Saturday, December 15, 2012

Scala for Java 7 Developers: Overview

Target audience: Beginner
Estimated reading time: 10'


The purpose of this post is to introduce the Scala programming language from the perspective of an experienced Java developer. It is assumed that the reader has a good command of design patterns and multi-threaded applications.
In a nutshell, Scala is:
- Object-oriented
- Functional
- Scalable (Actors and immutability)

The elements of the Scala programming language presented in this post are a small subset of the rich set of features developers would benefit from the language.

Note: For the sake of readability of the implementation of algorithms, all non-essential code such as error checking, comments, exception, validation of class and method arguments, scoping qualifiers or import is omitted


Immutability
Scala relies on immutable object and collections to reduce complexity associated with multi-threaded applications, eliminate race condition, simplify debugging.
The CRU (Create, Read, Update) object lifecycle in Java is reduced to CR (Create, Read) in Scala.
The Scala standard library offers a large variety of immutable collections.

val x = 23.5  // Read only
val mp = new immutable.Map[String, String]
val xs = List[Any]()


Traits vs. interface
Java interfaces are purely declarative. Scala traits allows:
* method implementation
* member value initialization
* abstract and lazy values (to be initialized through sub-class or on demand)
Scala traits cannot be instantiated and do not have constructors. Scala emulates multiple inheritance by stacking traits (Cake pattern).

trait MyObject[T] {
  val tobeDefined: Int
  def add(t: T): Option[T] = {}
  def head: T = { }
}


Lazy values and methods
Scala supports lazy values that are declared as member of a class but instantiated once and only once when they are accessed

class MyObject[T] {
  lazy val n = (n<<2) + Random.next(n)
  ...
}


Implicits
3rd party libraries and frameworks cannot always be extended through composition or inheritance. Scala implicit class/types and methods allows a "behind the scene" type conversion. The following code snippet converts a double value into a vector/array of double of size 1. The implicit is automatically invoke whenever it is imported into another section of the code base: for instance, the vector v is constructed through the implicit conversion dbl2Vec.

type DblVector = Array[Double]
  // Definition of implicit
implicit def dbl2Vec(x: Double): DblVector = Array.fill(1)(x)
  ...
val v: DblVector = 4.5

The following example extends the concept of Exception handling, Try with a concept of Either converting any instance of Try to an instance of Either using toEither method of the implicit class Try2Either. The implicit class conversion is used in handling a rounding error in the normalize function.

implicit class Try2Either[T](_try: Try[T]) {
  def toEither[U](r: ()=>U)(f: T=>T): Either[U, T] = _try match {
    case Success(res) => Right(f(res))
    case Failure(e) => Console.println(e.toString); Left(r())  
  }
}

/**
 * Method fails and recover if the denominator is too small
*/
def normalize(x: DblVector): Either[Recover, DblVector] = Try {
  val sum = x.sum 
  if(sum > 0.01) x.map( _ / sum) 
  else throw new IllegalStateException("prob out of range")  
}
.toEither(() => new Recover)((v: DblVector) => v.map( Math.log(_)))


Tail recursion
Although convenient, traditional recursive implementation of computation can be ineffective, leaving developers to rely on iterative algorithms.
Scala's tail recursion is a very effective way to iterate through a data sequence without incurring a performance overhead because it avoids the creation of new stack frames during the recursion.
The following code snippet implements the Viterbi algorithm used to extract the most likely sequence of latent states in an hidden Markov model given a lambda model and a set of observations.

