Avoiding destructive updates and keeping history of data using event sourcing approaches has large advantages for data analytics. This talk describes how Cassandra can be used as event journal as part of CQRS/Lambda Architecture using event sourcing and further used for data mining and machine learning purposes in a big data pipeline.
All the principles are demonstrated on an application called Muvr that we built. It uses data from wearable devices such as accelerometer in a watch or heartbeat monitor to classify user's exercises in near real time. It uses mobile devices and clustered Akka actor framework to distribute computation and then stores events as immutable facts in journal backed by Cassandra. The data are then read by Apache Spark and used for more expensive analytics and machine learning tasks such as suggests improvements to user's exercise routine or improves machine learning models for better real time exercise classification that can be used immediately. The talk mentions some of the internals of Spark when working with Cassandra and focuses on its machine learning capabilities enabled by Cassandra. A lot of the analytics are done for each user individually so the whole pipeline must handle potentially large amount of concurrent users and a lot of raw data so we need to ensure attributes such as responsiveness, elasticity and resilience.
2. ● Introduction
● Event sourcing and CQRS
● An emerging technology stack to handle data
● A reference application and it’s architecture
● A few use cases of the reference application
● Conclusion
3. ● Increasing importance of data analytics
● Current state
○ Destructive updates
○ Analytics tools with poor scalability and integration
○ Manual processes
○ Slow iterations
○ Not suitable for large amounts of data
4. ● Whole lifecycle of data
● Data processing
● Data stores
● Integration and messaging
● Distributed computing primitives
● Cluster managers and task schedulers
● Deployment, configuration management and DevOps
● Data analytics and machine learning
● Spark, Mesos, Akka, Cassandra, Kafka (SMACK, Infinity)
9. ● Append only data store
● No updates or deletes (rewriting history)
● Immutable data model
● Decouples data model of the application and storage
● Current state not persisted, but derived. A sequence of updates that led to it.
● History, state known at any point in time
● Replayable
● Source of truth
● Optimisations possible
● Works well in distributed environment - easy partitioning, conflicts
● Helps avoiding transactions
● Works well with DDD
11. ● Command Query Responsibility Segregation
● Read and write logically and physically separated
● Reasoning about the application
● Clear separation of concerns (business logic)
● Often different technology, scalability
● Often lower consistency - eventual, causal
12. Command
● Write side
● Messages, requests to mutate state
● Behaviour, serialized method call essentially
● Don’t expose state
● Validated and may be rejected or emit one or more events (e.g. submitting a form)
Event
● Write side
● Immutable
● Indicating something that has happened
● Atomic record of state change
● Audit log
Query
● Read side
● Precomputed
21. ● Actor backed by data store
● Immutable event sourced journal
● Supports CQRS (write and read side)
● Persistence, replay on failure, rebalance, at least once delivery
29. ● Akka 2.4
● Potentially infinite stream of data
● Ordered, replayable, resumable
● Aggregation, transformation, moving data
● EventsByPersistenceId
● AllPersistenceids
● EventsByTag
30. val readJournal =
PersistenceQuery(system).readJournalFor(CassandraJournal.Identifier)
val source = readJournal.query(
EventsByPersistenceId(UserPersistenceId(name).persistenceId, 0, Long.MaxValue), NoRefresh)
.map(_.event)
.collect{ case s: EntireResistanceExerciseSession => s }
.mapConcat(_.deviations)
.filter(condition)
.map(process)
implicit val mat = ActorMaterializer()
val result = source.runFold(List.empty[ExercisePlanDeviation])((x, y) => y :: x)
31. ● Potentially infinite stream of events
Source[Any].map(process).filter(condition)
Publisher Subscriber
process
condition
backpressure
32. ● In Akka we have the read and write sides separated,
in Cassandra we don’t
● Different data model
● Avoid using operational datastore
● Eventual consistency
● Streaming transformations to different format
● Unify journalled and other data
33. ● Computations and analytics queries on the data
● Often iterative, complex, expensive computations
● Prepared and interactive queries
● Data from multiple sources, joins and transformations
● Often directly on a stream of data
● Whole history of events
● Historical behaviour
● Works retrospectively, can answer questions in the future that we don’t
know exist yet
● Various data types from various sources
● Large amounts of fast data
● Automated analytics
34. ● Cassandra 3.0 - user defined functions, functional indexes, aggregation
functions, materialized views
● Server side denormalization
● Eventual consistency
● Copy of data with different partitioning
userId
performance
35. ● In memory dataflow distributed data processing framework, streaming
and batch
● Distributes computation using a higher level API
● Load balancing
● Moves computation to data
● Fault tolerant
39. ● Cassandra can store
● Spark can process
● Gathering large amounts of heterogeneous data
● Queries
● Transformations
● Complex computations
● Machine learning, data mining, analytics
● Now possible
● Prepared and interactive queries
40. lazy val sparkConf: SparkConf =
new SparkConf()
.setAppName(...).setMaster(...).set("spark.cassandra.connection.host", "127.0.0.1")
val sc = new SparkContext(sparkConf)
val data = sc.cassandraTable[T]("keyspace", "table").select("columns")
val processedData = data.flatMap(...)...
