Dimensions are key to navigating the data warehouse / business intelligence system, and hierarchies are the key to navigating dimensions. Any time a business user talks about wanting to drill up, down or into the data, they are implicitly referring to a dimension hierarchy. In order for those drill paths to work properly, and for a large DW/BI system to perform well, those hierarchies must be correctly designed, cleaned, and maintained.
Hierarchies are important not just for usability. They play a huge role in query performance for a modern DW/BI system: aggregations are often precomputed and stored for intermediate hierarchy levels and transparently used in queries. Precomputed aggregations are one of the most valuable tools to improve query performance, but in order for them to work, your hierarchies have to be clean.
Start with the Design
The solution to the problem of maintaining hierarchies begins during the design phase. For every substantial dimension, spend time thinking through the hierarchical relationships. Business-user input is absolutely imperative, as is time spent exploring the data.
The first question to resolve is what are the drilldown paths or hierarchies in each dimension? Most dimensions have a hierarchy, even if it's not coded in the transaction system. A core dimension such as customer, product, account, or even date may have many hierarchies. Date provides a good example that we all understand.
The date dimension often has three or more hierarchies. Novice dimensional modelers will try to create a single hierarchy that goes from day to week, month, quarter, and year. But that just doesn't work! Weeks do not roll up smoothly to months or even years. There is usually a separate fiscal calendar, and sometimes several others.
Display the hierarchies graphically to review them with the business users. The diagram below shows clearly the different hierarchies and levels that will be available. Notice the attributes that apply at different levels. This picture is a graphical display suitable for communicating with users and among the DW/BI team; it does not represent the table's physical structure. Get user buy-in on the hierarchies, levels, and names. Equally important, test how much transformation you need to apply to the actual data in order to populate these hierarchical structures.
The familiar date dimension contains lessons that are applicable to the administration of all dimensions:
- • You can have multiple hierarchies. Most interesting dimensions have several alternative hierarchies. Work with business users to name columns and hierarchies so that the meaning of each is clear.
- • You must have many-to-one referential integrity between each level: a day rolls up to one and only one month, month to quarter, and quarter to year.
- • If the data warehouse environment (as opposed to the original source) maintains referential integrity with explicit physical tables for each level, then a unique primary key must be identified at each level. If these keys are artificial surrogate keys, then they should be hidden from the business users in the final single, flat denormalized dimension table in the presentation layer of the data warehouse. A common error is to think of the key for the month level as month name (January) or month number. The correct primary key is year and month. This is a very common mistake which we encounter in many kinds of dimensions. In a geography dimension, for example, city name alone is not an identifier column; it needs to be some combination of city, state, and perhaps country.
- • Think carefully during the design phase about whether columns can be reused between hierarchies. You might think that the week hierarchy could share the year column with the calendar hierarchy, but what about the first and last weeks of the year? If our business rule is to have week 1 for a new year start on the first Monday of the year, Week 1 of 2009 starts on January 5. January 1-4 will fall in 2008 for the week hierarchy. You need a separate year-of-week column. Sometimes you do want hierarchies to intersect, but you must be certain that the data will support that intersection.