Distributed Data

Distributed Data

Akka Distributed Data is useful when you need to share data between nodes in an Akka Cluster. The data is accessed with an actor providing a key-value store like API. The keys are unique identifiers with type information of the data values. The values are Conflict Free Replicated Data Types (CRDTs).

All data entries are spread to all nodes, or nodes with a certain role, in the cluster via direct replication and gossip based dissemination. You have fine grained control of the consistency level for reads and writes.

The nature CRDTs makes it possible to perform updates from any node without coordination. Concurrent updates from different nodes will automatically be resolved by the monotonic merge function, which all data types must provide. The state changes always converge. Several useful data types for counters, sets, maps and registers are provided and you can also implement your own custom data types.

It is eventually consistent and geared toward providing high read and write availability (partition tolerance), with low latency. Note that in an eventually consistent system a read may return an out-of-date value.

Warning

This module is marked as “experimental” as of its introduction in Akka 2.4.0. We will continue to improve this API based on our users’ feedback, which implies that while we try to keep incompatible changes to a minimum the binary compatibility guarantee for maintenance releases does not apply to the contents of the akka.persistence package.

Using the Replicator

The akka.cluster.ddata.Replicator actor provides the API for interacting with the data. The Replicator actor must be started on each node in the cluster, or group of nodes tagged with a specific role. It communicates with other Replicator instances with the same path (without address) that are running on other nodes . For convenience it can be used with the akka.cluster.ddata.DistributedData extension.

Cluster members with status WeaklyUp, if that feature is enabled, will currently not participate in Distributed Data, but that is something that should be possible to add in a future release.

Below is an example of an actor that schedules tick messages to itself and for each tick adds or removes elements from a ORSet (observed-remove set). It also subscribes to changes of this.

import scala.concurrent.forkjoin.ThreadLocalRandom
import akka.actor.Actor
import akka.actor.ActorLogging
import akka.cluster.Cluster
import akka.cluster.ddata.DistributedData
import akka.cluster.ddata.ORSet
import akka.cluster.ddata.ORSetKey
import akka.cluster.ddata.Replicator
import akka.cluster.ddata.Replicator._

object DataBot {
  private case object Tick
}

class DataBot extends Actor with ActorLogging {
  import DataBot._

  val replicator = DistributedData(context.system).replicator
  implicit val node = Cluster(context.system)

  import context.dispatcher
  val tickTask = context.system.scheduler.schedule(5.seconds, 5.seconds, self, Tick)

  val DataKey = ORSetKey[String]("key")

  replicator ! Subscribe(DataKey, self)

  def receive = {
    case Tick =>
      val s = ThreadLocalRandom.current().nextInt(97, 123).toChar.toString
      if (ThreadLocalRandom.current().nextBoolean()) {
        // add
        log.info("Adding: {}", s)
        replicator ! Update(DataKey, ORSet.empty[String], WriteLocal)(_ + s)
      } else {
        // remove
        log.info("Removing: {}", s)
        replicator ! Update(DataKey, ORSet.empty[String], WriteLocal)(_ - s)
      }

    case _: UpdateResponse[_] => // ignore

    case c @ Changed(DataKey) =>
      val data = c.get(DataKey)
      log.info("Current elements: {}", data.elements)
  }

  override def postStop(): Unit = tickTask.cancel()

}

Update

To modify and replicate a data value you send a Replicator.Update message to the local Replicator.

The current data value for the key of the Update is passed as parameter to the modify function of the Update. The function is supposed to return the new value of the data, which will then be replicated according to the given consistency level.

The modify function is called by the Replicator actor and must therefore be a pure function that only uses the data parameter and stable fields from enclosing scope. It must for example not access sender() reference of an enclosing actor.

Update is intended to only be sent from an actor running in same local ActorSystem as
  • the Replicator, because the modify function is typically not serializable.

