Mapped types break `for-in` loops with intersections

[gcnew]

TypeScript Version: nightly

The following used to work but is broken now.

Code

function extend<T, U>(base: T, ext: U): T & U {
    const retval = <T & U> {};

    for (const x in base) {
        // Error: Type 'T[keyof T]' is not assignable to type '(T & U)[keyof T]'.
        retval[x] = base[x];
    }

    for (const x in ext) {
        // Error: Type 'U[keyof U]' is not assignable to type '(T & U)[keyof U]'.
        retval[x] = ext[x];
    }

    return retval;
}

[ahejlsberg]

This is working as intended, the fact that it was previously allowed was a bug. A (T & U)[keyof T] is assignable to a T[keyof T], but not vice versa. Consider:

type A = {
    a: string;
    x: { s: string };
}
type B = {
    b: string;
    x: { t: string };
}

The intersection A & B corresponds to:

type AB = {
    a: string;
    b: string;
    x: { s: string, t: string };
}

Neither property x from A or B is assignable to property x in A & B because of the deep merging.

[gcnew]

I agree that in a deep setting the new behaviour is the correct one, however it is also a breaking change introduced in version 2.1 and should be marked as such.

[gcnew]

PS: shallow intersections seem to be more natural. Will they ever be supported? Is the deep behaviour actually used in RWC?

[gcnew]

The What's new documentation on intersection types features the very same extend example - nothing is mentioned about a deep nature of & and the provided code doesn't handle it either. It was also used for the old definitions of Object.assign (which is shallow). I find it a massive breaking change from a logical perspective. Maybe intersection types should be made shallow? It will be a fix for the current misleading/buggy behaviour.

[weswigham]

@gcnew If you want "shallow intersections" following the logic of Object.assign, you want spread types, and not really intersections.

[gcnew]

@weswigham Thank you!

The major problem I'm after is that basically all the information and marketing of intersection types are incomplete at best (leaning to wrong). This has created a flawed mental model for many of us.

The unfortunate state we are currently in should be very explicitly explained and fixed in the new documentation (even if the behaviour is changed to match the status-quo explanation). Also the TypeScript book and What's new should be retroactively updated.

I believe the revelation @ahejlsberg made is world breaking not only for me.

[ahejlsberg]

The original PR on intersection types, #3622, has all of the semantics described in a fair amount of detail.

[gcnew]

@ahejlsberg Yes, #3622 has the semantics described, however they are:

1. not really intelligible for average programmers
2. they are not covered in What's new / the TS book
3. even there the extends example is wrong
4. the TS team continuously used intersections in shallow situations

The XYZ example is the only hint, except for Properties and Index Signatures.

My proposal is to state the intended semantics clearly and add an appropriate notice on the pertinent resources. Otherwise people reading the official resources will still be led astray.

[aluanhaddad]

I'm not sure I understand the notion of shallow intersections.
I would expect the intersection to be as deep as the deepest common path between between the types in question.
Maybe I'm missing something fundamental, but I don't see what else it could be.

[gcnew]

It seems Flow has taken the shallow path with the left operand being a base type and the right one an extension. link

It is also what I had inferred form the TS documentation as nothing is said about intersections being deep and all the motivating examples being shallow. The inclusion of extend which is infamously shallow (and also implemented this way) doesn't help either. I believe this has been the reason for the many issues asking why the right hand intersection type doesn't override the properties on the left one.

From another perspective this is breaking the code in the official documentation. There are two routes of action, either the behaviour is changed to match the documentation, or the documentation is updated. In both cases there should be a notice clearing up the confusion and all the relevant parts updated.