Cluster Sharding
Loading

Cluster Sharding

Cluster sharding is useful when you need to distribute actors across several nodes in the cluster and want to be able to interact with them using their logical identifier, but without having to care about their physical location in the cluster, which might also change over time.

It could for example be actors representing Aggregate Roots in Domain-Driven Design terminology. Here we call these actors "entries". These actors typically have persistent (durable) state, but this feature is not limited to actors with persistent state.

Cluster sharding is typically used when you have many stateful actors that together consume more resources (e.g. memory) than fit on one machine. If you only have a few stateful actors it might be easier to run them on a Cluster Singleton node.

In this context sharding means that actors with an identifier, so called entries, can be automatically distributed across multiple nodes in the cluster. Each entry actor runs only at one place, and messages can be sent to the entry without requiring the sender() to know the location of the destination actor. This is achieved by sending the messages via a ShardRegion actor provided by this extension, which knows how to route the message with the entry id to the final destination.

An Example in Java

This is how an entry actor may look like:

public class Counter extends UntypedPersistentActor {

  public static enum CounterOp {
    INCREMENT, DECREMENT
  }

  public static class Get {
    final public long counterId;

    public Get(long counterId) {
      this.counterId = counterId;
    }
  }

  public static class EntryEnvelope {
    final public long id;
    final public Object payload;

    public EntryEnvelope(long id, Object payload) {
      this.id = id;
      this.payload = payload;
    }
  }

  public static class CounterChanged {
    final public int delta;

    public CounterChanged(int delta) {
      this.delta = delta;
    }
  }

  int count = 0;
  
  // getSelf().path().parent().name() is the type name (utf-8 URL-encoded) 
  // getSelf().path().name() is the entry identifier (utf-8 URL-encoded)
  @Override
  public String persistenceId() {
    return getSelf().path().parent().name() + "-" + getSelf().path().name();
  }

  @Override
  public void preStart() throws Exception {
    super.preStart();
    context().setReceiveTimeout(Duration.create(120, TimeUnit.SECONDS));
  }

  void updateState(CounterChanged event) {
    count += event.delta;
  }

  @Override
  public void onReceiveRecover(Object msg) {
    if (msg instanceof CounterChanged)
      updateState((CounterChanged) msg);
    else
      unhandled(msg);
  }

  @Override
  public void onReceiveCommand(Object msg) {
    if (msg instanceof Get)
      getSender().tell(count, getSelf());

    else if (msg == CounterOp.INCREMENT)
      persist(new CounterChanged(+1), new Procedure<CounterChanged>() {
        public void apply(CounterChanged evt) {
          updateState(evt);
        }
      });

    else if (msg == CounterOp.DECREMENT)
      persist(new CounterChanged(-1), new Procedure<CounterChanged>() {
        public void apply(CounterChanged evt) {
          updateState(evt);
        }
      });

    else if (msg.equals(ReceiveTimeout.getInstance()))
      getContext().parent().tell(
          new ShardRegion.Passivate(PoisonPill.getInstance()), getSelf());

    else
      unhandled(msg);
  }
}

The above actor uses event sourcing and the support provided in UntypedPersistentActor to store its state. It does not have to be a persistent actor, but in case of failure or migration of entries between nodes it must be able to recover its state if it is valuable.

Note how the persistenceId is defined. You may define it another way, but it must be unique.

When using the sharding extension you are first, typically at system startup on each node in the cluster, supposed to register the supported entry types with the ClusterSharding.start method. ClusterSharding.start gives you the reference which you can pass along.

ActorRef startedCounterRegion = ClusterSharding.get(system).start("Counter", Props.create(Counter.class),
          messageExtractor);

The messageExtractor defines application specific methods to extract the entry identifier and the shard identifier from incoming messages.

