Awesome

Kafka Plugins for Akka Persistence

Replicated Akka Persistence journal and snapshot store backed by Apache Kafka.

Dependency

To include the Kafka plugins into your sbt project, add the following lines to your build.sbt file:

resolvers += "krasserm at bintray" at "http://dl.bintray.com/krasserm/maven"

libraryDependencies += "com.github.krasserm" %% "akka-persistence-kafka" % “0.4”

This version of akka-persistence-kafka depends on Kafka 0.8.2.1, Akka 2.3.11 and is cross-built against Scala 2.10.4 and 2.11.6. A complete list of released versions is here.

Usage hints

Kafka does not permanently store log entries but rather deletes them after a configurable retention time which defaults to 7 days in Kafka 0.8.x. Therefore, applications need to take snapshots of their persistent actors at intervals that are smaller than the configured retention time (for example, every 3 days). This ensures that persistent actors can always be recovered successfully.

Alternatively, the retention time can be set to a maximum value so that Kafka will never delete old entries. In this case, all events written by a single persistent actor must fit on a single node. This is a limitation of the current implementation which may be removed in later versions. However, this limitation is likely not relevant when running Kafka with default (or comparable) retention times and taking snapshots.

The latest snapshot of a persistent actor is never deleted if log compaction is enabled. See also section Configuration hints for details how to properly configure Kafka for being used with the storage plugins.

Journal plugin

Activation

To activate the journal plugin, add the following line to application.conf:

akka.persistence.journal.plugin = "kafka-journal"

This will run the journal plugin with default settings and connect to a Zookeeper instance running on localhost:2181. The Zookeeper connect string can be customized with the kafka-journal.zookeeper.connect configuration key (see also section Kafka cluster). Recommended Kafka broker configurations are given in section Configuration hints.

Use cases

• Akka Persistence journal plugin (obvious).
• Event publishing to user-defined topics.
• Event consumption from user-defined topics by external consumers.

Journal topics

For each persistent actor, the plugin creates a Kafka topic where the topic name equals the actor's persistenceId (only if it contains alphanumeric, ., - or _ characters, otherwise, all other characters are replaced by _). Events published to these topics are serialized akka.persistence.PersistentRepr objects (see journal plugin API). Serialization of PersistentRepr objects can be customized. Journal topics are mainly intended for internal use (for recovery of persistent actors) but can also be consumed externally.

User-defined topics

The journal plugin can also publish events to user-defined topics. By default, all events generated by all persistent actors are published to a single events topic. This topic is intended for external consumption only. Events published to user-defined topics are serialized Event objects

package akka.persistence.kafka

/**
 * Event published to user-defined topics.
 *
 * @param persistenceId Id of the persistent actor that generates event `data`.
 * @param sequenceNr Sequence number of the event.
 * @param data Event data generated by a persistent actor.
 */
case class Event(persistenceId: String, sequenceNr: Long, data: Any)

where data is the actual event written by a persistent actor (by calling persist or persistAsync), sequenceNr is the event's sequence number and persistenceId the id of the persistent actor. Event objects are serialized with a protobuf serializer and event data serialization can be customized with a user-defined serializer in the same way as for journal topics. Custom serializer configurations always apply to both, journal topics and user-defined topics.

For publishing events to user-defined topics the journal plugin uses an EventTopicMapper:

package akka.persistence.kafka

/**
 * Defines a mapping of events to user-defined topics.
 */
trait EventTopicMapper {
  /**
   * Maps an event to zero or more topics.
   *
   * @param event event to be mapped.
   * @return a sequence of topic names.
   */
  def topicsFor(event: Event): immutable.Seq[String]
}

The default mapper is DefaultEventTopicMapper which maps all events to the events topic. It is configured in the reference configuration as follows:

kafka-journal.event.producer.topic.mapper.class = "akka.persistence.kafka.DefaultEventTopicMapper"

To customize the mapping of events to user-defined topics, applications can implement and configure a custom EventTopicMapper. For example, in order to publish

