Map, Flatmap and Reduce are Your New Best Friends: Simpler Collections, Concurrency, and Big Data (#oscon)
Post on 27-Aug-2014
806 Views
Preview:
DESCRIPTION
Transcript
Map(), flatMap() and reduce() are your new best friends:
simpler collections, concurrency, and big data
Chris Richardson
Author of POJOs in ActionFounder of the original CloudFoundry.com
@crichardsonchris@chrisrichardson.nethttp://plainoldobjects.com
@crichardson
Presentation goal
How functional programming simplifies your code
Show that map(), flatMap() and reduce()
are remarkably versatile functions
@crichardson
About Chris
@crichardson
About Chris
Founder of a buzzword compliant (stealthy, social, mobile, big data, machine learning, ...) startup
Consultant helping organizations improve how they architect and deploy applications using cloud, micro services, polyglot applications, NoSQL, ...
@crichardson
Agenda
Why functional programming?
Simplifying collection processing
Simplifying concurrency with Futures and Rx Observables
Tackling big data problems with functional programming
@crichardson
Functional programming is a programming paradigm
Functions are the building blocks of the application
Best done in a functional programming language
@crichardson
Functions as first class citizens
Assign functions to variables
Store functions in fields
Use and write higher-order functions:
Pass functions as arguments
Return functions as values
@crichardson
Avoids mutable state
Use:
Immutable data structures
Single assignment variables
Some functional languages such as Haskell don’t allow side-effects
@crichardson
Why functional programming?
"the highest goal of programming-language design to enable good ideas to be elegantly expressed"
http://en.wikipedia.org/wiki/Tony_Hoare
@crichardson
Why functional programming?More expressive
More intuitive - declarative code matches problem definition
Functional code is usually much more composable
Immutable state:
Less error-prone
Easy parallelization and concurrency
But be pragmatic
@crichardson
An ancient idea that has recently become popular
@crichardson
Mathematical foundation:
λ-calculus
Introduced byAlonzo Church in the 1930s
@crichardson
Lisp = an early functional language invented in 1958
http://en.wikipedia.org/wiki/Lisp_(programming_language)
1940
1950
1960
1970
1980
1990
2000
2010
garbage collection dynamic typing
self-hosting compiler tree data structures
(defun factorial (n) (if (<= n 1) 1 (* n (factorial (- n 1)))))
@crichardson
My final year project in 1985: Implementing SASL
sieve (p:xs) = p : sieve [x | x <- xs, rem x p > 0];
primes = sieve [2..]
A list of integers starting with 2
Filter out multiples of p
Mostly an Ivory Tower technology
Lisp was used for AI
FP languages: Miranda, ML, Haskell, ...
“Side-effects kills kittens and puppies”
@crichardson
http://steve-yegge.blogspot.com/2010/12/haskell-researchers-announce-discovery.html
!*
!*
!*
@crichardson
But today FP is mainstreamClojure - a dialect of Lisp
A hybrid OO/functional language
A hybrid OO/FP language for .NET
Java 8 has lambda expressions
@crichardson
Java 8 lambda expressions are functions
x -> x * x
x -> { for (int i = 2; i < Math.sqrt(x); i = i + 1) { if (x % i == 0) return false; } return true; };
(x, y) -> x * x + y * y
An instance of an anonymous inner class that implements a functional interface (kinda)
@crichardson
Agenda
Why functional programming?
Simplifying collection processing
Simplifying concurrency with Futures and Rx Observables
Tackling big data problems with functional programming
@crichardson
Lot’s of application code=
collection processing:
Mapping, filtering, and reducing
@crichardson
Social network examplepublic class Person {
enum Gender { MALE, FEMALE }
private Name name; private LocalDate birthday; private Gender gender; private Hometown hometown;
private Set<Friend> friends = new HashSet<Friend>(); ....
public class Friend {
private Person friend; private LocalDate becameFriends; ...}
public class SocialNetwork { private Set<Person> people; ...
