Scala Type Inference and Static Linking - An Accident Waiting to Happen?

Type inference in Scala is great. Anything that means I have to type less code is great! I'll also openly admit that I'm a die hard fan of static typing. When I think dynamic about typing, all I can think about is an object being validly sent to a block of code that was never meant to handle that object, and the compiler thinking that's okay. For me, that's just one more crack for bugs to slip through that our team doesn't need.

So, basically, I want it both ways - I want type security, but I don't want to have to specify it. But could this cause problems? If my answer was no, this would be a rather pointless blog entry.

An issue that comes to bear with Java and Scala is that along with the static typing comes static linking. I'm going to admit that I'm nowhere near enough of a language theorist to be able to tell you that these two are intrinsically paired, but I see signs that this is the case: Static typing means the compiler will check that the type of every expression is guaranteed to be the same type, or at least assignable to the same type, as the location to which you apply the result of the expression, be it a variable, a method parameter or part of another expression. Static linking is the way in which the compiler records in byte code which methods it will call. Where there are overloaded methods on the called object, the compiler will choose which of those overloaded method to call at compile time, based on the static type of the method parameters, rather than letting the JVM decide at run time which method to call based on the actual type of the parameter. So, the way I see it, if you didn't have static linking, then the static typing would only provide guarantees at compile time and not run time.

Here's a Scala example of the effect that static linking has on a program:

object StaticLinkingExample {
  def main(args : Array[String]) {
    val stringReference : String = "foo";
    val objectReference : Object = stringReference;
    printit(stringReference);
    printit(objectReference);
  }

  def printit(value : String) { println("String version called") }
  def printit(value : Object) { println("Object version called !!!") }
}

The output of this application is:

String version called
Object version called !!!

Notice that, even though 'objectReference' is clearly a String, the printit(Object) method is called. Why? Because the static type of objectReference is Object, and that's the only information that the compiler uses when deciding which method to link. This is static linking in action.

What has this got to do with type inference? Well, the thing about type inference is that, even though you generally don't have to specify the type that values have or that methods return, the type that is used is still very important to the JVM, because of static linking.

The potential for trouble comes to a head when you start changing code, because when you change code that's making use of type inference, you can very easily change the type of a variable or method without even realising. In Java, if you changed the type, you would realise straight away, because your compiler would tell you that the resulting type doesn't match your specified type. With type inference, this doesn't happen - the compiler will just invisibly change the type of the member for you.

Below is an example of how the maintenance of code using type inference could cause errors in your application. Imagine that I have several applications and each of my applications make use of a Foo, so I have all my Foo-related classes in a separate project (and JAR) which is used by each of my applications. As a classically-trained GoF Pattern programmer, I make sure all my applications get their Foo from a FooFactory. Here's the first version of my FooFactory:

trait Foo
class StandardFoo extends Foo

object FooFactory {
  def newFoo() = new StandardFoo
}

Note that the type of newFoo() is not declared, but is inferred by Scala.

Now here's one of my applications that makes use of the FooFactory:

object FooApplication {
  def main(args : Array[String]) {
    val result : Foo = FooFactory.newFoo()
    println(result);
  }
}

I run my application, and the result is as expected:

StandardFoo@7ca3d4cf

But when I deploy this code into production, I find the performance is rather bad. I sit and think about it for a little while and realise there is a much better way to implement Foo, so I write a new, AdvancedFoo and change my FooFactory:

trait Foo
class StandardFoo extends Foo
class AdvancedFoo extends Foo

object FooFactory {
  def newFoo() = new AdvancedFoo
}

The performance problem has to be fixed before the weekend, but it's Friday night and I really want to get home to my wife and kids. I know my application won't break, because I've used a Gang of Four pattern to hide the implementation details from my applications, so I just build the FooFactory library version 2, chuck it into production and go home.