@scala.annotation.tailrec
private def viterbi(t: Int): Double = {

  if( t == 0) initial   
  else { 
    Range(0, lambda.getN).foreach( updateMaxDelta(t, _) )   
    if( t ==  obs.size-1) {
        val idxMaxDelta = Range(0, lambda.getN).map(i => 
                    (i, state.delta(t, i))).maxBy(_._2)
        state.QStar.update(t+1, idxMaxDelta._1)
        idxMaxDelta._2
    }
    else viterbi(t+1)
  }
}


Actor model
Traditional multi-threaded applications rely on accessing data located in shared memory. The mechanism relies on synchronization monitors such as locks, mutexes or semaphores to avoid deadlocks and inconsistent mutable states. Those applications are difficult to debug because of race condition and incur the cost of a large number of context switches.
The Actor model addresses those issues by using immutable data structures (messages) and asynchronous (non-blocking) communication.
The actors exchange messages and data using immutable messages as described in the example below.

sealed abstract class Message(val id: Int)
case class Activate(i: Int, xt: Array[Double]) extends Message(i)
case class Completed(i: Int, xt: Array[Double]) extends Message(i)
case class Start(i: Int =0) extends Message(i)

An actor processes messages queued in its mailbox asynchronously, FIFO. The method, receive which returns a PartialFunction is the message processing loop, which correspond to the Runnable.run() in Java

final class Worker(
     id: Int, 
     fct: DblSeries => DblSeries) 
extends Actor {
    // Actor event/message loop
  override def receive = {
    case msg: Activate => {
      val msgId = msg.id+id
      val output = fct(msg.xt)
      sender ! Completed(msgId, output)
  }
}


Syntactic similarities

Java developers will find comfort knowing that Scala preserves a lot of constructs and idioms which make Java 7 such a wide spread and supported programming language.
Construct ScalaJava 7
Variable         var x: Float = 1.0F  float x = 1.0F;
Value val n = 4L
Constants final val MAX_TEMP: Float = 232.0F final float MAX_TEMP = 232.0F;
Class class Person { ...} public class Person { }
Constructor class Person(name: String, age: Int) {} public class Person {
....private String name;
....private int age;
....public Person(String name, int age) {}
}
Setters Not applicable
Immutability
public class Person {
....public void setName(String name) {}
}
Getters class Person(val name: String, val age: Int{}
....//Public access to member values.
public class Person {
....public final String getName() {}
}
Method Arguments validation def add(index: Int): Unit = {
....require(index>=0 && index<size, .... s"$index out of bounds")
...
public void add(int index) {
....if( index < 0 || index >= size) {
........ throw new IllegalArgumentException(...);
....
Singleton object Scientist extends Person {} public class scientist extends Person {
.... private static Scientist instance = new Scientist();
.... private Scientist();
.... public static Scientist getInstance() {
........ return instance;
... }
}
Overriding override def getName(id : Int) : String {.. } @override
public final getName(int id) { ... }
Interfaces trait Searchable {
.....def search(id : Int) : String
}
interface Searchable {
....String search(int id);
}
Type Casting n : Int = o.asInstanceOf[Int] Integer n = (Integer)o;
Type Checking n.isInstanceOf[Int] n instanceOf Integer
Generics class Stack[T](val maxSize : Int) {
....def push(t : T) : Unit = { ... }
....def pop : T = { ... }
}
public class Stack <T> {
........private int maxSize;
........public Stack(int maxSize) { ... }
........public void push(T t) {.... }
........public <T> pop() {... }
}
Abstract Class abstract class Collection[T](val max : Int) {
....def add(t : T) : Unit
....def get(index : Int) : Int
}
abstract public class Collection <T> {
........abstract public void add(T t);
........abstract public <T> get(int index);
}
Exception Try { .... }
match { 
....case Success(res) =>{}
... case Failure(e) => {}
}
try { .. }
catch( NullPointerException ex) { ..}
catch( Exception ex) { ...}
Bounded type
Parameters
class List[T <: Float] {  ... }
class List[T >: Node] { ...}
public class List<T extend Float> { ... }
public class List <T super Node> { .... }
Wildcard
Parameters
class List[+T] {}
class List[-Node] { ...}
public class List<? extends T> { ... }
public class List<? super Node> { .... }
Index-based
For loop
for( i <= 0 until n) { .. } for( int i = 0; i < n; i++) { .. }
Iteration
Collection
for(index <- indices) { ... }for(int index : indices) { .. }


