[Pitch] Allow coding of non-`String`/`Int` keyed `Dictionary` into a `KeyedContainer`

Hi all,
I would like to present a pitch to solving an issue with the current state of Dictionary encoding and decoding.

Many thanks to @itaiferber for providing input and feedback, for revising the pitch and for helping me shape the overall direction.


Pitch: Allow coding of non- String / Int keyed Dictionary into a KeyedContainer

Introduction

The current conformance of Swift's Dictionary to the Codable protocols has a somewhat-surprising limitation in that dictionaries whose key type is not String or Int (values directly representable in CodingKey types) encode not as KeyedContainers but as UnkeyedContainers. This behavior has caused much confusion for users and I would like to offer a way to improve the situation.

Motivation

The primary motivation for this pitch lays in the much-discussed confusion of this default behavior:

The common situations where people have found the behavior confusing include:

  • Using enums as keys (especially when RawRepresentable, and backed by String or Int types)
  • Using String wrappers (like the generic Tagged library or custom wrappers) as keys
  • Using Int8 or other Int* flavours as keys

In the various discussions, there are clear and concise explanations for this behavior, but it is also mentioned that supporting encoding of RawRepresentable String and Int keys into keyed containers may indeed be considered to be a bug, and is an oversight in the implementation (JSON Encoding / Decoding weird encoding of dictionary with enum values, reply by Itai Ferber).

There's a bug at bugs.swift.org tracking the issue: SR-7788

Unfortunately, it is too late to change the behavior now:

  1. It is a breaking change with respect to existing behavior, with backwards-compatibility ramifications (new code couldn't decode old data and vice versa), and
  2. The behavior is tied to the Swift stdlib, so the behavior would differ between consumers of the code and what OS versions they are on

Instead, I propose the addition of a new protocol to the standard library. Opting in to this protocol for the key type of a Dictionary will allow the Dictionary to encode/decode to/from a KeyedContainer.

Proposed Solution

I propose adding a new protocol to the standard library: CodingKeyRepresentable

Types conforming to CodingKeyRepresentable indicate that they can be represented by a CodingKey instance (which they can offer), allowing them to opt in to having dictionaries use their CodingKey representations in order to encode into KeyedContainers.

The opt-in can only happen for a version of Swift where the protocol is available, so the user will be in full control of the situation. For instance I am currently using my own workaround, but once I only support iOS versions running a specific future Swift version with this feature, I could skip my own workaround and rely on this behavior instead.

I have a draft PR for the proposed solution: #34458

Examples

// Same as stdlib's _DictionaryCodingKey
struct _AnyCodingKey: CodingKey {
    let stringValue: String
    let intValue: Int?
    
    init?(stringValue: String) {
        self.stringValue = stringValue
        self.intValue = Int(stringValue)
    }
    
    init?(intValue: Int) {
        self.stringValue = "\(intValue)"
        self.intValue = intValue
    }
}

struct ID: Hashable, CodingKeyConvertible {
    static let knownID1 = ID(stringValue: "<some-identifier-1>")
    static let knownID2 = ID(stringValue: "<some-identifier-2>")
    
    let stringValue: String
    
    var codingKey: CodingKey {
        return _AnyCodingKey(stringValue: stringValue)
    }
    
    init?(codingKey: CodingKey) {
        stringValue = codingKey.stringValue
    }
    
    init(stringValue: String) {
        self.stringValue = stringValue
    }
}

let data: [ID: String] = [
    .knownID1: "...",
    .knownID2: "...",
]

let encoder = JSONEncoder()
try String(data: encoder.encode(data), encoding: .utf8)

/*
{
    "<some-identifier-1>": "...",
    "<some-identifier-2>": "...",
}
*/

Detailed Design

The proposed solution adds a new protocol, CodingKeyRepresentable:

/// Indicates that the conforming type can provide a `CodingKey` to be used when
/// encoding into a keyed container.
public protocol CodingKeyRepresentable {
    var codingKey: CodingKey { get }
    init?(codingKey: CodingKey)
}

In the conditional Encodable conformance on Dictionary, the following extra case can handle such conforming types:

    } else if let _ = Key.self as? CodingKeyRepresentable.Type {
      // Since the keys are CodingKeyRepresentable, we can use the `codingKey`
      // to create `_DictionaryCodingKey` instances.
      var container = encoder.container(keyedBy: _DictionaryCodingKey.self)
      for (key, value) in self.dict {
        let codingKey = (key as! CodingKeyRepresentable).codingKey
        let dictionaryCodingKey = _DictionaryCodingKey(codingKey: codingKey)
        try container.encode(value, forKey: dictionaryCodingKey)
      }
    } else {
      // Keys are Encodable but not Strings or Ints, so we cannot arbitrarily

