2015-07-19

Introducing Emblem - TypeKeyMaps

This is the third post in the Introducing Emblem series, where I discuss a reflection-based utility library I have developed called emblem. If you want to try emblem out on your project, follow these instructions. Or read this if you just want to explore try it out in the Scala REPL.

In the previous post, we looked at the core building block of the emblem library: the TypeKey. In this post, we begin to explore the utility of the TypeKey using TypeKeyMaps.

In the previous post in this series, we saw how emblem's TypeKeys can be used in much the same way as TypeTags. Unlike TypeTags, TypeKeys compare as equal (as in ==) for equivalent types. We saw how this allowed us to create sets of types, and use types as keys in a map. The map example we presented, however, was not very interesting, as the values stored were just Ints. What if we wanted to do something more interesting than that, where the types of the values actually depended on the types of the keys?

Let's take a look at an example to see what I'm getting at. Suppose we are developing an application for a company that builds computers according to the customers' specifications. We want to model the different kinds of computer parts, and how they fit together to form a computer. We might come up with something like this:

sealed trait ComputerPart
case class Memory(gb: Int) extends ComputerPart
case class CPU(mhz: Double) extends ComputerPart
case class Display(resolution: Int) extends ComputerPart

case class Computer(memory: Memory, cpu: CPU, display: Display)

In the workshop where we build the computers, we have an inventory of computer parts that customers can select. We'll store each choice for Memory in a List[Memory], each available CPU in a List[CPU], and so on:

val memoryList = Memory(2) :: Memory(4) :: Memory(8) :: Nil
val cpuList = CPU(2.2) :: CPU(2.4) :: CPU(2.6) :: Nil
val displayList = Display(720) :: Display(1080) :: Nil

We anticipate adding more kinds of ComputerParts in the future, and we are not terribly comfortable with having individual Lists around for each kind of part. We really want to store the entire inventory in a single data structure. We settle on building a Map where the keys to the map are the type of computer part, and the values are the list of parts of that type. The keys have this type:

  TypeKey[_ <: ComputerPart]

And the values have this type:

  List[_ <: ComputerPart]

The standard way to bring TypeKey into your namespace is as follows:

import emblem.imports._

We can construct our inventory using the hardcoded part lists we built above:

val inventory =
  Map[TypeKey[_ <: ComputerPart], List[_ <: ComputerPart]](
    typeKey[Memory] -> memoryList,
    typeKey[CPU] -> cpuList,
    typeKey[Display] -> displayList)

We can pull up our inventory of CPUs like so:

val cpus = inventory(typeKey[CPU])

But there is a problem. The type of the resulting value is not right. It is List[_ <: ComputerPart], when we are expecting List[CPU]. If we need to use it as a List[CPU], we will have to cast it to the right type. Consider what a method would look like that pulls a parts list out of the inventory by type. We want it to take in the type as parameter, and return a properly typed list. In other words, we want a method signature that looks something like this:

def partList[P <: ComputerPart]: List[P]

We would call this method like so:

val cpus = partList[CPU]

Since we need to look up the part list by TypeKey, we can add a TypeKey as an implicit parameter, like so:

def partList[P <: ComputerPart : TypeKey]: List[P]

Within the body of this method, we have to cast from List[_ <: ComputerPart] to List[P]:

def partList[P <: ComputerPart : TypeKey]: List[P] =
  inventory(typeKey[P]).asInstanceOf[List[P]]

And of course, there is nothing preventing this typecast from failing, as inventory may have been mistakenly initialized like so:

val inventory =
  Map[TypeKey[_ <: ComputerPart], List[_ <: ComputerPart]](
    typeKey[Memory] -> memoryList,
    typeKey[CPU] -> displayList,
    typeKey[Display] -> displayList)

At this point, calling val cpus = partList[CPU] does fail, but not in the way you might think. Due to type erasure, the typecast is successful, and we end up with something with type List[CPU], and value List(Display(720), Display(1080))! The 
java.lang.ClassCastException does not occur until we actually access one of the elements of the list, say with cpus(0) or cpus.head.

