Bring Your Own Effect

Chris Birchall

http://lambdale.org/

@lambd_ale

1st September 2018

The good old days

interface Cache<K, V> {

  public Optional<V> get(K key);

  public void put(K key, V value);

}
trait Cache[K, V] {

  def get(key: K): Future[Option[V]]

  def put(key: K, value: V): Future[Unit]

}
trait AsyncCache[K, V] {

  def get(key: K): Future[Option[V]]

  def put(key: K, value: V): Future[Unit]

}

trait SyncCache[K, V] {

  def get(key: K): Option[V]

  def put(key: K, value: V): Unit

}

class RedisCache extends AsyncCache { ...}

class CaffeineCache extends SyncCache { ... }
trait Cache[K, V] {

  def get(key: K): Future[Option[V]]

  def put(key: K, value: V): Future[Unit]

}

def getSync[K, V](cache: Cache[K, V])
                 (key: K): Option[V] =
  Await.result(cache.get(key), Duration.Inf)

def putSync[K, V](cache: Cache[K, V])
                 (key: K, value: V): Option[V] =
  Await.result(cache.put(key, value), Duration.Inf)

Source: Wikipedia

Source: pxhere.com

trait Cache[K, V] {

  def get[F[_]](key: K): F[Option[V]]

  def put[F[_]](key: K, value: V): F[Unit]

}
val cache: Cache[String, Int] = ...

val future: Future[Option[Int]] = cache.get[Future]("foo")

val task: Task[Option[Int]] = cache.get[Task]("foo")
val plainOldValue: Option[Int] = cache.get[???]("foo")
val plainOldValue: Option[Int] = cache.get[Id]("foo")
type Id[A] = A
class MemcachedCache[K, V] extends Cache[K, V] {

  val memcachedClient = ...

  def get[F[_]](key: K): F[Option[V]] = ???
 
  def put[F[_]](key: K, value: V): F[Unit] = ???

}
class MemcachedCache[K, V] extends Cache[K, V] {

  val memcachedClient = ...

  def get[F[_]](key: K)
               (implicit tc: TypeClass[F]): F[Option[V]] = {
    // TODO cache lookup!
    tc.pure(None)
  }
 
  def put[F[_]](key: K, value: V)
               (implicit tc: TypeClass[F]): F[Unit] = {
    // TODO cache write!
    tc.pure(())
  }
}

What typeclass do I need?

  • Depends on what your code needs to do
  • Use the least powerful tool for the job

Functor

Applicative

Monad

MonadError

Sync

Async

def map[A,B](fa: F[A])(f: A => B): F[B]
def pure[A](a: A): F[A]
def flatMap[A,B](fa: F[A])(f: A => F[B]): F[B]
def raiseError[A](e: Throwable): F[A]
def handleError[A](fa: F[A])(f: Throwable => A): F[A]
def delay[A](thunk: => A): F[A]
def async[A](
  register: (Either[Throwable, A] => Unit) => Unit): F[A]

Aside: "non-blocking"/"async" I/O

Blocking the current thread:

val x: Result = makeRemoteApiCall()
val x: Future[Result] = Future { makeRemoteApiCall() }

Still blocking, but on a different thread:

val x: Unit = makeRemoteApiCall(callback = {
  case Left(err) => println("Oh no!")
  case Right(result) => println("Yay!")
})

Callback-based async I/O:

Aside: "non-blocking"/"async" I/O

Callback-based async I/O

Future

Blocking the current thread

(wrap in Await)

(wrap in Future)

(use callback to complete a Promise)

ScalaCache's Async typeclass

trait Async[F[_]] {

  def pure[A](a: A): F[A]

  def map[A, B](fa: F[A])(f: A => B): F[B]

  def flatMap[A, B](fa: F[A])(f: A => F[B]): F[B]

  def raiseError[A](t: Throwable): F[A]

  def handleNonFatal[A](fa: => F[A])(f: Throwable => A): F[A]

  def delay[A](thunk: => A): F[A]

  def suspend[A](thunk: => F[A]): F[A]

  def async[A](register: (Either[Throwable, A] => Unit) => Unit): F[A]

}
class AsyncForFuture(implicit ec: ExecutionContext) 
  extends Async[Future] {

  def pure[A](a: A): Future[A] = Future.successful(a)

  def map[A, B](fa: Future[A])(f: A => B): Future[B] = fa.map(f)

  ...

  def async[A](register: (Either[Throwable, A] => Unit) => Unit)
      : Future[A] = {
    val promise = Promise[A]()
    register {
      case Left(e) => promise.failure(e)
      case Right(x) => promise.success(x)
    }
    promise.future
  }

}

Example instance: Future

class CaffeineCache[V](...) {

  def doGet[F[_]](key: String)
                 (implicit mode: Mode[F]): F[Option[V]] =
    mode.M.delay {
      // do the cache lookup ...
    }

}

Synchronous example: Caffeine

class MemcachedCache[V](...) {

  def doGet[F[_]](key: String)
                 (implicit mode: Mode[F]): F[Option[V]] =
    mode.M.async { cb =>
      val f = client.asyncGet(key)
      f.addListener(new GetCompletionListener {
        def onComplete(g: GetFuture[_]): Unit = {
          if (g.getStatus.isSuccess) {
            // ... call the callback with the result
          } else {
            // ... call the callback with the error
          }
        }
      })
    }

}

Callback example: Memcached

Why should I do this?

In a library - More flexibility for users
 
In an application - Extra dimension of abstraction
- More testable

Extra dimension of abstraction

Decouple:

  • Composition of operations (business logic)
  • Choice of effects (error handling, context passing)
  • Concrete implementation details

Composition of operations

trait Persistence[F[_]] {
  def saveNewUser(details: UserDetails): F[UserId]
  def findUserById(userId: UserId): F[User]
}

trait Events[F[_]] {
  def sendCreatedUserEvent(user: User): F[Unit]
}

class UserOps[F[_]: Monad](
  persistence: Persistence[F],
  events: Events[F]) {

  import persistence._, events._

  def createUser(details: UserDetails): F[User] =
    for {
      userId <- saveNewUser(details)
      user   <- findUserById(userId)
      _      <- sendCreatedUserEvent(user)
    } yield user

}

Choice of effects

sealed trait UserServiceFailure
case object NotFound extends UserServiceFailure
case class InternalError(e: Throwable) extends UserServiceFailure

type UserServiceResult[A] = Either[UserServiceFailure, A]
case class Context(traceToken: String)
type UserServiceOp[A] = Kleisli[UserServiceResult, Context, A]

Error handling

Context passing

Concrete impl details

class DbPersistence(db: Database) extends Persistence[UserServiceOp] {

  def saveNewUser(userDetails: UserDetails): UserServiceOp[UserId] = {
    ...
  }

  def findUserById(userId: UserId): UserServiceOp[User] = {
    ...
  }

}
class KafkaEvents(config: KafkaConfig) extends Events[UserServiceOp] {

  def sendCreatedUserEvent(user: User): UserServiceOp[Unit] = {
    ...
  }

}

Bring it all together

type UserService = UserOps[UserServiceOps]

val userService = new UserService(dbPersistence, kafkaEvents)

val userDetails: UserDetails = ...

val program: UserServiceOp[Unit] = userService.createUser(userDetails)

// remember, UserServiceOp[A] = Kleisli[UserServiceResult, Context, A]
program.run(Context(...)) // returns a UserServiceResult[Unit]

Conclusion

Made with Slides.com