Last Updated: April 12, 2016
·
6.186K
· mortimerpa

Adding Semantic to Base Types Parameters in Scala

Here is another "fun" dive in the complex type system of Scala. This was inspired by Eric Torreborre's post that gives a very good use case for what I am going to show, but I wanted to share another one.

I am giving a simple example here, it's a bit naive and there would be other ways to implement it, but I hope you will still understand the overarching concept: being able to annotate the meaning of basic type parameters (and variables) such as Int, List[String], etc.
This will help the compiler can automatically detect more errors and a code reviewer to understand better what's going on.

A Use Case

Imagine that you are trying to get users to score products. Each product can be scored along different dimensions: design, usability, durability.

You could represent a score with a simple object:

case class User(id: Long)
case class Product(id: Long)

case class Scoring(user: User, product: Product, design: Int, usability: Int, durability: Int)

Now, you can create Scoring instances in your code Scoring(u1, p1, 10, 15, 20).

You will probably end up creating such objects in different parts of your code, from raw Ints coming from diverse sources, such as a REST request, or from parsing a line from your database.

The issue here is that the Scoring constructor is simple, but not very safe. If you are like me, you will have to go check in the documentation to know in what order to set the parameters.

For instance, you might write something like this, to create an instance with Anorm from a database query:

def parseProduct(row: Row): Product = …

def parseUser(row: Row): User = …

def parseScoring(row: Row): Scoring = Scoring(
        parseProduct(row), 
        parseUser(row),
        row[Int]("design"),
        row[Int]("durability"),
        row[Int]("usability")
)

There are two errors in this code:

  1. the scala compiler performing type check would straightaway flag the inverted product/user parameters as a compilation error: type mismatch; found : Product required: User,
  2. however, the second error, inverting the durability score and the usability score will not be spotted by the compiler and even a code reviewer would have to be very concentrated to spot such a trivial error.

Wrappers

The simple solution would be to create wrapper classes for each type of score:

case class Design(val score: Int) extends AnyVal
case class Usability(val score: Int) extends AnyVal
case class Durability(val score: Int) extends AnyVal

case class Scoring(user: User, product: Product, design: Design, usability: Usability, durability: Durability)

def parseScoring(row: Row): Scoring = Scoring(
        parseUser(row),
        parseProduct(row), 
        Design(row[Int]("design")),
        Usability(row[Int]("usability")),
        Durability(row[Int]("durability"))
)

This makes a bit more code to write, but it explicitly marks the semantic of the Ints that you are passing in; making the code easier to understand and the scala compiler will now complain when you create Scoring with the scores in the wrong order. If you instantiate a Usability score with the "durability" column, the compiler won't know the different, but an attentive reviewer would quickly spot the confusion.

Note the AnyVal extension on the case classes, this is a new feature of scala 2.10, called value classes. This tells the compiler that these types are just wrappers for typechecking, but that they can be optimized at runtime to their inner value, without the expensive creation of the wrapper object.

This is already quite a good solution and it will go a long way. But imagine that you now want to compute the average of all the scores of a product, you might have a list of Scoring for a product and will foldLeft on it to sum up every scores. However, because you have included new types, you can't just sum prod1.design + prod2.design as the operation + is not defined on the Design class. You end up having to either write: Design(prod1.design.score + prod2.design.score) which makes your code more cumbersome than it should be.

You could also implement a proxy + operator in all new wrapper case class and replicate all other operations you might need within the wrapper.
When it's just one operation, for three case classes, that's fine. But if you deal with more complex basic types, such as String or composed types, such as List[Int] or Future[Int], you might end up having to proxy a lot of operations in your wrappers. If you have loads of such wrappers, it might be a lot of extra code.

Unboxed Tagged Types

The idea that I have decided to follow, inspired by Eric's article, is to use "Unboxed Tagged Types".
The idea is that you can "tag" types (and not just basic types) with a keyword, without loosing all its operations as you would with a simple wrapper.

The principle is simple and it is basic object oriented inheritance: to retain all the features of Int you can define a class:

class DesignScore extends Int

However, we use a scala trick to make this definition easier and to avoid actually creating new objects: type aliases.
It takes a very simple code scaffold:

 //code from Eric's article
type Tagged[U] = { type Tag = U }
type @@[T, U] = T with Tagged[U]
  • Tagged[U] defines a new type, which just adds a type property containing the parameter type U. I.E it's a type that has an annotation U. So if you wrote type TaggedInt = Int with Tagged[Design], you would just define a new type alias, which extends Int and tags it with the new inner type property type Tag = Design.
  • @@[T, U] is only a convenient shortcut to extend types with the Tagged interface. So you could write the previous example with: type TaggedInt = Int @@ Design

You can add this snippet to your own code or use the implementation provided by scalaz, which is pretty much equivalent.

