Collection {coerceKey, coerceValue}

This proposal seeks to add a coerceKey and coerceValue parameter to collection creation.

Rendered Spec

Motivation

Various data structures are expected to have consistent constraints and or normalization applied to their keys and/or values when being used after creation. In order to facilitate normalization be guaranteed for such data types it needs to be applied prior to being inserted into the data structure. The current workflow is to manually ensure that at all locations which insert data into the structure properly normalize the data.

This leads to various issues such as potentially fragile boilerplate problems such as; normalization not applied to all insertion locations or normalization not applied in consistent manners across insertion locations.

This proposal does add a new capability of guarantees around the stored values in these data structures. Currently calling various built ins methods can avoid any mechanism to intercept and normalize data as it is inserted by calling methods on a structure like Map.prototype.set.call(mapOfStringsToStrings, number) which prevents guarantees about a Map having a consistent set of guarantees about the data it contains.

Various examples of structures that normalize data as it is inserted exist in the wild:

Converts data to strings for both parameter name and value.

Validates and converts data to strings for both header name and value.

Converts data to strings for both header name and value.

Clamps the key when appending to length + 1 by using append.

Performs validation on key.

All of these do similar effects using differing ways of applying data type coercion and validation of data being inserted.

In order to unify how to define these general operations on dealing with incoming data and to avoid boilerplate issues and fragile base class issues as mentioned in the FAQ a hook to intercept data would alleviate some issues.

Use cases

Specialized maps

Given an application with User Objects it may be desirable to create collections based upon username and email for separate purposes.

new Map(undefined, {
  coerceKey({email}) {
    return email;
  },
  coerceValue(state) {
    return state instanceof AccountState ? 
      state :
      new AccountState(state);
  }
});

new Set(undefined, {
  coerceValue({username}) {
    return username;
  }
});

Checked keys

It is a common occurrence to want to check types when performing operations on collections. This can be done during keying.

new Map(undefined, {
  coerceKey(user) {
    if (user instanceof User !== true) {
      throw new TypeError('Expected User for key');
    }
    return user;
  }
});

DOM Example: Headers..set

  1. Normalize value.
  1. If name is not a name or value is not a value, then throw a TypeError.
  1. If this’s guard is "immutable", then throw a TypeError.
  1. Otherwise, if this’s guard is "request" and name is a forbidden header name, return.
  1. Otherwise, if this’s guard is "request-no-cors" and name/value is not a no-CORS-safelisted request-header, return.
  1. Otherwise, if this’s guard is "response" and name is a forbidden response-header name, return.
  1. Set name/value in this’s header list.
  1. If this’s guard is "request-no-cors", then remove privileged no-CORS request headers from this.

All of these can be interacted with if a hook can intercept name and value as it comes in. The general workflow is to store a name and value in the backing header list storage.

DOM Example: SearchParams..has

The has(name) method steps are to return true if there is a name-value pair whose name is name in this’s list, and false otherwise.

Note, all normalization is done in WebIDL layer here to convert name to a USVString, there is no actual text stating that normalization is done for name.

FAQ

How do other languages handle this customized keying?

A collection of references can be found via this document.

Generally it falls into using container types. If you wanted to create a Map of People by person.email. You would implement a wrapper class PersonByEmail to use as your key, and others for keying of other aspects. Static typing and compiler/language enforced coercion can alleviate problems with misusing collections, but wrapping and unwrapping is manual in scenarios with dynamic typing that cannot be coerced automatically.

This proposal would provide a hook to do that manual wrapping and unwrapping without requiring the user of a collection to remain vigilant about properly marshalling keys before providing them to the collection.

When are the normalization steps applied?

Normalization is applied when data is incoming to find the identity of the key location in [[MapData]] and when placing the value in [[SetData]] or [[MapData]]. e.g.

const map = new Map([], {
  coerceKey: String
});
// stored using { [[Key]]: "1", [[Value]]: "one" } in map.[[MapData]]
map.set(1, 'one');
// looks for corresponding { [[Key]]: "1" } in map.[[MapData]]
map.has(1); // true
// functions directly exposing the underlying entry list are unaffected
[...map.entries()]; // [["1", "one"]]

const set = new Set([], {coerceValue: JSON.stringify});
// stored using { [[Value]]: '{"path": "/foo"}' } in set.[[SetData]]
set.add({path: '/foo'});
// looks for corresponding { [[Value]]: '{"path": "/foo"}' } in set.[[SetData]]
set.has({path: '/foo')};
// functions directly exposing the underlying entry list are unaffected
[...set]; // ['{"path": "/foo"}']

Normalization is not done when iterating or returning internal data, it is only done on parameters.

Why would someone want to use coerceValue with Map?

A variety of use cases exist to normalize the values of map like structures in different APIs. Even if they do not directly use Map, we can se the utility of this pattern from existing DOM APIs.

  • URLSearchParams
  • DomStringMap
  • Header

Node also has APIs that also normalize values such as process.env.

Normalizing values can avoid certain situations as well such as putting invalid values into a Map by either validation errors, coercion, or other means. Existing map like structures such as require.cache can produce error if you put the incorrect values in them. A normalization step allows the map to properly handle situations when unexpected values are inserted.

Why are Sets only given coerceValue?

Sets are collections of values, and do not have a mapping operation from one value to another.

Why not call it coerceKey for Sets?

An audit of other languages was done with their documentation and APIs concerning Sets. The majority of terminology used was "elements"; however, the terms "keys" and "values" were also used. It was noted that whenever "keys" was used "values" was also used, but when "values" was used it did not always also use "keys". To match existing JS usage of terms "values" was chosen as the base for this name.

Why not value[Symbol.coerceKey]?

Having specialized identity conflicts with the idea of having multiple kinds of specialized maps per type of value. It also would cause conflicts when wanting to specialize keys that are based upon primitives.

Why not encourage extending collections?

  1. This would be succeptible to prototype crawling such as:
myCustomMap.__proto__.get.call(myCustomMap, key);

which would somewhat invalidate the idea of checking types of keys.

  1. It prevents needing to synchronize all of the methods which is a fair amount of boilerplate and potential place for code going out of sync. It also means that your custom implementation will work even if new methods are added to collections in the JS standard library:
class MyMap extends Map {
  constructor([...entries]) {
    super(entries.map(...));
  }
  delete(k) { ... }
  get(k) { ... }
  has(k) { ... }
  set(k, v) { ... }
}

If we add something like emplace() this code now needs to be updated or it will have bugs if people expect it to work like a standard Map.

This is roughly the fragile base class problem, where Map is the base class.

  1. Even if this is a userland solution, it seems prudent to allow easier usage of maps. We should aim to alleviate developers without requiring that all new features have new kernel semantics. I spoke of this with respect to expanding the standard library.

  2. Composition, while extending is nice it doesn't always allow for simple chaining and composition of features. If we introduce RekeyableMap as a concrete base class it may conflict with other base classes that may be introduced like if there was InsertIfMissingMap. Since both are base classes it would not allow both features to be combined easily.