You supply a write consistency level which has the following meaning:

  • WriteLocal the value will immediately only be written to the local replica, and later disseminated with gossip
  • WriteTo(n) the value will immediately be written to at least n replicas, including the local replica
  • WriteMajority the value will immediately be written to a majority of replicas, i.e. at least N/2 + 1 replicas, where N is the number of nodes in the cluster (or cluster role group)
  • WriteAll the value will immediately be written to all nodes in the cluster (or all nodes in the cluster role group)
implicit val node = Cluster(system)
val replicator = DistributedData(system).replicator

val Counter1Key = PNCounterKey("counter1")
val Set1Key = GSetKey[String]("set1")
val Set2Key = ORSetKey[String]("set2")
val ActiveFlagKey = FlagKey("active")

replicator ! Update(Counter1Key, PNCounter(), WriteLocal)(_ + 1)

val writeTo3 = WriteTo(n = 3, timeout = 1.second)
replicator ! Update(Set1Key, GSet.empty[String], writeTo3)(_ + "hello")

val writeMajority = WriteMajority(timeout = 5.seconds)
replicator ! Update(Set2Key, ORSet.empty[String], writeMajority)(_ + "hello")

val writeAll = WriteAll(timeout = 5.seconds)
replicator ! Update(ActiveFlagKey, Flag.empty, writeAll)(_.switchOn)

As reply of the Update a Replicator.UpdateSuccess is sent to the sender of the Update if the value was successfully replicated according to the supplied consistency level within the supplied timeout. Otherwise a Replicator.UpdateFailure subclass is sent back. Note that a Replicator.UpdateTimeout reply does not mean that the update completely failed or was rolled back. It may still have been replicated to some nodes, and will eventually be replicated to all nodes with the gossip protocol.

case UpdateSuccess(Counter1Key, req) => // ok
case UpdateSuccess(Set1Key, req)  => // ok
case UpdateTimeout(Set1Key, req)  =>
// write to 3 nodes failed within 1.second

You will always see your own writes. For example if you send two Update messages changing the value of the same key, the modify function of the second message will see the change that was performed by the first Update message.

In the Update message you can pass an optional request context, which the Replicator does not care about, but is included in the reply messages. This is a convenient way to pass contextual information (e.g. original sender) without having to use ask or maintain local correlation data structures.

implicit val node = Cluster(system)
val replicator = DistributedData(system).replicator
val writeTwo = WriteTo(n = 2, timeout = 3.second)
val Counter1Key = PNCounterKey("counter1")

def receive: Receive = {
  case "increment" =>
    // incoming command to increase the counter
    val upd = Update(Counter1Key, PNCounter(), writeTwo, request = Some(sender()))(_ + 1)
    replicator ! upd

  case UpdateSuccess(Counter1Key, Some(replyTo: ActorRef)) =>
    replyTo ! "ack"
  case UpdateTimeout(Counter1Key, Some(replyTo: ActorRef)) =>
    replyTo ! "nack"
}

Get

To retrieve the current value of a data you send Replicator.Get message to the Replicator. You supply a consistency level which has the following meaning:

  • ReadLocal the value will only be read from the local replica
  • ReadFrom(n) the value will be read and merged from n replicas, including the local replica
  • ReadMajority the value will be read and merged from a majority of replicas, i.e. at least N/2 + 1 replicas, where N is the number of nodes in the cluster (or cluster role group)
  • ReadAll the value will be read and merged from all nodes in the cluster (or all nodes in the cluster role group)
val replicator = DistributedData(system).replicator
val Counter1Key = PNCounterKey("counter1")
val Set1Key = GSetKey[String]("set1")
val Set2Key = ORSetKey[String]("set2")
val ActiveFlagKey = FlagKey("active")

replicator ! Get(Counter1Key, ReadLocal)

val readFrom3 = ReadFrom(n = 3, timeout = 1.second)
replicator ! Get(Set1Key, readFrom3)

val readMajority = ReadMajority(timeout = 5.seconds)
replicator ! Get(Set2Key, readMajority)

val readAll = ReadAll(timeout = 5.seconds)
replicator ! Get(ActiveFlagKey, readAll)

As reply of the Get a Replicator.GetSuccess is sent to the sender of the Get if the value was successfully retrieved according to the supplied consistency level within the supplied timeout. Otherwise a Replicator.GetFailure is sent. If the key does not exist the reply will be Replicator.NotFound.

case g @ GetSuccess(Counter1Key, req) =>
  val value = g.get(Counter1Key).value
case NotFound(Counter1Key, req) => // key counter1 does not exist
case g @ GetSuccess(Set1Key, req) =>
  val elements = g.get(Set1Key).elements
case GetFailure(Set1Key, req) =>
// read from 3 nodes failed within 1.second
case NotFound(Set1Key, req)   => // key set1 does not exist

You will always read your own writes. For example if you send a Update message followed by a Get of the same key the Get will retrieve the change that was performed by the preceding Update message. However, the order of the reply messages are not defined, i.e. in the previous example you may receive the GetSuccess before the UpdateSuccess.

