Akka Analytics

Large-scale event processing with Akka Persistence and Apache Spark. At the moment you can

• batch-process events from an Akka Persistence Cassandra journal as Spark RDD.
• stream-process events from an Akka Persistence Kafka journal as Spark Streaming DStream.

This project is in early preview state.

Dependencies

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

libraryDependencies ++= Seq(
  "com.github.krasserm" %% "akka-analytics-cassandra" % "0.2",
  "com.github.krasserm" %% "akka-analytics-kafka" % "0.2"
)

Event batch processing

With akka-analytics-cassandra you can expose and process events written by all persistent actors as resilient distributed dataset (RDD). It uses the Spark Cassandra Connector to fetch data from the Cassandra journal. Here's a primitive example (details here):

import org.apache.spark.rdd.RDD
import org.apache.spark.SparkConf
import org.apache.spark.SparkContext
import org.apache.spark.SparkContext._

import akka.analytics.cassandra._

val conf = new SparkConf()
 .setAppName("CassandraExample")
 .setMaster("local[4]")
 .set("spark.cassandra.connection.host", "127.0.0.1")

val sc = new SparkContext(conf)

// expose journaled Akka Persistence events as RDD
val rdd: RDD[(JournalKey, Any)] = sc.eventTable(keyspace = "akka", table = "messages").cache()

// and do some processing ... 
rdd.sortByKey().map(...).filter(...).collect().foreach(println)

The dataset generated by eventTable() is of type RDD[(JournalKey, Any)] where Any represents the persisted event (see also Akka Persistence API) and JournalKey is defined as

package akka.analytics.cassandra

case class JournalKey(persistenceId: String, partition: Long, sequenceNr: Long)

Events for a given persistenceId are partitioned across nodes in the Cassandra cluster where the partition is represented by the partition field in the key. The eventTable() method returns an RDD in which events with the same persistenceId - partition combination (= cluster partition) are ordered by increasing sequenceNr but the ordering across cluster partitions is not defined. If needed the RDD can be sorted with sortByKey() by persistenceId, partition and sequenceNr in that order of significance. Btw, the default size of a cluster partition in the Cassandra journal is 5000000 events (see akka-persistence-cassandra).

Event stream processing

With akka-analytics-kafka you can expose and process events written by all persistent actors (more specific, from any user-defined topic) as discretized stream (DStream). Here's a primitive example (details here):

import org.apache.spark.SparkConf
import org.apache.spark.streaming.{Seconds, StreamingContext}
import org.apache.spark.streaming.dstream.DStream

import akka.analytics.kafka._
import akka.persistence.kafka.Event

val sparkConf = new SparkConf()
  .setAppName("events-consumer")
  .setMaster("local[4]")

// read from user-defined events topic 
// with 2 threads (see also Kafka API) 
val topics = Map("events" -> 2)
val params = Map[String, String](
  "group.id" -> "events-consumer",
  "auto.commit.enable" -> "false",  
  "auto.offset.reset" -> "smallest",
  "zookeeper.connect" -> "localhost:2181",
  "zookeeper.connection.timeout.ms" -> "10000")

val ssc = new StreamingContext(sparkConf, Seconds(1))
val es: DStream[Event] = ssc.eventStream(params, topics)

es.foreachRDD(rdd => rdd.map(...).filter(...).collect().foreach(println))

ssc.start()
ssc.awaitTermination()

The stream generated by eventStream(...) is of type DStream[Event] where Event is defined in akka-persistence-kafka as

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)

The stream of events (written by all persistent actors) is partially ordered i.e. events with the same persistenceId are ordered by sequenceNr whereas the ordering of events with different persistenceId is not defined. Details about Kafka consumer params are described here.

Custom deserialization

See this ticket for details.