David TaiebSTSM - IBM Cloud Data ServicesDeveloper advocate [email protected]
My boss wants me to work on Apache Spark: should I use Scala, Java or Python or a combination of the three?Java One 2016, San Francisco
©2016 IBM Corporation 
Objectives
By the end of this session, you should be able to:– Have Basic knowledge of Apache Spark (at least enough to get
you started)– Understand key differences between Java, Scala and Python– Better informed about deciding which language to use– Understand the role of the Notebook in quickly building data
solution.
©2016 IBM Corporation 
Let’s start with a story…
DisclaimerAll characters and events depicted in this Story are entirely fictitious. Any similarity to actual use cases, events or persons is actually intentional
©2016 IBM Corporation 
Meet Ben “the developer”
• Hold a master degree in computer science• 10 year experience, 6 years with the company• Full stack Web developer• Languages of choice: Java, Node.js, HTML5/CSS3• Data: No SQL (Cloudant, Mongo), relational• Protocols: REST, JSON, MQTT• No major experience with Big data
Favorite Quote:“The Best Line of Code is the One I Didn't Have to
Write!”
©2016 IBM Corporation 
Meet Natasha “the data scientist”
• Hold a PHD in data science• 5 year experience, 2 years with the company• Experienced in Python and R• Expert in Machine Learning and Data visualization• Software engineering is not her thing
Favorite Quote:“In God we trust.
All others bring data”W. Edwards Deming
©2016 IBM Corporation 
Surprise meeting with VP of development
We have an urgent need for our marketing departmentto build an application that can provide real-time sentiment analysis on Twitter data
courtesy of http://linkq.com.vn/
©2016 IBM Corporation 
Key Constraints
• You only have 6 weeks to build the application• Target consumer is the LOB user: must be easy to use even for
non technical people• Web interface: should be accessible from a standard browser
(desktop or mobile)• It must scale out of the box: I want you to use Apache Spark
©2016 IBM Corporation 
Some learning to do
What exactly is Apache Spark?
©2016 IBM Corporation 
What is Apache spark ?
Spark is an open sourcein-memory
computing framework for distributed data processing and
iterative analysis on massive data volumes
©2016 IBM Corporation 
Spark Core Libraries
general compute engine, handles distributed task dispatching, scheduling and basic I/O functions
Spark SQL
executes SQL statements
performs streaming
analytics using micro-batches
common machine learning and
statistical algorithmsdistributed graph
processing framework
Spark Streaming
MllibMachine Learning
GraphXGraph
Spark Core
©2016 IBM Corporation 
Spark is evolved quickly and has strong tractionSpark is one of the most active open source projects
Interest over time (Google Trends)
Source: https://www.google.com/trends/explore#q=apache%20spark&cmpt=q&tz=http://www.indeed.com/jobanalytics/jobtrends?q=apache+spark&l=
Job Trends (Indeed.com)
©2016 IBM Corporation 
Resilient Distributed Datasets (RDDs)• A collection of elements that Spark works on in parallel• May be kept in memory or on disk• Applications can also explicitly tell Spark to
cache an RDD, which is great for iterative algorithms• An RDD contains the “raw data”, plus the function
to compute it • Fault-tolerance: if any partition of an RDD is lost, it will automatically be
recomputed using the transformations that originally created it
RDD built from a Java collection
RDD built from an external dataset(local FS, HDFS, Hbase,…)
©2016 IBM Corporation 
RDD’s fault tolerance scenario
13
RDD1
Map Transformation
Data Node A
File Input Split 1
File Input Split 2
File Input Split 3
RDD1
RDD1
RDD2
RDD2
ReduceTransformation
RDD3
RDD3
filterTransformation
Data Node B
©2016 IBM Corporation 
Spark SQL and Data Frames• Spark’s interface for working with structured and semi-structured data
• Ability to load from a verity of data sources (parquet, JSON, Hive)• Supports SQL syntax for both JDBC/ODBC connectors• Special RDD tooling (Join between RDD and an SQL table, Expose UDF’s)
• DataFrames• A Data Frame is a distributed collection of data organized into named columns.• It is conceptually equivalent to a table in a relational database• DataFrames can be constructed from a wide array of sources such as: structured data
files, tables in Hive, external databases, or existing RDDs• Provide better performance than RDD thanks to Spark SQL’s catalyst query optimizer
©2016 IBM Corporation 
Consuming Spark
15
Spark Application(driver)
Master(cluster Manager)
Worker Node
Worker Node
Worker Node
Worker Node
…
Spark Cluster
Kernel
Master(cluster Manager)
Worker Node
Worker Node
…Spark Cluster
Notebook Server
Browser
Http/WebSockets
Kernel Protocol
Batch Job(Spark-Submit)
Interactive Notebook
• RDD Partitioning• Task packaging and
dispatching• Worker node
scheduling
©2016 IBM Corporation 
IBM’s commitment to Spark
• Contribute to the Core• Spark Technology Cluster (STC)• Open source SystemML
• Foster Community• Educate 1M+ data scientists and engineers• Sponsor AMPLab
IBM Analytics for Apache Spark
• IBM Analytics for Apache Spark Managed Service• www.ibm.com/analytics/us/en/technology/cloud-data-services/spark-as-a-service
IBM Data Science Experience
• DataScience Experience: • datascience.ibm.com
IBM Packages for Apache Spark
• IBM Packages for Apache Spark: • ibm.biz/spark-kit
©2016 IBM Corporation 
Ben and Natasha start brainstorming
• I’ll work on data acquisition from Twitter and enrichment with sentiment analysis scores using Spark Streaming
• I know Java very well, but I don’t have time to learn Python.
