Design Meeting Notes, 11/16/2018

[DanielRosenwasser]

Arbitrary Index Signatures

#26797

• Today we only support string and number index signatures.

• We didn't have symbol index signatures (though it used to be allowed in string index signatures)

• Now we have bigint - but we don't support those either.

• People asked for index signatures on enums; but instead we said you could have a mapped type over an enum.

  • However, once you learn about index signatures, that's the intuitive thing to use (even if they're technically less-powerful than a mapped type).

• As far as implementation goes, it is far reaching but it is reasonable.

  • Conceptually, they're quite similar to regular signatures; we just need to enforce a sort of priority.

• When you index with something x[y], you have to try to apply the type of y against every index signature of x and union it to get the result.

  • As the duality, when writing an assignment x[y] = z, each type is intersected in the assignment.

• There is some annoying behavior with these index signatures on literal types.

  interface Foo {
      [index: "a" | "b"]: string;
      [index: "c" | "d"]: number;
  }
  
  // Error! But what's the deal!
  class Bar implements Foo {
      a = "a";
      b = "b";
      c = 10;
      d = 20;
  }
  • Bar hasn't declared explicit index signatures.
  • But when you have index signatures over a closed union, people will expect these to be equivalent. Seems bad.

• What about "enum-yness"

  • Today, a mapped type over enums just flattens down to the numeric values that an enum maps to.

  • But with the current PR, we keep them distinct.

    enum SomeEnum {
        Zero,
        One,
    }
    interface Foo {
        [index: number]: number;
        [index: SomeEnum]: number;
    }

  • Here, we enforce that both of those index signatures are related to each other.

• This also doesn't perfectly map to index signatures; you can't toggle optionality with ?.

• If we started TypeScript over today, would we have both mapped types and index signatures?

  • Why have two ways to represent very similar concepts? Could we try to consolidate on mapped types?

• [[Discussion around syntax and computed property names]]

• "I think there's a story around index signatures for creating a set of known property names, and mapped types just being a template."