Clearest blog post I've read on the Reader monad in Scala. This is something I can see coming to our code base very soon.
The next step for us is monad transformers in Scalaz. We are going to end up with something like Future[Reader[Writer[Problem \/ Result]]] (a few type parameters are missing there; hopefully you get the idea). Monad transformers give a way to flatten this stack of monads into a single monad, which saves a lot of unnecessary wrapping and unwrapping in code. I haven't yet worked out the all the details of using them, though. Should probably do a blog post when I do :-)
I've been thinking about using scalaz for a while now. My main concern about the above pattern is the type complexity that will arise. I would love to learn more about monad transformers and how they help out with that issue.
A bit of searching will find a few talks, blog posts, and Stackoverflow answers. I haven't found anything I really like, which is why I'm writing the post.
I look forward to that blog post. In order to use Scalaz's monad transformers with scala.concurrent.Future, would I have to write my own monad instance for it, or is there a way to re-use the one from scalaz.concurrent?
Great use of the same example under three different dependency injection mechanisms.
One issue I see with the Reader pattern, however, is that anyone using one of these injected methods has to deal with the monad stuff of using for comprehensions to extract the values, et cetera, whereas with the other two patterns they get to use normal functions.
Put another way: having to distinguish between the "normal" functions in a class versus the injected Reader functions, and not be able to use them in the same way, feels like it would get old fast. I wonder if the author has found this annoying or not.
I do really like how purely functional the Reader approach is though, no state!
I actually like using comprehensions in Scala so I don't find it annoying at all. Also I think having a clear distinction between the "normal" and "injected" functions is a good thing.
I do find having to add implicit parameters to every function slightly annoying, especially because when I forget, it's not always clear from the compiler error what I did wrong.
in general your repository functions will be wrapped in some monad, whether it's Reader/Maybe/List doesn't really matter as far as the for comprehension is concerned.
Yeah, I guess that helps a little since Scala's for comprehensions let you mix monads. (Which, for the record, Haskell's otherwise similar do notation doesn't seem to allow.)
Technically, Scala's for comprehensions don't let you mix monads either, but implicit conversions make it work in some cases. For example, there's an implicit conversion from Option to Iterable so the Option (Maybe) and List monads can be mixed. It's really more a matter of Scala's type system allowing it. The Scala compiler just re-writes comprehensions into equivalent higher-order function application.
It feels like this isn't actually solving the dependency injection problem - we've moved all our dependencies into this global (ish) Config object, but that object is effectively acting as a service registry. Constructing the Config is still going to require a DI strategy (as the end of the post acknowledges), so what do we actually gain from this reader pattern? I don't think it makes testing any easier (if anything it makes it harder, since we have to build up a whole dummy Config, rather than building a small test cake that only includes the dependencies a specific test needs).
It does actually solve the problem that DI solves in general, because you create readers that expect the dependency to be injected into them. You still have to decide how to do the injection and that's true for all of the non-framework approaches AFAICT. With the cake pattern you still have to create an object somewhere that mixes in all of the concrete implementations. What you gain from the reader pattern is being able to limit the biolerplate to edge where the injection actually occurs.
As far as having to create a whole dummy Config for testing, that's a fair point, although the reader approach makes it easy to inject it since every reader is a function of the Config. I usually only have a small number of dependencies so it hasn't really been an issue for me. If you're mocking the dependencies each one is only one line of code, and for dependencies you don't need in the you can just use ???.
I'm not sure I understand the concern regarding the approach using implicits. Sure, you need to add an extra implicit method parameter to each method that requires the dependency, but in the reader approach, every same method needs to be written in the reader monad. In fact, they're really the same concern, since the reader monad instance is just doing the work of threading what would have been implicit in the first place.
Actually it's not. Only the primitive readers have similar "overhead" as with the implicits approach. Most of the readers are defined by mapping or flatMapping over the primitives (either explicitly or with comprehensions). With the implicits approach all of the "injected" methods need to declare the dependency via the implicit parameter. With the reader approach, only the primitives have to declare it. Compare the implicits version of UserInfo to the reader version. UserRepository doesn't appear anywhere in the signatures of the reader version.
The next step for us is monad transformers in Scalaz. We are going to end up with something like Future[Reader[Writer[Problem \/ Result]]] (a few type parameters are missing there; hopefully you get the idea). Monad transformers give a way to flatten this stack of monads into a single monad, which saves a lot of unnecessary wrapping and unwrapping in code. I haven't yet worked out the all the details of using them, though. Should probably do a blog post when I do :-)