ShardRegion.MessageExtractor messageExtractor = new ShardRegion.MessageExtractor() {

  @Override
  public String entryId(Object message) {
    if (message instanceof Counter.EntryEnvelope)
      return String.valueOf(((Counter.EntryEnvelope) message).id);
    else if (message instanceof Counter.Get)
      return String.valueOf(((Counter.Get) message).counterId);
    else
      return null;
  }

  @Override
  public Object entryMessage(Object message) {
    if (message instanceof Counter.EntryEnvelope)
      return ((Counter.EntryEnvelope) message).payload;
    else
      return message;
  }

  @Override
  public String shardId(Object message) {
    if (message instanceof Counter.EntryEnvelope) {
      long id = ((Counter.EntryEnvelope) message).id;
      return String.valueOf(id % 10);
    } else if (message instanceof Counter.Get) {
      long id = ((Counter.Get) message).counterId;
      return String.valueOf(id % 10);
    } else {
      return null;
    }
  }

};

This example illustrates two different ways to define the entry identifier in the messages:

  • The Get message includes the identifier itself.
  • The EntryEnvelope holds the identifier, and the actual message that is sent to the entry actor is wrapped in the envelope.

Note how these two messages types are handled in the entryId and entryMessage methods shown above.

A shard is a group of entries that will be managed together. The grouping is defined by the shardResolver function shown above. Creating a good sharding algorithm is an interesting challenge in itself. Try to produce a uniform distribution, i.e. same amount of entries in each shard. As a rule of thumb, the number of shards should be a factor ten greater than the planned maximum number of cluster nodes.

Messages to the entries are always sent via the local ShardRegion. The ShardRegion actor for a named entry type can be retrieved with ClusterSharding.shardRegion. The ShardRegion will lookup the location of the shard for the entry if it does not already know its location. It will delegate the message to the right node and it will create the entry actor on demand, i.e. when the first message for a specific entry is delivered.

ActorRef counterRegion = ClusterSharding.get(system).shardRegion("Counter");
counterRegion.tell(new Counter.Get(100), getSelf());

counterRegion.tell(new Counter.EntryEnvelope(100,
    Counter.CounterOp.INCREMENT), getSelf());
counterRegion.tell(new Counter.Get(100), getSelf());

An Example in Scala

This is how an entry actor may look like:

case object Increment
case object Decrement
case class Get(counterId: Long)
case class EntryEnvelope(id: Long, payload: Any)

case object Stop
case class CounterChanged(delta: Int)

class Counter extends PersistentActor {
  import ShardRegion.Passivate

  context.setReceiveTimeout(120.seconds)

  // self.path.parent.name is the type name (utf-8 URL-encoded) 
  // self.path.name is the entry identifier (utf-8 URL-encoded)
  override def persistenceId: String = self.path.parent.name + "-" + self.path.name

  var count = 0

  def updateState(event: CounterChanged): Unit =
    count += event.delta

  override def receiveRecover: Receive = {
    case evt: CounterChanged  updateState(evt)
  }

  override def receiveCommand: Receive = {
    case Increment       persist(CounterChanged(+1))(updateState)
    case Decrement       persist(CounterChanged(-1))(updateState)
    case Get(_)          sender() ! count
    case ReceiveTimeout  context.parent ! Passivate(stopMessage = Stop)
    case Stop            context.stop(self)
  }
}

The above actor uses event sourcing and the support provided in PersistentActor to store its state. It does not have to be a persistent actor, but in case of failure or migration of entries between nodes it must be able to recover its state if it is valuable.

Note how the persistenceId is defined. You may define it another way, but it must be unique.

When using the sharding extension you are first, typically at system startup on each node in the cluster, supposed to register the supported entry types with the ClusterSharding.start method. ClusterSharding.start gives you the reference which you can pass along.

val counterRegion: ActorRef = ClusterSharding(system).start(
  typeName = "Counter",
  entryProps = Some(Props[Counter]),
  idExtractor = idExtractor,
  shardResolver = shardResolver)

The idExtractor and shardResolver are two application specific functions to extract the entry identifier and the shard identifier from incoming messages.

val idExtractor: ShardRegion.IdExtractor = {
  case EntryEnvelope(id, payload)  (id.toString, payload)
  case msg @ Get(id)               (id.toString, msg)
}

val shardResolver: ShardRegion.ShardResolver = msg  msg match {
  case EntryEnvelope(id, _)  (id % 12).toString
  case Get(id)               (id % 12).toString
}

This example illustrates two different ways to define the entry identifier in the messages:

  • The Get message includes the identifier itself.
  • The EntryEnvelope holds the identifier, and the actual message that is sent to the entry actor is wrapped in the envelope.