In the Get message you can pass an optional request context in the same way as for the Update message, described above. For example the original sender can be passed and replied to after receiving and transforming GetSuccess.

implicit val node = Cluster(system)
val replicator = DistributedData(system).replicator
val readTwo = ReadFrom(n = 2, timeout = 3.second)
val Counter1Key = PNCounterKey("counter1")

def receive: Receive = {
  case "get-count" =>
    // incoming request to retrieve current value of the counter
    replicator ! Get(Counter1Key, readTwo, request = Some(sender()))

  case g @ GetSuccess(Counter1Key, Some(replyTo: ActorRef)) =>
    val value = g.get(Counter1Key).value.longValue
    replyTo ! value
  case GetFailure(Counter1Key, Some(replyTo: ActorRef)) =>
    replyTo ! -1L
  case NotFound(Counter1Key, Some(replyTo: ActorRef)) =>
    replyTo ! 0L
}

Consistency

The consistency level that is supplied in the Update and Get specifies per request how many replicas that must respond successfully to a write and read request.

For low latency reads you use ReadLocal with the risk of retrieving stale data, i.e. updates from other nodes might not be visible yet.

When using WriteLocal the update is only written to the local replica and then disseminated in the background with the gossip protocol, which can take few seconds to spread to all nodes.

WriteAll and ReadAll is the strongest consistency level, but also the slowest and with lowest availability. For example, it is enough that one node is unavailable for a Get request and you will not receive the value.

If consistency is important, you can ensure that a read always reflects the most recent write by using the following formula:

(nodes_written + nodes_read) > N

where N is the total number of nodes in the cluster, or the number of nodes with the role that is used for the Replicator.

For example, in a 7 node cluster this these consistency properties are achieved by writing to 4 nodes and reading from 4 nodes, or writing to 5 nodes and reading from 3 nodes.

By combining WriteMajority and ReadMajority levels a read always reflects the most recent write. The Replicator writes and reads to a majority of replicas, i.e. N / 2 + 1. For example, in a 5 node cluster it writes to 3 nodes and reads from 3 nodes. In a 6 node cluster it writes to 4 nodes and reads from 4 nodes.

Here is an example of using WriteMajority and ReadMajority:

private val timeout = 3.seconds
private val readMajority = ReadMajority(timeout)
private val writeMajority = WriteMajority(timeout)
def receiveGetCart: Receive = {
  case GetCart 
    replicator ! Get(DataKey, readMajority, Some(sender()))

  case g @ GetSuccess(DataKey, Some(replyTo: ActorRef)) 
    val data = g.get(DataKey)
    val cart = Cart(data.entries.values.toSet)
    replyTo ! cart

  case NotFound(DataKey, Some(replyTo: ActorRef)) 
    replyTo ! Cart(Set.empty)

  case GetFailure(DataKey, Some(replyTo: ActorRef)) 
    // ReadMajority failure, try again with local read
    replicator ! Get(DataKey, ReadLocal, Some(replyTo))
}
def receiveAddItem: Receive = {
  case cmd @ AddItem(item) 
    val update = Update(DataKey, LWWMap.empty[LineItem], writeMajority, Some(cmd)) {
      cart  updateCart(cart, item)
    }
    replicator ! update
}

In some rare cases, when performing an Update it is needed to first try to fetch latest data from other nodes. That can be done by first sending a Get with ReadMajority and then continue with the Update when the GetSuccess, GetFailure or NotFound reply is received. This might be needed when you need to base a decision on latest information or when removing entries from ORSet or ORMap. If an entry is added to an ORSet or ORMap from one node and removed from another node the entry will only be removed if the added entry is visible on the node where the removal is performed (hence the name observed-removed set).

