(Almost no) Dependency Injection for Scala

anodi is a Scala micro library for initialization of application components and dependency injection, featuring:

• application component definitions in plain Scala (e.g. simple constructor invocations)
• automatic resolution of dependency graph and initialization order, with early cycle detection
• opt-in parallel initialization of independent components
• opt-in dependency injection with Scala implicits/givens (compile-time "autowiring")

Project setup (sbt)

libraryDependencies += "com.github.ghik" %% "anodi" % "<version>"

Compatibility

anodi is currently available for Scala 2.13 and Scala 3

Application components

An application is made of components. In short, a component is an object which has a lifecycle, i.e. defines some starting and destroying logic. This is usually associated with some of the following traits:

• a component has an internal, mutable state
• a component acquires and releases some resources
• a component has dependencies on other components that do so

Typical examples of components include a web server, database client, HTTP client, etc.

Defining components

In Anodi, components are defined by defining an object or class extending the Components trait. Each component is then defined as a member whose implementation uses component or singleton method to wrap a plain Scala expression (e.g. a constructor invocation) into a Component definition.

import scala.concurrent._
import scala.concurrent.duration._
import com.github.ghik.anodi._

class Database {
  // DB initialization goes here...
}
class Server(database: Database) {
  // server initialization goes here...
  
  def join(): Unit = { /* wait for shutdown */ }
}

object MyApp extends Components {
  def server: Component[Server] =
    singleton(new Server(database.ref))

  def database: Component[Database] =
    singleton(new Database)
    
  def main(args: Array[String]): Unit = {
    // use appropriate execution context of your choice for application initialization
    import ExecutionContext.Implicits.global
    val srv = Await.result(server.init, Duration.Inf)
    srv.join()
  }
}

Component

The Component type can be thought of as an enhanced version of Scala's built-in lazy val construct. It is associated with a lazy-evaluated expression that creates and initializes the component (often a simple constructor invocation, but it may be an arbitrary piece of code).

Wrapping component definitions into Component instances relieves the programmer from manually figuring out the initialization order of interdependent components. Note that simply using lazy vals would already do that job, but using Component offers a nicer and safer experience:

• in case of failures, exceptions are wrapped in a way that lets you track the initialization path that
  leads to the faulty component (i.e. a stack trace but nicer to read than a raw one)
• dependency cycles are detected early (before any initialization is done) and come with a readable error message
  that lets you pinpoint the cycle

Component instances are aware of their source position. If they are assigned to a val or def, the name of this val or def is interpreted as the component name. This name may be later useful for debugging. Here's an example of a dependency cycle error that uses component names and source positions:

Exception in thread "main" com.github.ghik.anodi.DependencyCycleException: component dependency cycle detected:
  server(MyApp.scala:13) ->
  service(MyApp.scala:16) ->
  database(MyApp.scala:19) ->
  server(MyApp.scala:13)

Entry points

In order to start our application in the previous example, we explicitly requested initialization of the server component. This way we assumed server to be an entry point of our application - the primary service that exposes core system functionality (e.g. an HTTP server). All the other components are initialized only as direct or indirect dependencies of this toplevel entry point.

If your application consists of multiple entry points, e.g. an UI server, REST API server, FTP server, etc. then it is recommended to create a single toplevel component to aggregate them:

import com.github.ghik.anodi._

class UiServer
class ApiServer
class FtpServer

class FullApplication(ui: UiServer, api: ApiServer, ftp: FtpServer)

object MyApp extends Components {
  def ui: Component[UiServer] = singleton(new UiServer)
  def api: Component[ApiServer] = singleton(new ApiServer)
  def ftp: Component[FtpServer] = singleton(new FtpServer)
  
  def fullApplication: Component[FullApplication] =
    singleton(new FullApplication(ui.ref, api.ref, ftp.ref))
}

This way your application always has a single entry point, which makes it easier to navigate through code and understand its structure (knowing where to start).

Dependency injection

Anodi is designed so that the compiler can help you with dependency injection, if desired. There are multiple flavors of doing so, depending on how explicit or implicit you want to be with passing dependencies.

Explicit dependency references

A Component may refer to another Component using .ref, e.g.

class Database
class Server(db: Database)

object MyApp extends Components {
  def database: Component[Database] = singleton(new Database)
  def server: Component[Server] = singleton(new Server(database.ref))
}

.ref is not a real method. It exists only during compilation and is interpreted by the component or singleton macro. A dependency reference is extracted by the macro out of the component initialization expression. This way the macro separates initialization of a component from initialization of its dependencies. This technique makes it possible to inspect the dependency graph before initializing any components. This allows early cycle detection and lets us parallelize component initialization.

NOTE: If you're familiar with sbt, then .ref works somewhat similarly to .value in sbt settings & task definitions.

Implicit dependency injection

.ref is used to refer to dependencies explicitly. This isn't necessarily bad because it gives you the most control and often produces the most readable code (it's the least "magic" way). However, passing a lot of constructor parameters manually may be tedious. That's why most DI frameworks feature some form of "autowiring" that figures out dependencies automatically by their type. In traditional Java DI frameworks, this autowiring is done with runtime reflection. In Scala, we can leverage implicits for this purpose and have everything wired in compile time.

In order for your Component to be implicitly injectable, simply declare it as implicit/given. Then you can declare your dependencies (e.g. constructor parameters) also as implicits. Anodi takes care of the implicit Component[Thing] => implicit Thing conversion.

class Database
class Server(implicit db: Database)

object MyApp extends Components {
  implicit def database: Component[Database] = singleton(new Database)
  def server: Component[Server] = singleton(new Server))
}

Explicitly-declared implicit dependencies

A middle ground between implicit and explicit dependencies is a situation where a component definition is declared as implicit/given but a dependency to that component is expressed with an explicit parameter.

class Database
class Server(db: Database)

object MyApp extends Components {
  implicit def database: Component[Database] = singleton(new Database)
  def server: Component[Server] = singleton(fromImplicits[Server])
}

The core utility is the fromImplicits macro, which instantiates a component using constructor invocation, and resolves all parameters as implicits, even though they are not declared as implicits.

This is a nice sweet spot between fully implicit and fully explicit approach:

• you avoid the boilerplate of passing all parameters explicitly
• constructor parameters remain explicit, so that component classes stay pure and agnostic of the DI mechanism

Complete example

See ComponentsExample.scala