Build your own Kafka in ZIO – Queues & Fibers

Last update: 20 January 2021

Queueing and messaging platforms have been gaining in popularity in recent years. They solve numerous problems based on asynchronous message passing or consumer and producer patterns. In this blog post, we’re going to build a basic message broker functionality with ZIO for our internal clinic messaging system, specifically with ZIO Queues and ZIO Fibers.

In our clinic, we have x-ray rooms which produce x-ray photographs of hips and knees, which are sent via a messaging system. For any given body part, some physicians can perform a photographic analysis. Additionally, we want to be able to perform message logging for selected body parts.

This example accurately describes a message broker with topics: sending messages to defined topics, subscribing to them in two ways – the ‘one message one consumer’ type pattern and the multicast type pattern. We will be performing this subscription via consumer groups to which consumers subscribe within any particular topic.

Each topic’s message is delivered to every consumer group (like multicast), but within each group, only one consumer can digest the message (like producers and consumers). Here’s an image showing this:

Of course, there are plenty of distributed platforms that can achieve this, e.g. RabbitMQ provides us with a so-called exchange – a broker between a producer and queues that decides which queues to send the message to. Broadcast is supplied via a funout exchange, as opposed to direct and topic exchange types that require a match to the message’s topic.

So let’s try to implement this concept one more time, but this time with ZIO Queues and ZIO Fibers in an effective way.

ZIO Queues & ZIO Fibers

But first things first – let’s briefly introduce Fibers and Queues in ZIO.

So Fibers are data types for expressing concurrent computations. Fibers are loosely related to threads – a single Fiber can be executed on multiple threads by shifting between them – all with full resource safety!

What makes Fibers stronger is the seamless setting in ZIO. Having some effect e.g. UIO("work") we only need to call .fork on it to make it run on Fiber. Then it’s up to us what to do next: interrupt – stop Fiber by force, join – block current Fiber until it returns the result or races with another Fiber – runs two ZIO Fibers and returns the first that succeeded.

I should mention that the underlying implementation of race is done via raceWith – a powerful method that allows you to provide any logic for managing two separate Fibers. raceWith is used not only in race but also zipPar – for running two Fibers in parallel and returning both results as a tuple.

On the other hand, Queues in ZIO addresses issues that we can encounter while using BlockingQueue. The effective, back-pressured ZIO Queue makes it easy to avoid blocked threads on Queues core operations such as offer and take.

Apart from a bounded back-pressured queue, ZIO Queues deliver other overflow behaviors such as sliding – for removing the last inserted element, or dropping – for discarding the newly received elements. All this in a non-blocking manner.

So the moment we use queue.offer(sth).fork on a filled back-pressured queue, we are sure that running a separate fiber will make it non-blocking for the main one. Other ZIO Queue assets are interruption (as fibers are) and safe shutdown.

Domain

We’ll start with defining our domain and request class with a topic field.

Additionally, we will implement RequestGenerator for generating Requests:

sealed trait Diagnostic

case object HipDiagnostic extends Diagnostic

case object KneeDiagnostic extends Diagnostic

case class Request[A](topic: Diagnostic, XRayImage: A)

trait RequestGenerator[R, A] {
  def generate(topic: Diagnostic): URIO[R, Request[A]]
} 

Imports required by our project:

import zio._
import zio.random._
import zio.console._
import zio.duration._

For the sake of simplicity let’s assume our x-ray images are simply Ints:

case class IntRequestGenerator() extends RequestGenerator[Random, Int] {
  override def generate(topic: Diagnostic): URIO[Random, Request[Int]] =
    nextIntBounded(1000) >>= (n => UIO(Request(topic, n)))
}

Before getting started with the first part, let’s take a look at the architecture diagram. It might look strange at first so let’s leave it this way for now:

Consumer

The first component of our system is a Consumer[A]. Here we are providing two API methods – create for constructing a consumer wrapped in UIO and run that starts a new fiber that continuously waits for elements in its queue to process. The processing is rather dull but following console logs are definitely not!

It’s worth stressing that run returns (Queue, Fiber) in effect so apart from connecting the consumer to the system we can also interrupt or join the customer:

case class Consumer[A](title: String) {
  def run = for {
    queue <- Queue.bounded[A](10)
    loop = for {
      img  <- queue.take
      _    <- putStrLn(s"[$title] worker: Starting analyzing task $img")
      rand <- nextIntBounded(4)
      _    <- ZIO.sleep(rand.seconds)
      _    <- putStrLn(s"[$title] worker: Finished task $img")
    } yield ()
    fiber <- loop.forever.fork
  } yield (queue, fiber)
}

object Consumer {
  def create[A](title: String) = UIO(Consumer[A](title))
}

As we are more used to an imperative approach, let’s focus for a moment on the advantages of using ZIO effects here.