The following example illustrates how to do that:

def receiveRemoveItem: Receive = {
  case cmd @ RemoveItem(productId) 
    // Try to fetch latest from a majority of nodes first, since ORMap
    // remove must have seen the item to be able to remove it.
    replicator ! Get(DataKey, readMajority, Some(cmd))

  case GetSuccess(DataKey, Some(RemoveItem(productId))) 
    replicator ! Update(DataKey, LWWMap(), writeMajority, None) {
      _ - productId
    }

  case GetFailure(DataKey, Some(RemoveItem(productId))) 
    // ReadMajority failed, fall back to best effort local value
    replicator ! Update(DataKey, LWWMap(), writeMajority, None) {
      _ - productId
    }

  case NotFound(DataKey, Some(RemoveItem(productId))) 
  // nothing to remove
}

Warning

Caveat: Even if you use WriteMajority and ReadMajority there is small risk that you may read stale data if the cluster membership has changed between the Update and the Get. For example, in cluster of 5 nodes when you Update and that change is written to 3 nodes: n1, n2, n3. Then 2 more nodes are added and a Get request is reading from 4 nodes, which happens to be n4, n5, n6, n7, i.e. the value on n1, n2, n3 is not seen in the response of the Get request.

Subscribe

You may also register interest in change notifications by sending Replicator.Subscribe message to the Replicator. It will send Replicator.Changed messages to the registered subscriber when the data for the subscribed key is updated. Subscribers will be notified periodically with the configured notify-subscribers-interval, and it is also possible to send an explicit Replicator.FlushChanges message to the Replicator to notify the subscribers immediately.

The subscriber is automatically removed if the subscriber is terminated. A subscriber can also be deregistered with the Replicator.Unsubscribe message.

val replicator = DistributedData(system).replicator
val Counter1Key = PNCounterKey("counter1")
// subscribe to changes of the Counter1Key value
replicator ! Subscribe(Counter1Key, self)
var currentValue = BigInt(0)

def receive: Receive = {
  case c @ Changed(Counter1Key) =>
    currentValue = c.get(Counter1Key).value
  case "get-count" =>
    // incoming request to retrieve current value of the counter
    sender() ! currentValue
}

Delete

A data entry can be deleted by sending a Replicator.Delete message to the local local Replicator. As reply of the Delete a Replicator.DeleteSuccess is sent to the sender of the Delete if the value was successfully deleted according to the supplied consistency level within the supplied timeout. Otherwise a Replicator.ReplicationDeleteFailure is sent. Note that ReplicationDeleteFailure does not mean that the delete completely failed or was rolled back. It may still have been replicated to some nodes, and may eventually be replicated to all nodes.

A deleted key cannot be reused again, but it is still recommended to delete unused data entries because that reduces the replication overhead when new nodes join the cluster. Subsequent Delete, Update and Get requests will be replied with Replicator.DataDeleted. Subscribers will receive Replicator.DataDeleted.

val replicator = DistributedData(system).replicator
val Counter1Key = PNCounterKey("counter1")
val Set2Key = ORSetKey[String]("set2")

replicator ! Delete(Counter1Key, WriteLocal)

val writeMajority = WriteMajority(timeout = 5.seconds)
replicator ! Delete(Set2Key, writeMajority)

Data Types

The data types must be convergent (stateful) CRDTs and implement the ReplicatedData trait, i.e. they provide a monotonic merge function and the state changes always converge.

You can use your own custom ReplicatedData types, and several types are provided by this package, such as:

  • Counters: GCounter, PNCounter
  • Sets: GSet, ORSet
  • Maps: ORMap, ORMultiMap, LWWMap, PNCounterMap
  • Registers: LWWRegister, Flag

Counters

GCounter is a "grow only counter". It only supports increments, no decrements.

It works in a similar way as a vector clock. It keeps track of one counter per node and the total value is the sum of these counters. The merge is implemented by taking the maximum count for each node.

If you need both increments and decrements you can use the PNCounter (positive/negative counter).