On Monday morning, I look at the results from the weekend's FooApplication run and, to my absolute surprise and horror, I find this:

java.lang.NoSuchMethodError: FooFactory$.newFoo()LStandardFoo;

What happened? No such method? My FooFactory still has a newFoo() method. But the last part reveals the problem: StandardFoo. If we have a look at the byte code, we'll see the source of the problem. First, let's have a look at the output of running javap on on the first version of FooFactory:

Compiled from "FooFactory.scala"
public final class FooFactory extends java.lang.Object{
    public static final StandardFoo newFoo();
    public static final int $tag()       throws java.rmi.RemoteException;
}

And let's compare that to the byte code from version 2:

Compiled from "FooFactory.scala"
public final class FooFactory extends java.lang.Object{
    public static final AdvancedFoo newFoo();
    public static final int $tag()       throws java.rmi.RemoteException;
}

Okay, so newFoo() was returning a StandardFoo, and in version two the return type is AdvancedFoo. It's tempting to think that shouldn't matter. The type of the 'result' val in FooApplication was still just Foo, so it should happily accept whatever newFoo() returns, shouldn't it? Let's have a look at the byte code of FooApplication's main method:

public void main(java.lang.String[]);
  Code:
   0: getstatic #26; //Field FooFactory$.MODULE$:LFooFactory$;
   3: invokevirtual #30; //Method FooFactory$.newFoo:()LStandardFoo;
   6: astore_2
   7: getstatic #35; //Field scala/Predef$.MODULE$:Lscala/Predef$;
   10: aload_2
   11: invokevirtual #39; //Method scala/Predef$.println:(Ljava/lang/Object;)V
   14: return

And there's the problem, clear as day, on the line starting at byte 3. FooApplication is statically linked to a version of newFoo() that returns a StandardFoo. Did you know that the JVM considers the return type as part of the method signature? A method with the exact same name and same arguments but with a different return type is a different method. So version 2 of the FooFactory doesn't actually contain the method to which FooApplication was statically linked - and hence my NoSuchMethodError.

The quick-fix solution to this is pretty simple: FooApplication needs to be recompiled. No change to the source is necessary, but the recompilation will re-write the byte code with a static link to the new version of newFoo() that has an AdvancedFoo return type.

So, that's a fix to this instance of the problem, but it won't stop it from happening again (when I upgrade to GridFoo!). How would we stop it again? Sadly, for "less is more" coders like me, the solution is to not use type inference, at least in locations where you want to hide implementation details. Basically we need to follow Item 34 from Effective Java (1st Edition): "Refer to objects by their interfaces". While in Java this principle mainly protects you from having to make source changes if you change the return type, in Scala the source would typically not require any change but, without recompilation, the statically-linked byte code is still incompatible. If our original implementation of newFoo() had a return type of Foo like such...

object FooFactory {
  def newFoo() : Foo = new StandardFoo
}

… then the method signature of FooFactory version 2 would have been identical to version 1 and the statically linked FooApplication wouldn't have known the difference.

In some ways this is obviously a bit of a straw man: I changed the version of a library and didn't recompile my code. But in the world of large projects, complex dependency graphs and (gasp) transitive dependencies, there's always a chance for bizarre things like this to crop up as other people change things and don't let you know.

I think an argument could even be made here for not using type inference at all, or at least not on public methods in projects that are depended on by others. If this sounds a bit Draconian, consider this: the point of type inference is mostly to relieve you from having to write or think about type declarations. But if you now have to think about when you do and don't want to rely on type inference, half that advantage is lost, and it may be easier to just to declare your types every time.

Scala: Beyond Hello World

I overheard someone at work last week asking another guy how he would go about implementing string-wrapping in Java. Memories came flooding back of university functional programming assignments where we had to use Miranda to shape nursery rhymes into a trapezoid.

I listened to the continuing conversation and realised that the solution being discussed - which involved seeking through the string, remembering where the last whitespace was and cutting off substrings at appropriate points - was obviously imperative in nature. "Not that there's anything wrong with that." If I wasn't reading a book about Scala programming at the moment I would have joined right in. But my functionally-leaning mind instantly started to wonder what the functional solution would be.