• However, I am willing to learn Scala if that helps improve my productivity
• I’ll perform the data exploration and analysis
• I know Python and R, but am not familiar enough with Java or Scala
• I like pandas and numpy. I’m ok to learn Spark but expect the same level of apis
• I need to work iteratively with the data
©2016 IBM Corporation 
More learning
What exactly is a Notebook?
©2016 IBM Corporation https://www.pinterest.com/pin/309763280589997658/&psig=AFQjCNGrBtz45g-g-yU_j9Wj3eOuqhIz8w&ust=1472442566813212
©2016 IBM Corporation http://ia.media-imdb.com/images/M/MV5BMTk3OTM5Njg5M15BMl5BanBnXkFtZTYwMzA0ODI3._V1_SX640_SY720_.jpg
©2016 IBM Corporation 
What is a Notebook
Text, Annotations
Code, Data
Visualizations, Widgets, Output
• Web based UI for running apache spark console commands
• Easy, no install spark accelerator
• Best way to start working with spark
©2016 IBM Corporation 
What is Jupyter?
with a “y”, clever ah? Browser
Kernel
Code Output
https://www.bluetrack.com/uploads/items_images/kernel-of-corn-stress-balls1_thumb.jpg?r=1
Look! Apache Toree for Spark
"Open source, interactive data science and scientific computing"
– Formerly IPython– Large, open, growing community and ecosystem
Very popular– “~2 million users for IPython” [1]
– $6m in funding in 2015 [3]
– 200 contributors to notebook subproject alone [4]
– 275,000 public notebooks on GitHub [2]
©2016 IBM Corporation 
Language war?
Likes ScalaLikes Python
Easy to learn Scala has many powerful features that may be hard to learn, but I only want to use the features that are similar to Java
Fairly Easy once you get familiar with Python Idiosyncrasies
Performance Scala is super fast, because it runs in the JVMYeah, Python is interpreted and slower, but there are ways to optimize when needed
Ecosystem Scala ecosystem is small but can reuse Java ecosystem as it is fully interoperable with Java
Statistics: Pandas, Numpy, matplotlibMachine Learning: scikit-learn, PyML
Tooling IntelliJ, Scala IDE for Eclipse PyCharm, PyDev (Eclipse)
Interactive Scala repl (Console), Scala Notebook for Spark (Limited viz libraries)
Python interpreter, Python Notebook (Jupyter, Zeppelin) provide great viz libs
©2016 IBM Corporation 
Scala vs Java• Less Verbosepublic class Circle{ private double radius; private double xCenter; private double yCenter; public Circle(double radius,double xCenter,
double yCenter){this.radius = radius;this.xCenter = xCenter;this.yCenter = yCenter;
} public void setRadius(double radius){
this.radius=radius;} public double getRadius(){return radius;} public void setXCenter(double xCenter){
this.xCenter=xCenter;} public double getXCenter(){return xCenter;} public void setYCenter(double yCenter){
this.yCenter=yCenter;} public double getYCenter(){return yCenter;}}
Java
class Circle(var radius:Double, var xCenter:Double, var yCenter:Double
)
Scala
Define a Class with a few fields
©2016 IBM Corporation 
Scala vs Java• Less Verbose• Mixes OOP and FP
Sort a list of circles by radius size and by center abscissa Collections.sort(circles, new Comparator<Circle>() { public int compare(Circle a, Circle b) { return a.getRadius()-b.getRadius() }});Collections.sort(circles, new Comparator<Circle>() { public int compare(Circle a, Circle b) { return a.getXCenter()-b.getXCenter(); }});