It is tracking the increments (P) separate from the decrements (N). Both P and N are represented as two internal GCounter. Merge is handled by merging the internal P and N counters. The value of the counter is the value of the P counter minus the value of the N counter.

implicit val node = Cluster(system)
val c0 = PNCounter.empty
val c1 = c0 + 1
val c2 = c1 + 7
val c3: PNCounter = c2 - 2
println(c3.value) // 6

Several related counters can be managed in a map with the PNCounterMap data type. When the counters are placed in a PNCounterMap as opposed to placing them as separate top level values they are guaranteed to be replicated together as one unit, which is sometimes necessary for related data.

implicit val node = Cluster(system)
val m0 = PNCounterMap.empty
val m1 = m0.increment("a", 7)
val m2 = m1.decrement("a", 2)
val m3 = m2.increment("b", 1)
println(m3.get("a")) // 5
m3.entries.foreach { case (key, value) => println(s"$key -> $value") }

Sets

If you only need to add elements to a set and not remove elements the GSet (grow-only set) is the data type to use. The elements can be any type of values that can be serialized. Merge is simply the union of the two sets.

val s0 = GSet.empty[String]
val s1 = s0 + "a"
val s2 = s1 + "b" + "c"
if (s2.contains("a"))
  println(s2.elements) // a, b, c

If you need add and remove operations you should use the ORSet (observed-remove set). Elements can be added and removed any number of times. If an element is concurrently added and removed, the add will win. You cannot remove an element that you have not seen.

The ORSet has a version vector that is incremented when an element is added to the set. The version for the node that added the element is also tracked for each element in a so called "birth dot". The version vector and the dots are used by the merge function to track causality of the operations and resolve concurrent updates.

implicit val node = Cluster(system)
val s0 = ORSet.empty[String]
val s1 = s0 + "a"
val s2 = s1 + "b"
val s3 = s2 - "a"
println(s3.elements) // b

Maps

ORMap (observed-remove map) is a map with String keys and the values are ReplicatedData types themselves. It supports add, remove and delete any number of times for a map entry.

If an entry is concurrently added and removed, the add will win. You cannot remove an entry that you have not seen. This is the same semantics as for the ORSet.

If an entry is concurrently updated to different values the values will be merged, hence the requirement that the values must be ReplicatedData types.

It is rather inconvenient to use the ORMap directly since it does not expose specific types of the values. The ORMap is intended as a low level tool for building more specific maps, such as the following specialized maps.

ORMultiMap (observed-remove multi-map) is a multi-map implementation that wraps an ORMap with an ORSet for the map's value.

PNCounterMap (positive negative counter map) is a map of named counters. It is a specialized ORMap with PNCounter values.

LWWMap (last writer wins map) is a specialized ORMap with LWWRegister (last writer wins register) values.

implicit val node = Cluster(system)
val m0 = ORMultiMap.empty[Int]
val m1 = m0 + ("a" -> Set(1, 2, 3))
val m2 = m1.addBinding("a", 4)
val m3 = m2.removeBinding("a", 2)
val m4 = m3.addBinding("b", 1)
println(m4.entries)

Note that LWWRegister and therefore LWWMap relies on synchronized clocks and should only be used when the choice of value is not important for concurrent updates occurring within the clock skew.

Instead of using timestamps based on System.currentTimeMillis() time it is possible to use a timestamp value based on something else, for example an increasing version number from a database record that is used for optimistic concurrency control.

When a data entry is changed the full state of that entry is replicated to other nodes, i.e. when you update a map the whole map is replicated. Therefore, instead of using one ORMap with 1000 elements it is more efficient to split that up in 10 top level ORMap entries with 100 elements each. Top level entries are replicated individually, which has the trade-off that different entries may not be replicated at the same time and you may see inconsistencies between related entries. Separate top level entries cannot be updated atomically together.

Flags and Registers

Flag is a data type for a boolean value that is initialized to false and can be switched to true. Thereafter it cannot be changed. true wins over false in merge.

val f0 = Flag.empty
val f1 = f0.switchOn
println(f1.enabled)

LWWRegister (last writer wins register) can hold any (serializable) value.

Merge of a LWWRegister takes the register with highest timestamp. Note that this relies on synchronized clocks. LWWRegister should only be used when the choice of value is not important for concurrent updates occurring within the clock skew.

