EasyDI – Who wants some cake?

by Eric Torreborre - 7 Apr 2016

Scala developers have lots of options when it comes to doing Dependency Injection (or DI). The usual Java libraries can be used, like Spring, or Guice for Play developers.

But Scala being Scala, there are other options. You can use libraries leveraging macros, like MacWire, or use the Scala type system and the infamous “Cake pattern” and its many variations using traits, with or without self-types.

Unfortunately, many teams have been burnt by the Cake Pattern and are looking for other solutions, often returning to more or less involved DI libraries.

I want to take a step back and propose another approach for DI that is both simple and flexible, using basic Scala features and … a special library, which is not a DI library at all!

Back to requirements

What are the basic things we can expect from DI?

  • A way to define components
  • A way to configure them from files
  • The possibility to substitute some of them for testing, regardless of their nesting
  • The possibility to declare some of them as singletons

Can we do that with some simple Scala? With just case classes and traits used as interfaces? Or to put it differently: What’s wrong with constructor injection again?

Define components

With case classes and constructor injection we can declare our components like this:

case class ZalandoNotifier(config: NotifierConfig, email: EmailService) {

// we only send an email for orders having a high level of
// priority as defined in the config file
def orderReady(order: Order): Future[Status] =
if (config.sendFor(order.getPriority)) email.send(order.emailAddress, "Your order is ready!")
else Future.delay(Status.ok) // scalaz Future


// a trait used as a simple interface
trait EmailService {

// send an email to a given address
// return Ok if the email could be sent
def send(address: String, body: String): Future[Status]


// a specific implementation of the EmailService
case class JavaMailEmailService(smtpHost: String, smtpPort: Int) extends EmailService {
def send(address: String, body: String): Future[Status] =
??? // use the JavaMail api to implement this

Create components

Creating the ZalandoNotifier service from a configuration file is not very hard. Let’s pretend we have the following classes to read from configuration files:

// a configuration file
trait Config {
def get[A](key: String): ConfigError Xor A

// Abstract data type for possible errors
// happening when reading configuration files
sealed trait ConfigError
case class MissingKey(key: String) extends ConfigError

object ConfigError {

def render(e: ConfigError): String =
e match {
case MissingKey(key) => s"Key $key not found"


The Xor type is an Either type from the cats library. See here for an introduction to cats, Xor and the |@| notation used below.

Creating the ZalandoNotifier service from the configuration file requires reading the NotifierConfig, the EmailService, and creating a ZalandoNotifier from those 2 instances:

case class NotifierConfig(priority: Int) {
def sendFor(p: Int): Boolean = p >= priority

// create a NotifierConfig from file
object NotifierConfig {

def fromConfig(config: Config): ConfigError Xor NotifierConfig =


// create an EmailService from file
object EmailService {

// The default EmailService instance is using the Java email API
def fromConfig(config: Config): ConfigError Xor EmailService =
(config.get[String]("smtp-host") |@|

// create the ZalandoNotifier from its components
object ZalandoNotifier {

def fromConfig(config: Config): ConfigError Xor EmailService =
(NotifierConfig.fromConfig(config) |@|

Assembling the ZalandoNotifier from its components is very simple here, we just combine the two parts into one. We could have a more complex fromConfig method using a for comprehension and injecting different EmailService instances based on a configuration parameter.

We have now assembled components from values in a configuration file, taking care of possible configuration errors. What is the next difficulty?


The next difficulty is this one. Suppose we have a bigger application using the ZalandoNotifier and we want to switch out the JavaEmailService with a mock implementation for testing. Using a naive approach for constructor injection, we might want to do this:

def createApplication(
orderService: OrderService,
notifierConfig: NotifierConfig,
emailService: EmailService): Application =
Application(orderService, ZalandoNotifier(notifierConfig, emailService))

def testApplication = {
// use a mocked email service
val app = createApplication(OrderService(), NotifierConfig(10), MockedEmailService())
// test the application now

This is really problematic because we think we need to expose all the components and their dependencies to the top-level of the application to be able to build the exact component graph we want and substitute other components. This is tedious and breaks all encapsulation.

Fortunately, there exists a library precisely aimed at replacing objects in graph: Kiama.


Kiama is a library for language processing, a toolbox for parsing computer languages, analysing and interpreting them.

We are going to use one feature of Kiama: Tree rewriting. As you probably know, computer languages are being parsed into Abstract Syntax Trees (ASTs). Most of the time these trees are being rewritten to simpler trees, in order to remove some syntactic sugar or to optimise some constructs. For example, a tree representing collection operations might rewrite two consecutive map operations into one for efficiency (a fusion operation).

