Monday, 23 February 2015

Zero-overhead Recursive ADT Coproducts

Zero-product Recursive AMD what???

Ok. Imagine this: you're building some app, and in certain parts users can type text with special tokens. (It's like in Twitter, when typing “Hello @japgolly” the “@japgolly” part gets special treatment.) You parse various types of tokens. You might have different locations to type text, and rules about which tokens are allowed in each location. You want the compiler to enforce those rules but you also want to handle tokens generically sometimes. How would you do such a thing in Scala?

Initial Attempts

Ideally you'd define an ADT (algebraic data type) for all tokens possible, then create new types for each location that form a subset of tokens allowed. If that were possible, here's what that would look like.

sealed trait Token
case class  PlainText(t: String)  extends Token
case object NewLine               extends Token
case class  Link(url: String)     extends Token
case class  Quote(q: List[Token]) extends Token

// type BlogTitle   = PlainText | Link    | Quote[BlogTitle]

// type BlogComment = PlainText | NewLine | Quote[BlogComment]

Now Scala won't let us create our BlogTitle type like shown above. It doesn't have a syntax for coproducts (which is what BlogTitle and BlogComment would be, also called “disjoint unions” and “sum-types”) over existing types. Seeing as we have control over the definition of the generic tokens, we can be tricky and inverse the declarations like this:

sealed trait BlogTitle
sealed trait BlogComment

sealed trait Token
case class  PlainText(t: String) extends Token with BlogTitle with BlogComment
case object NewLine              extends Token                with BlogComment
case class  Link(url: String)    extends Token with BlogTitle
case class  Quote(q: List[????]) extends Token with BlogTitle with BlogComment
//                        ↑ hmmm...

...but as you can see, we hit a wall when we to Quote, which is recursive. We want a Quote in a BlogTitle to only contain BlogTitle tokens, not just any type of token. We can continue our poor hack as follows.

abstract class Quote[A <: Token] extends Token { val q: List[A] }
case class QuoteInBlogTitle  (q: List[BlogTitle])   extends Quote[BlogTitle]
case class QuoteInBlogComment(q: List[BlogComment]) extends Quote[BlogComment]

Not pleasant. And we're not really sharing types anymore. What else could we do?

We could create separate ADTs for BlogTitle and BlogComment, that would mirror and wrap their matching generic tokens, then write converters from specific to generic. That's a lot of duplicated logic and tedium, plus we now double the allocations and memory needed. Let's try something else...


NOTE: This bit about Shapeless is an interesting detour, but it can be skipped.

We could use Shapeless! Shapeless is an ingeniously-sculpted library that facilitates abstractions that a panel of sane experts would deem impossible in Scala, one such abstraction being Coproducts. Here's what a solution looks like using Shapeless.

(Sorry I thought I had this working but just realised recursive coproducts don't work. I've commented-out Quote[A] for now. There is probably a way of doing this – Shapeless often doesn't take no from scalac for an answer – I'll update this if some kind 'netizen shares how.)

sealed trait Token
case class  PlainText(text: String) extends Token
case object NewLine                 extends Token
case class  Link(url: String)       extends Token
case class  Quote[A](q: List[A])    extends Token

type BlogTitle   = PlainText :+: Link         :+: /*Quote[BlogTitle]   :+: */ CNil
type BlogComment = PlainText :+: NewLine.type :+: /*Quote[BlogComment] :+: */ CNil

/* compiles → */ val title = Coproduct[BlogTitle](PlainText("cool"))
// error    → // val title = Coproduct[BlogTitle](NewLine)

So far so good. What would a Token ⇒ Text function look like?

object ToText extends Poly1 {
  implicit def caseText    = at[PlainText   ](_.text)
  implicit def caseNewLine = at[NewLine.type](_ => "\n")
  implicit def caseLink    = at[Link        ](_.url)
  // ...

val text: String =

Ok I'm a little unhappy because I'm very fold of pattern-matching in these situations, but the above does work effectively. One thing to be aware of with Shapeless, is how it works. To achieve its awesomeness, it must build up a hierarchy of proofs which incurs time and space costs at both compile- and run-time – its awesomeness ain't free. The val title = ... statement above creates at least 7 new classes at runtime, where I want 1. Depending on your usage and needs, that overhead might be nothing, but it might be significant. It's something to be aware of when you decide on your solution.

Zero-overhead Recursive ADT Coproducts

There's another way. I mentioned “zero-overhead” and it can be done. Here is a different solution that relies solely on standard Scala features, one such feature being path-dependent types.

You can create an abstract ADT, putting each constituent in a trait, then simply combine those traits into an object to have it reify a new, concrete, sealed ADT. Sealed! Let's see the new definition:

// Generic

sealed trait Base {
  sealed trait Token

sealed trait PlainTextT extends Base {
  case class PlainText(text: String) extends Token

sealed trait NewLineT extends Base {
  case class NewLine() extends Token

sealed trait LinkT extends Base {
  case class Link(url: String) extends Token

sealed trait QuoteT extends Base {
  case class Quote(content: List[Token]) extends Token

// Specific

object BlogTitle   extends PlainTextT with LinkT    with QuoteT
object BlogComment extends PlainTextT with NewLineT with QuoteT

Now let's use it:

   List[BlogTitle.Token](BlogTitle.PlainText("Hello")) // success
// List[BlogTitle.Token](BlogTitle.NewLine)   ← error: BlogTitle.NewLine doesn't exist
// List[BlogTitle.Token](BlogComment.NewLine) ← error: BlogComment tokens aren't BlogTitle tokens

// Specific
val blogTitleToText: BlogTitle.Token => String = {
  // case BlogTitle.NewLine   => ""     ← error: BlogTitle.NewLine doesn't exist
  // case BlogComment.NewLine => ""     ← error: BlogComment tokens not allowed
  case BlogTitle.PlainText(txt) => txt
  case BlogTitle.Link(url)      => url
  // Compiler warns missing BlogTitle.Quote(_) ✓

// General
val anyTokenToText: Base#Token => String = {
  case a: PlainTextT#PlainText => a.text
  case a: LinkT     #Link      => a.url
  case a: NewLineT  #NewLine   => "\n"
  // Compiler warns missing QuoteT#Quote ✓

// Recursive types
val t: BlogTitle  .Quote => List[BlogTitle  .Token] = _.content
val c: BlogComment.Quote => List[BlogComment.Token] = _.content
val g: QuoteT     #Quote => List[Base       #Token] = _.content

Look at that! That is awesome. These are some things that we get:

  • No duplicated definitions or logic.
  • Generic & specific hierarchies are sealed, meaning the compiler will let you know when you forget to cater for a case, or try to cater for a case not allowed.
  • Children of recursive types have the same specialisation.
    Eg. a BlogTitle can only quote using BlogTitle tokens.
  • Tokens can be processed generically.
  • Zero-overhead. No additional computation or new memory allocation needed to store tokens, or move them into a generic context. No implicits.
  • Nice, neat pattern-matching which makes me happy.
  • It's just plain ol' Scala traits so you're free to encode more constraints & organisation. You can consolidate traits, add type aliases, all that jazz.

There you go. Seems like a great solution to this particular scenario.

Nothing is without downsides though. Creation will likely be a little hairy; imagine writing a serialisation codec – the Generic ⇒ Binary part will be easy but Binary ⇒ Specific will be more effort. In my case I will only create this data thrice {serialisation, parsing, random data generation} but read and process it many, many times. Good tradeoff.