Any potentially dangerous side effects here are kept inside the ZIO monad. This makes a unit println method more substantial and, referentially transparent. Also, having a physical grasp on everything is really beneficial when it comes to parallelism.

Here, we were able to build an arbitrary chain of computations and make it run forever on a separate ZIO Fiber with a pleasing .forever.fork.

Topic Queue

TopicQueue is kind of the most complicated part. It’s in charge of the proper distribution of messages among subscribers. The subscribe method receives a subscriber’s queue and the consumerGroup number. As you will no doubt recall, each message is passed to each consumerGroup and then to a random subscriber within each group. The run method follows the pattern from previous components – a continuous loop of acquiring messages and distributing them within the described scheme:

case class TopicQueue[A](queue: Queue[A], subscribers: Ref[Map[Int, List[Queue[A]]]]) {
  def subscribe(sub: Queue[A], consumerGroup: Int): UIO[Unit] =
    subscribers.update { map =>
      map.get(consumerGroup) match {
        case Some(value) =>
          map + (consumerGroup -> (value :+ sub))
        case None =>
          map + (consumerGroup -> List(sub))
      }
    }

  private val loop =
    for {
      elem <- queue.take
      subs <- subscribers.get
      _    <- ZIO.foreach(subs.values) { group =>
        for {
          idx <- nextIntBounded(group.length)
          _   <- group(idx).offer(elem)
        } yield ()
      }
    } yield ()

  def run = loop.forever.fork
}

object TopicQueue {
  def create[A](queue: Queue[A]): UIO[TopicQueue[A]] =
    Ref.make(Map.empty[Int, List[Queue[A]]]) >>= (map => UIO(TopicQueue(queue, map)))
}

In this part, immutability is what strikes us first. No explicit, side-effect modifications of a subscribers map can occur without our knowledge. Here we’re using Ref from ZIO to store the map and perform updates.

It’s worth mentioning that wrapping the constructor method in UIO is essential for consistency, as instantiating a new ZIO Queue should always be a part of our effect chain.

Exchange

Our Exchange is pretty similar to the RabbitMQ exchange. The constructor simply creates three queues – the input queue for incoming jobs (jobQueue) and two output queues for routing (queueHip and queueKnee). What our exchange is also doing is unwrapping XRayImage from Request:

case class Exchange[A]() {
  def run = for {
    jobQueue       <- Queue.bounded[Request[A]](10)
    queueHip       <- Queue.bounded[A](10)
    queueKnee      <- Queue.bounded[A](10)
    hipTopicQueue  <- TopicQueue.create(queueHip)
    kneeTopicQueue <- TopicQueue.create(queueKnee)
    loop = for {
      job <- jobQueue.take
      _   <- job.topic match {
        case HipDiagnostic =>
          queueHip.offer(job.XRayImage)
        case KneeDiagnostic =>
          queueKnee.offer(job.XRayImage)
      }
    } yield ()
    fiber <- loop.forever.fork
  } yield (jobQueue, hipTopicQueue, kneeTopicQueue, fiber)
}

object Exchange {
  def create[A] = UIO(Exchange[A]())
}

Producer

Producing is simply done by supplying a provided queue with Requests. You might have noticed that the run method follows some patterns. Building asynchronous computations with self-explanatory schedules and a lazy execution is easy:

case class Producer[R, A](queue: Queue[Request[A]], generator: RequestGenerator[R, A]) {
  def run = {
    val loop = for {
      _    <- putStrLn("[XRayRoom] generating hip and knee request")
      hip  <- generator.generate(HipDiagnostic)
      _    <- queue.offer(hip)
      knee <- generator.generate(KneeDiagnostic)
      _    <- queue.offer(knee)
      _    <- ZIO.sleep(2.seconds)
    } yield ()
    loop.forever.fork
  }
}

object Producer {
  def create[R, A](queue: Queue[Request[A]], generator: RequestGenerator[R, A]) = UIO(Producer(queue, generator))
}

