Routers

You are viewing the documentation for the new actor APIs, to view the Pekko Classic documentation, see Classic Routing.

Dependency

To use Pekko Actor Typed, you must add the following dependency in your project:

sbt

val PekkoVersion = "1.0.0"
libraryDependencies += "org.apache.pekko" %% "pekko-actor-typed" % PekkoVersion

Maven

<properties>
  <scala.binary.version>2.13</scala.binary.version>
</properties>
<dependencyManagement>
  <dependencies>
    <dependency>
      <groupId>org.apache.pekko</groupId>
      <artifactId>pekko-bom_${scala.binary.version}</artifactId>
      <version>1.0.0</version>
      <type>pom</type>
      <scope>import</scope>
    </dependency>
  </dependencies>
</dependencyManagement>
<dependencies>
  <dependency>
    <groupId>org.apache.pekko</groupId>
    <artifactId>pekko-actor-typed_${scala.binary.version}</artifactId>
  </dependency>
</dependencies>

Gradle

def versions = [
  ScalaBinary: "2.13"
]
dependencies {
  implementation platform("org.apache.pekko:pekko-bom_${versions.ScalaBinary}:1.0.0")

  implementation "org.apache.pekko:pekko-actor-typed_${versions.ScalaBinary}"
}

Introduction

In some cases it is useful to distribute messages of the same type over a set of actors, so that messages can be processed in parallel - a single actor will only process one message at a time.

The router itself is a behavior that is spawned into a running actor that will then forward any message sent to it to one final recipient out of the set of routees.

There are two kinds of routers included in Pekko Typed - the pool router and the group router.

Pool Router

The pool router is created with a routee Behavior and spawns a number of children with that behavior which it will then forward messages to.

If a child is stopped the pool router removes it from its set of routees. When the last child stops the router itself stops. To make a resilient router that deals with failures the routee Behavior must be supervised.

As actor children are always local the routees are never spread across a cluster with a pool router.

Let’s first introduce the routee:

Scala

sourceobject Worker {
  sealed trait Command
  case class DoLog(text: String) extends Command

  def apply(): Behavior[Command] = Behaviors.setup { context =>
    context.log.info("Starting worker")

    Behaviors.receiveMessage {
      case DoLog(text) =>
        context.log.info("Got message {}", text)
        Behaviors.same
    }
  }
}

Java

sourceclass Worker {
  interface Command {}

  static class DoLog implements Command {
    public final String text;

    public DoLog(String text) {
      this.text = text;
    }
  }

  static final Behavior<Command> create() {
    return Behaviors.setup(
        context -> {
          context.getLog().info("Starting worker");

          return Behaviors.receive(Command.class)
              .onMessage(DoLog.class, doLog -> onDoLog(context, doLog))
              .build();
        });
  }

  private static Behavior<Command> onDoLog(ActorContext<Command> context, DoLog doLog) {
    context.getLog().info("Got message {}", doLog.text);
    return Behaviors.same();
  }
}

static class Proxy {

  public final ServiceKey<Message> registeringKey =
      ServiceKey.create(Message.class, "aggregator-key");

  public String mapping(Message message) {
    return message.getId();
  }

  static class Message {

    public Message(String id, String content) {
      this.id = id;
      this.content = content;
    }

    private String content;
    private String id;

    public final String getContent() {
      return content;
    }

    public final String getId() {
      return id;
    }
  }

  static Behavior<Message> create(ActorRef<String> monitor) {
    return Behaviors.receive(Message.class)
        .onMessage(Message.class, in -> onMyMessage(monitor, in))
        .build();
  }

  private static Behavior<Message> onMyMessage(ActorRef<String> monitor, Message message) {
    monitor.tell(message.getId());
    return Behaviors.same();
  }
}

After having defined the routee, we can now concentrate on configuring the router itself. Note again that the router is an Actor in itself:

Scala

sourceimport org.apache.pekko
import pekko.actor.testkit.typed.scaladsl.{ LogCapturing, ScalaTestWithActorTestKit }
import pekko.actor.typed.{ ActorRef, Behavior, SupervisorStrategy }
import pekko.actor.typed.receptionist.{ Receptionist, ServiceKey }
import pekko.actor.typed.scaladsl.{ Behaviors, Routers }

// This would be defined within your actor object
Behaviors.setup[Unit] { ctx =>
  val pool = Routers.pool(poolSize = 4) {
    // make sure the workers are restarted if they fail
    Behaviors.supervise(Worker()).onFailure[Exception](SupervisorStrategy.restart)
  }
  val router = ctx.spawn(pool, "worker-pool")

