Val
Val

Reputation: 22817

Typescript: deep keyof of a nested object, with related type

I'm looking for a way to have all keys / values pair of a nested object.

(For the autocomplete of MongoDB dot notation key / value type)

interface IPerson {
    name: string;
    age: number;
    contact: {
        address: string;
        visitDate: Date;
    }
}

Here is what I want to achieve, to make it becomes:

type TPerson = {
    name: string;
    age: number;
    contact: { address: string; visitDate: Date; }
    "contact.address": string;
    "contact.visitDate": Date;
}

What I have tried:

In this answer, I can get the key with Leaves<IPerson>. So it becomes 'name' | 'age' | 'contact.address' | 'contact.visitDate'.

And in another answer from @jcalz, I can get the deep, related value type, with DeepIndex<IPerson, ...>.

Is it possible to group them together, to become type like TPerson?

Modified 9/14: The use cases, need and no need:

When I start this question, I was thinking it could be as easy as something like [K in keyof T]: T[K];, with some clever transformation. But I was wrong. Here is what I need:

1. Index Signature

So the interface

interface IPerson {
    contact: {
        address: string;
        visitDate: Date;
    }[]
}

becomes

type TPerson = {
    [x: `contact.${number}.address`]: string;
    [x: `contact.${number}.visitDate`]: Date;
    contact: {
        address: string;
        visitDate: Date;
    }[];
}

No need to check for valid number, the nature of Array / Index Signature should allow any number of elements.

2. Tuple

The interface

interface IPerson {
    contact: [string, Date]
}

becomes

type TPerson = {
    [x: `contact.0`]: string;
    [x: `contact.1`]: Date;
    contact: [string, Date];
}

Tuple should be the one which cares about valid index numbers.

3. Readonly

readonly attributes should be removed from the final structure.

interface IPerson {
    readonly _id: string;
    age: number;
    readonly _created_date: Date;
}

becomes

type TPerson = {
    age: number;
}

The use case is for MongoDB, the _id, _created_date cannot be modified after the data has been created. _id: never is not working in this case, since it will block the creation of TPerson.

4. Optional

interface IPerson {
    contact: {
        address: string;
        visitDate?: Date;
    }[];        
}

becomes

type TPerson = {
    [x: `contact.${number}.address`]: string;
    [x: `contact.${number}.visitDate`]?: Date;
    contact: {
        address: string;
        visitDate?: Date;
    }[];
}

It's sufficient just to bring the optional flags to transformed structure.

5. Intersection

interface IPerson {
    contact: { address: string; } & { visitDate: Date; }
}

becomes

type TPerson = {
    [x: `contact.address`]: string;
    [x: `contact.visitDate`]?: Date;
    contact: { address: string; } & { visitDate: Date; }
}

6. Possible to Specify Types as Exception

The interface

interface IPerson {
    birth: Date;
}

becomes

type TPerson = {
    birth: Date;
}

not

type TPerson = {
    age: Date;
    "age.toDateString": () => string;
    "age.toTimeString": () => string;
    "age.toLocaleDateString": {
    ...
}

We can give a list of Types to be the end node.

Here is what I don't need:

  1. Union. It could be too complex with it.
  2. Class related keyword. No need to handle keywords ex: private / abstract .
  3. All the rest I didn't write it here.

Upvotes: 17

Views: 11027

Answers (2)

In order to achieve this goal we need to create permutation of all allowed paths. For example:

type Structure = {
    user: {
        name: string,
        surname: string
    }
}

type BlackMagic<T>= T

// user.name | user.surname
type Result=BlackMagic<Structure>

Problem becomes more interesting with arrays and empty tuples.

Tuple, the array with explicit length, should be managed in this way:

type Structure = {
    user: {
        arr: [1, 2],
    }
}

type BlackMagic<T> = T

// "user.arr" | "user.arr.0" | "user.arr.1"
type Result = BlackMagic<Structure>

Logic is straitforward. But how we can handle number[]? There is no guarantee that index 1 exists.

I have decided to use user.arr.${number}.

type Structure = {
    user: {
        arr: number[],
    }
}

type BlackMagic<T> = T

// "user.arr" | `user.arr.${number}`
type Result = BlackMagic<Structure>

We still have 1 problem. Empty tuple. Array with zero elements - []. Do we need to allow indexing at all? I don't know. I decided to use -1.

type Structure = {
    user: {
        arr: [],
    }
}

type BlackMagic<T> = T

//  "user.arr" | "user.arr.-1"
type Result = BlackMagic<Structure>

I think the most important thing here is some convention. We can also use stringified `"never". I think it is up to OP how to handle it.