• events from persistent actor a to topics topic-a-1 and topic-a-2 and
• events from persistent actor b to topic topic-b

and to turn of publishing of events from all other actors, one would implement the following ExampleEventTopicMapper

package akka.persistence.kafka.example

class ExampleEventTopicMapper extends EventTopicMapper {
  def topicsFor(event: Event): Seq[String] = event.persistenceId match {
    case "a" => List("topic-a-1", "topic-a-2")
    case "b" => List("topic-b")
    case _   => Nil
  }

and configure it in application.conf:

kafka-journal.event.producer.topic.mapper.class = "akka.persistence.kafka.example.ExampleEventTopicMapper"

To turn off publishing events to user-defined topics, the EmptyEventTopicMapper should be configured.

kafka-journal.event.producer.topic.mapper.class = "akka.persistence.kafka.EmptyEventTopicMapper"

External consumers

The following example shows how to consume Events from a user-defined topic with name topic-a-2 (see previous example) using Kafka's high-level consumer API:

import java.util.Properties

import akka.persistence.kafka.{EventDecoder, Event}

import kafka.consumer.{Consumer, ConsumerConfig}
import kafka.serializer.StringDecoder

val props = new Properties()
props.put("group.id", "consumer-1")
props.put("zookeeper.connect", "localhost:2181")
// ...

val system = ActorSystem("consumer")

val consConn = Consumer.create(new ConsumerConfig(props))
val streams = consConn.createMessageStreams(Map("topic-a-2" -> 1),
  keyDecoder = new StringDecoder, valueDecoder = new EventDecoder(system))

streams("topic-a-2")(0).foreach { mm =>
  val event: Event = mm.message
  println(s"consumed ${event}")
}

Applications may also consume serialized PersistentRepr objects from journal topics and deserialize them with Akka's serialization extension:

import java.util.Properties

import akka.actor._
import akka.persistence.PersistentRepr
import akka.serialization.SerializationExtension

import com.typesafe.config.ConfigFactory

import kafka.consumer.{Consumer, ConsumerConfig}
import kafka.serializer.{DefaultDecoder, StringDecoder}

val props = new Properties()
props.put("group.id", "consumer-2")
props.put("zookeeper.connect", "localhost:2181")
// ...

val system = ActorSystem("example")
val extension = SerializationExtension(system)

val consConn = Consumer.create(new ConsumerConfig(props))
val streams = consConn.createMessageStreams(Map("a" -> 1),
  keyDecoder = new StringDecoder, valueDecoder = new DefaultDecoder)

streams("a")(0).foreach { mm =>
  val persistent: PersistentRepr = extension.deserialize(mm.message, classOf[PersistentRepr]).get
  println(s"consumed ${persistent}")
}

There are many other libraries that can be used to consume (event) streams from Kafka topics, such as Spark Streaming, to mention only one example.

Implementation notes

• During initialization, the journal plugin fetches cluster metadata from Zookeeper which may take up to a few seconds.
• The journal plugin always writes PersistentRepr entries to partition 0 of journal topics. This ensures that all events written by a single persistent actor are stored in correct order. Later versions of the plugin may switch to a higher partition after having written a configurable number of events to the current partition.
• The journal plugin distributes Event entries to all available partitions of user-defined topics. The partition key is the event's persistenceId so that a partial ordering of events is preserved when consuming events from user-defined topics. In other words, events written by a single persistent actor are always consumed in correct order but the relative ordering of events from different persistent actors is not defined.

Current limitations

• The journal plugin does not support features that have been deprecated in Akka 2.3.4 (channels and single event deletions).
• Range deletions are not persistent (which may not be relevant for applications that configure Kafka with reasonably small retention times).

Example source code

The complete source code of all examples from previous sections is in Example.scala, the corresponding configuration in example.conf.

Snapshot store plugin

Activation

To activate the snapshot store plugin, add the following line to application.conf:

akka.persistence.snapshot-store.plugin = "kafka-snapshot-store"

This will run the snapshot store plugin with default settings and connect to a Zookeeper instance running on localhost:2181. The Zookeeper connect string can be customized with the kafka-snapshot-store.zookeeper.connect configuration key (see also section Kafka cluster). Recommended Kafka broker configurations are given in section Configuration hints.

