The repository pattern became popular a couple of years ago by the hand of Eric Evans with the DDD (Domain Driven Design) movement and Martin Fowler with his catalog of Enterprise Application Patterns.
The idea of having persistence ignorant entities inferred from the business domain and a repository as a simple intermediary for abstracting the underline persistence storage details had a great acceptance in the development community. No matter if you decided to stick to DDD or not, the repository pattern brought an incredible value to the table by decoupling your business domain code from details of the database access code (or any other underline storage). This was particularly important for unit testing your business domain classes. I personally don’t think in a case of completely replacing the underline storage, from a database to a http service or a file for example, because those scenarios are not very common and usually involves more changes than just switch the repository layer.
The idea of using an abstraction layer at that level was not new at all, but it was mutating with different names and shapes over the time. I am pretty sure many of you still remember the data access layer in the old times of COM when the N Tier architecture was a popular idea pushed by Microsoft. Same idea but different name, design and implementation details.
However, a time after the repository became popular, a great game changer was introduced in the .NET ecosystem with something with all know as Linq.
Linq already provides an abstraction layer with common query capabilities on top of any data source so why bother with yet another abstraction layer like a repository. It results that the entry point to the query system in Linq, IQueryable, is not enough most of the times and only provides a read-only view with query support over the data source. We also need operations to persist changes in the repository, and that’s not something we get out of the box with IQueryable.
Microsoft on his side has provided some popular Linq implementations as part of the .NET framework to address particular scenarios or common problems like object relation mapping for databases with Entity Framework, xml management with Linq to XML or data management over http with WCF data services to name a few.
An initial problem with some of these implementations is that they didn’t make certain abstractions implicit making really hard to replicate their behavior with mocks or stubs as part of an unit test. A typical example was the the “eager loading” capability of EF in the initial versions. It was not possible to use lazy loading for associations, and the Include method for loading those was not something you could easily abstract as part of the repository. While this issue was partially addressed with POCOs, only the latest EF code first bits makes the approach of making an unit testable repository something possible.
I agree with Daniel that a repository should only provide IQueryable for the aggregated roots and methods for persisting changes in the underline store. This is how the ideal repository abstraction should look to me,
Before you say something, this is not the typical repository pattern, yes, because it also uses the unit of work pattern to submit the changes. To be honest, I don’t really care having the unit of work as part of this abstraction because it makes the repository easier to test and use too. However, there are some disadvantages of exposing IQueryable directly in your repository.
- You can not easily reuse queries. At that point, you might want to have some pre-defined extension methods for creating IQueryable instances with all the filters already set. If you are still not a big fan of using this approach, you might want to switch to an approach based on query specs as this one, LinqSpecs.
- Other developers can alter the original purpose of the queries, or make things wrong with it. I think this is not a problem of this approach at all, but it mostly related to “developer protection”.
What about methods for filtering data that receive an expression like this one,
Unless Microsoft provides a good support for comparing expression trees in the .NET framework, I would stay away from that approach. I made a mistake myself of using it in the past for a repository over WCF Data Services. What I learned from that experience is that most of the mocking frameworks don’t support that approach very well, so it works some times.
In conclusion, try to provide a single entry point to your unit of work and expose all the aggregate roots with IQueryable as part of it whenever it is possible.