Reader Monad for Dependency Injection: multiple de

Question

When asked about Dependency Injection in Scala, quite a lot of answers point to the using the Reader Monad, either the one from Scalaz or just rolling your own. There are a number of very clear articles describing the basics of the approach (e.g. Runar's talk, Jason's blog), but I didn't manage to find a more complete example, and I fail to see the advantages of that approach over e.g. a more traditional "manual" DI (see the guide I wrote). Most probably I'm missing some important point, hence the question.

Just as an example, let's imagine we have these classes:

trait Datastore { def runQuery(query: String): List[String] }
trait EmailServer { def sendEmail(to: String, content: String): Unit }

class FindUsers(datastore: Datastore) {
  def inactive(): Unit = ()
}

class UserReminder(findUser: FindUsers, emailServer: EmailServer) {
  def emailInactive(): Unit = ()
}

class CustomerRelations(userReminder: UserReminder) {
  def retainUsers(): Unit = {}
}

Here I'm modelling things using classes and constructor parameters, which plays very nicely with "traditional" DI approaches, however this design has a couple of good sides:

• each functionality has clearly enumerated dependencies. We kind of assume that the dependencies are really needed for the functionality to work properly
• the dependencies are hidden across functionalities, e.g. the UserReminder has no idea that FindUsers needs a datastore. The functionalities can be even in separate compile units
• we are using only pure Scala; the implementations can leverage immutable classes, higher-order functions, the "business logic" methods can return values wrapped in the IO monad if we want to capture the effects etc.

How could this be modelled with the Reader monad? It would be good to retain the characteristics above, so that it is clear what kind of dependencies each functionality needs, and hide dependencies of one functionality from another. Note that using classes is more of an implementation detail; maybe the "correct" solution using the Reader monad would use something else.

I did find a somewhat related question which suggests either:

• using a single environment object with all the dependencies
• using local environments
• "parfait" pattern
• type-indexed maps

However, apart from being (but that's subjective) a bit too complex as for such a simple thing, in all of these solutions e.g. the retainUsers method (which calls emailInactive, which calls inactive to find the inactive users) would need to know about the Datastore dependency, to be able to properly call the nested functions - or am I wrong?

In what aspects would using the Reader Monad for such a "business application" be better than just using constructor parameters?

Answer 1

How to model this example

How could this be modelled with the Reader monad?

I'm not sure if this should be modelled with the Reader, yet it can be by:

1. encoding the classes as functions which makes the code play nicer with Reader
2. composing the functions with Reader in a for comprehension and using it

Just right before the start I need to tell you about small sample code adjustments that I felt beneficial for this answer. First change is about FindUsers.inactive method. I let it return List[String] so the list of addresses can be used in UserReminder.emailInactive method. I've also added simple implementations to methods. Finally, the sample will use a following hand-rolled version of Reader monad:

case class Reader[Conf, T](read: Conf => T) { self =>

  def map[U](convert: T => U): Reader[Conf, U] =
    Reader(self.read andThen convert)

  def flatMap[V](toReader: T => Reader[Conf, V]): Reader[Conf, V] =
    Reader[Conf, V](conf => toReader(self.read(conf)).read(conf))

  def local[BiggerConf](extractFrom: BiggerConf => Conf): Reader[BiggerConf, T] =
    Reader[BiggerConf, T](extractFrom andThen self.read)
}

object Reader {
  def pure[C, A](a: A): Reader[C, A] =
    Reader(_ => a)

  implicit def funToReader[Conf, A](read: Conf => A): Reader[Conf, A] =
    Reader(read)
}

Modelling step 1. Encoding classes as functions

Maybe that's optional, I'm not sure, but later it makes the for comprehension look better. Note, that resulting function is curried. It also takes former constructor argument(s) as their first parameter (parameter list). That way

class Foo(dep: Dep) {
  def bar(arg: Arg): Res = ???
}
// usage: val result = new Foo(dependency).bar(arg)

becomes

object Foo {
  def bar: Dep => Arg => Res = ???
}
// usage: val result = Foo.bar(dependency)(arg)

Keep in mind that each of Dep, Arg, Res types can be completely arbitrary: a tuple, a function or a simple type.

Here's the sample code after the initial adjustments, transformed into functions:

trait Datastore { def runQuery(query: String): List[String] }
trait EmailServer { def sendEmail(to: String, content: String): Unit }

object FindUsers {
  def inactive: Datastore => () => List[String] =
    dataStore => () => dataStore.runQuery("select inactive")
}

object UserReminder {
  def emailInactive(inactive: () => List[String]): EmailServer => () => Unit =
    emailServer => () => inactive().foreach(emailServer.sendEmail(_, "We miss you"))
}

object CustomerRelations {
  def retainUsers(emailInactive: () => Unit): () => Unit =
    () => {
      println("emailing inactive users")
      emailInactive()
    }
}

One thing to notice here is that particular functions don't depend on the whole objects, but only on the directly used parts. Where in OOP version UserReminder.emailInactive() instance would call userFinder.inactive() here it just calls inactive() - a function passed to it in the first parameter.