Snapshot topics

For each persistent actor, the plugin creates a Kafka topic where the topic name equals the actor's persistenceId, prefixed by the value of the kafka-snapshot-store.prefix configuration key which defaults to snapshot-. For example, if an actor's persistenceId is example, its snapshots are published to topic snapshot-example. For persistent views, the viewId is taken instead of the persistenceId.

Implementation notes

• During initialization, the journal plugin fetches cluster metadata from Zookeeper which may take up to a few seconds.
• The journal plugin always writes snapshots to partition 0 of snapshot topics.

Current limitations

• Deletions are not persistent (which may not be relevant for applications that configure Kafka with reasonably small retention times).

Kafka

Kafka cluster

To connect to an existing Kafka cluster, an application must set a value for the kafka-journal.zookeeper.connect key in its application.conf:

kafka-journal.zookeeper.connect = "<host1>:<port1>,<host2>:<port2>,..."

If you want to run a Kafka cluster on a single node, you may find this article useful.

Test server

To use the test server, the following additional dependencies must be added to build.sbt:

libraryDependencies ++= Seq(
  "com.github.krasserm" %% "akka-persistence-kafka" % "0.4" % "test" classifier "tests",
  "org.apache.curator" % "curator-test" % "2.7.1" % "test"
)

This makes the TestServer class available which can be used to start a single Kafka and Zookeeper instance:

import akka.persistence.kafka.server.TestServer

// start a local Kafka and Zookeeper instance
val server = new TestServer() 

// use the local instance
// ...

// and stop it
server.stop()

The TestServer configuration can be customized with the test-server.* configuration keys (see reference configuration for details).

Configuration hints

The following broker configurations are recommended for being used with the storage plugins:

• num.partitions should be set to 1 by default because the plugins only write to partition 0 of journal topics and snapshot topics. If a higher number of partitions is needed for user-defined topics (e.g. for scalability or throughput reasons) then this should be configured manually with the kafka-topics command line tool.
• default.replication.factor should be set to at least 2 for high-availability of topics created by the plugins.
• message.max.bytes and replica.fetch.max.bytes should be set to a value that is larger than the largest snapshot size. The default value is 1024 * 1024 which may be large enough for journal entries but likely to small for snapshots. When changing these settings make sure to also set kafka-snapshot-store.consumer.fetch.message.max.bytes and kafka-journal.consumer.fetch.message.max.bytes to this value.
• log.cleanup.policy must be set to "compact" otherwise the most recent snapshot may be deleted if the retention time is exceeded and complete state recovery of persistent actors is not possible any more.

See also section Usage hints.

Reference configuration

kafka-journal {

  # FQCN of the Kafka journal plugin
  class = "akka.persistence.kafka.journal.KafkaJournal"

  # Dispatcher for the plugin actor
  plugin-dispatcher = "kafka-journal.default-dispatcher"

  # Number of concurrent writers (should be <= number of available threads in
  # dispatcher).
  write-concurrency = 8

  # The partition to use when publishing to and consuming from journal topics.
  partition = 0

  # Default dispatcher for plugin actor.
  default-dispatcher {
    type = Dispatcher
    executor = "fork-join-executor"
    fork-join-executor {
      parallelism-min = 2
      parallelism-max = 8
    }
  }

  consumer {
    # -------------------------------------------------------------------
    # Simple consumer configuration (used for message replay and reading
    # metadata).
    #
    # See http://kafka.apache.org/documentation.html#consumerconfigs
    # See http://kafka.apache.org/documentation.html#simpleconsumerapi
    # -------------------------------------------------------------------

    socket.timeout.ms = 30000

    socket.receive.buffer.bytes = 65536

    fetch.message.max.bytes = 1048576
  }

  producer {
    # -------------------------------------------------------------------
    # PersistentRepr producer (to journal topics) configuration.
    #
    # See http://kafka.apache.org/documentation.html#producerconfigs
    #
    # The metadata.broker.list property is set dynamically by the journal.
    # No need to set it here.
    # -------------------------------------------------------------------