In the conditional Decodable conformance on Dictionary, we can similarly handle conforming types:

    } else if let codingKeyRepresentableType = Key.self as? CodingKeyRepresentable.Type {
      // The keys are CodingKeyRepresentable, so we should be able to expect a keyed container.
      let container = try decoder.container(keyedBy: _DictionaryCodingKey.self)
      for key in container.allKeys {
        let value = try container.decode(Value.self, forKey: key)
        let k = codingKeyRepresentableType.init(codingKey: key)
        self.dict[k as! Key] = value
      }
    } else {
      // We should have encoded as an array of alternating key-value pairs.

Impact on Existing Code

No direct impact, since adoption of this protocol is additive.

However, special care must be taken in adopting the protocol, since adoption on any type T which has previously been encoded as a dictionary key can introduce backwards incompatibility with archives. It is always safe to adopt CodingKeyConvertible on new types, or types newly-conforming to Codable.

Other Considerations

Conforming stdlib types to CodingKeyRepresentable

Along the above lines, we do not propose conforming any existing stdlib or Foundation type to CodingKeyRepresentable due to backwards-compatibility concerns. Should end-user code require this conversion on existing types, we recommend writing wrapper types which conform on those types' behalf (for example, a MyUUIDWrapper which contains a UUID and conforms to CodingKeyRepresentable to allow using UUIDs as dictionary keys directly).

Adding an AnyCodingKey type to the standard library

Since types that conform to CodingKeyRepresentable will need to supply a CodingKey, most likely generated dynamically from type contents, this may be a good time to introduce a general key type which can take on any String or Int value it is initialized from.

Dictionary already uses exactly such a key type internally (_DictionaryCodingKey), as do JSONEncoder / JSONDecoder with _JSONKey (and PropertyListEncoder / PropertyListDecoder with _PlistKey), so generalization could be useful. The implementation of this type could match the implementation of _AnyCodingKey provided above.

Alternatives Considered

Why not just make the type conform to CodingKey directly?

For two reasons:

  1. In the rare case in which a type already conforms to CodingKey, this runs the risk of behavior-breaking changes
  2. CodingKey requires exposure of a stringValue and intValue property, which are only relevant when encoding and decoding; forcing types to expose these properties arbitrarily seems unreasonable

Why not refine RawRepresentable, or use a RawRepresentable where RawValue == CodingKey constraint?

RawRepresentable conformance for types indicates a lossless conversion between the source type and its underlying RawValue type; this conversion is often the "canonical" conversion between a source type and its underlying representation, most commonly between enums backed by raw values, and option sets similarly backed by raw values.

In contrast, we expect conversion to and from CodingKey to be incidental , and representative only of the encoding and decoding process. We wouldn't suggest (or expect) a type's canonical underlying representation to be a CodingKey, which is what a protocol CodingKeyRepresentable: RawRepresentable where RawValue == CodingKey would require. Similarly, types which are already RawRepresentable with non- CodingKey raw values couldn't adopt conformance this way, and a big impetus for this feature is allowing Int- and String-backed enums to participate as dictionary coding keys.

Add workarounds to each Encoder / Decoder

Following a suggestion from @itaiferber, I have previously tried to provide a solution to this issue — not in general, but instead solving it by providing a DictionaryKeyEncodingStrategy for JSONEncoder : #26257

The idea there was to be able to express an opt-in to the new behavior directly in the JSONEncoder and JSONDecoder types by venting a new encoding/decoding 'strategy' configuration. I have since changed my personal opinion about this and I believe that the problem should not just be fixed for specific Encoder / Decoder pairs, but rather for all.

The implementation of this was not very pretty, involving casts and iterations over the dictionaries to be encoded/decoded.

Await design of newtype

I have heard mentions of a newtype design, that basically tries to solve the issue that the Tagged library solves: namely creating type safe wrappers around other primitive types.

I am in no way an expert in this, and I don't know how this would be implemented, but if it were possible to tell that SomeType is a newtype of String, then this could be used to provide a new implementation in the Dictionary Codable conformance, and since this feature does not exist in older versions of Swift (providing that this is a feature that requires changes to the Swift run-time), then adding this to the Dictionary Codable conformance would not be behavior breaking.

But those are an awful lot of ifs and buts, and it only solves one of the issues that people appear to run in to (the wrapping issue) — and not for instance String based enums or Int8-based keys.

Do nothing

It is of course possible to handle this situation manually during encoding.