Key differences

There are quite a few important and tricky programmatic difference between Scala 2.11 and Java 7
  • == In Scala == compares instances similarly to equals. In Java == compares the reference to instances
  • Comparable integer java.lang.Integer supports comparable interface while scala.Int does not
  • Static method Java supports explicit static methods. Scala does not, although static methods are implicitly defined as method of objects (Singletons)
  • Inner class In Scala, Inner class are accessed through the reference Self to the outer class. Java accesses Inner class through the Outer class
  • Checked exceptions Scala does not support Checked exceptions, Java does.
  • Primitive types As a pure object oriented, Scala does not support primitive types. Java does
  • Expression as a value As a functional language, Scala defines expressions such as if ... else ... as value, that can be evaluated and assigned to. Java does not
  • Immutable collections Scala supports immutable collections by default. Java does not.
  • Localized imports and type aliases Scala has scoped import and type aliases. Java does not
  • Explicit operator overloading Scala allows developers to overload operator through the definition of new methods. Java does not.
  • Actor Scala provides full parallelism with actor and immutable messages. Java does not.
  • Futures Scala futures are asynchronous and can be updated thorough promises. Java does not support the concept
  • Higher order kind Scala has higher order kind, such as Monad. Java does not
  • Tail recursion (or elimination) Scala supports tail recursion in order to avoid stack overflow. Java does not
  • Monad and functors are core concept in Scala, but not in Java.

Note This comparative analysis relies on version 2.11.x of Scala language and Java 7. Future version of Scala and Java may produce a slightly different result.


So much more....
This post barely scratches the surface of the capabilities of Scala which include
  • Futures
  • Partial functions and partially applied functions
  • Dependency injection
  • Monadic representation
  • Macros
  • View bounded type parameterization
  • F-Bounded polymorphism
  • Closures
  • Delimited continuation
  • Method argument currying

References

Monday, November 12, 2012

Lazy Evaluation in C++ & Scala

Target audience: Intermediate
Estimated reading time: 15'

Overview

Lazy evaluation is a not a new concept in software engineering. The concept is rather simple: initialize a variable or an object once it is needed. A typical use case is the instantiating of a large object in memory from the content of database: the lifecycle of the object is shorter than the application so it is convenient to allocate the memory when and if needed.
Both C++ and Scala support lazy evaluation using two very different approach: C++ relies on a class constructor to initialize iteratively its attribute member of a class while Scala supports lazy initialization natively.

Note: For the sake of readability of the implementation of algorithms, all non-essential code such as error checking, comments, exception, validation of class and method arguments, scoping qualifiers or import is omitted 


C++
Lazy evaluation or fetching can be implemented in C++ using a design pattern. Let consider a large object which state is stored in a remote database or initialized through a request REST API. In many cases, only some of the attributes of the object need to be initialized from the remote data source (table for instance): load the content of the entire table would not make much sense.
How C++ address the problem?
Let's consider a class that contains all the information regarding the visit of a user in a user . 
 
class SiteVisit {
private:
  int  id;
  mutable string*  url;
  mutable long  visitDate;
  mutable std::list* targets;
  mutable char[4] sources; 
  void loadFromDB(std::list*)

public:
  SiteVisit(int id) : id(id), url(null), visitDate(-1L), targets(null), sources(null) { }    
  const string& getName() const;
  long          getVisitDate() const;
  std::list&    getTargets() const;
 const char[4]& getSources();
}


It is conceivable that not all the data describing the visit has to be retrieved for analysis or reporting. For instance, there is no need to load the targeted URLs, for all the visits from the database until a report needs to be created. In this case, the role of the constructor in C++ is to initialize the unique identifier for the object (or table unique primary key) and nothing else. The getTargets method retrieves the list of pages visited at a later stage in the application lifecycle. Therefore, this list is generated only once; the first time it is invoked. The method retrieveAllVisitedPages, iterates through all the pages of the site that have been visited.

