Understanding Monix Observable

Piotr Gawryś

About me

  • An open source contributor for fun
  • One of the maintainers of Monix
  • Kraków Scala User Group co-organizer

https://github.com/Avasil

twitter.com/p_gawrys

Monix

  • Scala / Scala.js library for asynchronous programming
  • Multiple modules exposing Task, IO[E, A], Observable, Iterant, Coeval, Local, and many concurrency primitives
  • Favors purely functional programming but provides for all
  • Big focus on being both Future, and 

twitter.com/p_gawrys

Monix Observable

twitter.com/p_gawrys

High Level Example

twitter.com/p_gawrys

val result: Task[Long] = 
  Observable.fromIterable(allElements)
    .bufferTumbling(bufferSize)
    .mapEval(seq => Task(seq.sum))
    .filter(_ > 0)
    .map(_.toLong)
    .foldLeftL(0L)(_ + _)

High Level Example

twitter.com/p_gawrys

val playerInputs: ConcurrentQueue[MovementCommand] = ???

val gameStateObservable: Observable[GameState] =
  Observable(initialState) ++
    Observable
      .repeatEvalF(playerInputs.poll) // get inputs
      .groupBy(_.playerId) // create a sub-stream per playerId
      // emit only the first element every 150.millis per sub-stream
      // and merge them concurrently to one stream
      .mergeMap(_.throttleFirst(150.millis)) 
      .bufferTimed(150.millis) // collect results every 150.millis
      .scan0(initialState) { // a state machine to update the latest GameState
        case (GameState(players, bullets, environment), commands) =>
          val (updatedPlayers, updatedBullets) =
            moveTank(players, bullets, environment, commands)
          val newGameState =
            resolveCollisions(
              GameState(updatedPlayers, updatedBullets, environment)
            )

          newGameState
      }

Today, we're going to talk about internals!

twitter.com/p_gawrys

Definition

twitter.com/p_gawrys

trait Observer[-A] {
  def onNext(elem: A): Future[Ack]

  def onError(ex: Throwable): Unit

  def onComplete(): Unit
}


// Needs some kind of ExecutionContext to do 
// anything with onNext (which returns Future)
trait Subscriber[-A] extends Observer[A] {
  implicit def scheduler: Scheduler
}

abstract class Observable[+A] {
  def unsafeSubscribeFn(subscriber: Subscriber[A]): Cancelable
}

Observable

Observer

Observer

subscribe

subscribe

Observable

Observer

Observer

onNext

onNext

Observer#onNext protocol

twitter.com/p_gawrys

trait Observer[-A] {
  def onNext(elem: A): Future[Ack]
}

sealed abstract class Ack extends Future[Ack]
case object Continue extends Ack
case object Stop extends Ack
  • Grammar: onNext CAN be called zero, one or multiple times until onComplete, or onError
  • Back-pressure: each onNext call MUST wait on a Continue result
  • Cancellation: after receiving Stop the data-source MUST no longer send any events

Observer protocol

twitter.com/p_gawrys

trait Observer[-A] {
  def onError(ex: Throwable): Unit
  
  def onComplete(): Unit
}
  • Grammar: either onComplete or onError at most one time, can't call both.
  • Back-pressure: optional, not required to wait for the last onNext

Observer protocol

twitter.com/p_gawrys

trait Observer[-A] {
  def onNext(elem: A): Future[Ack]

  def onError(ex: Throwable): Unit

  def onComplete(): Unit
}

Observable

twitter.com/p_gawrys

  • Subscriber/Observer subscribes to Observable and it starts emitting events
  • subscribe returns Cancelable which allows to stop the computation from the outside
abstract class Observable[+A] {
  def unsafeSubscribeFn(subscriber: Subscriber[A]): Cancelable
}