Note how these two messages types are handled in the idExtractor function shown above.

A shard is a group of entries that will be managed together. The grouping is defined by the shardResolver function shown above. Creating a good sharding algorithm is an interesting challenge in itself. Try to produce a uniform distribution, i.e. same amount of entries in each shard. As a rule of thumb, the number of shards should be a factor ten greater than the planned maximum number of cluster nodes.

Messages to the entries are always sent via the local ShardRegion. The ShardRegion actor for a named entry type can be retrieved with ClusterSharding.shardRegion. The ShardRegion will lookup the location of the shard for the entry if it does not already know its location. It will delegate the message to the right node and it will create the entry actor on demand, i.e. when the first message for a specific entry is delivered.

val counterRegion: ActorRef = ClusterSharding(system).shardRegion("Counter")
counterRegion ! Get(100)
expectMsg(0)

counterRegion ! EntryEnvelope(100, Increment)
counterRegion ! Get(100)
expectMsg(1)

A more comprehensive sample is available in the Typesafe Activator tutorial named Akka Cluster Sharding with Scala!.

How it works

The ShardRegion actor is started on each node in the cluster, or group of nodes tagged with a specific role. The ShardRegion is created with two application specific functions to extract the entry identifier and the shard identifier from incoming messages. A shard is a group of entries that will be managed together. For the first message in a specific shard the ShardRegion request the location of the shard from a central coordinator, the ShardCoordinator.

The ShardCoordinator decides which ShardRegion that owns the shard. The ShardRegion receives the decided home of the shard and if that is the ShardRegion instance itself it will create a local child actor representing the entry and direct all messages for that entry to it. If the shard home is another ShardRegion instance messages will be forwarded to that ShardRegion instance instead. While resolving the location of a shard incoming messages for that shard are buffered and later delivered when the shard home is known. Subsequent messages to the resolved shard can be delivered to the target destination immediately without involving the ShardCoordinator.

Scenario 1:

  1. Incoming message M1 to ShardRegion instance R1.
  2. M1 is mapped to shard S1. R1 doesn't know about S1, so it asks the coordinator C for the location of S1.
  3. C answers that the home of S1 is R1.
  4. R1 creates child actor for the entry E1 and sends buffered messages for S1 to E1 child
  5. All incoming messages for S1 which arrive at R1 can be handled by R1 without C. It creates entry children as needed, and forwards messages to them.

Scenario 2:

  1. Incoming message M2 to R1.
  2. M2 is mapped to S2. R1 doesn't know about S2, so it asks C for the location of S2.
  3. C answers that the home of S2 is R2.
  4. R1 sends buffered messages for S2 to R2
  5. All incoming messages for S2 which arrive at R1 can be handled by R1 without C. It forwards messages to R2.
  6. R2 receives message for S2, ask C, which answers that the home of S2 is R2, and we are in Scenario 1 (but for R2).

To make sure that at most one instance of a specific entry actor is running somewhere in the cluster it is important that all nodes have the same view of where the shards are located. Therefore the shard allocation decisions are taken by the central ShardCoordinator, which is running as a cluster singleton, i.e. one instance on the oldest member among all cluster nodes or a group of nodes tagged with a specific role.

The logic that decides where a shard is to be located is defined in a pluggable shard allocation strategy. The default implementation ShardCoordinator.LeastShardAllocationStrategy allocates new shards to the ShardRegion with least number of previously allocated shards. This strategy can be replaced by an application specific implementation.

To be able to use newly added members in the cluster the coordinator facilitates rebalancing of shards, i.e. migrate entries from one node to another. In the rebalance process the coordinator first notifies all ShardRegion actors that a handoff for a shard has started. That means they will start buffering incoming messages for that shard, in the same way as if the shard location is unknown. During the rebalance process the coordinator will not answer any requests for the location of shards that are being rebalanced, i.e. local buffering will continue until the handoff is completed. The ShardRegion responsible for the rebalanced shard will stop all entries in that shard by sending PoisonPill to them. When all entries have been terminated the ShardRegion owning the entries will acknowledge the handoff as completed to the coordinator. Thereafter the coordinator will reply to requests for the location of the shard and thereby allocate a new home for the shard and then buffered messages in the ShardRegion actors are delivered to the new location. This means that the state of the entries are not transferred or migrated. If the state of the entries are of importance it should be persistent (durable), e.g. with akka-persistence, so that it can be recovered at the new location.

