5 Ways to Use Spark to Enrich your Cassandra Environment
- 1. Jim Hatcher DFW Cassandra Users - Meetup 5/18/2016 5 Ways to use Spark to Enrich your Cassandra Environment C*
- 2. Agenda • Introduction • Data Systems • What is Cassandra? • Tradeoffs vs. RDBMS • Addressing Limitations • What is Spark? • Five Ways to Use Spark in a Cassandra Environment • ETL • Data Migrations • Consistency Checking / Syncing Denormalized Data • Analytics • Machine Learning • Spark Resources
- 3. Introduction Jim Hatcher james_hatcher@hotmail.com At IHS, we take raw data and turn it into information and insights for our customers. Automotive Systems (CarFax) Defense Systems (Jane’s) Oil & Gas Systems (Petra) Maritime Systems Technology Systems (Electronic Parts Database, Root Metrics) Sources of Raw Data Structure Data Add Value Customer-facing Systems
- 4. Data Systems Big Data Analytics Hadoop, MapReduce, Hive, Spark NoSQL Cassandra, Hbase, MongoDB Data Warehousing SQL Server, Oracle, SAS, Tableau Relational Database SQL Server, Oracle, DB2 Analytical Operational ConventionalScaleBigData Batch Processing Minutes-to-Hours Range Queries Visualization / Dashboards “Commodity” Hardware Scale Out Real-time Processing Milliseconds Discrete Seeks/Updates Line of Business Apps Large Servers using Shared Storage Scale Up
- 5. Hadoop, MapReduce, Hive, Spark Cassandra, Hbase, MongoDb SQL Server, Oracle, SAS, Tableau SQL Server, Oracle, DB2 Analytical Operational ConventionalScaleBigDataData Systems Factors: • Size/Scale of Data • Multi-Data Center (with writes) • Rate of Data Ingest • Massive Concurrency • Uptime Requirements • Operational Complexity
- 6. What is Cassandra? Cassandra Cluster B C D E F Client -9223372036854775808 through -6148914691236517207 -6148914691236517206 through -3074457345618258605 -3074457345618258604 through -3 -2 through 3074457345618258599 3074457345618258600 through 6148914691236517201 6148914691236517202 through 9223372036854775808 CREATE KEYSPACE orders WITH replication = { 'class': 'SimpleStrategy', 'replication_factor': 3 }; CREATE TABLE orders.customer ( customer_id uuid, customer_name varchar, customer_age int, PRIMARY KEY ( customer_id ) ) INSERT INTO customer (feb2b9e6-613b-4e6b-b470-981dc4d42525, ‘Bob’, 35) SELECT customer_name, customer_age FROM customer WHERE customer_id = feb2b9e6-613b-4e6b-b470-981dc4d42525 A
- 7. What is Cassandra? Cassandra is • A NoSQL (i.e., non-relational) operational database • Distributed (the data lives on many nodes) • Highly Scalable (no scale ceiling) • Highly Available (no single point of failure) • Open Source • Fast (optimized for fast reads and fast writes) Cassandra uses: • Commodity Hardware (no SAN/NAS or high-end hardware) • Ring Architecture (not master/slave) • Flexible Data Model • CQL (abstraction layer for data access; not tied to a particular language) DataStax provides consulting, support, and additional software around Cassandra.
