Page 1
Intro to Apache Spark:Fast cluster computing engine for Hadoop
Intro to Scala:Object-oriented and functional language for the Java Virtual Machine
ACM SIGKDD, 7/9/2014
Roger Huang
Lead System Architect
[email protected]
[email protected]
@BigDataWrangler
Page 2
2Intro to Spark: Intro to Scala | 7/9/2014
About me: Roger Huang• Visa
– Digital & Mobile Products Architecture, Strategic Projects & infrastructure
– Search infrastructure
– Customer segmentation
– Logging Framework
– Splunk on Hadoop (Hunk)
– Real-time monitoring
– Data
• PayPal– Java Infrastructure
Page 3
3Intro to Spark: Intro to Scala | 7/9/2014
Different perspectives on an elephant Scala
Page 4
4Intro to Spark: Intro to Scala | 7/9/2014
Outline• Spark
– Hadoop eco system
• Scala– Background
• Why Scala?– For the computer scientist
– For the Java / OO programmer
– For the Spark developer
– For the Big Data developer
– For the Big Data scientist / mathematician
– For the system architect
Page 5
5Intro to Spark: Intro to Scala | 7/9/2014
Spark in the Hadoop ecosystem
Page 6
6Intro to Spark: Intro to Scala | 7/9/2014
Spark Ecosystem of Software Projects
• Spark [Ognen]– APIs: Scala, Python [Robert], Java
• “SQL”– Shark (Hive + Spark) [Roger]
– SparkSQL (alpha)
• Machine Learning Library (MLlib) [Omar]– Clustering
– Classification• binary classification
• Linear regression
– recommendations
• Spark Streaming [Chance]
• GraphX [Srini]
• …
Page 7
7Intro to Spark: Intro to Scala | 7/9/2014
Resilient Distributed Dataset• Fault tolerant collection of elements partitioned across the
nodes of the cluster that can be operated on in parallel
• Data sources for RDDs– Parallelized collections
• From Scala collections
– Hadoop datasets• From HDFS, any Hadoop supported storage system (Hbase, Amazon
S3, …)
• Text files, SequenceFile, any Hadoop InputFormat
• Two types of operations– Transformation
• takes an existing dataset and creates a new one
– Action• takes a dataset, run a computation, and return value to driver program
Page 8
8Intro to Spark: Intro to Scala | 7/9/2014
(Some) RDD Operations• Transformations
– map(func)
– filter(func)
– flatMap(func)
– mapPartitions(func)
– mapPartitionsWithIndex(func)
– sample(withReplacement, fraction, seed)
– union(otherDataset)
– distinct()
– groupByKey()
– reduceByKey(func)
– sortByKey()
– Join(otherDataset)
– cogroup(otherDataset)
– cartesian(otherDataset)
• Actions– reduce(func)
– collect()
– count()
– first()
– take(n)
– takeSample(withReplacement, num, seed)
– saveAsTextFile(path)
– saveAsSequenceFile(path)
– countByKey()
– foreach(func)
– …
Page 9
9Intro to Spark: Intro to Scala | 7/9/2014
Scala background• Scalable, Object oriented, functional language
– Version 2.11 (4/2014)
• Runs on the Java Virtual Machine
• Martin Odersky
– javac
– Java generics
• http://scala-lang.org/, REPL
• http://www.scala-lang.org/api/current
• http://scala-ide.org/
• http://www.scala-sbt.org/, Simple build tool
• Who’s using Scala?
– Twitter, LinkedIn, …
• Powered by Scala
– Apache Spark, Apache Kafka, Akka,…
Page 10
10Intro to Spark: Intro to Scala | 7/9/2014
Outline• Spark
– Hadoop eco system
• Scala– Background
• Why Scala?– For the computer scientist
– For the Java / OO programmer
– For the Hadoop/Spark developer
– For the Big Data developer
– For the Big Data scientist / mathematician
– For the system architect
Page 11
11Intro to Spark: Intro to Scala | 7/9/2014
Scala for the computer scientist: functional programming (FP)
Page 12
12Intro to Spark: Intro to Scala | 7/9/2014
Scala for the computer scientist: functional programming (FP)
• Math functions, e.g., f(x) = y– A function has a single responsibility
– A function has no side effects
– A function is referentially transparent• A function outputs the same value for the same inputs.