Simple Observable

twitter.com/p_gawrys

final class NowObservable[+A](elem: A) extends Observable[A] {
  def unsafeSubscribeFn(subscriber: Subscriber[A]): Cancelable = {
    // No need to back-pressure for onComplete
    subscriber.onNext(elem)
    subscriber.onComplete()
    // There's no specific action needed in case the connection is canceled
    Cancelable.empty 
  }
}

twitter.com/p_gawrys

final class PrintSubscriber[-A] extends Subscriber[A] {
  override def scheduler: Scheduler = Scheduler.global

  override def onNext(elem: A): Future[Ack] = {
    println(s"Received $elem")
    Continue
  }

  override def onError(ex: Throwable): Unit = {
    println(s"Received error $ex")
  }

  override def onComplete(): Unit = {
    println(s"Received final event")
  }
}

Simple Subscriber

twitter.com/p_gawrys

val source: Observable[Int] = new NowObservable(10)

val cancelable: Cancelable =
  source.unsafeSubscribeFn(new PrintSubscriber)
  
// => Received 10
// => Received final event

Running Observable

twitter.com/p_gawrys

new Observable[Int] {
  def unsafeSubscribeFn(subscriber: Subscriber[Int]): Cancelable = {
    subscriber.onNext(elem)
    subscriber.onComplete()
    Cancelable.empty 
  }
}.unsafeSubscribeFn(new Subscriber[Int] {
  override def scheduler: Scheduler = Scheduler.global

  override def onNext(elem: A): Future[Ack] = {
    println(s"Received $elem")
    Continue
  }

  override def onError(ex: Throwable): Unit = {
    println(s"Received error $ex")
  }

  override def onComplete(): Unit = {
    println(s"Received final event")
  }
})

// => Received 10
// => Received final event

More complicated example

twitter.com/p_gawrys

import monix.eval.Task
import monix.reactive.Observable
import scala.concurrent.duration._
import scala.util.Random

val result: Task[List[Int]] =
  Observable.repeatEval(Random.nextInt(10))
    .takeByTimespan(10.second)
    .toListL

Observable.repeatEval

twitter.com/p_gawrys

object Observable {    
  def repeatEval[A](task: => A): Observable[A] =
    new RepeatEvalObservable(task)
}
final class RepeatEvalObservable[+A](eval: => A) extends Observable[A] {
  def unsafeSubscribeFn(subscriber: Subscriber[A]): Cancelable = {
    val s = subscriber.scheduler
    val cancelable = BooleanCancelable()
    fastLoop(subscriber, cancelable, s.executionModel, 0)(s)
    cancelable
  }
  
  @tailrec
  def fastLoop(
    o: Subscriber[A], 
    // We might check it periodically to
    // see if the subscription is not cancelled
    c: BooleanCancelable, 
    // Scheduler has ExecutionModel, e.g. Synchronous, Batched, AlwaysAsync
    // We could add async boundaries according to it
    em: ExecutionModel,
    // BatchedExecution model inserts async boundary
    // after N synchronous operations
    syncIndex: Int
  )(implicit s: Scheduler): Unit = ???
}

twitter.com/p_gawrys

@tailrec
def fastLoop(
  o: Subscriber[A],
  c: BooleanCancelable, 
  em: ExecutionModel, 
  syncIndex: Int
)(implicit s: Scheduler): Unit = {
  val ack =
    try o.onNext(eval)
    catch {
      case ex if NonFatal(ex) =>
        Future.failed(ex)
    }

  val nextIndex =
    if (ack == Continue) em.nextFrameIndex(syncIndex)
    else if (ack == Stop) -1
    else 0

  if (nextIndex > 0)
    fastLoop(o, c, em, nextIndex)
  else if (nextIndex == 0 && !c.isCanceled)
    reschedule(ack, o, c, em)
}
    
def reschedule(
  ack: Future[Ack], 
  o: Subscriber[A], 
  c: BooleanCancelable, 
  em: ExecutionModel
)(implicit s: Scheduler): Unit = ???
def reschedule(
  ack: Future[Ack], 
  o: Subscriber[A], 
  c: BooleanCancelable, 
  em: ExecutionModel
)(implicit s: Scheduler): Unit =
  ack.onComplete {
    case Success(success) =>
      if (success == Continue) fastLoop(o, c, em, 0)
    case Failure(ex) =>
      s.reportFailure(ex)
    case _ => () // this was a Stop, do nothing
  }

