Lodash's mapValues and empty object type interact confusingly

[seansfkelley]

I had trouble coming up with a good name and search terms for this issue. It looks like it's triggering some heuristic that is misbehaving and is very strange. Please feel free to rename if you can come up with a better summary for it.

TypeScript Version: 3.3.1

Search Terms: mapvalues ternary side effect unused inference empty object type

Code

// Begin types from Lodash.
interface Dictionary<T> {
 [index: string]: T;
}

interface NumericDictionary<T> {
 [index: number]: T;
}

type ObjectIterator<TObject, TResult> = (value: TObject[keyof TObject], key: string, collection: TObject) => TResult;

type DictionaryIterator<T, TResult> = ObjectIterator<Dictionary<T>, TResult>;

// In lodash.d.ts this function has many overloads, but this seems to be the problematic one.
function mapValues<T, TResult>(obj: Dictionary<T> | NumericDictionary<T> | null | undefined, callback: DictionaryIterator<T, TResult>): Dictionary<TResult> {
 return null as any;
} 
// End types from Lodash.

interface Foo {
 foo: string;
} 

interface Bar {
 bar: string;
}

export function fooToBar(
 foos: Record<string, Foo>,
): Record<string, Bar | null> {
 // The next line, despite being unrelated to the other two lines, causes them to fail.
 // Comment it out to see the function compile erroneously.
 //
 // This behavior only happens when `result` is used as an intermediate: if you return 
 // the ternary directly, the function always compiles.
 //
 // I couldn't find another Lodash function for which this issue reproduced after sampling
 // a few. 
 const wat = mapValues(foos, f => f.foo);
 const result = foos == null ? {} : mapValues(foos, f => f.foo);
 // This line _should_ fail, because `result` is not the right type.
 return result;
}

Expected behavior: fails to compile without the const wat line present.

Actual behavior: only fails to compile when const wat is present.

Playground Link: link

Related Issues:

[jack-williams]

I think this might caused by the same issue in #29393.

[seansfkelley]

Jack Williams (@jack-williams) that seems likely, but I'll keep the issue open and leave it up to the maintainers to judge if that is the case.

[weswigham]

Oh boy. So this strangeness really is strange. All that differs from the wat line existing and not is the type ids we assign to the object type and the Dictionary instantiation. With wat, Dictionary<string> gets a lower type id, which makes the result union Dictionary<string> | {}, which when subtype reduces (correctly) to {}. When the wat line is gone, the ordering of the assigned type ids inverts, making the union we make {} | Dictionary<string>, which we then seem to say that subtype reduces to Dictionary<string> (which is clearly incorrect, {} should not be considered a subtype of a Dictionary<anything>, since a Dictionary has an index signature).

[jack-williams]

Ahh the type ids. An amateur mistake by me there to overlook that one, haha.

[weswigham]

It's probably also worth noting that I believe by our normal subtype reduction rules, result's type should always simplify to type {}, which is assignable to the return type, despite not having an index signature, and despite having something with potentially incompatible member types assigned to it - which is a manifestation of the aliasing problem we show elsewhere. :(

[seansfkelley]

Hah, what a bug. Thanks for the explanation Wesley Wigham (@weswigham)!