• Functional programming– expresses computation as the evaluation and composition of
mathematical functions
– Avoid side effects and mutating state data
Page 13
13Intro to Spark: Intro to Scala | 7/9/2014
Why functional programming?
• Multi core processors
• Concurrency– Computation as a series of independent data transformations
– Parallel data transformations without side effects
• Referential transparency
Page 14
14Intro to Spark: Intro to Scala | 7/9/2014
Scala for the computer scientist: functional programming
• Functions– Lambda, closure
• For-comprehensions
• Type inference
• Pattern matching
• Higher order functions– map, flatMap, foldLeft
• And more …
Page 15
15Intro to Spark: Intro to Scala | 7/9/2014
FP: functions
• Anonymous function– Function without a name
– lambda function
• Example– scala> List(100, 200, 300) map { _ * 10/100}
– res0: List[Int] = List(10, 20, 30)
• Closure (Wikipedia)– Closure = A function, together with a referencing environment – a
table storing a reference to each of the non-local variables of that function.
– A closure allows a function to access those non-local variables even when invoked outside its immediate lexical scope.
Page 16
16Intro to Spark: Intro to Scala | 7/9/2014
FP: functions
• applyPercentage is an example of a closure– scala> var percentage = 10
– percentage: Int = 10
– scala> val applyPercentage = (amount: Int) => amount * percentage / 100
– applyPercentage: Int => Int = <function1>
– scala> percentage = 20
– percentage: Int = 20
– scala> List (100, 200, 300) map applyPercentage
– res1: List[Int] = List(20, 40, 60)
– scala>
Page 17
17Intro to Spark: Intro to Scala | 7/9/2014
FP: functions
• Anonymous function
• Closure
Page 18
18Intro to Spark: Intro to Scala | 7/9/2014
FP: Higher order functionsscala> :load Person.scala
Loading Person.scala...
defined class Person
scala> val jd = new Person("John", "Doe", 17)
jd: Person = Person@372a6e85
scala> val rh = new Person("Roger", "Huang", 34)
rh: Person = Person@611c4041
scala> val people = Array(jd, rh)
people: Array[Person] = Array(Person@372a6e85, Person@611c4041)
scala> val (minors, adults) = people partition (_.age < 18)
minors: Array[Person] = Array(Person@372a6e85)
adults: Array[Person] = Array(Person@611c4041)
scala>
Page 19
19Intro to Spark: Intro to Scala | 7/9/2014
FP: Higher order functions
• HOF– takes a function as an argument
– Returns a function
Page 20
20Intro to Spark: Intro to Scala | 7/9/2014
FP: Higher order functions: map
• Creates a new collection from an existing collection by applying a function
• Anonymous functionscala> List(1, 2, 3 ) map { (x: Int) => x + 1 }
res0: List[Int] = List(2, 3, 4)
• Function literalscala> List(1, 2, 3) map { _ + 1 }
res1: List[Int] = List(2, 3, 4)
• Passing an existing functionscala> def addOne(num: Int) = num + 1
addOne: (num: Int)Int
scala> List(1, 2, 3) map addOne
res2: List[Int] = List(2, 3, 4)
Page 21
21Intro to Spark: Intro to Scala | 7/9/2014
FP: Higher order functions: map
Page 22
22Intro to Spark: Intro to Scala | 7/9/2014
FP: Higher order functions: flatmap
Page 23
23Intro to Spark: Intro to Scala | 7/9/2014
FP: for-comprehension
• Syntax– for ( <generator> | <guard> ) <expression> [yield] <expression>
• Types– Imperative form. Does not return a value.
