商业智能研究(十六)materialized view+dimension提高mondrian性能(二)

商业智能研究(十六) 用materialized view + dimension 来提高mondrian 的性能(二)

接着上一篇的定义我们定义如下两个dimension :
CREATE DIMENSION PRODUCT_DIM
LEVEL "product_id" IS "product"."product_id"
LEVEL "brand_name" IS "product"."brand_name"
LEVEL "product_class_id" IS "product_class"."product_class_id"
LEVEL "product_category" IS "product_class"."product_category"
LEVEL "product_department" IS "product_class"."product_department"
LEVEL "product_family" IS "product_class"."product_family"
HIERARCHY PRODUCT_ROLLUP (
"product_id" CHILD OF
"brand_name" CHILD OF
"product_class_id" CHILD OF
"product_category" CHILD OF
"product_department" CHILD OF
"product_family"
JOIN KEY ("product"."product_class_id") REFERENCES "product_class_id"
)
ATTRIBUTE "product_id" DETERMINES ("product_name")
ATTRIBUTE "product_class_id" DETERMINES ("product_subcategory");

CREATE DIMENSION TIME_DIM
LEVEL time IS "time_by_day"."time_id"
LEVEL month IS "time_by_day"."month_of_year"
LEVEL quarter IS "time_by_day"."quarter"
LEVEL year IS "time_by_day"."the_year"
HIERARCHY TIME_ROLLUP (
time CHILD OF
month CHILD OF
quarter CHILD OF
year
)
ATTRIBUTE time DETERMINES ("time_by_day"."the_date");

然后我们建立materialized view , 注意QUERY_REWRITE_INTEGRITY 和 QUERY_REWRITE_ENABLED 应该已经正确的设置了.
CREATE MATERIALIZED VIEW PRODUCT_SUM
BUILD IMMEDIATE
REFRESH ON DEMAND
ENABLE QUERY REWRITE
AS
SELECT "time_by_day"."time_id" ,
"product"."product_id",
"product_class"."product_class_id" ,
SUM("sales_fact_1997"."store_sales") ,
SUM("sales_fact_1997"."store_cost")
FROM "time_by_day" "time_by_day",
"product" "product" ,
"product_class" "product_class" ,
"sales_fact_1997" "sales_fact_1997"
WHERE "sales_fact_1997"."time_id" = "time_by_day"."time_id"
AND "sales_fact_1997"."product_id" = "product"."product_id"
AND "product"."product_class_id" = "product_class"."product_class_id"
GROUP BY "time_by_day"."time_id",
"product"."product_id" ,
"product_class"."product_class_id";

现在我们 set autotrace on .

执行
SELECT "time_by_day"."the_date" ,
"product_class"."product_family" ,
SUM("sales_fact_1997"."store_sales") ,
SUM("sales_fact_1997"."store_cost")
FROM "time_by_day" "time_by_day",
"product" "product" ,
"product_class" "product_class" ,
"sales_fact_1997" "sales_fact_1997"
WHERE "sales_fact_1997"."time_id" = "time_by_day"."time_id"
AND "sales_fact_1997"."product_id" = "product"."product_id"
AND "product"."product_class_id" = "product_class"."product_class_id"
GROUP BY "time_by_day"."the_date",
"product_class"."product_family";

从图一中我们看到，当我们把product 聚合到了最高的level product_family,oracle 的执行计划是从product_sum 中来做聚合的.这就是因为我们建立的dimension 告诉了oracle product有这种层次的关系.product的dimension 即告诉了product_id 能够决定product_name，也告诉了product_id能够聚合product_family , 同样的我们把Time 聚合到最高的level
SELECT "time_by_day"."the_year" ,
"product"."product_name" ,
SUM("sales_fact_1997"."store_sales") ,
SUM("sales_fact_1997"."store_cost")
FROM "time_by_day" "time_by_day",
"product" "product" ,
"product_class" "product_class" ,
"sales_fact_1997" "sales_fact_1997"
WHERE "sales_fact_1997"."time_id" = "time_by_day"."time_id"
AND "sales_fact_1997"."product_id" = "product"."product_id"
AND "product"."product_class_id" = "product_class"."product_class_id"
GROUP BY "time_by_day"."the_year",
"product"."product_name";
从图二中的执行计划同样可以看出我们只选取Time 来做聚合的时候，oracle 仍然是从product_sum 表中来做聚合, 用time_id 来决定the_date ,time_id同样可以聚合year.

最后一个是同时聚合product 和 time
SELECT "time_by_day"."the_year" ,
"product_class"."product_family" ,
SUM("sales_fact_1997"."store_sales") ,
SUM("sales_fact_1997"."store_cost")
FROM "time_by_day" "time_by_day",
"product" "product" ,
"product_class" "product_class" ,
"sales_fact_1997" "sales_fact_1997"
WHERE "sales_fact_1997"."time_id" = "time_by_day"."time_id"
AND "sales_fact_1997"."product_id" = "product"."product_id"
AND "product"."product_class_id" = "product_class"."product_class_id"
GROUP BY "time_by_day"."the_year",
"product_class"."product_family";
同样的，oracle 还是从product_sum 中取数据 .

因为materialized view 的使用，我们可以把我们要分析的Cube 作成一个或几个非常大的materialized view , 建立正确的dimension 之后，当你查询的时候,由于数据已经提前计算过了，所以查询的速度比较快，在加上dimension可以告诉oracle 数据之间的层级关系，减少了我们建立不必要的materialized view , 所以使数据能够得到更加充分的利用.
关于如何建立dimension 倒是比较简单，只要你弄懂数据之间的层级关系就可以了.
对于如何建立materialized view 倒是比较麻烦，
我举个简单的例子吧:
在mondrian 的 foodmart 的例子中，我们可以任意的选取指标 , product , customers , education leve , gender ,marital sttus ,promotin media ,promotions , store , store size in SQFT , store type , time , yearly income 这十三个要分析的数据来建立cube ,用户有可能使用product 来做分析的维度，也有可能把product 来做Measure 或者不选，所以我们不可能建立所有情况考虑到的Cube .
ps : 如果你非要搞一个出来的话，我可以给你点提示
总的方案有2 的 13 次方：8096 种方案.也就是你要建立8096 个materialized view 就可以解决所有情况.
C 13 3 ：数学里面的概率问题，十三个里面选3个出来，不论顺序的. C 13 3 = 13*12*12/(1*2*3)
代表的意思是从13个里面选3个出来做fact table ,其余十个做dimension .不论你选不选这些dimension 都一样，
总的方案 = c 13 1 + c 13 2 + c 13 3 + ...... + c 13 13 = 2 的十三次方 = 8096.

如何建立materialized view 还是主要是看你如何建立你的分析的维度.如果你的fact table 本身很多.而维度也很多的情况下，不可能每个fact table 都建立一个关于所有dimension 的materialized view ,对于我们的product dimension数据还算是比较少的，如果达到像大型超市那么多的产品，可能还需要在brand_name 或者 subcategory 来建立materialized view ,所以还是建议根据用户的查询sql 来分析用户到底经常查询那些数据.

下一篇继续介绍 Mondrian 如何使用materialized view 来提高性能.

图一 : product 集合到最高层的执行计划.