- 8. What you gain with Cassandra: • Linear Horizontal Scale (HUGE!) • Multi-Data Center / Active-Active • Fast, Scalable writes • Fast Reads (by the key(s)) • Continuous Availability • High Concurrency • Schema Flexibility • Cheaper (commodity hardware)? What you give up with Cassandra: • Tables only queryable by key • 3rd Normal Form • Data Integrity Checks • Foreign Keys • Unique Indexes • Joins • Secondary Indexes • Grouping / Aggregation • ACID Tradeoffs (vs. RDBMS)
- 9. Limitations Solutions DenormalizeData IdempotentDataModel IndexinAnotherTool ConsistencyChecker BatchAnalytics BatchETL Tables only Queryable by Key X X No Foreign Keys / Unique Indexes X X No JOINs X X No GROUP BYs / Aggregation X Keeping Denormalized Data in Sync X Creating New Tables for New Queries X Addressing Limitations
- 10. What is Spark? Spark is a processing framework designed to work with distributed data. “up to 100X faster than MapReduce” according to spark.apache.org Used in any ecosystem where you want to work with distributed data (Hadoop, Cassandra, etc.) Includes other specialized libraries: • SparkSQL • Spark Streaming • MLLib • GraphX Spark Facts Conceptually Similar To MapReduce Written In Scala Supported By DataBricks Supported Languages Scala, Java, or Python
- 11. Spark Architecture Spark Client Driver Spark Context Spark Master Spark Worker Spark Worker Spark Worker Executor Executor Executor 1. Request Resources 2. Allocate Resources 3.StartExecutors 4.PerformComputation Credit: https://academy.datastax.com/courses/ds320-analytics-apache-spark/introduction-spark-architecture
- 12. Spark Terms / Concepts Resilient Distributed Dataset (RDD) Represents an immutable, partitioned collection of elements that can be operated on in parallel. Dataframe RDD + schema This is the “way that everything in Spark is going” Actions and Transformations Transformations – create a new RDD but are executed in a lazy fashion (i.e., when an action fires) Actions – cause a computation to be run and return a response to the driver program Executing Spark Code Spark Shell – run Spark commands interactively via the Spark REPL Spark Submit – execute Spark jobs (i.e., JAR files); you can build a JAR file in the Java IDE of your choice – Eclipse, IntelliJ, etc.
- 13. Spark with Cassandra Credit: https://academy.datastax.com/courses/ds320- analytics-apache-spark/introduction-spark- architecture Cassandra Cluster A CB Spark Worker Spark WorkerSpark Worker Spark Master Spark Client Spark Cassandra Connector – open source, supported by DataStax https://github.com/datastax/spark-cassandra-connector
- 14. ETL (Extract, Transform, Load) Text File JDBC Data Source Cassandra Hadoop Extract Data Spark: Create RDD Data Source(s) Spark Code Transform Data Spark: Map function Spark Code Cassandra Data Source(s) Load Data Spark: Save Spark Code
- 15. ETL import org.apache.spark.{SparkConf, SparkContext} //Create a SparkConfig and a SparkContext val sparkConf = new SparkConf(true) .setAppName("MyEtlApp") .setMaster("spark://10.1.1.1:7077") .set("spark.cassandra.connection.host", "10.2.2.2") ) val sc = new SparkContext(sparkConf) //EXTRACT: Using the SparkContext, read a text file and expose it as an RDD val logfile = sc.textFile("/weblog.csv") //TRANSFORM: split the CSV into fields and then put the fields into a tuple val split = logfile.map { line => line.split(",") } val transformed = split.map { record => ( record(0), record(1) ) } //LOAD: write the tuple structure into Cassandra transformed.saveToCassandra("test", "weblog")
- 16. Data Migrations Cassandra Extract Data Spark: Create RDD Data Source(s) Spark Code Transform Data Spark: Map function Spark Code Cassandra Data Source(s) Load Data Spark: Save Spark Code
