fs2 groupBy / KeyedEnqueue

fs2utils.scala

import cats.effect.{Concurrent, Timer}
import fs2.{Chunk, Pipe, Stream}

import scala.concurrent.duration.FiniteDuration

package object fs2utils {

  /**
    * Grouping logic used to split a stream into sub-streams identified by a unique key.
    * Be aware that when working with an unbounded number of keys K, if streams never
    * terminate, there can potentially be unbounded memory usage.
    *
    * The emitted streams terminate when the input terminates or when the stream
    * returned by the pipe itself terminates. Termination is graceful and all input elements are emitted
    *
    * @param selector Function to retrieve grouping key from type A
    * @tparam A Elements in the stream
    * @tparam K A key used to group elements into substreams
    * @return Streams grouped by a unique key identifier `K`.
    */
  def groupByUnbounded[F[_], A, K](selector: A => K)(
      implicit F: Concurrent[F]
  ): Pipe[F, A, (K, Stream[F, A])] = { in =>
    Stream
      .resource(KeyedEnqueue.unbounded[F, K, A])
      .flatMap { ke =>
        in.through(KeyedEnqueue.pipe(ke)(selector))
      }
  }

  /** Like `groupByUnbounded` but back pressures the stream when `maxItems` are inside */
  def groupBy[F[_], A, K](maxItems: Long)(selector: A => K)(
      implicit F: Concurrent[F]
  ): Pipe[F, A, (K, Stream[F, A])] = { in =>
    Stream
      .resource(KeyedEnqueue.itemBounded[F, K, A](maxItems))
      .flatMap { ke =>
        in.through(KeyedEnqueue.pipe(ke)(selector))
      }
  }

  /**
    * Like `groupBy` but each substream is concurrently merged, emitting chunks when the substream has
    * `maxChunkSize` pending or when the substream has waited `maxChunkTimeout` without emitting elements,
    * similar to the standard `groupWithin` combinator
    *
    * @param maxTotalItems Backpressure when this many items are "in flight" concurrently
    * @param maxChunkSize Output chunks satisfy: 0 < emittedChunk.size <= maxChunkSize
    * @param maxChunkTimeout Emit chunks smaller than `maxChunkSize` if `maxChunkTimeout` time has elapsed without
    *                        emitting any chunks for a given key `K` and we have elements that match that selector waiting
    * @param selector Output elements satisfy: (key, chunk) => chunk.forall(a => selector(a) == key)
    */
  def groupWithinBy[F[_], A, K](
      maxTotalItems: Long,
      maxChunkSize: Int,
      maxChunkTimeout: FiniteDuration
  )(
      selector: A => K
  )(implicit F: Concurrent[F], timer: Timer[F]): Pipe[F, A, (K, Chunk[A])] =
    _.through(groupBy(maxTotalItems)(selector)).map {
      case (key, stream) =>
        stream
          .groupWithin(maxChunkSize, maxChunkTimeout)
          .map(chunk => key -> chunk)
    }.parJoinUnbounded
}

KeyedEnqueue.scala

import cats.Monad
import cats.effect.concurrent.{Ref, Semaphore}
import cats.effect.{Concurrent, Resource}
import cats.implicits._
import fs2.{Pipe, Stream}
import fs2.concurrent.{NoneTerminatedQueue, Queue}

/** Represents the ability to enqueue keyed items into a stream of queues that emits homogenous keyed streams.
  *
  * This allows construction of a "keyed fan-out" behavior for a stream, which may be used for
  * homogenously batching items that arrive via a heterogenous input
  *
  * Somewhat analogous to [[fs2.concurrent.Enqueue]]
  */
private[fs2utils] trait KeyedEnqueue[F[_], K, A] {

  /** Enqueue a single item for a given key, possibly creating and returning a new substream for that key
    *
    * @return <ul><li> None if the item was published to an already-live substream </li>
    *         <li>Some if a new queue was created for this element. This can happen multiple times for the same
    *         key (for example, if the implementation automatically terminates old/quiet substreams).</li></ul>
    *         The returned stream may eventually terminate, but it won't be cancelled by this.
    */
  def enqueue1(key: K, item: A): F[Option[(K, Stream[F, A])]]