Let's sum up the problems we are having with this Map:
  1. We are forced to typecast when accessing data in the map, and this makes us mildly nauseous.
  2. Nothing is enforcing that the key and value in the map are working with the same kind of ComputerPart.
I created TypeKeyMap for situations like these. They behave similarly to the Map above, but the types of the key and the value in any key-value pair are forced to agree. To build a TypeKeyMap, we have to specify two type parameters: TypeBound, which serves as an upper bound on the type parameters for the key and value types; And Val, which describes the type of the values in the map. In our example, we want to use ComputerPart for TypeBound, and List for Val. We initialize our new inventory as follows:

val inventory = TypeKeyMap[ComputerPart, List]() +
  memoryList + cpuList + displayList
                                                               
There are a couple points of interest I'd like to mention here. First, notice that when we add a key-value pair to the map, we only need to mention the value, as the key can be inferred from the value. If you want to be explicit about the keys as well, the following code is equivalent:

val inventory = TypeKeyMap[ComputerPart, List]() +
  (typeKey[Memory] -> memoryList) +
  (typeKey[CPU] -> cpuList) +
  (typeKey[Display] -> displayList)

Second, in contrast to scala.collection.immutable.Map, there is no way to initialize the TypeKeyMap with a single varargs method invocation. This is because every key-value pair must be type-checked individually, to make sure the type bounds match. I have an idea about how I might overcome this problem, but it's not high priority for me at the moment, since using the + operator as above to construct the map seems elegant enough.

Now, when we pull out a part list, it is well typed. In this example, we look up a value explicitly by key, just as we would with a normal Scala Map. No cast required:

val memParts: List[Memory] = inventory(typeKey[Memory])

But that's a little verbose. Here's a more fluid way to look up a value by key:

val memParts: List[Memory] = inventory[Memory]

We can update the TypeKeyMap using the + operator. TypeKeyMaps are immutable, so this operation produces a new map:

val moreMemory = Memory(16) :: memoryList
val updatedInventory = inventory + moreMemory

Once again, the TypeKey is resolved implicitly, but we can be explicit about it if we like:

val updatedInventory = inventory + (typeKey[Memory] -> moreMemory)

Of course, providing a mistyped key-value pair doesn't work. If we say:

val updatedInventory = inventory + (typeKey[CPU] -> moreMemory)

We get the following compile-time error:

Cannot prove that List[CPU] <:< List[Memory].

Ideally, the TypeKeyMap API would be contain analogs for everything in scala.collection.immutable.Map. And while I would really like to see that happen, so far I have only implemented the easy parts of the API, and the methods that I wanted to use myself. Here's a rough list of the methods I have implemented so far:
+, ++, apply, contains, equals, foreach, get, getOrElse, hashCode, isEmpty, iterator, keys, mapValues, size, toString, values
I've found reproducing some portions of the Map API to be a little bit tricky, and to require the introduction of supporting classes. For instance, consider foreach. In the Map API, the method has the following signature:

def foreach(f: ((A, B)) ⇒ Unit): Unit

A little thought shows this signature to be insufficient for a TypeKeyMap. In our example above, the function f passed to method foreach would have a signature like this:

def f(pair: (TypeKey[_ <: ComputerPart],
             List[_ <: ComputerPart])): Unit

But what we actually want is a TypeKey/ComputerPart pair where the type parameters match. In other words, we want to supply a function with a signature like this:

def f[P <: ComputerPart](pair: (TypeKey[P], List[P]): Unit

We run into a stumbling block at this point, because anonymous functions in Scala cannot yet have type parameters. So to make foreach happen, I had to introduce a supporting class, TypeBoundPair.

I would really be thrilled if you were to start using this code, so if you find yourself wanting to use a part of the Map API that is not reflected here, please just let me know and I will do my best to add it in a timely manner. (Of course, I would be even more thrilled to see a pull request!)

There's a lot more I could write about TypeKeyMap, but I figure this is a good introduction. If you want more TypeKeyMaps right away, I recommend starting with TypeKeyMapSpec.scala and TypeBoundMapSpec.scala. I am happy to write more on the subject, but for now, I'd rather continue on with my overview of the emblem library. TypeBoundMaps will be next. It's the same basic idea as a TypeKeyMap, except that the keys in a TypeBoundMap do not have to be TypeKeys. I'll get a post out on that as soon as I can.


No comments:

Post a Comment

Note: Only a member of this blog may post a comment.