@crichardson
Typical iterative code - e.g. filteringpublic class SocialNetwork {
private Set<Person> people;
...
public Set<Person> lonelyPeople() { Set<Person> result = new HashSet<Person>(); for (Person p : people) { if (p.getFriends().isEmpty()) result.add(p); } return result; }
Declare result variable
Modify result
Return result
Iterate
@crichardson
Problems with this style of programming
Low level
Imperative (how to do it) NOT declarative (what to do)
Verbose
Mutable variables are potentially error prone
Difficult to parallelize
@crichardson
Java 8 streams to the rescue
A sequence of elements
“Wrapper” around a collection (and other types: e.g. JarFile.stream(), Files.lines())
Streams can also be infinite
Provides a functional/lambda-based API for transforming, filtering and aggregating elements
Much simpler, cleaner and declarative code
@crichardson
public class SocialNetwork {
private Set<Person> people;
...
public Set<Person> peopleWithNoFriends() { Set<Person> result = new HashSet<Person>(); for (Person p : people) { if (p.getFriends().isEmpty()) result.add(p); } return result; }
Using Java 8 streams - filteringpublic class SocialNetwork {
private Set<Person> people;
...
public Set<Person> lonelyPeople() { return people.stream()
.filter(p -> p.getFriends().isEmpty())
.collect(Collectors.toSet()); }
predicate lambda expression
@crichardson
The filter() function
s1 a b c d e ...
s2 a c d ...
s2 = s1.filter(f)
Elements that satisfy predicate f
@crichardson
Using Java 8 streams - mapping
class Person ..
private Set<Friend> friends = ...;
public Set<Hometown> hometownsOfFriends() { return friends.stream() .map(f -> f.getPerson().getHometown()) .collect(Collectors.toSet()); }
@crichardson
The map() function
s1 a b c d e ...
s2 f(a) f(b) f(c) f(d) f(e) ...
s2 = s1.map(f)
@crichardson
Using Java 8 streams - friend of friends using flatMap
class Person ..
public Set<Person> friendOfFriends() { return friends.stream() .flatMap(friend -> friend.getPerson().friends.stream()) .map(Friend::getPerson) .filter(f -> f != this) .collect(Collectors.toSet()); }
maps and flattens
@crichardson
The flatMap() function
s1 a b ...
s2 f(a)0 f(a)1 f(b)0 f(b)1 f(b)2 ...
s2 = s1.flatMap(f)
@crichardson
Using Java 8 streams - reducingpublic class SocialNetwork {
private Set<Person> people;
...
public long averageNumberOfFriends() { return people.stream() .map ( p -> p.getFriends().size() ) .reduce(0, (x, y) -> x + y) / people.size(); } int x = 0;
for (int y : inputStream) x = x + yreturn x;
@crichardson
The reduce() function
s1 a b c d e ...
x = s1.reduce(initial, f)
f(f(f(f(f(f(initial, a), b), c), d), e), ...)
@crichardson
Adopting FP with Java 8 is straightforward
Simply start using streams and lambdasEclipse can refactor anonymous inner classes to lambdas
@crichardson
Agenda
Why functional programming?
Simplifying collection processing
Simplifying concurrency with Futures and Rx Observables
Tackling big data problems with functional programming
@crichardson
Let’s imagine that you are writing code to display the
products in a user’s wish list
@crichardson
The need for concurrencyStep #1
Web service request to get the user profile including wish list (list of product Ids)
Step #2
For each productId: web service request to get product info
But
Getting products sequentially ⇒ terrible response time
Need fetch productInfo concurrentlyComposing sequential + scatter/gather-style
operations is very common
@crichardson
Futures are a great abstraction for composing concurrent operations
http://en.wikipedia.org/wiki/Futures_and_promises
@crichardson
Worker thread or event-driven code
Main thread
Composition with futures
Outcome
Future 2
Clientget Asynchronous
operation 2
set
initiates
Asynchronous operation 1
Outcome
Future 1
getset
@crichardson
But composition with basic futures is difficult
Java 7 future.get([timeout]):
Blocking API ⇒ client blocks thread
Difficult to compose multiple concurrent operations
Futures with callbacks:
e.g. Guava ListenableFutures, Spring 4 ListenableFuture
Attach callbacks to all futures and asynchronously consume outcomes
But callback-based code = messy code
See http://techblog.netflix.com/2013/02/rxjava-netflix-api.html
We need functional futures!
@crichardson
Functional futures - Scala, Java 8 CompletableFuture
def asyncPlus(x : Int, y : Int) : Future[Int] = ... x + y ...
val future2 = asyncPlus(4, 5).map{ _ * 3 }
assertEquals(27, Await.result(future2, 1 second))
Asynchronously transforms future
def asyncSquare(x : Int) : Future[Int] = ... x * x ...
val f2 = asyncPlus(5, 8).flatMap { x => asyncSquare(x) }
assertEquals(169, Await.result(f2, 1 second))
Calls asyncSquare() with the eventual outcome of
asyncPlus()
@crichardson
Functions like map() are asynchronous
someFn(outcome1)
f2
f2 = f1 map (someFn) Outcome1
f1
Implemented using callbacks
@crichardson
class WishListService(...) { def getWishList(userId : Long) : Future[WishList] = {
userService.getUserProfile(userId).