processedData.saveToCassandra("keyspace", "table")
41. ● Akka Analytics project
● Handles custom Akka serialization
case class JournalKey(persistenceId: String, partition: Long, sequenceNr: Long)
lazy val sparkConf: SparkConf =
new SparkConf()
.setAppName(...).setMaster(...).set("spark.cassandra.connection.host", "127.0.0.1")
val sc = new SparkContext(sparkConf)
val events: RDD[(JournalKey, Any)] = sc.eventTable()
events.sortByKey().map(...).filter(...).collect().foreach(println)
42. ● Spark streaming
● Precomputing using spark or replication often aiming for different data
model
Operational cluster Analytics cluster
Precomputation /
replication
Integration with
other data sources
43. val events: RDD[(JournalKey, Any)] = sc.eventTable().cache().filterClass[EntireResistanceExerciseSession].flatMap(_.deviations)
val deviationsFrequency = sqlContext.sql(
"""SELECT planned.exercise, hour(time), COUNT(1)
FROM exerciseDeviations
WHERE planned.exercise = 'bench press'
GROUP BY planned.exercise, hour(time)""")
val deviationsFrequency2 = exerciseDeviationsDF
.where(exerciseDeviationsDF("planned.exercise") === "bench press")
.groupBy(
exerciseDeviationsDF("planned.exercise"),
exerciseDeviationsDF("time”))
.count()
val deviationsFrequency3 = exerciseDeviations
.filter(_.planned.exercise == "bench press")
.groupBy(d => (d.planned.exercise, d.time.getHours))
.map(d => (d._1, d._2.size))
44. def toVector(user: User): mllib.linalg.Vector =
Vectors.dense(
user.frequency, user.performanceIndex, user.improvementIndex)
val events: RDD[(JournalKey, Any)] = sc.eventTable().cache()
val users: RDD[User] = events.filterClass[User]
val kmeans = new KMeans()
.setK(5)
.set...
val clusters = kmeans.run(users.map(_.toVector))
45. val weight: RDD[(JournalKey, Any)] = sc.eventTable().cache()
val exerciseDeviations = events
.filterClass[EntireResistanceExerciseSession]
.flatMap(session =>
session.sets.flatMap(set =>
set.sets.map(exercise => (session.id.id, exercise.exercise))))
.groupBy(e => e)
.map(g =>
Rating(normalize(g._1._1), normalize(g._1._2),
normalize(g._2.size)))
val model = new ALS().run(ratings)
val predictions = model.predict(recommend)
bench
press
bicep
curl
dead
lift
user 1 5 2
user 2 4 3
user 3 5 2
user 4 3 1
46. val events = sc.eventTable().cache().toDF()
val lr = new LinearRegression()
val pipeline = new Pipeline().setStages(Array(new UserFilter(), new ZScoreNormalizer(),
new IntensityFeatureExtractor(), lr))
val paramGrid = new ParamGridBuilder()
.addGrid(lr.regParam, Array(0.1, 0.01))
.addGrid(lr.fitIntercept, Array(true, false))
getEligibleUsers(events, sessionEndedBefore)
.map { user =>
val trainValidationSplit = new TrainValidationSplit()
.setEstimator(pipeline)
.setEvaluator(new RegressionEvaluator)
.setEstimatorParamMaps(paramGrid)
val model = trainValidationSplit.fit(
events,
ParamMap(ParamPair(userIdParam, user)))
val testData = // Prepare test data.
val predictions = model.transform(testData)
submitResult(userId, predictions, config)
}
47. val events: RDD[(JournalKey, Any)] = sc.eventTable().cache()
val connections = events.filterClass[Connections]
val vertices: RDD[(VertexId, Long)] =
connections.map(c => (c.id, 1l))
val edges: RDD[Edge[Long]] = connections
.flatMap(c => c.connections
.map(Edge(c.id, _, 1l)))
val graph = Graph(vertices, edges)
val ranks = graph.pageRank(0.0001).vertices
53. ● Exercise domain as an example
● Analytics of both batch (offline) and streaming (online) data
● Analytics important in other areas (banking, stock market, network,
cluster monitoring, business intelligence, commerce, internet of things, ...)
● Enabling value of data
54. ● Event sourcing
● CQRS
● Technologies to handle the data
○ Spark
○ Mesos
○ Akka
○ Cassandra
○ Kafka
● Handling data
● Insights and analytics enable value in data
55.
56. ● Jobs at www.cakesolutions.net/careers
● Code at https://github.com/muvr
● Martin Zapletal @zapletal_martin
● Anirvan Chakraborty @anirvan_c