And what is an “application”, if not a tree of services and configuration objects? So, with the help of Kiama, we can “rewrite” the application to replace some of its parts:

// return true if A implements the list of types defined by a given class tag */
def implements(a: Any)(implicit ct: ClassTag[_]): Boolean = {
val types: List[Class[_]] =
ct.runtimeClass +: ct.runtimeClass.getInterfaces.toList

types.forall(t => t.isAssignableFrom(a.getClass))

// a Kiama Strategy to replace any node having the same type as T
// with another instance
def replaceStrategy[T : ClassTag](t: T): Strategy =
strategy[Any] {
case v if implements(v) => Some(t)
case other => None

A Strategy in Kiama is more or less a partial function taking one of the nodes in the tree and returning an Option[A] if it succeeds and None if it doesn’t. You can then use combinators to define where and how you want to apply this Strategy. For example:

// use the strategy everywhere you can, from top to down
def replaceWithStrategy[G](strategy: Strategy, graph: G): G =

replaceWithStrategy(replaceStrategy[EmailService](mock), application)

Or, with a bit of syntactic sugar:


It really isn’t that difficult to write tests!

You can also conceive much more targeted replacements:

case class Leg(foot: String)
case class Robot(left: Leg, right: Leg)
case class Application(robot: Robot, house: House)

// just replace one leg of the robot!
application.replace {
case Robot(left, right) => Robot(Leg(Foot("repaired")), right)

Now that we know how to deal with testing, what else do we want to do? Singletons!


Applications are not just simple trees, but rather directed acyclic graphs where some nodes are being shared. For example, many Scala applications need to share an ExecutionContext (and / or an ActionSystem, a Materializer if working with Akka). Duplicating these contexts would waste resources.

When we build the application from the configuration file, we use independent fromConfig methods where each component knows how to read the file and create itself. What if 2 components need an ExecutionContext?

// Delay the evaluation of the ExecutionContext
case class ExecutionService(ec: Eval[ExecutionContext])

object ExecutionService {
def create(config: Config): ExecutionService = {
lazy val system: ActorSystem = ActorSystem("xxx", config)
lazy val executionContext: ExecutionContext = system.dispatcher

// First component
case class Service1(config: C1, es: ExecutionService)

object Service1 {
def fromConfig(config: Config): ConfigError Xor Service1 =
C1.fromConfig(config).map(c1 => Service1(c1, ExecutionService.create)

// Second component
case class Service2(config: C2, es: ExecutionService)

object Service2 {
def fromConfig(config: Config): ConfigError Xor Service2 =
C2.fromConfig(config).map(c2 => Service2(c2, ExecutionService.create)

// The main Application
case Application(s1: Service1, s2: Service2)

object Application {
def fromConfig(config: Config): ConfigError Xor Application =
(Service1.fromConfig(config) |@|


Here we are duplicating the ExecutionService component in the Application instance, this can’t be good.

This could also be worse because at this stage we haven’t consumed any resource. The ExecutionContext is encapsulated in an ExecutionService using the cats.Eval type to delay any evaluation of the ExecutionContext. Is there a way to use only one instance of the ExecutionService? Yes, indeed, with one more Strategy:

// store the first instance of the target type and replace all nodes of the same type with that one
def singletonStrategy[T : ClassTag]: Strategy = {
var t: Option[T] = None
strategy[Any] {
case v if implements(v)(implicitly[ClassTag[T]]) =>
t match {
case Some(singleton) => Some(singleton)
case None => t = Some(v.asInstanceOf[T])

case other => None

And with a bit of syntactic sugar, creating the full application becomes:

object Application {
def fromConfig(config: Config): ConfigError Xor Application =
(Service1.fromConfig(config) |@|



The technique presented here is really minimal: No annotations, no modules/bindings, no type system trickery. In addition, we can use other tools in the Kiama toolbox, like attribute grammars, to implement a topological sort in a few lines and start services in order.

The main drawback is the necessity to instantiate a full graph of objects before being able to modify it (the fromConfig methods). On the other hand, specifying how every component can be instantiated from the configuration file (and testing that!) needs to be done anyway.

I wouldn’t be surprised if large applications had new, unforeseen requirements that I’m not addressing right now, but simple applications can definitely use this technique and larger applications could explore more Kiama strategies.

I also hope that this post gave you an incentive to explore Kiama (in version 2.0.0 as of now), as it’s an awesome library full of possibilities.

Similar blog posts