- 17. Data Migrations import org.apache.spark.{SparkConf, SparkContext} //Create a SparkConfig and a SparkContext val sparkConf = new SparkConf(true) .setAppName("MyEtlApp") .setMaster("spark://10.1.1.1:7077") .set("spark.cassandra.connection.host", "10.2.2.2") ) val sc = new SparkContext(sparkConf) //EXTRACT: Using the SparkContext, read a C* table and expose it as an RDD val weblogRecords = sc.cassandraTable("test", "weblog").select("logtime", "page") //TRANSFORM: pull fields out of the CassandraRow and put the fields into a tuple val transformed = weblogRecords.map { row => ( row.getString(1), row.getLong(0) ) } //LOAD: write the tuple structure into Cassandra into a different table transformed.saveToCassandra("test", "weblog_bypage")
- 18. Consistency Checking / Syncing Denormalized Data Cassandra Extract Data Spark: Create RDD of missing records Data Source(s) Spark Code Base Table DenormalizedTable1 DenormalizedTable2 Transform Data Spark: Map Function Spark Code Spark Code Load Data Spark: Save missing records to Cassandra
- 19. Consistency Checking / Syncing Denormalized Data import org.apache.spark.sql.hive.HiveContext val hc = new HiveContext(sc) val query1 = """ SELECT w1.logtime, w1.page FROM test.weblog w1 LEFT JOIN test.weblog_bypage w2 ON w1.page = w2.page WHERE w2.page IS NULL""" val results1 = hc.sql(query1) results1.collect.foreach(println) val newRecord = Array(("2016-05-17 2:00:00", "page6.html")) val newRecordRdd = sc.parallelize(newRecord) newRecordRdd.saveToCassandra("test", "weblog") results1.collect.foreach(println) val transformed = results1.map { row => ( row.getString(1), row.get(0) ) } transformed.saveToCassandra("test", "weblog_bypage")
- 20. Analytics //EXAMPLE of a JOIN val query2 = """ SELECT w.page, w.logtime, p.owner FROM test.weblog w INNER JOIN test.webpage p ON w.page = p.page""" val results2 = hc.sql(query2) results2.collect.foreach(println) //EXAMPLE of a GROUP BY val query3 = """ SELECT w.page, COUNT(*) AS RecordCount FROM test.weblog w GROUP BY w.page ORDER BY w.page""" val results3 = hc.sql(query3) results3.collect.foreach(println)
- 21. Machine Learning import org.apache.spark.ml.Pipeline import org.apache.spark.ml.classification.LogisticRegression import org.apache.spark.ml.feature.{HashingTF, Tokenizer} import org.apache.spark.mllib.linalg.Vector import org.apache.spark.sql.{DataFrame, Row, SQLContext} case class LabeledDocument(id: Long, text: String, label: Double) case class DataDocument(id: Long, text: String) lazy val sqlContext = new SQLContext(sc) import sqlContext.implicits._ // Load the training data val modelTrainingRecords = sc.cassandraTable("test", "ml_training") .select("id", "text", "label") val labeledDocuments = modelTrainingRecords.map { record => LabeledDocument(record.getLong("id") , record.getString("text"), record.getDouble("label")) }.toDF
- 22. Machine Learning // Create the pipeline val pipeline = { val tokenizer = new Tokenizer() .setInputCol("text") .setOutputCol("words") val hashingTF = new HashingTF() .setNumFeatures(1000) .setInputCol(tokenizer.getOutputCol) .setOutputCol("features") val lr = new LogisticRegression() .setMaxIter(10) .setRegParam(0.001) new Pipeline() .setStages(Array(tokenizer, hashingTF, lr)) } // Fit the pipeline to training documents. val model = pipeline.fit(labeledDocuments)
- 23. Machine Learning // Load the data to run against the model val modelTestRecords = sc.cassandraTable("test", "ml_text") val dataDocuments = modelTestRecords.map { record => DataDocument(record.getLong(0), record.getString(1)) }.toDF model.transform(dataDocuments) .select("id", "text", "probability", "prediction") .collect() .foreach { case Row(id: Long, text: String, prob: Vector, prediction: Double) => println(s"($id, $text) --> prob=$prob, prediction=$prediction") }
- 24. Resources Spark • Books http://shop.oreilly.com/product/0636920028512.do Scala (Knowing Scala with really help you progress in Spark) • Functional Programming Principles in Scala (videos) https://www.youtube.com/user/afigfigueira/playlists?shelf_id=9&view=50&sort=dd • Books http://www.scala-lang.org/documentation/books.html Spark and Cassandra • DataStax Academy http://academy.datastax.com/ • Self-paced course: DS320: DataStax Enterprise Analytics with Apache Spark – Really Good! • Tutorials • Spark Cassandra Connector website – lots of good examples https://github.com/datastax/spark-cassandra-connector