  (0 to 10).foreach { n =>
    router ! Worker.DoLog(s"msg $n")
  }

  val poolWithBroadcast = pool.withBroadcastPredicate(_.isInstanceOf[DoBroadcastLog])
  val routerWithBroadcast = ctx.spawn(poolWithBroadcast, "pool-with-broadcast")
  // this will be sent to all 4 routees
  routerWithBroadcast ! DoBroadcastLog("msg")
  Behaviors.empty
}

Java

sourceimport org.apache.pekko.actor.typed.ActorRef;
import org.apache.pekko.actor.typed.Behavior;
import org.apache.pekko.actor.typed.SupervisorStrategy;
import org.apache.pekko.actor.typed.javadsl.ActorContext;
import org.apache.pekko.actor.typed.javadsl.Behaviors;
import org.apache.pekko.actor.typed.javadsl.GroupRouter;
import org.apache.pekko.actor.typed.javadsl.PoolRouter;
import org.apache.pekko.actor.typed.javadsl.Routers;
import org.apache.pekko.actor.typed.receptionist.Receptionist;
import org.apache.pekko.actor.typed.receptionist.ServiceKey;

import org.apache.pekko.actor.testkit.typed.javadsl.TestKitJunitResource;
import org.apache.pekko.actor.testkit.typed.javadsl.TestProbe;

import org.junit.ClassRule;
import org.junit.Test;
import org.scalatestplus.junit.JUnitSuite;

// This would be defined within your actor class
Behaviors.setup(
    context -> {
      int poolSize = 4;
      PoolRouter<Worker.Command> pool =
          Routers.pool(
              poolSize,
              // make sure the workers are restarted if they fail
              Behaviors.supervise(Worker.create()).onFailure(SupervisorStrategy.restart()));
      ActorRef<Worker.Command> router = context.spawn(pool, "worker-pool");

      for (int i = 0; i < 10; i++) {
        router.tell(new Worker.DoLog("msg " + i));
      }
    });

Configuring Dispatchers

Since the router itself is spawned as an actor the dispatcher used for it can be configured directly in the call to spawn. The routees, however, are spawned by the router. Therefore, the PoolRouter has a property to configure the Props of its routees:

Scala

source// make sure workers use the default blocking IO dispatcher
val blockingPool = pool.withRouteeProps(routeeProps = DispatcherSelector.blocking())
// spawn head router using the same executor as the parent
val blockingRouter = ctx.spawn(blockingPool, "blocking-pool", DispatcherSelector.sameAsParent())

Java

source// make sure workers use the default blocking IO dispatcher
PoolRouter<Worker.Command> blockingPool =
    pool.withRouteeProps(DispatcherSelector.blocking());
// spawn head router using the same executor as the parent
ActorRef<Worker.Command> blockingRouter =
    context.spawn(blockingPool, "blocking-pool", DispatcherSelector.sameAsParent());

Broadcasting a message to all routees

Pool routers can be configured to identify messages intended to be broad-casted to all routees. Therefore, the PoolRouter has a property to configure its broadcastPredicate:

Scala

sourceval poolWithBroadcast = pool.withBroadcastPredicate(_.isInstanceOf[DoBroadcastLog])
val routerWithBroadcast = ctx.spawn(poolWithBroadcast, "pool-with-broadcast")
// this will be sent to all 4 routees
routerWithBroadcast ! DoBroadcastLog("msg")
Behaviors.empty

Java

sourcePoolRouter<Worker.Command> broadcastingPool =
    pool.withBroadcastPredicate(msg -> msg instanceof DoBroadcastLog);

Group Router

The group router is created with a ServiceKey and uses the receptionist (see Receptionist) to discover available actors for that key and routes messages to one of the currently known registered actors for a key.

Since the receptionist is used this means the group router is cluster-aware out of the box. The router sends messages to registered actors on any node in the cluster that is reachable. If no reachable actor exists the router will fallback and route messages to actors on nodes marked as unreachable.

That the receptionist is used also means that the set of routees is eventually consistent, and that immediately when the group router is started the set of routees it knows about is empty, until it has seen a listing from the receptionist it stashes incoming messages and forwards them as soon as it gets a listing from the receptionist.

When the router has received a listing from the receptionist and the set of registered actors is empty the router will drop them (publish them to the event stream as org.apache.pekko.actor.Dropped).

