valibot icon indicating copy to clipboard operation
valibot copied to clipboard

using `union` as key of `record` adds `undefined` to the value type

Open bug-brain opened this issue 11 months ago • 13 comments

import * as v from 'valibot';

const record = v.record(v.union([v.string()]), v.string());

const t: v.Output<typeof record> = { a: undefined }
//    ^ { [x: string]: string | undefined; }
//                              ^ should not be here
const result = v.safeParse(record, t);
console.log(result); // errors correctly

bug-brain avatar Mar 19 '24 11:03 bug-brain

This is the intended behavior when string is not used directly as the key schema. In your case, I recommend removing union and using only string. This should solve your problem. union can be used with literal to describe certain record keys, but record does not force you to specify them and therefore we mark it as optional in this case. Feel free to contact me if you have any further questions.

fabian-hiller avatar Mar 19 '24 15:03 fabian-hiller

But please leave this issue open. I agree that it would be better to specify it as undefined only when literal is used inside of union. I will see if I can improve the types.

fabian-hiller avatar Mar 19 '24 15:03 fabian-hiller

How would I model the type Record<"a" | "b", number>? I would have expected v.record(v.union([v.literal("a"), v.literal("b")]), v.number()) but that makes the keys optional and adds undefined as you mentioned. The function required only works on ObjectSchema.

bug-brain avatar Mar 19 '24 20:03 bug-brain

I would use object and not record:

import * as v from 'valibot';

const Schema = v.object({ a: v.number(), b: v.number() });

But maybe you are right and our current typing and implementation is wrong and we should only mark it as optional when optional is used for the value of the record.

fabian-hiller avatar Mar 20 '24 16:03 fabian-hiller

@bug-brain I investigated this issue further. It is true that the behavior of Valibot is different from TypeScript and it would be nice to unify it. But there is a tradeoff. To support this behavior, record has to "understand" the schemas used and apply additional logic to ensure that every required entry is defined. This would increase the raw bundle size of the record function from 408 bytes to about 460 bytes (13%). Furthermore, this change would require us to limit the record key schema to string and picklist, as it would require much more code to handle other more "unpredictable" schemas such as union. That's probably also why Zod implemented it this way.

fabian-hiller avatar Apr 16 '24 04:04 fabian-hiller

Since many Zod users are also complaining about this problem in #2623, I will investigate this further to find an appropriate solution.

fabian-hiller avatar Apr 16 '24 23:04 fabian-hiller

I think I fixed it! Will test and investigate my code tomorrow. Thanks for bringing this up!

fabian-hiller avatar Apr 17 '24 03:04 fabian-hiller

Even if I found a solution, I am not sure if we should go this way as a library focused on bundle size. Supporting this behavior increases the bundle size of record, while the same behavior can already be achieved with object. Maybe we should limit the key argument of record to just string (or stick to the current behaviour with undefined) and provide a util function to convert a list of keys into the entries argument of object.

// `record` does only allow `string` as first argument
const Schema1 = v.record(v.string(), v.number());

// For explicit keys, `object` with `entriesFrom` is used instead
const Schema2 = v.object(v.entriesFrom(['foo', 'bar'], v.number()));

// Or without the util function
const Schema2 = v.object({ foo: v.number(), bar: v.number() });

fabian-hiller avatar Apr 17 '24 15:04 fabian-hiller

@fabian-hiller may I ask how much record would increase in bundle size?

macmillen avatar Apr 26 '24 10:04 macmillen

I am using a graphql-codegen-typescript-validation-schema to generate zod enums like this:

export const PosTypeEnumSchema = z.enum(['external', 'online', 'standard', 'telephone']);

That way I don't need to define all of my enum schemas manually.

Now I need a record with these enums as a key type and I don't want to re-define these keys by hand but instead use the generated z.enum. (the alternatives you showed unfortunately don't solve that issue)

macmillen avatar Apr 26 '24 11:04 macmillen

The next version of Valibot will allow you to write the following code to solve your problem. Feel free to give me feedback on this in the long run.

import * as v from 'valibot';

const ListSchema = v.picklist(['external', 'online', 'standard', 'telephone']);

const ObjectSchema = v.object(v.entriesFromList(ListSchema.options, v.number()));

fabian-hiller avatar Apr 26 '24 15:04 fabian-hiller

I am not a maintainer of Zod, so I cannot speak for it. I recommend asking the Zod team directly.

fabian-hiller avatar Apr 26 '24 21:04 fabian-hiller

Thank you, I'll do that :)

macmillen avatar Apr 26 '24 21:04 macmillen

The next version of Valibot will allow you to write the following code to solve your problem

This is now implemented and available.

fabian-hiller avatar May 22 '24 00:05 fabian-hiller

Good time! @fabian-hiller, the problem is still relevant for me. We cannot legally make numbers or symbols the keys of an object. entriesFromList accepts an array of strings. Therefore, you have to make crutches (I have attached the minimum code below). But in this case, the keys become optional. How can I fix this?

Sorry, but my opinion is that there is no need to come up with additional constructions, such as entriesFromList, when a similar utility from typescript contradicts the behavior of "record". No matter how difficult it may be, you need to try to do it correctly.

import { InferOutput, number, picklist, pipe, record, string, transform } from 'valibot';

export const rows = [0, 1, 2, 3, 4] as const;

// eslint-disable-next-line @typescript-eslint/no-explicit-any, @typescript-eslint/no-unsafe-argument
export const isRow = (input: any): input is (typeof rows)[number] => rows.includes(input);

export const RowScheme = picklist(rows);

export type Row = InferOutput<typeof RowScheme>;
// type Rows = 0 | 1 | 2 | 3 | 4

/*
  The input type for a "record" must be a string, although the output type can be a number or a symbol.
  Therefore, you have to do redundant transformation.
*/
export const RowNumberSchema = pipe(
  string(),
  transform<string, (typeof rows)[number]>((input) => {
    if (isRow(input)) return input;
    throw new Error(`${input} cannot be assigned to rows`);
  }),
);

// similar type to Row
export type RowNumber = InferOutput<typeof RowNumberSchema>;
// type RowsNumber = 0 | 1 | 2 | 3 | 4

export const RowOptionsSchema = record(RowNumberSchema, number());

export type RowOptions = InferOutput<typeof RowOptionsSchema>;
/*
  type RowOptions = {
    0?: number | undefined;
    1?: number | undefined;
    2?: number | undefined;
    3?: number | undefined;
    4?: number | undefined;
  }
*/

ksv90 avatar Jul 04 '24 04:07 ksv90

We cannot legally make numbers or symbols the keys of an object. entriesFromList accepts an array of strings.

I think we can extend it to numbers and symbols. I will investigate this.

Sorry, but my opinion is that there is no need to come up with additional constructions, such as entriesFromList, when a similar utility from typescript contradicts the behavior of "record". No matter how difficult it may be, you need to try to do it correctly.

This is an open source project. You are free to investigate the implementation and provide a proof of concept that aligns with our philosophy. If everything looks good, I will be happy to merge your changes.

fabian-hiller avatar Jul 04 '24 09:07 fabian-hiller

v0.36.0 is available

fabian-hiller avatar Jul 05 '24 14:07 fabian-hiller

Thank you very much for your promptness. Now this also works with numbers and symbols. I'll try to study the issue. If I have a proposal, I will publish it.

ksv90 avatar Jul 05 '24 15:07 ksv90