    request.required.acks = 1

    # DO NOT CHANGE!
    producer.type = "sync"

    # DO NOT CHANGE!
    partitioner.class = "akka.persistence.kafka.StickyPartitioner"

    # DO NOT CHANGE!
    key.serializer.class = "kafka.serializer.StringEncoder"

    # Increase if hundreds of topics are created during initialization.
    message.send.max.retries = 5

    # Increase if hundreds of topics are created during initialization.
    retry.backoff.ms = 100

    # Add further Kafka producer settings here, if needed.
    # ...
  }

  event.producer {
    # -------------------------------------------------------------------
    # Event producer (to user-defined topics) configuration.
    #
    # See http://kafka.apache.org/documentation.html#producerconfigs
    # -------------------------------------------------------------------

    producer.type = "sync"

    request.required.acks = 0

    topic.mapper.class = "akka.persistence.kafka.DefaultEventTopicMapper"

    key.serializer.class = "kafka.serializer.StringEncoder"

    # Add further Kafka producer settings here, if needed.
    # ...
  }

  zookeeper {
    # -------------------------------------------------------------------
    # Zookeeper client configuration
    # -------------------------------------------------------------------

    connect = "localhost:2181"

    session.timeout.ms = 6000

    connection.timeout.ms = 6000

    sync.time.ms = 2000
  }
}

kafka-snapshot-store {

  # FQCN of the Kafka snapshot store plugin
  class = "akka.persistence.kafka.snapshot.KafkaSnapshotStore"

  # Dispatcher for the plugin actor.
  plugin-dispatcher = "kafka-snapshot-store.default-dispatcher"

  # The partition to use when publishing to and consuming from snapshot topics.
  partition = 0

  # Topic name prefix (which prepended to persistenceId)
  prefix = "snapshot-"

  # Default dispatcher for plugin actor.
  default-dispatcher {
    type = Dispatcher
    executor = "fork-join-executor"
    fork-join-executor {
      parallelism-min = 2
      parallelism-max = 8
    }
  }

  consumer {
    # -------------------------------------------------------------------
    # Simple consumer configuration (used for loading snapshots and
    # reading metadata).
    #
    # See http://kafka.apache.org/documentation.html#consumerconfigs
    # See http://kafka.apache.org/documentation.html#simpleconsumerapi
    # -------------------------------------------------------------------

    socket.timeout.ms = 30000

    socket.receive.buffer.bytes = 65536

    fetch.message.max.bytes = 1048576
  }

  producer {
    # -------------------------------------------------------------------
    # Snapshot producer configuration.
    #
    # See http://kafka.apache.org/documentation.html#producerconfigs
    #
    # The metadata.broker.list property is set dynamically by the journal.
    # No need to set it here.
    # -------------------------------------------------------------------

    request.required.acks = 1

    producer.type = "sync"

    # DO NOT CHANGE!
    partitioner.class = "akka.persistence.kafka.StickyPartitioner"

    # DO NOT CHANGE!
    key.serializer.class = "kafka.serializer.StringEncoder"

    # Add further Kafka producer settings here, if needed.
    # ...
  }

  zookeeper {
    # -------------------------------------------------------------------
    # Zookeeper client configuration
    # -------------------------------------------------------------------

    connect = "localhost:2181"

    session.timeout.ms = 6000

    connection.timeout.ms = 6000

    sync.time.ms = 2000
  }
}

test-server {
  # -------------------------------------------------------------------
  # Test Kafka and Zookeeper server configuration.
  #
  # See http://kafka.apache.org/documentation.html#brokerconfigs
  # -------------------------------------------------------------------

  zookeeper {

    port = 2181

    dir = "data/zookeeper"
  }

  kafka {

    broker.id = 1

    port = 6667

    num.partitions = 2

    log.cleanup.policy = "compact"

    log.dirs = data/kafka

    log.index.size.max.bytes = 1024
  }
}

akka {
  actor {
    serializers {
      kafka-snapshot = "akka.persistence.kafka.snapshot.KafkaSnapshotSerializer"
    }

    serialization-bindings {
      "akka.persistence.kafka.snapshot.KafkaSnapshot" = kafka-snapshot
    }
  }
}