Right, this is still pretty abstract; so don't worry if you are not getting it yet. Just see it as a tool to annotate types. Let's see how to use it now with our previous example.

First, we define "tags" that will allow us to annotate the Int basic type with some contextual information:

trait Design
trait Usability
trait Durability

These traits don't do anything except existing as a known type for the compiler, you can see them as constants, or values of an enum, at the type level. We can now use them to define our tagged types:

type DesignScore = Int @@ Design
type UsabilityScore = Int @@ Usability
type DurabilityScore = Int @@ Durability

These type aliases just represent an extension of the type Int with an embedded type parameter Tag. We can use these as we would use any other type, so as for our previous example:

case class Scoring(user: User, product: Product,
                   design: DesignScore,
                   usability: UsabilityScore,
                   durability: DurabilityScore)

Because DesignScore is just a subclass of Int, it inherits all its operations, and we can now use the sum operator: prod1.design+prod2.design.

However, we are still missing something: scala doesn't know how to create a DesignScore, as when you write val score = 1, you get an Int and scala doesn't know apriori how to convert this to a subtype of Int. This actually plays in our favor, as it will force us to explicitely write the conversion when we need to assign an integer to something that needs a DesignScore.

Let's write an extension of Int that knows how to do this conversion:

implicit class TaggedInt(val i: Int) extends AnyVal {
     def design = i.asInstanceOf[DesignScore]
     def usability = i.asInstanceOf[UsabilityScore]
     def durability = i.asInstanceOf[DurabilityScore]
}

Note a few things:

  1. we use a cast to transform a superclass (Int) in a subclass (DesignScore), it might seem dirty, but it's safe as we will never encounter a case where we cannot perform the case,
  2. we are using a new feature of scala 2.10: an implicit class that extends Int with the new methods,
  3. we shouldn't use a direct implicit conversion between Int and DesignScore as we would loose the fact that now developers have to be explicit about how they want to use the Int type. Nothing would then stop you from using the Usability Int in a DesignScore slot.
  4. while I like the postfix operator style, you could just define unary operators if you prefer, without using an implicit class at all: def design(i: Int): DesignScore = i.asInstanceOf[DesignScore].

We can now create the Scoring object in the following manner Scoring(u1, p1, 10.design, 15.usability, 20.durability). And in our previous database parsing example:

def parseScoring(row: Row): Scoring = Scoring(
        parseUser(row),
        parseProduct(row), 
        row[Int]("design").design,
        row[Int]("usability").durability,
        row[Int]("durability").usability
)

As you can see, it's already a bit easier to spot the error, and if you don't the compiler will complain anyway.

I hope that through this simplistic example, you have been able to understand how Unboxed Tagged Types can help. Explicitly annotating the use of non semantic basic types (Int, String, List[Int], …) in your code will have two benefits:

  1. in many cases, the compiler will warm you when you are using a value in the wrong slot.
  2. by annotating method signatures (in and out types), and the uses of raw types in your code, this one will be easier to understand and you will hopefully make less errors and quickly spot bugs.

Parameter Validation

As George Leontiev points out, you can also use this feature to perform validation.

In our example, the scores are unbounded, you could pass in any value without anyone complaining. Ideally, you would probably want to have them in a bounded range, let's say between 0 and 5.

The usual approach would be to validate the input when the user inputs a value. But then, in the rest of the code, there is nothing stopping you from making bad operations on the values.

Clearly, this is not something you can deal with at compile time as you do not know what user inputs you will get or what the values in the database
will be. However, you can build in some gards to warn you when something is wrong.

As we have seen, we cannot build the tagged types without going through the explicit call to the design, durability or usability functions. So we can just extend these functions to add checks on the ranges:

implicit class TaggedInt(val i: Int) extends AnyVal {
     def design = {
         require(i >= 0 && i <= 5, "the design score has to be between 0 and 5")
         i.asInstanceOf[DesignScore]
     }
     ...
}

Now, the validation of your parameters is built-in the "type" that defines it, if you call Scoring(u1, p1, 10.design, 15.usability, 20.durability), an IllegalArgumentException will be thrown. If you check for such exception when receiving a score value, you can deal with it properly, report an error to the user, etc.

If you prefer to use the validation pattern instead of catching exceptions, you can use scalaz Validation monad, or scala default Option or Either constructs:

implicit class TaggedInt(val i: Int) extends AnyVal {
     def design: Either[String, DesignScore] = {
         if(i >= 0 && i <= 5){
            Right(i.asInstanceOf[DesignScore])
         } else {
            Left("the design score has to be between 0 and 5")
         }
     }
     ...
}

1 Response
Add your response

Which version of scala were you using to test this. Use of tagged types and case classes still does not seem to work (as of 10.2 which I am using) : https://issues.scala-lang.org/browse/SI-5183

over 1 year ago ·