Learning From Mistakes: Leaky Abstractions

On the project I’m working on I’ve had a requirement to store and read files from the file system. Alse the files had to be accessible from the web.

Having a gut feeling that the infrastructure may change as the business will grow, I decided to hide operations on the file system behind an interface:

1
2
3
4
5
6
public interface IFilesStorage {
    string StoreFile(Stream stream, string fileName);
    Stream GetFile(string virtualPath);
    string GetFileUrl(string virtualPath);
    string GetFilePath(string virtualPath);
}

As it looks, if someday I’ll need to switch from the file system to another storage mechanism, I’ll be able to do get the job done by writing another implementation of the interface. Right? Wrong! The requirement did come in - I’ve had to store the files in S3. And only then I realised that IFilesStorage is a leaky abstraction.

The problem lies in the last method, GetFilePath. This method leaks out the implementation detail, that each file has a path which can be used to access the file from the file system. Of course, other storage mechanisms can’t provide such functionality. This bummer makes switching storage mechanism nearly impossible.

From a SOLID point of view, this issue can be seen as a violation of the Liskov substitution principle: the file system based implementation of the interface cannot be replaced by another implementation, as it will break the correctness of the application.

The solution was to drop the problematic method and get rid of the dependancy on file paths in the system.

Lesson learned: having abstractions will make the code more testable, but if you’re aiming for supple design, make sure your abstractions don’t leak out details about the underlying architecture.

Comments