Decoupling Domain Model from Persistence

For the last several months, I've been working on a Windows desktop application. This application has a number of pretty common aspects: file manipulation, local data repository, GUI, user settings, collection/dataset manipulation (think drag-and-drop listview type stuff), communication with peripheral devices, and communication with remote services (such as a web service). I think just about every desktop application has a good cross-section of these aspects, and maybe a few others that are slipping my mind at the moment. As a result, there are patterns of application architecture that will arise, and which the development efforts of well-designed, robust applications will have in common. I'm not talking about the typical "design patterns" here à la Gang of Four (GoF), but rather "application architecture patterns". Big patterns. A set of 4 or 5 patterns that, when meshed together, encompass the big, meaty chunks accounting for 95% or better of your application.

I know this is true. I know these patterns exist. But as this is my first effort in developing a desktop app of this complexity, I don't really know what the patterns are. One that GoF did account for was MVC/MVP, and sure I know those... But honestly, up until recently (and maybe even still) my knowledge of that pattern was very hazy and academic. Actually implementing it was something I'd never had to do before. It took a lot of blood, sweat, and tears--lots of trial and error--to get to a point where I really feel like I'm starting to grasp the "right" way to put together an MVP structure.

So now I've moved on to persistence. I have my domain model. But I need to be able to persist the objects in my domain model to and from disk (not a database!). I've been struggling to find ways to add in this functionality without tainting the domain model with persistence logic. Granted, this is a valid way of doing things, as is evidenced by Martin Fowler's ActiveRecord pattern. But it really doesn't make for ease of unit testing. Your business logic and persistence logic get all mixed up in the same classes and you can't easily isolate one from the other for testing purposes.

Then I saw Fowler's Repository pattern. I thought, okay, maybe that will help. I can create an interface that I can implement which will take care of all the logic for disk access to all the right places, based on what objects or collections are being retrieved. And I can just mock the interface for purposes of testing the domain model and business logic. I have a simple three-level hierarchy of domain interfaces, so it would be fairly simple to implement all the querying and CRUD I need for each domain class on the repository interface. But then I thought about how we plan to have multiple implementations of these interfaces in the future, with different permutations of under-the-covers data to be persisted, and thought maybe it would be best to have the repository be metadata-driven, rather than hard-coding the query logic for the different domain classes. This of course naturally leads to a record-based design, and that would mean I'd want data mappers to translate from my domain classes to repository records. And of course, the repository class should be strictly independent of the persistence mechanism itself (in this case the disk) so that I can unit test the repository logic without worrying about maintaining a test repository just for that purpose.

And suddenly I realized that I had mentally re-created ADO .NET. DataSets and DataTables are the repository mechanism, (with typed datasets and the ADO designer even supporting generating data mappers for you). DataAdapters are the persistence service.

So where this leaves me is with the question of whether I should just design my on-disk storage schema, build myself an app-specific data adapter mechanism, and let ADO .NET do all the work. If I really wanted to minimize the amount of mapping and repository-access code I have to write, I might even be able to use the Entity Framework to do my dirty work for me. Admittedly I have no idea how much work it would take to implement my own data adapter. The interfaces seem straightforward enough, but I have a nagging feeling that's really just the tip of the iceberg.

After all my radio silence here, I don't know if anyone is still listening... But if anyone has any thoughts on the wisdom of this approach, I'm listening intently.