Scala

sourceval serviceKey = ServiceKey[Worker.Command]("log-worker")

Behaviors.setup[Unit] { ctx =>
  // this would likely happen elsewhere - if we create it locally we
  // can just as well use a pool
  val worker = ctx.spawn(Worker(), "worker")
  ctx.system.receptionist ! Receptionist.Register(serviceKey, worker)

  val group = Routers.group(serviceKey)
  val router = ctx.spawn(group, "worker-group")

  // the group router will stash messages until it sees the first listing of registered
  // services from the receptionist, so it is safe to send messages right away
  (0 to 10).foreach { n =>
    router ! Worker.DoLog(s"msg $n")
  }

  Behaviors.empty
}

Java

sourceServiceKey<Worker.Command> serviceKey = ServiceKey.create(Worker.Command.class, "log-worker");

Behaviors.setup(
    context -> {

      // this would likely happen elsewhere - if we create it locally we
      // can just as well use a pool
      ActorRef<Worker.Command> worker = context.spawn(Worker.create(), "worker");
      context.getSystem().receptionist().tell(Receptionist.register(serviceKey, worker));

      GroupRouter<Worker.Command> group = Routers.group(serviceKey);
      ActorRef<Worker.Command> router = context.spawn(group, "worker-group");

      // the group router will stash messages until it sees the first listing of
      // registered
      // services from the receptionist, so it is safe to send messages right away
      for (int i = 0; i < 10; i++) {
        router.tell(new Worker.DoLog("msg " + i));
      }

      return Behaviors.empty();
    });

Routing strategies

There are three different strategies for selecting which routee a message is forwarded to that can be selected from the router before spawning it:

Scala

sourceval alternativePool = pool.withPoolSize(2).withRoundRobinRouting()

Java

sourcePoolRouter<Worker.Command> alternativePool = pool.withPoolSize(2).withRoundRobinRouting();

Round Robin

Rotates over the set of routees making sure that if there are n routees, then for n messages sent through the router, each actor is forwarded one message.

Round robin gives fair routing where every available routee gets the same amount of messages as long as the set of routees stays relatively stable, but may be unfair if the set of routees changes a lot.

This is the default for pool routers as the pool of routees is expected to remain the same.

An optional parameter preferLocalRoutees can be used for this strategy. Routers will only use routees located in local actor system if preferLocalRoutees is true and local routees do exist. The default value for this parameter is false.

Random

Randomly selects a routee when a message is sent through the router.

This is the default for group routers as the group of routees is expected to change as nodes join and leave the cluster.

An optional parameter preferLocalRoutees can be used for this strategy. Routers will only use routees located in local actor system if preferLocalRoutees is true and local routees do exist. The default value for this parameter is false.

Consistent Hashing

Uses consistent hashing to select a routee based on the sent message. This article gives good insight into how consistent hashing is implemented.

Currently you have to define hashMapping of the router to map incoming messages to their consistent hash key. This makes the decision transparent for the sender.

Consistent hashing delivers messages with the same hash to the same routee as long as the set of routees stays the same. When the set of routees changes, consistent hashing tries to make sure, but does not guarantee, that messages with the same hash are routed to the same routee.

Scala

sourceval router = spawn(Routers.group(Proxy.RegisteringKey).withConsistentHashingRouting(10, Proxy.mapping))

router ! Proxy.Message("123", "Text1")
router ! Proxy.Message("123", "Text2")

router ! Proxy.Message("zh3", "Text3")
router ! Proxy.Message("zh3", "Text4")
// the hash is calculated over the Proxy.Message first parameter obtained through the Proxy.mapping function

Java

sourceActorRef<Proxy.Message> router =
    testKit.spawn(
        Routers.group(proxy.registeringKey)
            .withConsistentHashingRouting(10, command -> proxy.mapping(command)));

final String id1 = "123";
router.tell(new Proxy.Message(id1, "Text1"));
router.tell(new Proxy.Message(id1, "Text2"));

final String id2 = "abcdef";
router.tell(new Proxy.Message(id2, "Text3"));
router.tell(new Proxy.Message(id2, "Text4"));
// the hash is calculated over the Proxy.Message first parameter obtained through the
// Proxy.mapping function

See also Pekko Cluster Sharding which provides stable routing and rebalancing of the routee actors.

Routers and performance

Note that if the routees are sharing a resource, the resource will determine if increasing the number of actors will actually give higher throughput or faster answers. For example if the routees are CPU bound actors it will not give better performance to create more routees than there are threads to execute the actors.

Since the router itself is an actor and has a mailbox this means that messages are routed sequentially to the routees where it can be processed in parallel (depending on the available threads in the dispatcher). In a high throughput use cases the sequential routing could become a bottle neck. Pekko Typed does not provide an optimized tool for this.

Edit this page

1.0.0