twitter.com/p_gawrys

Observable#toListL

twitter.com/p_gawrys

abstract class Observable[+A] {    
  final def toListL: Task[List[A]] =
    foldLeft(mutable.ListBuffer.empty[A])(_ += _)
    // We know for sure that there will be only one element
    .firstOrElseL(mutable.ListBuffer.empty[A])
    .map(_.toList)
        
  final def foldLeft[R](seed: => R)(op: (R, A) => R): Observable[R] = ???
  
  final def firstOrElseL[B >: A](default: => B): Task[B] = ???
}

twitter.com/p_gawrys

final class FoldLeftObservable[A, R](
  source: Observable[A], 
  initial: () => R,
  f: (R, A) => R
) extends Observable[R] {
  def unsafeSubscribeFn(out: Subscriber[R]): Cancelable = {
    var streamErrors = true
    try {
      val initialState = initial()
      streamErrors = false

      source.unsafeSubscribeFn(new Subscriber[A] { ... })
    } catch {
      // If an error was thrown in source.unsafeSubscribeFn(...)
      // it is a breach of the protocol and the behavior is undefined
      // but we don't want to call out.onError in case it already happened there
      case NonFatal(ex) if streamErrors =>
        out.onError(ex)
        Cancelable.empty
    }
  }
}
      source.unsafeSubscribeFn(new Subscriber[A] {
        implicit val scheduler = out.scheduler
        // We might call onError in onNext so we need this 
        // flag to protect from potentially calling it twice
        // (once from onNext, once by upstream)
        private[this] var isDone = false
        private[this] var state: R = initialState

        def onNext(elem: A): Ack = {
          try {
            // User-supplied function
            // could throw exception
            state = f(state, elem)
            Continue
          } catch {
            case ex if NonFatal(ex) =>
              onError(ex)
              Stop
          }
        }

        def onComplete(): Unit =
          if (!isDone) {
              isDone = true
              out.onNext(state)
              out.onComplete()
            }

          def onError(ex: Throwable): Unit =
            if (!isDone) {
              isDone = true
              out.onError(ex)
            }
        })

Are those vars thread-safe?

twitter.com/p_gawrys

private[this] var isDone = false
private[this] var state: R = initialState
  • They can be modified and read from a different thread, after all...

Are those vars thread-safe?

twitter.com/p_gawrys

private[this] var isDone = false
private[this] var state: R = initialState
  • They can be modified and read from a different thread, after all...
  • but the protocol guarantees that and we'll see how!

Are those vars thread-safe?

twitter.com/p_gawrys

out.onNext(next).flatMap(_ => out2.onNext).flatMap(_ => out3.onNext) ...
out.onNext(next).flatMap(_ => out2.onNext).flatMap(_ => Continue) ...
out.onNext(next).flatMap(_ => Continue) ...
Continue

If we follow onNext calls, it goes like that:

And then the next element is sent after Continue is received (remember onComplete in repeatEval?)

Internally, each Future might be scheduled on a potentially different Thread with ec.execute():

Which establishes a happens-before relation between writing and reading isDone from potentially different threads.

var isDone = false

ec.execute(() => {
  isDone = true

  // second thread
  ec.execute(() => {
    assert(isDone)
  })
})