scala> val aList = List(1, 2, 3)
aList: List[Int] = List(1, 2, 3)
scala> val bList = List(4, 5, 6)
bList: List[Int] = List(4, 5, 6)
scala> for { a <- aList; if (a < 2); b <- bList; if (b < 7) } println( a + b )
5
6
7
Page 24
24Intro to Spark: Intro to Scala | 7/9/2014
FP: for-comprehension
• Syntax– for ( <generator> | <guard> ) <expression> [yield] <expression>
• Types– Functional form (a.k.a., sequence comprehension) . Returns/yields
a value
scala> for { a <- aList; b <- bList} yield a + b
res0: List[Int] = List(5, 6, 7, 6, 7, 8, 7, 8, 9)
scala> res0.take(1)
res1: List[Int] = List(5)
scala> for { a <- aList; if (a < 2); b <- bList } yield a + b
res2: List[Int] = List(5, 6, 7)
scala>
Page 25
25Intro to Spark: Intro to Scala | 7/9/2014
FP: for-comprehension
Page 26
26Intro to Spark: Intro to Scala | 7/9/2014
FP: foldLeft• scala> val numbers = 1.to(10)
• numbers: scala.collection.immutable.Range.Inclusive = Range(1, 2, 3, 4, 5, 6, 7, 8, 9, 10)
• scala> def add( a:Int, b:Int ): Int = { a + b }
• add: (a: Int, b: Int)Int
• scala> numbers.foldLeft(0){ add }
• res0: Int = 55
• scala> numbers.foldLeft(0){ (acc, b) => acc + b }
• res1: Int = 55
• scala>
Page 27
27Intro to Spark: Intro to Scala | 7/9/2014
FP: foldLeft
Page 28
28Intro to Spark: Intro to Scala | 7/9/2014
FP: find the last item in an array
• scala> val ns = Array(20, 40, 60)
• ns: Array[Int] = Array(20, 40, 60)
• scala> ns.foldLeft(ns.head) {(acc, b) => b}
• res0: Int = 60
• scala>
Page 29
29Intro to Spark: Intro to Scala | 7/9/2014
FP: reverse an array w/ foldLeft
• scala> val ns = Array(20, 40, 60)
• ns: Array[Int] = Array(20, 40, 60)
• scala> ns.foldLeft( Array[Int]() ) { (acc, b) => b +: acc}
• res1: Array[Int] = Array(60, 40, 20)
• scala>
Page 30
30Intro to Spark: Intro to Scala | 7/9/2014
FP: reverse an array w/ foldLeft
Page 31
31Intro to Spark: Intro to Scala | 7/9/2014
Outline• Spark
– Hadoop eco system
• Scala– Background
• Why Scala?– For the computer scientist
– For the Java / OO programmer
– For the Spark developer
– For the Big Data developer
– For the Big Data scientist / mathematician
– For the system architect
Page 32
32Intro to Spark: Intro to Scala | 7/9/2014
Scala for the Java / OO developer: • Interoperable w/ Java
• Case classes
• Mixins with traits
Page 33
33Intro to Spark: Intro to Scala | 7/9/2014
Scala for the Java / OO developer: • case class
– Implements equals(), hashCode(), toString()
– Can be used in Pattern Matching
Page 34
34Intro to Spark: Intro to Scala | 7/9/2014
Scala for the Java / OO developer: • http://
docs.oracle.com/javase/8/docs/api/java/util/stream/Stream.html
• map– <R> Stream<R> map(Function<? super T,? extends
R> mapper)Returns a stream consisting of the results of applying the given function to the elements of this stream.This is an intermediate operation.
• flatMap– <R> Stream<R> flatMap(Function<? super T,? extends Stream<?
extends R>> mapper)Returns a stream consisting of the results of replacing each element of this stream with the contents of a mapped stream produced by applying the provided mapping function to each element. Each mapped stream is closed after its contents have been placed into this stream. (If a mapped stream is null an empty stream is used, instead.)This is an intermediate operation.
`
Page 35
35Intro to Spark: Intro to Scala | 7/9/2014
Outline• Spark
– Hadoop eco system
• Scala– Background
• Why Scala?– For the computer scientist
– For the Java / OO programmer
– For the Spark developer
– For the Big Data developer
– For the Big Data scientist / mathematician
– For the system architect
Page 36
36Intro to Spark: Intro to Scala | 7/9/2014
Scala for the Spark developer• ResilientDistributedDataset (RDD)
• A Resilient Distributed Dataset (RDD), the basic abstraction in Spark. Represents an immutable, partitioned collection of elements that can be operated on in parallel. This class contains the basic operations available on all RDDs, such as map, filter, and persist.
• http://spark.apache.org/docs/latest/api/scala/index.html#org.apache.spark.rdd.RDD
Page 37
37Intro to Spark: Intro to Scala | 7/9/2014
Outline• Spark
– Hadoop eco system
• Scala– Background
• Why Scala?– For the computer scientist
– For the Java / OO programmer
– For the Spark developer
– For the Big Data developer
– For the Big Data scientist / mathematician
– For the system architect
Page 38
38Intro to Spark: Intro to Scala | 7/9/2014
Scala for the Big Data developer• Spark
– Programming API in Scala
– Implemented in Scala
• Scalding– Scala DSL on top of Cascading
– data processing API and processing query planner used for defining, sharing, and executing data-processing workflows
– Abstractions: tuples, pipes, source/sink taps
• Algebird
• Summingbird– Library that lets you write MapReduce programs that look like
native Scala or Java collection transformations
– Execute them on a number of well-known distributed MapReduce platforms, including Storm and Scalding.
