A few points, Sasman. <br>
<br>
You have surmised correctly that I am talking about a high level ERD. What I am talking about is not what you have described as a "subject area model". I believe that we both agree that a star schema can be characterized as a radically denormalised relational database design. This begs the question, "denormalized from what?" The obvious answer is, "from a normalized model". There is an advantage to identifying dimensions early rather than late. There is also an advantage in basing the definition of these dimensions on the broadest possible analysis.<br>
<br>
No doubt you are familiar with articles written by fellow practicioners in this field that explain the denormalization techniques and analysis techniques that can be used to derive a star schema from a detailed, normalized logical data model. The same techniques can be applied to a less detailed model, even one that is not, technically, normalized, in order to derive dimensions that are reusable. "Conforming dimensions" in Ralph Kimball's parlance. Even if the central warehouse is to be a normalized database, conforming dimensions are necessary in order to assure consistent query results from one data mart to the next where a dimensional approach is selected.<br>
<br>
The subject matter for this high level data model is the business area that forms the context of data warehouse plus whatever external data are necessary to satisfy the various business cases that can be anticipated at project initiation. There is no need to be finicky at this point. It is best to paint with a broad brush. This model becomes the framework from which new sbject matter will hang. As new subject matter is added, this model is the source of fundamental entities that will seed the detailed logical model. The high level model will, in return, be enhanced with any new fundamental entities discovered during creation of the detailed model. The scope of each detailed model is determined by the business case or cases that it is intended to represent. Each of the detailed models is then used to enhance the physical design of the central warehouse, as well as the design of the data mart that will actually be used to deliver the information.<br>
<br>
About the central data warehouse and normalization: Normalization is good. It is better for the central warehouse to be rigorously normalized, in textbook third normal form at the very least. This is an article of faith. However, the size of a data warehouse may make it difficult, if not impossible, to maintain the data within the available operational window. In practice, most medium to large data warehouses are denormalized to some extent. I agree with you that this should be kept at a minimum.