Observable#toListL

twitter.com/p_gawrys

abstract class Observable[+A] {    
  final def toListL: Task[List[A]] =
    foldLeft(mutable.ListBuffer.empty[A])(_ += _)
    // We know for sure that there will be only one element
    .firstOrElseL(mutable.ListBuffer.empty[A])
    .map(_.toList)
        
  final def foldLeft[R](seed: => R)(op: (R, A) => R): Observable[R] = 
    new FoldLeftObservable(source, seed, op)
  
  final def firstOrElseL[B >: A](default: => B): Task[B] = ???
}

Observable#firstOrElseL

final def firstOrElseL[B >: A](default: => B): Task[B] = 
  Task.create { (s, cb) =>
    unsafeSubscribeFn(new Subscriber[A] {
      implicit val scheduler: Scheduler = s
      private[this] var isDone = false

      def onNext(elem: A): Ack = {
        cb.onSuccess(elem)
        isDone = true
        Stop
      }

      def onError(ex: Throwable): Unit =
        if (!isDone) {
          isDone = true
          cb.onError(ex)
        }

      def onComplete(): Unit =
        if (!isDone) {
          isDone = true
          cb(Try(default))
        }
    })
  }

Observable#firstOrElseL Bonus!

final def firstOrElseLZIOOO[B >: A](default: => B): zio.Task[B] = {
  ZIO.descriptorWith { desc =>
    ZIO.effectAsync { cb =>
      unsafeSubscribeFn(new Subscriber[A] {
        implicit val scheduler: Scheduler =
          Scheduler(desc.executor.asEC)

        private[this] var isDone = false

        def onNext(elem: A): Ack = {
          cb(ZIO.succeed(elem))
          isDone = true
          Stop
        }

        def onError(ex: Throwable): Unit =
          if (!isDone) {
            isDone = true
            cb(ZIO.fail(ex))
          }

        def onComplete(): Unit =
          if (!isDone) {
            isDone = true
            cb(ZIO(default))
          }
      })
    }
  }
}

TakeLeftByTimespanObservable

twitter.com/p_gawrys

abstract class Observable[+A] {    
  final def takeByTimespan(timespan: FiniteDuration): Observable[A] =
    new TakeLeftByTimespanObservable(this, timespan)
}
  • Takes the elements until timespan passes
  • We could run source as usual but run it concurrently with a timeoutTask that will stop the source gracefully
final class TakeLeftByTimespanObservable[A](
  source: Observable[A], 
  timespan: FiniteDuration
) extends Observable[A] {

  def unsafeSubscribeFn(out: Subscriber[A]): Cancelable = {
    val composite = CompositeCancelable()

    composite += source.unsafeSubscribeFn(new Subscriber[A] with Runnable {
      implicit val scheduler = out.scheduler
      private[this] var isActive = true
      private[this] val timeoutTask: Cancelable = {
        val ref = scheduler.scheduleOnce(timespan.length, timespan.unit, this)
        composite += ref
        ref
      }

      def run(): Unit = onComplete()

      private def deactivate(): Unit = synchronized {
        isActive = false
        timeoutTask.cancel()
      }

      def onNext(elem: A): Future[Ack] = synchronized {
        if (isActive) out.onNext(elem).syncOnStopOrFailure(_ => deactivate())
        else Stop
      }

      def onError(ex: Throwable): Unit = synchronized {
        if (isActive) {
          deactivate()
          out.onError(ex)
        }
      }

      def onComplete(): Unit = synchronized {
        if (isActive) {
          deactivate()
          out.onComplete()
        }
      }
    })
  }
}

Notable implementation details

  • Subscriber extends Runnable - common optimization to minimize allocations
  • Access to isActive flag is synchronized because onComplete can be called from an asynchronous timeoutTask 
  • timeoutTask is added to the subscription cancelable
  • syncOnStopOrFailure optimization

twitter.com/p_gawrys

syncOnStopOrFailure

// F-bounded polymorphism, see 
// https://github.com/ghik/opinionated-scala/blob/master/chapters/Generics-and-type-members.md#f-bounded-polymorphism
implicit class AckExtensions[Self <: Future[Ack]](val source: Self) extends AnyVal {

  def syncOnStopOrFailure(
    cb: Option[Throwable] => Unit
  )(implicit r: UncaughtExceptionReporter): Self = {
    if (source eq Stop)
      try cb(None)
      catch { case e if NonFatal(e) => r.reportFailure(e) }
    else if (source ne Continue)
      source.onComplete { ack =>
        try ack match {
          case Success(Stop) => cb(None)
          case Failure(e) => cb(Some(e))
          case _ => ()
        } catch {
            case e if NonFatal(e) => r.reportFailure(e)
          }
      }(immediate)
      
    source
  }
}

Complete example

twitter.com/p_gawrys

val result: Task[List[Int]] =
  Observable.repeatEval(Random.nextInt(10))
    .takeByTimespan(10.second)
    .toListL