Since we know how we need to handle different cases we can start our implementation. Before we continue, we need to define several helpers.

type Values<T> = T[keyof T]
{
    // 1 | "John"
    type _ = Values<{ age: 1, name: 'John' }>
}

type IsNever<T> = [T] extends [never] ? true : false;
{
    type _ = IsNever<never> // true 
    type __ = IsNever<true> // false
}

type IsTuple<T> =
    (T extends Array<any> ?
        (T['length'] extends number
            ? (number extends T['length']
                ? false
                : true)
            : true)
        : false)
{
    type _ = IsTuple<[1, 2]> // true
    type __ = IsTuple<number[]> // false
    type ___ = IsTuple<{ length: 2 }> // false
}

type IsEmptyTuple<T extends Array<any>> = T['length'] extends 0 ? true : false
{
    type _ = IsEmptyTuple<[]> // true
    type __ = IsEmptyTuple<[1]> // false
    type ___ = IsEmptyTuple<number[]> // false

}

I think naming and tests are self explanatory. At least I want to believe :D

Now, when we have all set of our utils, we can define our main util:

/**
 * If Cache is empty return Prop without dot,
 * to avoid ".user"
 */
type HandleDot<
    Cache extends string,
    Prop extends string | number
    > =
    Cache extends ''
    ? `${Prop}`
    : `${Cache}.${Prop}`

/**
 * Simple iteration through object properties
 */
type HandleObject<Obj, Cache extends string> = {
    [Prop in keyof Obj]:
    // concat previous Cacha and Prop
    | HandleDot<Cache, Prop & string>
    // with next Cache and Prop
    | Path<Obj[Prop], HandleDot<Cache, Prop & string>>
}[keyof Obj]

type Path<Obj, Cache extends string = ''> =
    // if Obj is primitive
    (Obj extends PropertyKey
        // return Cache
        ? Cache
        // if Obj is Array (can be array, tuple, empty tuple)
        : (Obj extends Array<unknown>
            // and is tuple
            ? (IsTuple<Obj> extends true
                // and tuple is empty
                ? (IsEmptyTuple<Obj> extends true
                    // call recursively Path with `-1` as an allowed index
                    ? Path<PropertyKey, HandleDot<Cache, -1>>
                    // if tuple is not empty we can handle it as regular object
                    : HandleObject<Obj, Cache>)
                // if Obj is regular  array call Path with union of all elements
                : Path<Obj[number], HandleDot<Cache, number>>)
            // if Obj is neither Array nor Tuple nor Primitive - treat is as object    
            : HandleObject<Obj, Cache>)
    )

// "user" | "user.arr" | `user.arr.${number}`
type Test = Extract<Path<Structure>, string>

There is small issue. We should not return highest level props, like user. We need paths with at least one dot.

There are two ways:

  • extract all props without dots
  • provide extra generic parameter for indexing the level.

Two options are easy to implement.

Obtain all props with dot (.):

type WithDot<T extends string> = T extends `${string}.${string}` ? T : never

While above util is readable and maintainable, second one is a bit harder. We need to provide extra generic parameter in both Path and HandleObject. See this example taken from other question / article:

type KeysUnion<T, Cache extends string = '', Level extends any[] = []> =
  T extends PropertyKey ? Cache : {
    [P in keyof T]:
    P extends string
    ? Cache extends ''
    ? KeysUnion<T[P], `${P}`, [...Level, 1]>
    : Level['length'] extends 1 // if it is a higher level - proceed
    ? KeysUnion<T[P], `${Cache}.${P}`, [...Level, 1]>
    : Level['length'] extends 2 // stop on second level
    ? Cache | KeysUnion<T[P], `${Cache}`, [...Level, 1]>
    : never
    : never
  }[keyof T]

Honestly, I don't think it will be easy for any one to read this.