circles.sortBy( c=>(c.radius,c.xCenter))
ScalaJava 7 and below
circles .sort(Comparator.comparing(Circle::getRadius()) .thenComparing(Circle::getXCenter())
Java 8
©2016 IBM Corporation 
Scala vs Java• Less Verbose• Mixes OOP and FP• Concurrency: Actors and Futures
class JavaOneActor extends Actor { def receive = { case ”version" => println(”JavaOne 2016")
case “location” => println(“San Francisco”) case _ => println(”Sorry, I didn’t get that") }}
object JavaOneActorMain extends App { val system = ActorSystem(”JavaOne”) val actor= system.actorOf(Props[JavaOneActor],name=”JavaOne") actor ! ”version" actor ! ”location” actor ! “Is it raining?”}
©2016 IBM Corporation 
Scala vs Java• Less Verbose• Mixes OOP and FP• Actors• Better type system
• Parameterized Types• Abstract Types• Compound Types• Singleton Types• Tuples Types
• Function Types• Lambdas Types• Self-Recursive Types• Functors• Monads
©2016 IBM Corporation 
Scala vs Java• Less Verbose• Mixes OOP and FP• Actors• Better type system• Runs as fast
• Scala compiles into Java Bytecode • Full interoperability with Java• Runs in the JVM• Of course, algorithm and optimizations matters
©2016 IBM Corporation 
Scala vs Java• Less Verbose• Mixes OOP and FP• Actors• Better type system• Runs as fast• Some features can be confusing
• SBT looks like black magic• Everything is a function: easy to write cryptic code• Implicits• Currying• Partial functions• Partially applied functions
courtesy of http://www.codeodor.com/
©2016 IBM Corporation 
Let’s look at an example Spark Application
The world famous Word Count!
©2016 IBM Corporation 
Java
JavaRDD<String> textFile = sc.textFile("hdfs://...");JavaRDD<String> words = textFile.flatMap(new FlatMapFunction<String, String>() { public Iterable<String> call(String s) { return Arrays.asList(s.split(" ")); }});JavaPairRDD<String, Integer> pairs = words.mapToPair(new PairFunction<String, String, Integer>() { public Tuple2<String, Integer> call(String s) { return new Tuple2<String, Integer>(s, 1); }});JavaPairRDD<String, Integer> counts = pairs.reduceByKey(new Function2<Integer, Integer, Integer>() { public Integer call(Integer a, Integer b) { return a + b; }});counts.saveAsTextFile("hdfs://...");
Create RDD from a file on the hadoop cluster
New RDD made of all words in the files
For each word create a tuple (word,1)
Aggregate all the results, using the word as the key
Save the results back to the hadoop cluster
©2016 IBM Corporation 
Scala
val textFile = sc.textFile("hdfs://...")val counts = textFile.flatMap(line => line.split(" ")) .map(word => (word, 1)) .reduceByKey(_ + _)counts.saveAsTextFile("hdfs://...")
Create RDD from a file on the hadoop cluster
New RDD made of all words in the files
For each word create a tuple (word,1)
Aggregate all the results, using the word as the key
Save the results back to the hadoop cluster
©2016 IBM Corporation 
Python
text_file = sc.textFile("hdfs://...")counts = text_file.flatMap(lambda line: line.split(" ")) \ .map(lambda word: (word, 1)) \ .reduceByKey(lambda a, b: a + b)counts.saveAsTextFile("hdfs://...")