Merge takes the register updated by the node with lowest address (UniqueAddress is ordered) if the timestamps are exactly the same.

implicit val node = Cluster(system)
val r1 = LWWRegister("Hello")
val r2 = r1.withValue("Hi")
println(s"${r1.value} by ${r1.updatedBy} at ${r1.timestamp}")

Instead of using timestamps based on System.currentTimeMillis() time it is possible to use a timestamp value based on something else, for example an increasing version number from a database record that is used for optimistic concurrency control.

case class Record(version: Int, name: String, address: String)

implicit val node = Cluster(system)
implicit val recordClock = new LWWRegister.Clock[Record] {
  override def apply(currentTimestamp: Long, value: Record): Long =
    value.version
}

val record1 = Record(version = 1, "Alice", "Union Square")
val r1 = LWWRegister(record1)

val record2 = Record(version = 2, "Alice", "Madison Square")
val r2 = LWWRegister(record2)

val r3 = r1.merge(r2)
println(r3.value)

For first-write-wins semantics you can use the LWWRegister#reverseClock instead of the LWWRegister#defaultClock.

Custom Data Type

You can rather easily implement your own data types. The only requirement is that it implements the merge function of the ReplicatedData trait.

A nice property of stateful CRDTs is that they typically compose nicely, i.e. you can combine several smaller data types to build richer data structures. For example, the PNCounter is composed of two internal GCounter instances to keep track of increments and decrements separately.

Here is s simple implementation of a custom TwoPhaseSet that is using two internal GSet types to keep track of addition and removals. A TwoPhaseSet is a set where an element may be added and removed, but never added again thereafter.

case class TwoPhaseSet(
  adds: GSet[String] = GSet.empty,
  removals: GSet[String] = GSet.empty)
  extends ReplicatedData {
  type T = TwoPhaseSet

  def add(element: String): TwoPhaseSet =
    copy(adds = adds.add(element))

  def remove(element: String): TwoPhaseSet =
    copy(removals = removals.add(element))

  def elements: Set[String] = adds.elements -- removals.elements

  override def merge(that: TwoPhaseSet): TwoPhaseSet =
    copy(
      adds = this.adds.merge(that.adds),
      removals = this.removals.merge(that.removals))
}

Data types should be immutable, i.e. "modifying" methods should return a new instance.

Serialization

The data types must be serializable with an Akka Serializer. It is highly recommended that you implement efficient serialization with Protobuf or similar for your custom data types. The built in data types are marked with ReplicatedDataSerialization and serialized with akka.cluster.ddata.protobuf.ReplicatedDataSerializer.

Serialization of the data types are used in remote messages and also for creating message digests (SHA-1) to detect changes. Therefore it is important that the serialization is efficient and produce the same bytes for the same content. For example sets and maps should be sorted deterministically in the serialization.

This is a protobuf representation of the above TwoPhaseSet:

option java_package = "docs.ddata.protobuf.msg";
option optimize_for = SPEED;

message TwoPhaseSet {
  repeated string adds = 1;
  repeated string removals = 2;
}

The serializer for the TwoPhaseSet:

import java.util.ArrayList
import java.util.Collections
import scala.collection.JavaConverters._
import akka.actor.ExtendedActorSystem
import akka.cluster.ddata.GSet
import akka.cluster.ddata.protobuf.SerializationSupport
import akka.serialization.Serializer
import docs.ddata.TwoPhaseSet
import docs.ddata.protobuf.msg.TwoPhaseSetMessages

class TwoPhaseSetSerializer(val system: ExtendedActorSystem)
  extends Serializer with SerializationSupport {

  override def includeManifest: Boolean = false

  override def identifier = 99999

  override def toBinary(obj: AnyRef): Array[Byte] = obj match {
    case m: TwoPhaseSet  twoPhaseSetToProto(m).toByteArray
    case _  throw new IllegalArgumentException(
      s"Can't serialize object of type ${obj.getClass}")
  }

  override def fromBinary(bytes: Array[Byte], clazz: Option[Class[_]]): AnyRef = {
    twoPhaseSetFromBinary(bytes)
  }