std::list<string> SiteVisit::getTargets() {
  if( targets == null) {
     targets = new std::list
     loadTargetsFromDB(targets);
  }
  return targets;
}   

int retrieveAllVisitedPages(std::list&amp; visits, std::map* visitsMap) const {                           
  std::list::const_iterator visitsIt;
  std::list::const_iterator urlsIt;

  int counter = 0, totalPages = 0;
  for( visitsIt =visits.beging()l visitsIt !=visits.end(); visitsIt++) {
    std::list* urls = visitsIt.getTargets();   
    for( urlsIt = urls.begin(); urlsIt != urls.end(); urlsIt++) {
       counter = visitsMap[urlsIt];
       visitsMap-&gt;insert(std::make_pair(urlsIt, ++counter));
    }           
    totalPages += urls.size();
  }
  return totalPages;
}

Although feasible, the implementation of the lazy initialization is quite cumbersome as it involves at a minimum, a constructor and a method (2 step initialization


Scala
Scala supports lazy values and lazy local functions. Under some circumstances, the user may want to initialize a value only when it is required. A database developer would qualify this pattern as a write once, read many times pattern.
Note: The developer needs to keep in mind that lazy values or expressions are checked whether the value has been indeed already initialized, each time it is accessed. Such validation is to be thread-safe, and consequently adds overhead to the run-time execution.

Let's evaluate or experience the concept by writing a simple example in Scala. The SiteVisits class implements the basic function to retrieve the list of URLs/pages a user identified by its id access on a specific web site.

class SiteVisits(userId: Int, webSite: String) {
  lazy val urls =  {
    val urlsList = new ArrayBuffer[String]
   
   // Initialize the urls list from the database
    query(userId, webSite)
    urlsList.toArray
  }
  ... 
  def isValidWebsite; Boolean = ....
}

The list of URLs is not loaded from the database when an Instance of SiteVisits is created but when the value urls is read by the client code. For example, the methods of the class SiteVisits such as isValidWebsite can be executed without having to access the list of URLs the user visited on this particular web site. Let's look how lazy evaluation can be applied.

val siteVisit = new SiteVisits(91841, "www.yahoo.com")
  ... 
  if( siteVisit.isValidWebSite )  
     Console.println(siteVisit.urls.mkString(", ") 
  ..
}

In the example above, the list of URLS is not loaded if the site is not valid.


References
  • Effective C++   - S. Meyers  - Addison Wesley 1991
  • More Effective C++  - S. Meyers  - Addison Wesley 1993
  • Programming in Scala  - M. Odersky, L. Spoon, B. Venners  - Artima 2007
  • https://github.com/prnicolas

Friday, October 5, 2012

Type Class Polymorphism and Monads

Target audience: Advanced
Estimated reading time: 45'

Overview

Most of software developers associate the concept of polymorphism to dynamic binding & class inheritance available in the most common object oriented programming languages such as C++ and Java. As part of its functional programming "DNA", Scala provides developers with  alternative polymorphism approaches (beside the ubiquitous method override).

Note: For the sake of readability of the implementation of algorithms, all non-essential code such as error checking, comments, exception, validation of class and method arguments, scoping qualifiers or import is omitted 


Object-Oriented Polymorphism

Java or C++ programmers are familiar with this implementation of polymorphism as described in the example below. A trait the key behavior of a collection by declaring two method += to add a new item into the collection and -= to retrieve the first item from the collection.

trait Collection[T <: Double] {                           
   def += (t:T) : Boolean
   def -= : T 
}

The class, OrderedCollection implements ordered collection of type inherited from Double by defining the two abstract methods += & -=. This class specializes the behavior of the trait by ordering the elements in the collection.

class OrderedCollection[T <: Double] extends Collection[T] {
   import scala.collection.mutable.TreeSet
 
   var c = TreeSet.empty(Ordering.fromLessThan[T]((x:Double, y:Double) => (x < y)))
   override def += (t:T) : Boolean = c.add(t)
   override def -= : T = { 
     val top = c.head
     c.remove(top)
     top 
   }
}