Scala wish list service
Java 8 Completable Futures let you write similar code
Future[UserProfile]
map { userProfile => userProfile.wishListProductIds}. flatMap { productIds => val listOfProductFutures = productIds map productInfoService.getProductInfo Future.sequence(listOfProductFutures) }. map { products => WishList(products) }
Future[List[Long]]
List[Future[ProductInfo]]
Future[List[ProductInfo]]
Future[WishList]
@crichardson
Your mouse is your database
Erik Meijer
http://queue.acm.org/detail.cfm?id=2169076
@crichardson
Introducing Reactive Extensions (Rx)
The Reactive Extensions (Rx) is a library for composing asynchronous and event-based programs using observable sequences and LINQ-style query operators. Using Rx, developers represent asynchronous data streams
with Observables , query asynchronous data streams using LINQ operators , and .....
https://rx.codeplex.com/
@crichardson
About RxJava
Reactive Extensions (Rx) for the JVM
Original motivation for Netflix was to provide rich Futures
Implemented in Java
Adaptors for Scala, Groovy and Clojure
Embraced by Akka and Spring Reactor: http://www.reactive-streams.org/
https://github.com/Netflix/RxJava
@crichardson
RxJava core concepts
trait Observable[T] { def subscribe(observer : Observer[T]) : Subscription ...}
trait Observer[T] {def onNext(value : T)def onCompleted()def onError(e : Throwable)
}
Notifies
An asynchronous stream of items
Used to unsubscribe
Comparing Observable to...Observer pattern - similar but adds
Observer.onComplete()
Observer.onError()
Iterator pattern - mirror image
Push rather than pull
Futures - similar
Can be used as Futures
But Observables = a stream of multiple values
Collections and Streams - similar
Functional API supporting map(), flatMap(), ...
But Observables are asynchronous
@crichardson
Fun with observables
val every10Seconds = Observable.interval(10 seconds)
-1 0 1 ...
t=0 t=10 t=20 ...
val oneItem = Observable.items(-1L)
val ticker = oneItem ++ every10Seconds
val subscription = ticker.subscribe { (value: Long) => println("value=" + value) }...subscription.unsubscribe()
@crichardson
def getTableStatus(tableName: String) : Observable[DynamoDbStatus]=
Observable { subscriber: Subscriber[DynamoDbStatus] =>
}
Connecting observables to the outside world
amazonDynamoDBAsyncClient.describeTableAsync(new DescribeTableRequest(tableName), new AsyncHandler[DescribeTableRequest, DescribeTableResult] {
override def onSuccess(request: DescribeTableRequest, result: DescribeTableResult) = { subscriber.onNext(DynamoDbStatus(result.getTable.getTableStatus)) subscriber.onCompleted() }
override def onError(exception: Exception) = exception match { case t: ResourceNotFoundException => subscriber.onNext(DynamoDbStatus("NOT_FOUND")) subscriber.onCompleted() case _ => subscriber.onError(exception) } }) }
Called once per subscriber
Asynchronously gets information about DynamoDB table
@crichardson
Transforming observables
val tableStatus : Observable[DynamoDbMessage] = ticker.flatMap { i => logger.info("{}th describe table", i + 1) getTableStatus(name) }
Status1 Status2 Status3 ...
t=0 t=10 t=20 ...
+ Usual collection methods: map(), filter(), take(), drop(), ...
@crichardson
Calculating rolling averageclass AverageTradePriceCalculator {
def calculateAverages(trades: Observable[Trade]): Observable[AveragePrice] = { ... }
case class Trade( symbol : String, price : Double, quantity : Int ...)
case class AveragePrice(symbol : String, price : Double, ...)
@crichardson
Calculating average pricesdef calculateAverages(trades: Observable[Trade]): Observable[AveragePrice] = {
trades.groupBy(_.symbol). map { case (symbol, tradesForSymbol) => val openingEverySecond =
Observable.items(-1L) ++ Observable.interval(1 seconds) def closingAfterSixSeconds(opening: Any) =
Observable.interval(6 seconds).take(1)
tradesForSymbol.window(openingEverySecond, closingAfterSixSeconds _).map { windowOfTradesForSymbol => windowOfTradesForSymbol.fold((0.0, 0, List[Double]())) { (soFar, trade) => val (sum, count, prices) = soFar (sum + trade.price, count + trade.quantity, trade.price +: prices) } map { case (sum, length, prices) => AveragePrice(symbol, sum / length, prices) } }.flatten }.flatten }
Create an Observable of per-symbol Observables
Create an Observable of per-symbol Observables
@crichardson
Agenda
Why functional programming?