A rather unintrusive way of handling the situation is by using a property wrapper as suggested here: CodableKey.

This solution needs to be applied for each Dictionary and is a quite elegant workaround. But it is still a workaround for something that could be fixed in the stdlib.

11 Likes

I feel the Codable property wrapper is a better way

Hi BigSur,
Thanks for your comment.
As I am gathering feedback for future directions, I would like to hear if you can elaborate on your opinion a bit?

And if any of the people liking the post have any comments, they too are more than welcome.

I am also taking the liberty of tagging some of the commenters of the previous discussions mentioned in the pitch with hopes of getting additional feedback:
@cherrywoods @zarghol @ahti @Jon_Shier @hartbit @acecilia @AlexanderM @Lantua @Martin @SDGGiesbrecht

@Karl @jjanke @ffried @WAHa.06x36

@Morten_Bek_Ditlevsen Thanks for the pitch and the ping!

I really like the idea of the CodingKeyRepresentable protocol - I've even done something similar in one of my projects.

I have a few remarks:

Associated Type for CodingKey

Given that CodingKeyRepresentable follows RawRepresentable, I think it should gain an associatedType for its CodingKey. This will break the direct cast, but it's still doable by e.g. making _DictionaryCodingKey generic over Dictionary.Key and then having a conditional conformance to a protocol that extracts the type-erased CodingKey internally. IMHO it's worth it given that implementations are then able to better decide whether to return nil from init(codingKey:).
To show this, here's an example:

enum MyKey: Int, CodingKey {
    case a = 1
    case b = 3
    case c = 5

    var intValue: Int? { rawValue }

    var stringValue: String {
        switch self {
        case .a: return "a"
        case .b: return "b"
        case .c: return "c"
        }
    }

    init?(intValue: Int) { self.init(rawValue: intValue) }

    init?(stringValue: String) {
        guard let rawValue = RawValue(stringValue) else { return nil }
        self.init(rawValue: rawValue)
    }
}

struct MyCustomType: CodingKeyRepresentable {
    typealias CodingKey = MyKey

    var useB = false

    var codingKey: CodingKey {
        useB ? .b : .a
    }

    init?(codingKey: CodingKey) {
        switch codingKey {
        case .a: useB = false
        case .b: useB = true
        case .c: return nil // .c is unsupported
        }
    }
}

If there's no associated type for CodingKey, things get a bit more complicated inside our init:

init?(codingKey: CodingKey) {
    // oh no, this fails because `codingKey` is actually `_DictionaryCodingKey`...
    guard let key = codingKey as? MyKey else { return nil }

    // This would work, but needs to be done manually every time.
    guard let key = MyKey(intValue: codingKey.intValue) else { return nil }

    switch key { /* ... */ }
}

IMHO it's counter-intuitive that you return MyKey from the codingKey accessor, but get a totally different codingKey into the init(codingKey:) initializer.

AnyCodingKey

I'm unsure about the AnyCodingKey type. One the one hand, I totally agree that there will be tons of similar implementations. On the other hand it could result in people missing out on the advantages of having their own key type (as in my example before) by simply using the AnyCodingKey type.
However, in the end, I think I'd still vote for adding AnyCodingKey.

CodableKey

The property wrapper solution is definitively a nice idea, but I think it has some major drawbacks:

  • Using Int8 (or any other numeric stdlib type for that matter) as key requires it to conform to CodingKey. This conformance would have to come from the stdlib to prevent conformance collisions across e.g. Swift packages. And IMHO those types shouldn't provide a CodingKey conformance per se...
  • It's not straightforward to simply encode/decode e.g. a Dictionary<Int8, String> that is not a property of another Codable type (also mentioned in the example in the linked post).
  • It's impossible to add a Codable conformance to an object that is already defined. So if I define a struct (MyType) having a Dictionary<Int8, String> in one file, I can't simply put an extension MyType: Codable { /* ... */ } into another file.
  • Currently, there can only be one property wrapper per property. While this limitation is likely to be lifted (I don't know the current state here), it currently means that every wrapper adds a potential limitation.

Sorry for the long post. Let me know what you think!

2 Likes

Hi there,
I think this is really a nice solution to the outlined problem. I think it is sensible adding an _AnyCodingKey type, because that is needed quite often as previously mentioned.

What's unfortunate is that this was not handled back then when it was still acceptable to change the encoding and decoding behavior. But since that opportunity has gone, I think this proposal makes sense. It will still be unintuitive why key types that are e.g. string representible are not directly encoded as string keys in keyed containers, but now there is at least an easy fix to this.

