Assuming that I'm writing a program for vector multiplication. Following the requirement in this article:
https://etrain.github.io/2015/05/28/type-safe-linear-algebra-in-scala
The multiplication should only compile successfully if the dimension of both vectors are equal. For this I define a generic type Axis
that uses a shapeless literal type (the number of dimension) as the type parameter:
import shapeless.Witness
trait Axis extends Serializable
case object UnknownAxis extends Axis
trait KnownAxis[W <: Witness.Lt[Int]] extends Axis {
def n: Int
def ++(that: KnownAxis[W]): Unit = {}
}
object KnownAxis {
val w1 = Witness(1)
val w2 = Witness(2)
case class K1(n: Witness.`1`.T) extends KnownAxis[w1.type]
case class K2(n: Witness.`2`.T) extends KnownAxis[w2.type]
// K2(2) ++ K1(1) // doesn't compile
K2(2) ++ K2(2)
}
So far so good, the problem however appears when I try to generalise it for all n:
case class KN[W <: Witness.Lt[Int]](n: W#T) extends KnownAxis[W]
KN(1)
The above code triggers a compilation error:
Axis.scala:36: type mismatch;
found : Int(1)
required: this.T
[ERROR] KN(1)
[ERROR] ^
[ERROR] one error found
My question is: why Spark is incapable of focusing on the more refined type Witness.`1`.T
, instead it is using type Int
? What does it take to override this behaviour so case class KN
can be successfully defined?
UPDATE 1: The follow up has been moved to a new question:
It isn't surprising that Scala can't infer W
given W#T
, because generally speaking it's possible for two different W
s to have the same W#T
. Not for witness types, but they aren't treated specially by the compiler.
What I expected to work (and not sure why it doesn't) is to specify the type parameter:
KN[Witness.`1`](1)
// error: type arguments [scala.this.Any] do not conform to method apply's type parameter bounds [W <: shapeless.this.Witness.Lt[scala.this.Int]]
or more likely
KN[w1.type](1)
// error: type mismatch;
// found : scala.this.Int(1)
// required: .this.T
What does work:
case class KN[W <: Witness.Lt[Int]](w: W) extends KnownAxis[W] {
val n = w.value
}
KN(Witness(1))
It seems to suit the requirements in your question, but I don't know if it'll work with the rest of your code.
You may also want to consider this alternative which doesn't require Shapeless in Scala 2.13:
trait Axis extends Serializable
case object UnknownAxis extends Axis
trait KnownAxis[W <: Int with Singleton] extends Axis {
def n: W
def ++(that: KnownAxis[W]): Unit = {}
}
case class KN[W <: Int with Singleton](n: W) extends KnownAxis[W]
KN(1)
For 2.12
import shapeless.syntax.singleton._
...
KN(1.narrow)