It sounded like a simple but interesting problem to try and solve with a new language - a nice "real" program to attempt after Hello World. So I sat down after work and whipped up a small Scala program to wrap words. I have to admit, it took a bit longer than I expected - over an hour. Though a large part of that was spent looking up docs for the List class and trying to solve new and bemusing syntax errors, I do recall getting distracted by an investigation into when tail call optimisation does and doesn't occur.

If you're new to Scala and looking for a little problem to help you learn some syntax and waste an hour, I can recommend word-wrapping as a suitable foe. My solution is below. If you come up with your own, I'd love to know what you did differently!

object Wrapper {

  type Line = List[String]

  def main(args: Array[String]) {
    for (i <- 1 to 31) print(if (i == 31) '\n' else '.')
    val s = "This is a really, really long line of text that, hopefully, " +
            "will be long enough to sufficiently test a function that must " +
            "divide a really, really long string into lines that are no " +
            "more than 30 (that's thirty) characters long."
    for (line <- wrap(s)) {
      for (string <- line) {
        print(string + " ")
      }
      println
    }
  }

  def wrap(stringToWrap : String) : List[Line] = 
        wrap(List.fromArray(stringToWrap.split(" ")), 30, List(List()))

  def wrap(stringsToWrap : List[String], 
           maxLineLength : Int, 
           lines : List[Line]) : List[Line] =
    stringsToWrap match {
      case Nil => lines.reverse
      case nextString :: remainingStrings => 
        if (lineLength(lines.head) + nextString.length <= maxLineLength)
          wrap(remainingStrings, maxLineLength, 
               (lines.head ::: List(nextString)) :: lines.tail)
        else {
          wrap(remainingStrings, maxLineLength, List(nextString) :: lines)

    }
  }

  def lineLength(s : Line) : Int = s match {
    case Nil => 0
    case head :: tail => 1 + head.length + lineLength(tail)
  }

}

Missing Parameter Type For Expanded Function?

I just made a silly Scala syntax mistake and got a fairly unhelpful compiler error message (often the case with newer languages). Google wasn't much help, so I thought I should post up the solution to help others save time in the future.

The code I wrote was essentially this:

def main(args: Array[String]) {
println(for {s <- List("One", "Two", "Three")} yield _) } 

And the error I got was this:

error: missing parameter type for expanded function ((x$1) => List("One", "Two", "Three").map(((s) => x$1)))
println(for {s <- List("One", "Two", "Three")} yield _) 

My novice mistake was in assuming that I could use an underscore to access the variable from the 'for' comprehension. Simply changing that underscore to 's' fixed it good:

def main(args: Array[String]) {
println(for {s <- List("One", "Two", "Three")} yield s) } 

Something else worth noting is that the command-line scalac was actually much more helpful than my IDE, because it printed a third line with a caret pointing straight at the underscore, while the IDE only told me what line the error was on.

Hey Scala, Where's My Ternary Operator?

While still playing around with Scala Hello World programs, I thought I'd do a little Java integration test so I wrote this awesome program:

println(Math.random() > 0.5 ? "Hello World!" : "Goodbye World")

The result, of course, was neither "Hello World!" nor "Goodbye World" (lest this be a very short blog), but...

error: identifier expected but string literal found

Huh? Why wouldn't that work? Is there no ternary operator? How primitive!

Of course, it dawned on me without too much more wasted neurons: this is a functional language - everything returns a value! (It has been some time, but the memories are slowly starting to come back.) So, of course, it only makes sense that if/else returns a value as well. Voila:

println(if (java.lang.Math.random() > 0.5) "Hello World!"
       else "Goodbye World")

And once you see it, you see how obvious it is - that this is just the way if/else should work. Why should any language need a quirky ternary operator when they all have a perfectly good if/else construct already?

On a related note, I'm feeling an increasing annoyance as I look at Java code at work all day. For example, today I was looking at code that essentially did this:

public List convert(List customers) {
   ArrayList dtos = new ArrayList();
   for (Customer c : customers) {
       dtos.add(convert(c));
   }
   return dtos;
}

public CustomerDTO convert(Customer c) {
   return new CustomerDTO(c.getCustomerId(), c.getName(), c.getBalance());
}

So, as soon as I saw that, I thought, "Man, if we were using Scala, I'd just have to write this:"

def convert(customers : List[Customer]) =
     customers.map(c => new CustomerDTO(c.customerId, c.name, c.balance))

In fact, the expression is so short, you most likely wouldn't even bother to put it in a function. (Except to test it, or mock it.)