The suggestion by @ffried to add an associated type for the CodingKey of a coding key representable sounds good to me. Still I would add an _AnyCodingKey type, I have also previously written something like this, if remember correctly.

Thanks for writing the pitch @Morten_Bek_Ditlevsen!

1 Like

Thank you both so much for the excellent feedback!

I had initially not imagined a use case for having the CodingKey type be an associated value, but you make a compelling argument - and the only ‘cost’ appears to be the slight extra work in the implementation, and here you provided a good solution, so thanks for that, @ffried ! :slight_smile:

Also thank you for the analysis of the possible drawbacks of using the propertyWrapper solution - I’ll add those to the pitch text. I do think that you can already compose property wrappers, however, so perhaps I’ll skip that argument. :slight_smile:

Thanks again for the comments! I’ll return with a modified pitch text and a revised PR.

This is some excellent feedback! One thing I wanted to respond to:

When @Morten_Bek_Ditlevsen and I discussed some of the topics here originally, this subject came up and I suggested that CodingKeyRepresentable not have an associatedtype — I wanted to share some of those reasons for posterity, and in case they might influence the decision of where to go here.

Because associatedtypes have non-zero cost on the consuming side (e.g. checking for CodingKeyRepresentable conformance, using the key type), I think that the associated type definition would need to carry its weight. Despite the name, I think that the key difference between CodingKeyRepresentable and RawRepresentable is that the identity of the RawValue type is crucial to RawRepresentable, but not so in the CodingKeyRepresentable case.

On the consuming side of CodingKeyRepresentable.codingKey (e.g. in Dictionary), I don't believe key type identity is necessarily useful enough:

  • The main use for the .codingKey value is immediate retrieval of the underlying String/Int values. Dictionary would either pull those values out for immediate use and throw away the original key
  • In a non-generic context (or even one not predicated on CodingKeyRepresentable conformance), you can't meaningfully get at the key type. The type-erasure song and dance you have to do to get the key values won't be able to hand you a typed key (and the pain of doing that dance is that because it doesn't make sense to expose a public protocol for doing the erasure, every consumer that wants to do this needs to reinvent the wheel and add another protocol for doing it; we had to do it a few times for Optionals and it's a bit of a shame)
  • Even if it were necessary to get a meaningful key type, the majority use-case for this feature, I believe, will be to provide dynamic-value keys for non-enumerable types (e.g. structs like UUID [though yes, we can't make it conform]); for these types, you can't necessarily define a CodingKeys enum and instead, you'd likely want to use a more generic key type like AnyCodingKey (which by definition doesn't have identity)

On the producing side (e.g. in MyCustomType), I'm also not sure the utility is necessarily enough: in general, the majority of CodingKeyRepresentable types (I believe) will only really care about the String/Int values of the keys, since they will be initialized dynamically (again, I think of UUID initialization from a CodingKey.stringValue — you can do this from any CodingKey).

I believe that the constrained MyKey example above will be the minority use-case, but expressed without the associatedtype constraint too:

enum MyKey: Int, CodingKey {
    case a = 1, b = 3, c = 5

    // There are several ways to express this, just an example:
    init?(codingKey: CodingKey) {
        if let key = codingKey.intValue.flatMap(Self.init(intValue:)) {
            self = key
        } else if let key = Self(stringValue: codingKey.stringValue) {
            self = key
        } else {
            return nil
        }
    }
}

struct MyCustomType: CodingKeyRepresentable {
    var useB = false

    var codingKey: CodingKey {
        useB ? MyKey.b : MyKey.a
    }

    init?(codingKey: CodingKey) {
        switch MyKey(codingKey: codingKey) {
        case .a: useB = false
        case .b: useB = true
        default: return nil
        }    
    }
}

I personally find this equally as expressive, and I think that not requiring the associated type gives more flexibility without a significant loss, especially with non-enum types in mind. What do you think? (If MyCustomType here was inspired by a real-world example, I'd love to know more about it!)

1 Like

Hi all,
Thank you for the input - I have updated the pitch text to reflect the discussion above.

  • Add discussion about using an associated type for the CodingKey of the CodingKeyRepresentable type.

  • Add drawbacks to the property wrapper workaround.

Here is the updated pitch text:

Pitch: Allow coding of non- String / Int keyed Dictionary into a KeyedContainer

Introduction

The current conformance of Swift's Dictionary to the Codable protocols has a somewhat-surprising limitation in that dictionaries whose key type is not String or Int (values directly representable in CodingKey types) encode not as KeyedContainers but as UnkeyedContainers. This behavior has caused much confusion for users and I would like to offer a way to improve the situation.