  def twoPhaseSetToProto(twoPhaseSet: TwoPhaseSet): TwoPhaseSetMessages.TwoPhaseSet = {
    val b = TwoPhaseSetMessages.TwoPhaseSet.newBuilder()
    // using java collections and sorting for performance (avoid conversions)
    val adds = new ArrayList[String]
    twoPhaseSet.adds.elements.foreach(adds.add)
    if (!adds.isEmpty) {
      Collections.sort(adds)
      b.addAllAdds(adds)
    }
    val removals = new ArrayList[String]
    twoPhaseSet.removals.elements.foreach(removals.add)
    if (!removals.isEmpty) {
      Collections.sort(removals)
      b.addAllRemovals(removals)
    }
    b.build()
  }

  def twoPhaseSetFromBinary(bytes: Array[Byte]): TwoPhaseSet = {
    val msg = TwoPhaseSetMessages.TwoPhaseSet.parseFrom(bytes)
    TwoPhaseSet(
      adds = GSet(msg.getAddsList.iterator.asScala.toSet),
      removals = GSet(msg.getRemovalsList.iterator.asScala.toSet))
  }
}

Note that the elements of the sets are sorted so the SHA-1 digests are the same for the same elements.

You register the serializer in configuration:

akka.actor {
  serializers {
    two-phase-set = "docs.ddata.protobuf.TwoPhaseSetSerializer"
  }
  serialization-bindings {
    "docs.ddata.TwoPhaseSet" = two-phase-set
  }
}

Using compression can sometimes be a good idea to reduce the data size. Gzip compression is provided by the akka.cluster.ddata.protobuf.SerializationSupport trait:

override def toBinary(obj: AnyRef): Array[Byte] = obj match {
  case m: TwoPhaseSet ⇒ compress(twoPhaseSetToProto(m))
  case _ ⇒ throw new IllegalArgumentException(
    s"Can't serialize object of type ${obj.getClass}")
}

override def fromBinary(bytes: Array[Byte], clazz: Option[Class[_]]): AnyRef = {
  twoPhaseSetFromBinary(decompress(bytes))
}

The two embedded GSet can be serialized as illustrated above, but in general when composing new data types from the existing built in types it is better to make use of the existing serializer for those types. This can be done by declaring those as bytes fields in protobuf:

message TwoPhaseSet2 {
  optional bytes adds = 1;
  optional bytes removals = 2;
}

and use the methods otherMessageToProto and otherMessageFromBinary that are provided by the SerializationSupport trait to serialize and deserialize the GSet instances. This works with any type that has a registered Akka serializer. This is how such an serializer would look like for the TwoPhaseSet:

import scala.collection.JavaConverters._
import akka.actor.ExtendedActorSystem
import akka.cluster.ddata.GSet
import akka.cluster.ddata.protobuf.ReplicatedDataSerializer
import akka.cluster.ddata.protobuf.SerializationSupport
import akka.serialization.Serializer
import docs.ddata.TwoPhaseSet
import docs.ddata.protobuf.msg.TwoPhaseSetMessages

class TwoPhaseSetSerializer2(val system: ExtendedActorSystem)
  extends Serializer with SerializationSupport {

  override def includeManifest: Boolean = false

  override def identifier = 99999

  val replicatedDataSerializer = new ReplicatedDataSerializer(system)

  override def toBinary(obj: AnyRef): Array[Byte] = obj match {
    case m: TwoPhaseSet  twoPhaseSetToProto(m).toByteArray
    case _  throw new IllegalArgumentException(
      s"Can't serialize object of type ${obj.getClass}")
  }

  override def fromBinary(bytes: Array[Byte], clazz: Option[Class[_]]): AnyRef = {
    twoPhaseSetFromBinary(bytes)
  }

  def twoPhaseSetToProto(twoPhaseSet: TwoPhaseSet): TwoPhaseSetMessages.TwoPhaseSet2 = {
    val b = TwoPhaseSetMessages.TwoPhaseSet2.newBuilder()
    if (!twoPhaseSet.adds.isEmpty)
      b.setAdds(otherMessageToProto(twoPhaseSet.adds).toByteString())
    if (!twoPhaseSet.removals.isEmpty)
      b.setRemovals(otherMessageToProto(twoPhaseSet.removals).toByteString())
    b.build()
  }

  def twoPhaseSetFromBinary(bytes: Array[Byte]): TwoPhaseSet = {
    val msg = TwoPhaseSetMessages.TwoPhaseSet2.parseFrom(bytes)
    val adds =
      if (msg.hasAdds)
        otherMessageFromBinary(msg.getAdds.toByteArray).asInstanceOf[GSet[String]]
      else
        GSet.empty[String]
    val removals =
      if (msg.hasRemovals)
        otherMessageFromBinary(msg.getRemovals.toByteArray).asInstanceOf[GSet[String]]
      else
        GSet.empty[String]
    TwoPhaseSet(adds, removals)
  }
}

