For certain abstractions, we are looking to maintain the state of a business process over time:
The relational concern is the storage of metadata sufficient to locate & retrieve the state. E.g.: integer primary key, name of process, current transition in process, some datetime info, active session id, last user, etc.
The 'non-relational' portion is simply a final 'Json' column per row that contains the actual serialized state. This state model can vary wildly depending on the particular process in play, so by having it serialized we can get a lot of reuse potential out of our solution.
In terms of the models, it's not 1 gigantic class with 1000 properties. Its more like a set of 100+ related models with 5-30 properties each.