The brevity here isn't just "cool" (even though it is). This is an extremely simple piece of code; and an extremely common operation. The brevity of the Scala code helps to express how simple the function is, whereas the Java code for the same operation takes longer to read and could leave you wondering whether there was something you missed.

The most ridiculous thing about this comparison is that Java has been around for almost 15 years now! And still we have to write 5 lines of code to do ridiculously simple operations like this. I can't imagine it would be much work at all for a few smart guys to add the functionality that Scala displays here to Java, but it appears more and more that Java is slowing to a halt under its own sheer weight and the burden of keeping 6 million-odd developers happy. (As if writing the above code over and over again is the way to do it.)

R.I.P. Java?

New to Scala? Study the Syntax. CAREFULLY.

So, it only took me about five seconds to go from having a Hello World Scala program working to having a 10% more complex Hello World program failing in a way that I had no idea how to fix.

Here's my program:

object HelloWorld {
def main(args: Array[String]) {
   println(testfunc)
}

def testfunc() {
   "Hello World!"
}
}

and the output of this awesome program? Just this...

()

If you're a seasoned Scala programmer, you can probably see already why I'm such an idiot, but there's an important point to be made here.

I was trying to test out two things with this example. First, this being the first time I've done functional programming in almost ten years, I wanted to "practice" the idea that the last expression in a function is the return value. Yes, I wanted to practice not writing "return". Secondly, I had read that Scala let's you invoke functions without parentheses, so I thought I'd give that a go by calling "testfunc" instead of "testfunc()".

But neither of these were the problem here. The problem was that, during all my Java-ingrained, time-poor reading of Scala tutorials, I'd had seen but never cared to take particular notice of one small Scala detail: function declarations and function definitions are separated by an equal sign:

object HelloWorld {
def main(args: Array[String]) {
   println(testfunc)
}

def testfunc() = {
   "Hello World!"
}
}

This code produces the output I was expecting at first: that magic phrase "Hello World!"

I'm actually really glad this happened. It wasted ten minutes of my time, but it taught me a real lesson that I've forgotten through not learning new languages often enough: sit at the computer and type the examples in. No amount of reading examples will actually make you understand a language or teach you how to write it. Only by sitting down, typing in every symbol from an example and understanding what it does (and figuring out where you stuffed it up) can you hope to become proficient.

There is a corollary lesson here, too: no matter how much people tell you Scala is like Java and how much you want to believe it, Scala is NOT Java. There are little differences which, as in this example, can make a HUGE difference. My approach from here on in will be to try and forget about "migrating" my knowledge from Java (the language) to Scala, and just approach Scala as an absolutely new language that needs learning from scratch.

Lastly: The thing that got me really confused is that my original code, complete with the nonsensical error, compiled and ran! To be honest, I still have absolutely no idea what that code is actually doing. If I change my 'println(testfunc)' to 'println(testfunc.getClass)', the output is:
class scala.Predef$$anon$1
So, I'm guessing that my function definition without the equal sign is somehow defining an anonymous type rather than a function, but I really am guessing - I actually have no idea. I'm looking forward to burying my head more and more into Scala syntax until one day I finally gain the knowledge to understand my mistake and jump up screaming "Aha!"

UPDATE:
Well, it only took a little more reading to discover what the original code is doing. Despite Scala being a primarily functional language, it still has the concept of "procedures", or what would be void methods in Java. These are methods that don't return a value, and they are defined using the exact same syntax as a function, except without the equal sign.

So, my original assumption - that my function was not being invoked but was instead being passed to println as some kind of anonymous type - was quite wrong. The function IS being invoked, but it doesn't return anything. But why does the code compile and run if the method returns nothing? That's because everything in Scala has a value. Even procedures that don't return anything still have a value! Crazy. (But fun.)