We need to implement one more thing. We need to obtain a value by computed path.


type Acc = Record<string, any>

type ReducerCallback<Accumulator extends Acc, El extends string> =
    El extends keyof Accumulator ? Accumulator[El] : Accumulator

type Reducer<
    Keys extends string,
    Accumulator extends Acc = {}
    > =
    // Key destructure
    Keys extends `${infer Prop}.${infer Rest}`
    // call Reducer with callback, just like in JS
    ? Reducer<Rest, ReducerCallback<Accumulator, Prop>>
    // this is the last part of path because no dot
    : Keys extends `${infer Last}`
    // call reducer with last part
    ? ReducerCallback<Accumulator, Last>
    : never

{
    type _ = Reducer<'user.arr', Structure> // []
    type __ = Reducer<'user', Structure> // { arr: [] }
}

You can find more information about using Reducein my blog.

Whole code:

type Structure = {
    user: {
        tuple: [42],
        emptyTuple: [],
        array: { age: number }[]
    }
}


type Values<T> = T[keyof T]
{
    // 1 | "John"
    type _ = Values<{ age: 1, name: 'John' }>
}

type IsNever<T> = [T] extends [never] ? true : false;
{
    type _ = IsNever<never> // true 
    type __ = IsNever<true> // false
}

type IsTuple<T> =
    (T extends Array<any> ?
        (T['length'] extends number
            ? (number extends T['length']
                ? false
                : true)
            : true)
        : false)
{
    type _ = IsTuple<[1, 2]> // true
    type __ = IsTuple<number[]> // false
    type ___ = IsTuple<{ length: 2 }> // false
}

type IsEmptyTuple<T extends Array<any>> = T['length'] extends 0 ? true : false
{
    type _ = IsEmptyTuple<[]> // true
    type __ = IsEmptyTuple<[1]> // false
    type ___ = IsEmptyTuple<number[]> // false
}

/**
 * If Cache is empty return Prop without dot,
 * to avoid ".user"
 */
type HandleDot<
    Cache extends string,
    Prop extends string | number
    > =
    Cache extends ''
    ? `${Prop}`
    : `${Cache}.${Prop}`

/**
 * Simple iteration through object properties
 */
type HandleObject<Obj, Cache extends string> = {
    [Prop in keyof Obj]:
    // concat previous Cacha and Prop
    | HandleDot<Cache, Prop & string>
    // with next Cache and Prop
    | Path<Obj[Prop], HandleDot<Cache, Prop & string>>
}[keyof Obj]

type Path<Obj, Cache extends string = ''> =
    (Obj extends PropertyKey
        // return Cache
        ? Cache
        // if Obj is Array (can be array, tuple, empty tuple)
        : (Obj extends Array<unknown>
            // and is tuple
            ? (IsTuple<Obj> extends true
                // and tuple is empty
                ? (IsEmptyTuple<Obj> extends true
                    // call recursively Path with `-1` as an allowed index
                    ? Path<PropertyKey, HandleDot<Cache, -1>>
                    // if tuple is not empty we can handle it as regular object
                    : HandleObject<Obj, Cache>)
                // if Obj is regular  array call Path with union of all elements
                : Path<Obj[number], HandleDot<Cache, number>>)
            // if Obj is neither Array nor Tuple nor Primitive - treat is as object    
            : HandleObject<Obj, Cache>)
    )

type WithDot<T extends string> = T extends `${string}.${string}` ? T : never


// "user" | "user.arr" | `user.arr.${number}`
type Test = WithDot<Extract<Path<Structure>, string>>



type Acc = Record<string, any>

type ReducerCallback<Accumulator extends Acc, El extends string> =
    El extends keyof Accumulator ? Accumulator[El] : El extends '-1' ? never : Accumulator

type Reducer<
    Keys extends string,
    Accumulator extends Acc = {}
    > =
    // Key destructure
    Keys extends `${infer Prop}.${infer Rest}`
    // call Reducer with callback, just like in JS
    ? Reducer<Rest, ReducerCallback<Accumulator, Prop>>
    // this is the last part of path because no dot
    : Keys extends `${infer Last}`
    // call reducer with last part
    ? ReducerCallback<Accumulator, Last>
    : never

{
    type _ = Reducer<'user.arr', Structure> // []
    type __ = Reducer<'user', Structure> // { arr: [] }
}

type BlackMagic<T> = T & {
    [Prop in WithDot<Extract<Path<T>, string>>]: Reducer<Prop, T>
}

type Result = BlackMagic<Structure>

Playground

This implementation is worth considering

Upvotes: 8

jcalz
jcalz

Reputation: 330436

Below is the full implementation I have of Flatten<T, O> which transforms a type possibly-nested T into a "flattened" version whose keys are the dotted paths through the original T. The O type is an optional type where you can specify a (union of) object type(s) to leave as-is without flattening them. In your example, this is just Date, but you could have other types.