Program

Finally, the Program. Now we will combine all the previous components to assemble a fully operational clinic messaging system. First, we instantiate Consumers and launch them (reminder: ZIO Fibers are lazy, unlike Futures). Then it’s time for Exchange and Producer. Notice that returning tuples gives a  possibility to ignore the fibers that we don’t need. Finally, we subscribe Consumers for the output queues and, importantly, define the ConsumerGroup with the launch:

val program = for {

  physicianHip             <- Consumer.create[Int]("Hip")
  ctxPhHip                 <- physicianHip.run
  (phHipQueue, phHipFiber) = ctxPhHip

  loggerHip           <- Consumer.create[Int]("HIP_LOGGER")
  ctxLoggerHip        <- loggerHip.run
  (loggerHipQueue, _) = ctxLoggerHip

  physicianKnee    <- Consumer.create[Int]("Knee1")
  ctxPhKnee        <- physicianKnee.run
  (phKneeQueue, _) = ctxPhKnee

  physicianKnee2    <- Consumer.create[Int]("Knee2")
  ctxPhKnee2        <- physicianKnee2.run
  (phKneeQueue2, _) = ctxPhKnee2


  exchange                                         <- Exchange.create[Int]
  ctxExchange                                      <- exchange.run
  (inputQueue, outputQueueHip, outputQueueKnee, _) = ctxExchange


  generator = IntRequestGenerator()
  xRayRoom  <- Producer.create(inputQueue, generator)
  _         <- xRayRoom.run


  _ <- outputQueueHip.subscribe(phHipQueue, consumerGroup = 1)
  _ <- outputQueueHip.subscribe(loggerHipQueue, consumerGroup = 2)

  _ <- outputQueueKnee.subscribe(phKneeQueue, consumerGroup = 1)
  _ <- outputQueueKnee.subscribe(phKneeQueue2, consumerGroup = 1)

  _ <- outputQueueHip.run
  _ <- outputQueueKnee.run

  _ <- phHipFiber.join

} yield ()

Also after launching TopicQueues with run, we can still subscribe to them.

Running the program

Phew… that was a lot, let’s put it into the ZIO application and run it:

object Main extends App {
  override def run(args: List[String]) = program.exitCode
}

Looking into the logs we see that:

1. Multicast for all the ConsumerGroups within the hip topic works as expected – hip physician and HIP_LOGGER receive the same messages.

2. Within a single ConsumerGroup the messages are routed in a random manner (definitely field for improvement!):

[XRayRoom] generating hip and knee request
[Knee1] worker: Starting analyzing task 474
[Hip] worker: Starting analyzing task 345
[Hip] worker: Finished task 345
[HIP_LOGGER] worker: Starting analyzing task 345
[HIP_LOGGER] worker: Finished task 345
[XRayRoom] generating hip and knee request
[Hip] worker: Starting analyzing task 179
[HIP_LOGGER] worker: Starting analyzing task 179
[Hip] worker: Finished task 179
[Knee1] worker: Finished task 474
[Knee1] worker: Starting analyzing task 154
[Knee1] worker: Finished task 154
[XRayRoom] generating hip and knee request
[Hip] worker: Starting analyzing task 763
[Knee1] worker: Starting analyzing task 562
[HIP_LOGGER] worker: Finished task 179
[HIP_LOGGER] worker: Starting analyzing task 763
[Hip] worker: Finished task 763
[Knee1] worker: Finished task 562
[HIP_LOGGER] worker: Finished task 763
[XRayRoom] generating hip and knee request
[Hip] worker: Starting analyzing task 474
[Knee2] worker: Starting analyzing task 997
[HIP_LOGGER] worker: Starting analyzing task 474
[Hip] worker: Finished task 474
[XRayRoom] generating hip and knee request
[Hip] worker: Starting analyzing task 184
[Knee1] worker: Starting analyzing task 578
[Knee2] worker: Finished task 997
[HIP_LOGGER] worker: Finished task 474

Conclusion

Our simple, yet operational, program shows how to implement a message broker with direct and multicast behaviors.

Having chosen ZIO we have managed to unearth only a fraction of its potential – by using ZIO Queues and ZIO Fibers within effects. Out of the box parallelism, immutability, referential transparency, and wrapped side effect managing are what has made this example painless and really very enjoyable to write.

To see complete example see gist link below.

Links

https://gist.github.com/mtsokol/0d6ab5473c04583899e3ffdcb7812959

https://github.com/zio/zio

More about ZIO on our BLOG:

• Why Developers Should Pay Attention to ZIO in 2023
• ZIO 1.0 is here! Join ZIO Hackathon by Scalac & Ziverge
• Making ZIO, Akka and Slick play together nicely
• Streaming microservices with ZIO and Kafka
• Introduction to Programming with ZIO Functional Effects
• What is ZIO?
• Optics beyond Lenses with Monocle