Some dimensions—if you push them to their physical extreme, something we never did in 2000—get very, very large at the bottom. As a simple example, you could have 50M customers. In 32-bit AS2K you would never dream of making that a dimension—or at least not a MOLAP dimension—and you wouldn’t want to make it into a ROLAP dimension either, because that would eliminate any MOLAP data storage anywhere in the cube. In 2005, I can easily imagine a situation where I would want the most aggregated attributes (Country, State, City) of a user hierarchy to be MOLAP—so that they can support good MOLAP aggregations—and the bottom attributes (such as Customer, or even Order ID) as ROLAP, so they don’t require MOLAP processing.
The developers imagined it as well, but apparently couldn’t get it to work in the Yukon time frame. So the whole dimension has to be either MOLAP or ROLAP. In the first place, my experience with large ROLAP dimensions used with a MOLAP cube is very limited (which is to say, not yet existent), so it may not be all that much of an issue.
If it turns out that a truly “hybrid” dimension is desirable, there are a couple of ways I can think of for creating the hybrid. One is to split the dimension in two, with the higher-level (MOLAP) one a reference dimension by way of the lower-level (ROLAP) one. Another is to create a many-to-many dimension to connect the two. It will be interesting to experiment with different alternatives in specific scenarios.