Create RDD from a file on the hadoop cluster
New RDD made of all words in the files
For each word create a tuple (word,1)
Aggregate all the results, using the word as the key
Save the results back to the hadoop cluster
©2016 IBM Corporation 
OK, let’s agree on the architecture
Watson Tone Analyzer
Input Stream
Enrich data with Emotion Tone Scores
Processed data
Notebook
Agree to disagree on the Language
©2016 IBM Corporation 
Dividing the tasks
• Implement a Spark Streaming connector to Twitter
• Call Watson Tone Analyzer for each tweets
• Return a Spark DataFrame with the tweets enriched with Tone scores
• Code written in Scala, delivered as a Jar• Will test in Scala Notebook
• Works in a Python Notebook• Load the twitter data with Tone score from
a persisted store• Perform the data exploration and analysis:
trending hashtags and sentiments• Produce visualizations to LOB Users
©2016 IBM Corporation 
A word on Watson Tone Analyzer• Uses linguistic analysis to detect 3 types of tones: Emotion, Social Tendencies and Language styles• Available as a cloud service on IBM Bluemix
Input
http://www.ibm.com/watson/developercloud/tone-analyzer.html
Results
©2016 IBM Corporation 
How is Ben doing?
Spark Streaming to Twitter
Get Watson Tone sentiment scores
Enrich Tweets with new Scores
Make spark Dataframe available
ssc = new StreamingContext( sc, Seconds(5) )ssc.addStreamingListener( new StreamingListener ) val keys = config.getConfig("tweets.key").split(",");val stream = org.apache.spark.streaming.twitter.TwitterUtils.createStream( ssc, None ); val tweets = stream.filter { status =>
Option(status.getUser).flatMap[String] { u => Option(u.getLang)
}.getOrElse("").startsWith("en") && CharMatcher.ASCII.matchesAllOf(status.getText) && ( keys.isEmpty || keys.exists{status.getText.contains(_)
})...
Create a Spark StremingContext with a 5 sec batch window
Create a Twitter Stream
Filter any tweet that are not in English or that do not match the word filters
©2016 IBM Corporation 
How is Ben doing?Spark Streaming to
TwitterGet Watson Tone sentiment scores
Enrich Tweets with new Scores
Make spark Dataframe available
case class DocumentTone( document_tone: Sentiment )case class Sentiment(tone_categories: Seq[ToneCategory]);…val sentimentResults: String = EntityEncoder[String].toEntity("{\"text\": " + JSONObject.quote( status.text ) + "}" ).flatMap { entity => val s = broadcastVar.value.get("watson.tone.url").get + "/v3/tone?version=" + broadcastVar.value.get("watson.api.version").get val toneuri: Uri = Uri.fromString( s ).getOrElse( null ) client( Request( method = Method.POST, uri = toneuri, headers = …, body = entity.body))
.flatMap { response => if (response.status.code == 200 ) { response.as[String] } else { println("Error received from Watson Tone Analyzer.”) null } } }.run //Return Sentiment object pickled from response
upickle.read[DocumentTone](sentimentResults).document_tone
Call the Watson Tone Analyzer Service using a POST Request
Pickle the JSON Response and return a DocumentTone Object
Define the Sentiment Data Model
©2016 IBM Corporation 
How is Ben doing?Spark Streaming to
TwitterGet Watson Tone sentiment scores
Enrich Tweets with new Scores
Make spark Dataframe available
lazy val client = PooledHttp1Client()val rowTweets = tweets.map(status=> {
val sentiment = ToneAnalyzer.computeSentiment( client, status, broadcastVar )
var colValues = Array[Any]( status.getUser.getName, //author status.getCreatedAt.toString, //date status.getUser.getLang, //Lang status.getText, //text Option(status.getGeoLocation).map{ _.getLatitude}.getOrElse(0.0), //lat Option(status.getGeoLocation).map{_.getLongitude}.getOrElse(0.0) //long
) var scoreMap = getScoreMap(sentiment)
colValues = colValues ++ ToneAnalyzer.sentimentFactors.map { f => round(f) * 100.0}//Return [Row, (sentiment, status)](Row(colValues.toArray:_*),(sentiment, status))
}) Enrich the Tweets with Sentiment Scores, returns a Row Object
©2016 IBM Corporation 
How is Ben doing?Spark Streaming to
TwitterGet Watson Tone sentiment scores
Enrich Tweets with new Scores
Make spark Dataframe available
def createTwitterDataFrames(sc: SparkContext) : (SQLContext, DataFrame) = { if ( workingRDD.count <= 0 ){ println("No data receive. Please start the Twitter stream again to collect data") return null } try{ val df = sqlContext.createDataFrame( workingRDD, schemaTweets ) df.registerTempTable("tweets") println("A new table named tweets with " + df.count() + " records has been correctly created and can be accessed through the SQLContext variable") println("Here's the schema for tweets") df.printSchema() (sqlContext, df) }catch{ case e: Exception => {logError(e.getMessage, e ); return null} } }