The logic that decides which shards to rebalance is defined in a pluggable shard allocation strategy. The default implementation ShardCoordinator.LeastShardAllocationStrategy picks shards for handoff from the ShardRegion with most number of previously allocated shards. They will then be allocated to the ShardRegion with least number of previously allocated shards, i.e. new members in the cluster. There is a configurable threshold of how large the difference must be to begin the rebalancing. This strategy can be replaced by an application specific implementation.

The state of shard locations in the ShardCoordinator is persistent (durable) with akka-persistence to survive failures. Since it is running in a cluster akka-persistence must be configured with a distributed journal. When a crashed or unreachable coordinator node has been removed (via down) from the cluster a new ShardCoordinator singleton actor will take over and the state is recovered. During such a failure period shards with known location are still available, while messages for new (unknown) shards are buffered until the new ShardCoordinator becomes available.

As long as a sender() uses the same ShardRegion actor to deliver messages to an entry actor the order of the messages is preserved. As long as the buffer limit is not reached messages are delivered on a best effort basis, with at-most once delivery semantics, in the same way as ordinary message sending. Reliable end-to-end messaging, with at-least-once semantics can be added by using channels in akka-persistence.

Some additional latency is introduced for messages targeted to new or previously unused shards due to the round-trip to the coordinator. Rebalancing of shards may also add latency. This should be considered when designing the application specific shard resolution, e.g. to avoid too fine grained shards.

Proxy Only Mode

The ShardRegion actor can also be started in proxy only mode, i.e. it will not host any entries itself, but knows how to delegate messages to the right location. A ShardRegion starts in proxy only mode if the roles of the node does not include the node role specified in akka.contrib.cluster.sharding.role config property or if the specified entryProps is None / null.

Passivation

If the state of the entries are persistent you may stop entries that are not used to reduce memory consumption. This is done by the application specific implementation of the entry actors for example by defining receive timeout (context.setReceiveTimeout). If a message is already enqueued to the entry when it stops itself the enqueued message in the mailbox will be dropped. To support graceful passivation without loosing such messages the entry actor can send ShardRegion.Passivate to its parent ShardRegion. The specified wrapped message in Passivate will be sent back to the entry, which is then supposed to stop itself. Incoming messages will be buffered by the ShardRegion between reception of Passivate and termination of the entry. Such buffered messages are thereafter delivered to a new incarnation of the entry.

Configuration

The ClusterSharding extension can be configured with the following properties:

# Settings for the ClusterShardingExtension
akka.contrib.cluster.sharding {
  # The extension creates a top level actor with this name in top level user scope,
  # e.g. '/user/sharding'
  guardian-name = sharding
  # If the coordinator can't store state changes it will be stopped
  # and started again after this duration.
  coordinator-failure-backoff = 10 s
  # Start the coordinator singleton manager on members tagged with this role.
  # All members are used if undefined or empty.
  # ShardRegion actor is started in proxy only mode on nodes that are not tagged
  # with this role.
  role = ""
  # The ShardRegion retries registration and shard location requests to the
  # ShardCoordinator with this interval if it does not reply.
  retry-interval = 2 s
  # Maximum number of messages that are buffered by a ShardRegion actor.
  buffer-size = 100000
  # Timeout of the shard rebalancing process.
  handoff-timeout = 60 s
  # Rebalance check is performed periodically with this interval.
  rebalance-interval = 10 s
  # How often the coordinator saves persistent snapshots, which are
  # used to reduce recovery times
  snapshot-interval = 3600 s
  # Setting for the default shard allocation strategy
  least-shard-allocation-strategy {
    # Threshold of how large the difference between most and least number of
    # allocated shards must be to begin the rebalancing.
    rebalance-threshold = 10
    # The number of ongoing rebalancing processes is limited to this number.
    max-simultaneous-rebalance = 3
  }
}

Custom shard allocation strategy can be defined in an optional parameter to ClusterSharding.start. See the API documentation of ShardAllocationStrategy (Scala) or AbstractShardAllocationStrategy (Java) for details of how to implement a custom shard allocation strategy.

Contents