Could be inlined to:

Task.create { (s, cb) =>
  val source = 
    new FoldLeftObservable(
      new TakeLeftByTimespanObservable(
        new RepeatEvalObservable(Random.nextInt(10)), 
        10.second
      ), 
      mutable.ListBuffer.empty[Int]
      )(_ += _).firstOrElse().map(_.toList)
    
  source.unsafeSubscribeFn(new Subscriber[A] {
      implicit val scheduler: Scheduler = s
      private[this] var isDone = false

      def onNext(elem: A): Ack = {
        cb.onSuccess(elem)
        isDone = true
        Stop
      }

      def onError(ex: Throwable): Unit =
        if (!isDone) {
          isDone = true
          cb.onError(ex)
        }

      def onComplete(): Unit =
        if (!isDone) {
          isDone = true
          cb(Try(default))
        }
    })
    .map(_.toList)
}

What we didn't cover

  • Subject (both Observable and Observer)
  • BufferedSubscriber 
  • Hot Observable (sharing one Observable between multiple Subscribers)

twitter.com/p_gawrys

Benchmarks

I'm about to show few micro-benchmarks.

Please, keep in mind that the results can be misleading - it's best to measure for your specific use case.

 

API/Ecosystem/Familiarity is usually better criteria, as long as the library meets the minimum performance requirements.

twitter.com/p_gawrys

ChunkedMapFilterSum

twitter.com/p_gawrys

def monixObservableNoChunks(): Int = {
  val stream = Observable
    .fromIterable(allElements)
    .map(_ + 1)
    .filter(_ % 2 == 0)

  sum(stream)
}

def akkaStreamNoChunks(): Long = {
  val stream = AkkaSource
    .fromIterator(() => allElements.iterator)
    .map(_ + 1)
    .filter(_ % 2 == 0)
    .toMat(AkkaSink.fold(0L)(_ + _))(Keep.right)

  Await.result(stream.run(), Duration.Inf)
}
def zioStream(): Int = {
  val stream = ZStream
    .fromChunks(zioChunks: _*)
    .map(_ + 1)
    .filter(_ % 2 == 0)
    .runSum

  zioUntracedRuntime.unsafeRun(stream)
}

def fs2Stream(): Int = {
  val stream = FS2Stream(fs2Chunks: _*)
    .flatMap(FS2Stream.chunk)
    .map(_ + 1)
    .filter(_ % 2 == 0)
    .compile
    .fold(0)(_ + _)

  stream
}

ChunkedMapFilterSum

twitter.com/p_gawrys

[info] Benchmark  (chunkCount)  (chunkSize)   Mode  Cnt   Score   Error  Units
[info] akka               1000         1000  thrpt   20  10.866 ± 0.059  ops/s
[info] fs2                1000         1000  thrpt   20  55.301 ± 0.495  ops/s
[info] monix              1000         1000  thrpt   20  95.506 ± 0.241  ops/s
[info] zio                1000         1000  thrpt   20  32.106 ± 0.228  ops/s