Transform the RDD of Enriched Tweets to a DataFrame
Display the DataFrame Schema
©2016 IBM Corporation 
How’s Natasha doing?Load the Tweets
from clusterCompute the distribution
of SentimentsCompute top 5
hashtagsCompute aggregate Tone
scores for each 5 top hashtags
Load the Tweets data from a parquet file in Object Storage
Register a Temp SQL Table so it can be queried later on
©2016 IBM Corporation 
How’s Natasha doing?Load the Tweets
from clusterCompute the distribution
of SentimentsCompute top 5
hashtagsCompute aggregate Tone
scores for each 5 top hashtags
Build an array that contains the number of tweets where score is greater than 60%
©2016 IBM Corporation 
How’s Natasha doing?Load the Tweets
from clusterCompute the distribution
of SentimentsCompute top 5
hashtagsCompute aggregate Tone
scores for each 5 top hashtags
Creates a flat Map of all the words in the tweets and filter to keep only the hashtags
Map then Reduce to count the occurrences of each hashtag
©2016 IBM Corporation 
How’s Natasha doing?Load the Tweets
from clusterCompute the distribution
of SentimentsCompute top 5
hashtagsCompute aggregate Tone
scores for each 5 top hashtags
Create RDD from tweets dataframe
Keep only the entries in the top 10
Index by Tag-Tone
©2016 IBM Corporation 
How’s Natasha doing?Load the Tweets
from clusterCompute the distribution
of SentimentsCompute top 5
hashtagsCompute aggregate Tone
scores for each 5 top hashtags
Count the occurrences for each score
Count the tone average for each tag and count
Final Reduce to get the data the way we want it for display
©2016 IBM Corporation 
How’s Natasha doing?Load the Tweets
from clusterCompute the distribution
of SentimentsCompute top 5
hashtagsCompute aggregate Tone
scores for each 5 top hashtags
©2016 IBM Corporation 
How good is Python at handling data?Given the following quarterly sales series
Year Quarter Revenue
Return a list that contains the revenue for a specific quarter, 0 if not defined. e.g.: 1st Quarter: [8977551.03, 6405575.14, 2401247.98, 1877666.79]4th Quarter: [9179464.4, 6717172.01, 2694937.3, 0]
Let’s look at a simple task
Scala
Python
©2016 IBM Corporation 
Let’s take a look at the application
©2016 IBM Corporation 
Meeting back with the VP
This is great, but why do I need to run 2 notebooks – one in Scala and one in Python? Please Fix it!
©2016 IBM Corporation 
Can’t we both get along?• It would be great if we could run Scala code from a Python Notebook• And be able to transfer variables between the 2 languages
Open Source Pixiedust Python library let’s us do just that
https://github.com/ibm-cds-labs/pixiedust
©2016 IBM Corporation 
New Version with Scala mix-in
©2016 IBM Corporation 
What about the LOB User?
courtesy: http://www.flickr.com
C-Suite executive need to be able to run the application from Notebook, select filters and see real-time charts without writing code!
©2016 IBM Corporation 
Embedded application with Pixiedust plugin
©2016 IBM Corporation 
Embedded application with Pixiedust plugin
©2016 IBM Corporation 
Conclusion
• Programming languages are just tools, you need to choose the right one for you (the programmer) and the task:• Scala is better suited for engineering work that
involves large, reusable components• Python is the language of choice for data scientists
• If you are starting with Spark and just want to play:• Start local: no need for a big cluster yet, get installed
and started in minutes• Use the language you are most familiar with, or is
easiest to learn• Use Notebooks to learn the APIs
©2016 IBM Corporation 
Resources• http://programming-scala.org• http://python.org• http://spark.apache.org• www.ibm.com/analytics/us/en/technology/cloud-data-services/spark-as-a-service • http://datascience.ibm.com • http://ibm.biz/spark-kit• www.ibm.com/watson/developercloud/tone-analyzer.html • developer.ibm.com
/clouddataservices/2016/01/15/real-time-sentiment-analysis-of-twitter-hashtags-with-spark/
• developer.ibm.com/clouddataservices/start-developing-with-spark-and-notebooks/ • www.ibm.com/analytics/us/en/technology/spark/• github.com/ibm-cds-labs/spark.samples• github.com/ibm-cds-labs/pixiedust