Simplifying collection processing
Simplifying concurrency with Futures and Rx Observables
Tackling big data problems with functional programming
@crichardson
Let’s imagine that you want to count word frequencies
@crichardson
Scala Word Count
val frequency : Map[String, Int] = Source.fromFile("gettysburgaddress.txt").getLines() .flatMap { _.split(" ") }.toList
frequency("THE") should be(11)frequency("LIBERTY") should be(1)
.groupBy(identity) .mapValues(_.length))
Map
Reduce
@crichardson
But how to scale to a cluster of machines?
@crichardson
Apache HadoopOpen-source software for reliable, scalable, distributed computing
Hadoop Distributed File System (HDFS)
Efficiently stores very large amounts of data
Files are partitioned and replicated across multiple machines
Hadoop MapReduce
Batch processing system
Provides plumbing for writing distributed jobs
Handles failures
...
@crichardson
Overview of MapReduceInputData
Mapper
Mapper
Mapper
Reducer
Reducer
Reducer
OutputDataShuffle
(K,V)
(K,V)
(K,V)
(K,V)*
(K,V)*
(K,V)*
(K1,V, ....)*
(K2,V, ....)*
(K3,V, ....)*
(K,V)
(K,V)
(K,V)
@crichardson
MapReduce Word count - mapperclass Map extends Mapper<LongWritable, Text, Text, IntWritable> { private final static IntWritable one = new IntWritable(1); private Text word = new Text(); public void map(LongWritable key, Text value, Context context) { String line = value.toString(); StringTokenizer tokenizer = new StringTokenizer(line); while (tokenizer.hasMoreTokens()) { word.set(tokenizer.nextToken()); context.write(word, one); } }}
(“Four”, 1), (“score”, 1), (“and”, 1), (“seven”, 1), ...
Four score and seven years⇒
http://wiki.apache.org/hadoop/WordCount
@crichardson
Hadoop then shuffles the key-value pairs...
@crichardson
MapReduce Word count - reducer class Reduce extends Reducer<Text, IntWritable, Text, IntWritable> {
public void reduce(Text key, Iterable<IntWritable> values, Context context) { int sum = 0; for (IntWritable val : values) { sum += val.get(); } context.write(key, new IntWritable(sum)); } }
(“the”, 11)
(“the”, (1, 1, 1, 1, 1, 1, ...))⇒
http://wiki.apache.org/hadoop/WordCount
@crichardson
About MapReduceVery simple programming abstract yet incredibly powerful
By chaining together multiple map/reduce jobs you can process very large amounts of data in interesting ways
e.g. Apache Mahout for machine learning
But
Mappers and Reducers = verbose code
Development is challenging, e.g. unit testing is difficult
It’s disk-based, batch processing ⇒ slow
@crichardson
Scalding: Scala DSL for MapReduceclass WordCountJob(args : Args) extends Job(args) { TextLine( args("input") ) .flatMap('line -> 'word) { line : String => tokenize(line) } .groupBy('word) { _.size } .write( Tsv( args("output") ) )
def tokenize(text : String) : Array[String] = { text.toLowerCase.replaceAll("[^a-zA-Z0-9\\s]", "") .split("\\s+") }}
https://github.com/twitter/scalding
Expressive and unit testable
Each row is a map of named fields
@crichardson
Apache SparkPart of the Hadoop ecosystem
Key abstraction = Resilient Distributed Datasets (RDD)
Collection that is partitioned across cluster members
Operations are parallelized
Created from either a Scala collection or a Hadoop supported datasource - HDFS, S3 etc
Can be cached in-memory for super-fast performance
Can be replicated for fault-tolerance
REPL for executing ad hoc queries
http://spark.apache.org
@crichardson
Spark Word Countval sc = new SparkContext(...)
sc.textFile("s3n://mybucket/...") .flatMap { _.split(" ")} .groupBy(identity) .mapValues(_.length) .toArray.toMap }}
Expressive, unit testable and very fast
@crichardson
Summary
Functional programming enables the elegant expression of good ideas in a wide variety of domains
map(), flatMap() and reduce() are remarkably versatile higher-order functions
Use FP and OOP together
Java 8 has taken a good first step towards supporting FP
@crichardson
Questions?
@crichardson chris@chrisrichardson.net
http://plainoldobjects.com
top related