MapAccumulate

twitter.com/p_gawrys

def monixMapAccumulate() = {
  Observable
    .fromIterable(0 until n)
    .mapAccumulate(0) { case (acc, i) =>
      val added = acc + i
      (added, added)
    }
    .completedL
    .runSyncUnsafe()
}
def zioMapAccumulate() = {
  val stream = ZStream
    .fromIterable(0 until n)
    .mapAccum(0) { case (acc, i) =>
      val added = acc + i
      (added, added)
    }
    .runDrain

  zioUntracedRuntime.unsafeRun(stream)
}

def fs2MapAccumulate() = {
  FS2Stream
    .emits(0 until n)
    .mapAccumulate(0) { case (acc, i) =>
      val added = acc + i
      (added, added)
    }
    .compile
    .drain
}

MapAccumulate

twitter.com/p_gawrys

[info] Benchmark  (n)   Mode  Cnt      Score     Error  Units
[info] fs2       1000  thrpt   20  66490.570 ± 211.840  ops/s
[info] fs2      10000  thrpt   20   8241.498 ±  52.588  ops/s
[info] monix     1000  thrpt   20  99300.153 ± 619.293  ops/s
[info] monix    10000  thrpt   20  10539.976 ± 203.321  ops/s
[info] zio       1000  thrpt   20   1819.379 ±  16.974  ops/s
[info] zio      10000  thrpt   20    201.752 ±   2.983  ops/s

ChunkedEvalFilterMapSum

twitter.com/p_gawrys

def fs2Stream = {
  val stream = FS2Stream
    .apply(allElements: _*)
    .chunkN(chunkSize)
    .evalMap[MonixTask, Int](chunk => MonixTask(sumIntScala(chunk.iterator)))
    .filter(_ > 0)
    .map(_.toLong)
    .compile
    .fold(0L)(_ + _)
}

def fs2StreamPreChunked = {
  val stream = FS2Stream(fs2Chunks: _*)
    .evalMap[MonixTask, Int](chunk => MonixTask(sumIntScala(chunk.iterator)))
    .filter(_ > 0)
    .map(_.toLong)
    .compile
    .fold(0L)(_ + _)
}

ChunkedEvalFilterMapSum

twitter.com/p_gawrys

[info] Benchmark (chunkCount)  (chunkSize)   Mode  Cnt    Score   Error  Units
[info] akka              1000         1000  thrpt   20   17.120 ± 0.418  ops/s
[info] akkaPreChunked    1000         1000  thrpt   20  214.725 ± 1.147  ops/s
[info] fs2               1000         1000  thrpt   20   63.284 ± 1.243  ops/s
[info] fs2PreChunked     1000         1000  thrpt   20  169.957 ± 7.040  ops/s
[info] monix             1000         1000  thrpt   20   77.922 ± 2.219  ops/s
[info] monixPreChunked   1000         1000  thrpt   20  364.595 ± 1.009  ops/s
[info] zio               1000         1000  thrpt   20  122.227 ± 5.387  ops/s
[info] zioPreChunked     1000         1000  thrpt   20  121.596 ± 2.566  ops/s

Tradeoffs

 

Cons:

  • Pure API, Dirty Internals - individual operators are hard to reason about in comparison to higher-level implementations of fs2/zio
  • Push Model - if you want to maximize throughput, you need to use buffers yourself
  • Shared Data Sources are not purely functional
  • Current implementation of flatMap is not stack-safe

twitter.com/p_gawrys

Tradeoffs

Pros:

  • Pure API, Dirty Internals - nice API and best-in-class performance 
  • Push Model - awesome for latency and time-based operators
  • Effect independent - Observable is fully capable of executing on its own, without any overhead of going through Task/IO Run-Loop, and could support all effect types natively
  • ReactiveX based - tons of related resources and a perfect step into FP for people coming from Java/JS :)

twitter.com/p_gawrys

Final words

  • If you have any questions or more ideas, make sure to let us know at https://github.com/monix/monix or https://gitter.im/monix/monix
  • Recently, I've released https://github.com/monix/monix-bio - Cats-Effect friendly IO[E, A] implementation
  • Contributions are very welcome!
  • ... Thank you for being here :)

twitter.com/p_gawrys

Understanding Monix Observable

By Piotr Gawryś

Understanding Monix Observable

  • 1,265