Page 39
39Intro to Spark: Intro to Scala | 7/9/2014
Outline• Spark
– Hadoop eco system
• Scala– Background
• Why Scala?– For the computer scientist
– For the Java / OO programmer
– For the Hadoop/Spark developer
– For the Big Data developer
– For the Big Data scientist / mathematician
– For the system architect
Page 40
40Intro to Spark: Intro to Scala | 7/9/2014
Scala for the Big Data scientist / mathematician
• Monoid– If you want to “attach” operations such as +, -, *, / or <= to data
objects (e.g., Bloom filters), then you want to provide monoid forms of those data objects
– Consists of• A set of objects
• Binary operation that satisfies the monoid axioms
• Monad– If you want to create a data processing pipeline that transforms the
state of a data object
– composition
Page 41
41Intro to Spark: Intro to Scala | 7/9/2014
Outline• Spark
– Hadoop eco system
• Scala– Background
• Why Scala?– For the computer scientist
– For the Java / OO programmer
– For the Hadoop/Spark developer
– For the Big Data developer
– For the Big Data scientist / mathematician
– For the system architect
Page 42
42Intro to Spark: Intro to Scala | 7/9/2014
Scala for the system architect• Concurrency
• Problem:
– Threads
– Shared mutable state
– Locks,
• Solution:
– message passing concurrency w/ Actors
– Future, Promise
• Abstractions
– Actor
• an object that processes a message
• encapsulates state (state not shared)
– ActorRef
– Message, usually sent asynchronously
– Mailbox
– ActorSystem
Page 43
43Intro to Spark: Intro to Scala | 7/9/2014
Scala for the system architect: Akka• Fault tolerance
– Supervision
– Strategies• Resume, restart, stop, escalate, …
• Scale out: remote actors– Via configuration
Page 44
44Intro to Spark: Intro to Scala | 7/9/2014
Scala for the system architect• Parallel collections
– scala> import scala.collection.parallel.immutable._
– import scala.collection.parallel.immutable._
– scala> ParVector(10, 20, 30, 40, 50, 60, 70, 80, 90) .map { x =>
– | println( Thread.currentThread.getName); x / 2 }
– ForkJoinPool-1-worker-13
– ForkJoinPool-1-worker-1
– ForkJoinPool-1-worker-1
– ForkJoinPool-1-worker-9
– ForkJoinPool-1-worker-11
– ForkJoinPool-1-worker-5
– ForkJoinPool-1-worker-3
– ForkJoinPool-1-worker-15
– ForkJoinPool-1-worker-7
– res0: scala.collection.parallel.immutable.ParVector[Int] = ParVector(5, 10, 15,
– 20, 25, 30, 35, 40, 45)
– scala>
Page 45
45Intro to Spark: Intro to Scala | 7/9/2014
Sequential collections
Page 46
46Intro to Spark: Intro to Scala | 7/9/2014
Parallel collections
Page 47
47Intro to Spark: Intro to Scala | 7/9/2014
Outline• Spark
– Hadoop eco system
• Scala– Background
• Why Scala?– For the computer scientist
– For the Java / OO programmer
– For the Spark developer
– For the Big Data developer
– For the Big Data scientist / mathematician
– For the system architect
Page 48
48Intro to Spark: Intro to Scala | 7/9/2014
Different perspectives on an elephant Scala
Page 49
49Intro to Spark: Intro to Scala | 7/9/2014
Spark in the Hadoop ecosystem
Page 50
50Intro to Spark: Intro to Scala | 7/9/2014
References• http://scala-lang.org/
• Scala in Action, Nilanjan Raychaudhuri
• Grokking Functional Programming, Aslam Khan
• Michael Noll
Page 51
Intro to Apache Spark:Fast cluster computing engine for Hadoop
Intro to Scala:Object-oriented and functional language for the Java Virtual Machine
ACM SIGKDD, 7/9/2014
Roger Huang
Lead System Architect
Digital & Mobile Products Architecture
[email protected]
[email protected]