 day child of

18 mmyyyy child of

19 qtr_yyyy child of

20 yyyy

21 )

22 attribute mmyyyy

23 determines mon_yyyy;



Dimension created.



再回到原来的查询,我们可以看到查询性能有了大幅的提升:



Roby@XUE> set autotrace on

Roby@XUE> select customer_hierarchy.region,

2 time_hierarchy.yyyy,

3 sum(sales.sales_amount) sales_amount

4 from sales, time_hierarchy, customer_hierarchy

5 where sales.trans_date = time_hierarchy.day

6 and sales.cust_id = customer_hierarchy.cust_id

7 group by customer_hierarchy.region, time_hierarchy.yyyy

8 /



REGION YYYY SALES_AMOUNT

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

0 2006 7.3144E+11

0 2007 4484956329

1 2006 7.8448E+11

2 2006 7.7257E+11

2 2007 4684418980

3 2006 7.7088E+11

4 2006 7.8004E+11

4 2007 3127953246

5 2006 7.3273E+11



9 rows selected.



Execution Plan

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

0 SELECT STATEMENT Optimizer=CHOOSE (Cost=15 Card=9 Bytes=576)

1 0 SORT (GROUP BY) (Cost=15 Card=9 Bytes=576)

2 1 HASH JOIN (Cost=10 Card=598 Bytes=38272)

3 2 VIEW (Cost=3 Card=100 Bytes=700)

4 3 SORT (UNIQUE) (Cost=3 Card=100 Bytes=700)

5 4 INDEX (FULL SCAN) OF 'SYS_IOT_TOP_7833' (UNIQUE)

6 2 HASH JOIN (Cost=7 Card=598 Bytes=34086)

7 6 VIEW (Cost=4 Card=19 Bytes=133)

8 7 SORT (UNIQUE) (Cost=4 Card=19 Bytes=133)

9 8 INDEX (FAST FULL SCAN) OF 'SYS_IOT_TOP_7828'

10 6 TABLE ACCESS (FULL) OF 'MV_SALES' (Cost=2 Card=409



Statistics

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

364 recursive calls

0 db block gets

88 consistent gets

0 physical reads



Roby@XUE> set autot trace

Roby@XUE> select customer_hierarchy.region,

2 time_hierarchy.qtr_yyyy,

3 sum(sales.sales_amount) sales_amount

4 from sales, time_hierarchy, customer_hierarchy

5 where sales.trans_date = time_hierarchy.day

6 and sales.cust_id = customer_hierarchy.cust_id

7 group by customer_hierarchy.region, time_hierarchy.qtr_yyyy;



27 rows selected.



Execution Plan

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

0 SELECT STATEMENT Optimizer=CHOOSE (Cost=23 Card=22 Bytes=154

1 0 SORT (GROUP BY) (Cost=23 Card=22 Bytes=1540)

2 1 HASH JOIN (Cost=11 Card=1447 Bytes=101290)

3 2 VIEW (Cost=3 Card=100 Bytes=700)

4 3 SORT (UNIQUE) (Cost=3 Card=100 Bytes=700)

5 4 INDEX (FULL SCAN) OF 'SYS_IOT_TOP_7833' (UNIQUE) (

6 2 HASH JOIN (Cost=7 Card=1447 Bytes=91161)

7 6 VIEW (Cost=4 Card=46 Bytes=598)

8 7 SORT (UNIQUE) (Cost=4 Card=46 Bytes=598)

9 8 INDEX (FAST FULL SCAN) OF 'SYS_IOT_TOP_7828' (UN

10 6 TABLE ACCESS (FULL) OF 'MV_SALES' (Cost=2 Card=409 B



Statistics

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

10 recursive calls

0 db block gets

19 consistent gets

0 physical reads



Roby@XUE> select customer_hierarchy.region,

2 time_hierarchy.mon_yyyy,

3 sum(sales.sales_amount) sales_amount

4 from sales, time_hierarchy, customer_hierarchy

5 where sales.trans_date = time_hierarchy.day

6 and sales.cust_id = customer_hierarchy.cust_id

7 group by customer_hierarchy.region, time_hierarchy.mon_yyyy;



75 rows selected.

Execution Plan

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

0 SELECT STATEMENT Optimizer=CHOOSE (Cost=41 Card=56 Bytes=386

1 0 SORT (GROUP BY) (Cost=41 Card=56 Bytes=3864)

2 1 HASH JOIN (Cost=11 Card=3775 Bytes=260475)

3 2 VIEW (Cost=4 Card=120 Bytes=1440)

4 3 SORT (UNIQUE) (Cost=4 Card=120 Bytes=1440)

5 4 INDEX (FAST FULL SCAN) OF 'SYS_IOT_TOP_7828' (UNIQ

6 2 HASH JOIN (Cost=6 Card=409 Bytes=23313)

7 6 VIEW (Cost=3 Card=100 Bytes=700)

8 7 SORT (UNIQUE) (Cost=3 Card=100 Bytes=700)

9 8 INDEX (FULL SCAN) OF 'SYS_IOT_TOP_7833' (UNIQUE)

10 6 TABLE ACCESS (FULL) OF 'MV_SALES' (Cost=2 Card=409 B



Statistics

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

0 recursive calls

0 db block gets

14 consistent gets

0 physical reads
分享到:
评论

相关推荐

Global site tag (gtag.js) - Google Analytics