Warning: it's hideously ugly and probably fragile. There are edge cases all over the place. The pieces that make it up involve weird type manipulations that either don't always do what one might expect, or are impenetrable to all but the most seasoned TypeScript veterans, or both.

In light of that, there is no such thing as a "canonical" answer to this question, other than possibly "please don't do this". But I'm happy to present my version.

Here it is:


type Flatten<T, O = never> = Writable<Cleanup<T>, O> extends infer U ?
    U extends O ? U : U extends object ?
    ValueOf<{ [K in keyof U]-?: (x: PrefixKeys<Flatten<U[K], O>, K, O>) => void }>
    | ((x: U) => void) extends (x: infer I) => void ?
    { [K in keyof I]: I[K] } : never : U : never;

The basic approach here is to take your T type, and return it as-is if it's not an object or if it extends O. Otherwise, we remove any readonly properties, and transform any arrays or tuples into a version without all the array methods (like push() and map()) and get U. We then flatten each property in that. We have a key K and a flattened property Flatten<U[K]>; we want to prepend the key K to the dotted paths in Flatten<U[K]>, and when we're done with all that we want to intersect these flattened objects (with the unflattened object too) all together to be one big object.

Note that convincing the compiler to produce an intersection involves conditional type inference in contravariant positions (see Transform union type to intersection type), which is where those (x: XXX) => void) and extends (x: infer I) => void pieces come in. It makes the compiler take all the different XXX values and intersect them to get I.

And while an intersection like {foo: string} & {bar: number} & {baz: boolean} is what we want conceptually, it's uglier than the equivalent {foo: string; bar: number; baz: boolean} so I do some more conditional type mapping with { [K in keyof I]: I[K] } instead of just I (see How can I see the full expanded contract of a Typescript type?).

This code generally distributes over unions, so optional properties may end up spawning unions (like {a?: {b: string}} could produce {"a.b": string; a?: {b: string}} | {"a": undefined, a?: {b: string}}, and while this might not be the representation you were going for, it should work (since, for example, "a.b" might not exist as a key if a is optional).


The Flatten definition depends on helper type functions that I will present here with various levels of description:

type Writable<T, O> = T extends O ? T : {
    [P in keyof T as IfEquals<{ [Q in P]: T[P] }, { -readonly [Q in P]: T[P] }, P>]: T[P]
}

type IfEquals<X, Y, A = X, B = never> =
    (<T>() => T extends X ? 1 : 2) extends
    (<T>() => T extends Y ? 1 : 2) ? A : B;

The Writable<T, O> returns a version of T with the readonly properties removed (unless T extends O in which case we leave it alone). It comes from TypeScript conditional types - filter out readonly properties / pick only required properties.

Next:

type Cleanup<T> =
    0 extends (1 & T) ? unknown :
    T extends readonly any[] ?
    (Exclude<keyof T, keyof any[]> extends never ?
        { [k: `${number}`]: T[number] } : Omit<T, keyof any[]>) : T;

The Cleanup<T> type turns the any type into the unknown type (since any really fouls up type manipulation), turns tuples into objects with just individual numericlike keys ("0" and "1", etc), and turns other arrays into just a single index signature.

Next:

type PrefixKeys<V, K extends PropertyKey, O> =
    V extends O ? { [P in K]: V } : V extends object ?
    { [P in keyof V as
        `${Extract<K, string | number>}.${Extract<P, string | number>}`]: V[P] } :
    { [P in K]: V };

PrefixKeys<V, K, O> prepends the key K to the path in V's property keys... unless V extends O or V is not an object. It uses template literal types to do so.

Finally:

type ValueOf<T> = T[keyof T]

turns a type T into a union of its properties. See Is there a `valueof` similar to `keyof` in TypeScript?.

Whew! 😅


So, there you go. You can verify how closely this conforms to your stated use cases. But it's very complicated and fragile and I wouldn't really recommend using it in any production code environment without a lot of testing.

Playground link to code

Upvotes: 10

Related Questions