Please note, that the code exhibits the three desirable properties from the question:

1. it is clear what kind of dependencies each functionality needs
2. hides dependencies of one functionality from another
3. retainUsers method should not need to know about the Datastore dependency

Modelling step 2. Using the Reader to compose functions and run them

Reader monad lets you only compose functions that all depend on the same type. This is often not a case. In our example FindUsers.inactive depends on Datastore and UserReminder.emailInactive on EmailServer. To solve that problem one could introduce a new type (often referred to as Config) that contains all of the dependencies, then change the functions so they all depend on it and only take from it the relevant data. That obviously is wrong from dependency management perspective because that way you make these functions also dependent on types that they shouldn't know about in the first place.

Fortunately it turns out, that there exist a way to make the function work with Config even if it accepts only some part of it as a parameter. It's a method called local, defined in Reader. It needs to be provided with a way to extract the relevant part from the Config.

This knowledge applied to the example at hand would look like that:

object Main extends App {

  case class Config(dataStore: Datastore, emailServer: EmailServer)

  val config = Config(
    new Datastore { def runQuery(query: String) = List("[e-mail]") },
    new EmailServer { def sendEmail(to: String, content: String) = println(s"sending [$content] to $to") }
  )

  import Reader._

  val reader = for {
    getAddresses <- FindUsers.inactive.local[Config](_.dataStore)
    emailInactive <- UserReminder.emailInactive(getAddresses).local[Config](_.emailServer)
    retainUsers <- pure(CustomerRelations.retainUsers(emailInactive))
  } yield retainUsers

  reader.read(config)()

}

Advantages over using constructor parameters

In what aspects would using the Reader Monad for such a "business application" be better than just using constructor parameters?

I hope that by preparing this answer I made it easier to judge for yourself in what aspects would it beat plain constructors. Yet if I were to enumerate these, here's my list. Disclaimer: I have OOP background and I may not appreciate Reader and Kleisli fully as I don't use them.

1. Uniformity - no mater how short/long the for comprehension is, it's just a Reader and you can easily compose it with another instance, perhaps only introducing one more Config type and sprinkling some local calls on top of it. This point is IMO rather a matter of taste, because when you use constructors nobody prevents you to compose whatever things you like, unless someone does something stupid, like doing work in constructor which is considered a bad practice in OOP.
2. Reader is a monad, so it gets all benefits related to that - sequence, traverse methods implemented for free.
3. In some cases you may find it preferable to build the Reader only once and use it for wide range of Configs. With constructors nobody prevents you to do that, you just need to build the whole object graph anew for every Config incoming. While I have no problem with that (I even prefer doing that on every request to application), it isn't an obvious idea to many people for reasons I may only speculate about.
4. Reader pushes you towards using functions more, which will play better with application written in predominantly FP style.
5. Reader separates concerns; you can create, interact with everything, define logic without providing dependencies. Actually supply later, separately. (Thanks Ken Scrambler for this point). This is often heard advantage of Reader, yet that's also possible with plain constructors.

I would also like to tell what I don't like in Reader.

1. Marketing. Sometimes I get impression, that Reader is marketed for all kind of dependencies, without distinction if that's a session cookie or a database. To me there's little sense in using Reader for practically constant objects, like email server or repository from this example. For such dependencies I find plain constructors and/or partially applied functions way better. Essentially Reader gives you flexibility so you can specify your dependencies at every call, but if you don't really need that, you only pay its tax.
2. Implicit heaviness - using Reader without implicits would make the example hard to read. On the other hand, when you hide the noisy parts using implicits and make some error, compiler will sometimes give you hard to decipher messages.
3. Ceremony with pure, local and creating own Config classes / using tuples for that. Reader forces you to add some code that isn't about problem domain, therefore introducing some noise in the code. On the other hand, an application that uses constructors often uses factory pattern, which is also from outside of problem domain, so this weakness isn't that serious.

What if I don't want to convert my classes to objects with functions?

You want. You technically can avoid that, but just look what would happen if I didn't convert FindUsers class to object. The respective line of for comprehension would look like:

getAddresses <- ((ds: Datastore) => new FindUsers(ds).inactive _).local[Config](_.dataStore)

which is not that readable, is that? The point is that Reader operates on functions, so if you don't have them already, you need to construct them inline, which often isn't that pretty.

Answer 2

I think the main difference is that in your example you are injecting all dependencies when objects are instantiated. The Reader monad basically builds a more and more complex functions to call given the dependencies, wich are then returned to the highest layers. In this case, the injection happens when the function is finally called.

One immediate advantage is flexibility, especially if you can construct your monad once and then want to use it with different injected dependencies. One disadvantage is, as you say, potentially less clarity. In both cases, the intermediate layer only need to know about their immediate dependencies, so they both work as advertised for DI.