  /** Gracefully terminate all sub-streams we have emitted so far */
  def shutdownAll: F[Unit]
}

private[fs2utils] object KeyedEnqueue {

  def unbounded[F[_]: Concurrent, K, A]: Resource[F, KeyedEnqueue[F, K, A]] =
    Resource.liftF(Ref[F].of(Map.empty[K, NoneTerminatedQueue[F, A]])).flatMap {
      st =>
        Resource.make(
          (new UnboundedKeyedEnqueue[F, K, A](st): KeyedEnqueue[F, K, A])
            .pure[F])(_.shutdownAll)

    }

  def pipe[F[_]: Concurrent, K, A](
      ke: KeyedEnqueue[F, K, A]
  )(selector: A => K): Pipe[F, A, (K, Stream[F, A])] = { in =>
    // Note this *must* be `++` specifically to allow for "input termination = output termination" behavior.
    // Using `onFinalize` will allow the finalizer to be rescoped to the output of this stream later, which
    // results in it not triggering because it's waiting for itself to terminate before it terminates itself
    in.evalMap(a => ke.enqueue1(selector(a), a)).unNone ++
      Stream.eval_(ke.shutdownAll)
  }

  def itemBounded[F[_]: Concurrent, K, A](
      maxItems: Long
  ): Resource[F, KeyedEnqueue[F, K, A]] =
    for {
      ke <- unbounded[F, K, A]
      limit <- Resource.liftF(Semaphore[F](maxItems))
    } yield new ItemBoundedKeyedEnqueue(ke, limit)

}

private class UnboundedKeyedEnqueue[F[_], K, A](
    queues: Ref[F, Map[K, NoneTerminatedQueue[F, A]]]
)(implicit F: Concurrent[F])
    extends KeyedEnqueue[F, K, A] {
  override def enqueue1(key: K, item: A): F[Option[(K, Stream[F, A])]] =
    withKey(key)(_.enqueue1(item.some))

  override val shutdownAll: F[Unit] =
    queues.get.flatMap(_.values.toList.traverse_(_.enqueue1(None)))

  private[this] def withKey(key: K)(
      use: NoneTerminatedQueue[F, A] => F[Unit]
  ): F[Option[(K, Stream[F, A])]] =
    queues.get.flatMap { qm =>
      qm.get(key)
        .fold {
          for { // No queue for key - create new one
            newQ <- Queue.noneTerminated[F, A]
            _ <- queues.update(x => x + (key -> newQ))
            _ <- use(newQ)
          } yield (key -> newQ.dequeue).some
        }(q => use(q).as(None))
    }

}

private class ItemBoundedKeyedEnqueue[F[_]: Monad, K, A](
    ke: KeyedEnqueue[F, K, A],
    limit: Semaphore[F]
) extends KeyedEnqueue[F, K, A] {

  override def enqueue1(key: K, item: A): F[Option[(K, Stream[F, A])]] =
    limit.acquire >> ke
      .enqueue1(key, item)
      .map(_.map {
        case (key, stream) =>
          // We only need to attach the "release" behavior to a given stream once because each stream is emitted once, and then reused
          key -> stream.chunks
            .evalTap(c => limit.releaseN(c.size.toLong))
            .flatMap(Stream.chunk)
      })

  override val shutdownAll: F[Unit] = ke.shutdownAll
}

You don't need the full power of a Semaphore, but you do want to swap that get + update with a modify

Thanks!

To me it looks like modify is not possible here as updated Map depends on effectful Queue.noneTerminated[F, A]:

queuesMap.get(key) match {
  case Some(queue) =>
    (queuesMap, use(queue).as(None))
  case None =>
    val action = for {
      newQ <- Queue.noneTerminated[F, A]
      updadted = queuesMap + (key -> newQ)      // We cannot return it
      result <- use(newQ).as((key -> newQ.dequeue).some)
    } yield result

    (???, action)
}

Would be happy to be wrong.

Yeah, it's not immediately obvious how to avoid that - it may involve refactoring. Possible using something like the keyed ref from davenverse

I feel like I reasoned out that the race condition was in practice not a problem, but I need to re-examine with fresh eyes. I wrote this quite a while ago.

o me it looks like modify is not possible here as updated Map depends on effectful Queue.noneTerminated[F, A]

you create the Queue in advance (which has negligible overhead), just like you do it when you need a Deferred. In any case I'm adding a similar combinator to the library directly

@SystemFw Hi, has such combinator already been added to fs2? Thanks :)

In any case I'm adding a similar combinator to the library directly

Have you done it? @SystemFw

No, I started on it at the time, and then got side-tracked in creating what has become Channel. I also felt unsatisfied by how large the possible api surface could be, but I will admit that's probably the perfect being the enemy of the good.