The following code snippet illustrate a simple use case for the ordered collection. Note the declaration of the type of the collection instance, thisCollection is not required and added for the sake of clarity.

val thisCollection: Collection[Double] = new OrderedCollection[Double]
thisCollection += 9.556
thisCollection += 45.7
thisCollection += 4.9
println(thisCollection)


Type Class Polymorphism 
Scala supports Type class polymorphism, also referred as Ah-hoc polymorphism in technical articles. Scala also supports Ah-Doc polymorphism which the ability to create and apply generic polymorphic operators or functions to arguments of different types. Ad hoc polymorphism lets you add features to a type on demand.
Let's define a Wrapper type (trait) which basic purpose is to wrap an element into a collection M[_] of this element. The Wrapper type class declares the operators apply and get that is to be implemented by wrapper associated to a predetermined collection.
- apply creates a collection M[T] of element of type T.
- get retrieves the first and only element of the collection.

Note: The type of the element in the collection is not a parameter type for the Wrapper: T is the type parameter for the collection and type M[_] is the parameter type for the wrapper. The Wrapper trait is defined as:

trait Container[M[_]] {
  def apply[T](t: T): M[T]
  def get[T](c: M[T]): T
}

We can create a type class ArrayBuffer to "wrap" any object by overriding the apply and get operators. The implicit conversion from ArrayBuffer to Wrapper is obviously independent from the type T of the contained element (validation of method argument is omitted).

implicit val arrayBuffer = new Container[ArrayBuffer] {
  override def apply[T](t: T) : ArrayBuffer[T] =  ArrayBuffer[T](t)
  override def get[T](ab: ArrayBuffer[T]) : T = ab.head 
}

def compose[M[_]: Container, X, Y](c1: M[X], c2: M[Y]) = {
   val c = implicitly[Container[M]]
   c(ab.get(c1), ab.get(c2))
}


Monads
One of the most known application of type class polymorphism is the monadic representation of collection. The concept and theoretical foundation of monads is discussed in the previous post. Monads define the minimum set of operations for a functional category of objects. Scala standard library includes quite a few of categories and their associated monadic representation: List[T], Option[T], ..
Let's extend our simple Wrapper to define a Monad for collection containing a single element. The monad must define a constructor, apply, a functor, map that converts a single element collection into a another single element collection by transforming its only element and a flatMap that generate a new single element collection from the element of an existing collection.

trait _Monad[M[_]] {
  def apply[T](t: T): M[T]
  def map[T, U](m: M[_])(f: T => U): M[U]
  def flatMap[T, U](m: M[_])(f: T => M[U]): M[U]
}

As with the trait wrapper we can create a implicit conversion from any single element ArrayBuffer into its associated monad to monadic operators can be used transparently by the client code.

implicit val mArrayBuffer = new _Monad[ArrayBuffer] {
  override def apply[T](t:T): ArrayBuffer[T] =  ArrayBuffer[T](t)
  override def map[T, U](m: ArrayBuffer[T])(f: T => U): ArrayBuffer[U] = ArrayBuffer[U](f(m.head))
  override def flatMap[T, U](m: ArrayBuffer[T])(f: T => ArrayBuffer[U]): ArrayBuffer[U] = f(m.head) 
}

Monads will be discussed in length in a future post.


References
* An Overview of the Scala Programming Language - 2nd Edition M. Odersky, P. Altherr, V. Cremet, I. Dragos, G. Dubochet, B. Emir, S. McDirmid, S. Micheloud, N. Mihaylov, M. Schinz, E. Stenman, L. Spoon, M. Zenger
* Programming in Scala - 2nd Edition M. Odersky, L. Spoon, B. Venners - Artima 2011
* Fabulous adventures in coding - Monads, part one E. Lippert's - 2013
* github.com/prnicolas