CRDT Garbage

One thing that can be problematic with CRDTs is that some data types accumulate history (garbage). For example a GCounter keeps track of one counter per node. If a GCounter has been updated from one node it will associate the identifier of that node forever. That can become a problem for long running systems with many cluster nodes being added and removed. To solve this problem the Replicator performs pruning of data associated with nodes that have been removed from the cluster. Data types that need pruning have to implement the RemovedNodePruning trait.

Samples

Several interesting samples are included and described in the Typesafe Activator tutorial named Akka Distributed Data Samples with Scala.

  • Low Latency Voting Service
  • Highly Available Shopping Cart
  • Distributed Service Registry
  • Replicated Cache
  • Replicated Metrics

Limitations

There are some limitations that you should be aware of.

CRDTs cannot be used for all types of problems, and eventual consistency does not fit all domains. Sometimes you need strong consistency.

It is not intended for Big Data. The number of top level entries should not exceed 100000. When a new node is added to the cluster all these entries are transferred (gossiped) to the new node. The entries are split up in chunks and all existing nodes collaborate in the gossip, but it will take a while (tens of seconds) to transfer all entries and this means that you cannot have too many top level entries. The current recommended limit is 100000. We will be able to improve this if needed, but the design is still not intended for billions of entries.

All data is held in memory, which is another reason why it is not intended for Big Data.

When a data entry is changed the full state of that entry is replicated to other nodes. For example, if you add one element to a Set with 100 existing elements, all 101 elements are transferred to other nodes. This means that you cannot have too large data entries, because then the remote message size will be too large. We might be able to make this more efficient by implementing Efficient State-based CRDTs by Delta-Mutation.

The data is only kept in memory. It is redundant since it is replicated to other nodes in the cluster, but if you stop all nodes the data is lost, unless you have saved it elsewhere. Making the data durable is a possible future feature, but even if we implement that it is not intended to be a full featured database.

Learn More about CRDTs

Dependencies

To use Distributed Data you must add the following dependency in your project.

sbt:

"com.typesafe.akka" %% "akka-distributed-data-experimental" % "2.4.0"

maven:

<dependency>
  <groupId>com.typesafe.akka</groupId>
  <artifactId>akka-distributed-data-experimental_2.11</artifactId>
  <version>2.4.0</version>
</dependency>

Configuration

The DistributedData extension can be configured with the following properties:

# Settings for the DistributedData extension
akka.cluster.distributed-data {
  # Actor name of the Replicator actor, /system/ddataReplicator
  name = ddataReplicator

  # Replicas are running on members tagged with this role.
  # All members are used if undefined or empty.
  role = ""

  # How often the Replicator should send out gossip information
  gossip-interval = 2 s

  # How often the subscribers will be notified of changes, if any
  notify-subscribers-interval = 500 ms

  # Maximum number of entries to transfer in one gossip message when synchronizing
  # the replicas. Next chunk will be transferred in next round of gossip.
  max-delta-elements = 1000
  
  # The id of the dispatcher to use for Replicator actors. If not specified
  # default dispatcher is used.
  # If specified you need to define the settings of the actual dispatcher.
  use-dispatcher = ""

  # How often the Replicator checks for pruning of data associated with
  # removed cluster nodes.
  pruning-interval = 30 s
  
  # How long time it takes (worst case) to spread the data to all other replica nodes.
  # This is used when initiating and completing the pruning process of data associated
  # with removed cluster nodes. The time measurement is stopped when any replica is 
  # unreachable, so it should be configured to worst case in a healthy cluster.
  max-pruning-dissemination = 60 s
  
}

Contents