Motivation

The primary motivation for this pitch lays in the much-discussed confusion of this default behavior:

The common situations where people have found the behavior confusing include:

  • Using enums as keys (especially when RawRepresentable, and backed by String or Int types)
  • Using String wrappers (like the generic Tagged library or custom wrappers) as keys
  • Using Int8 or other Int* flavours as keys

In the various discussions, there are clear and concise explanations for this behavior, but it is also mentioned that supporting encoding of RawRepresentable String and Int keys into keyed containers may indeed be considered to be a bug, and is an oversight in the implementation (JSON Encoding / Decoding weird encoding of dictionary with enum values, reply by Itai Ferber).

There's a bug at bugs.swift.org tracking the issue: SR-7788

Unfortunately, it is too late to change the behavior now:

  1. It is a breaking change with respect to existing behavior, with backwards-compatibility ramifications (new code couldn't decode old data and vice versa), and
  2. The behavior is tied to the Swift stdlib, so the behavior would differ between consumers of the code and what OS versions they are on

Instead, I propose the addition of a new protocol to the standard library. Opting in to this protocol for the key type of a Dictionary will allow the Dictionary to encode/decode to/from a KeyedContainer.

Proposed Solution

I propose adding a new protocol to the standard library: CodingKeyRepresentable

Types conforming to CodingKeyRepresentable indicate that they can be represented by a CodingKey instance (which they can offer), allowing them to opt in to having dictionaries use their CodingKey representations in order to encode into KeyedContainers.

The opt-in can only happen for a version of Swift where the protocol is available, so the user will be in full control of the situation. For instance I am currently using my own workaround, but once I only support iOS versions running a specific future Swift version with this feature, I could skip my own workaround and rely on this behavior instead.

I have a draft PR for the proposed solution: #34458

Examples

// Same as stdlib's _DictionaryCodingKey
struct _AnyCodingKey: CodingKey {
    let stringValue: String
    let intValue: Int?
    
    init?(stringValue: String) {
        self.stringValue = stringValue
        self.intValue = Int(stringValue)
    }
    
    init?(intValue: Int) {
        self.stringValue = "\(intValue)"
        self.intValue = intValue
    }
}

struct ID: Hashable, CodingKeyConvertible {
    static let knownID1 = ID(stringValue: "<some-identifier-1>")
    static let knownID2 = ID(stringValue: "<some-identifier-2>")
    
    let stringValue: String
    
    var codingKey: CodingKey {
        return _AnyCodingKey(stringValue: stringValue)
    }
    
    init?(codingKey: CodingKey) {
        stringValue = codingKey.stringValue
    }
    
    init(stringValue: String) {
        self.stringValue = stringValue
    }
}

let data: [ID: String] = [
    .knownID1: "...",
    .knownID2: "...",
]

let encoder = JSONEncoder()
try String(data: encoder.encode(data), encoding: .utf8)

/*
{
    "<some-identifier-1>": "...",
    "<some-identifier-2>": "...",
}
*/

Detailed Design

The proposed solution adds a new protocol, CodingKeyRepresentable:

/// Indicates that the conforming type can provide a `CodingKey` to be used when
/// encoding into a keyed container.
public protocol CodingKeyRepresentable {
    var codingKey: CodingKey { get }
    init?(codingKey: CodingKey)
}

In the conditional Encodable conformance on Dictionary, the following extra case can handle such conforming types:

    } else if let _ = Key.self as? CodingKeyRepresentable.Type {
      // Since the keys are CodingKeyRepresentable, we can use the `codingKey`
      // to create `_DictionaryCodingKey` instances.
      var container = encoder.container(keyedBy: _DictionaryCodingKey.self)
      for (key, value) in self.dict {
        let codingKey = (key as! CodingKeyRepresentable).codingKey
        let dictionaryCodingKey = _DictionaryCodingKey(codingKey: codingKey)
        try container.encode(value, forKey: dictionaryCodingKey)
      }
    } else {
      // Keys are Encodable but not Strings or Ints, so we cannot arbitrarily

In the conditional Decodable conformance on Dictionary, we can similarly handle conforming types:

    } else if let codingKeyRepresentableType = Key.self as? CodingKeyRepresentable.Type {
      // The keys are CodingKeyRepresentable, so we should be able to expect a keyed container.
      let container = try decoder.container(keyedBy: _DictionaryCodingKey.self)
      for key in container.allKeys {
        let value = try container.decode(Value.self, forKey: key)
        let k = codingKeyRepresentableType.init(codingKey: key)
        self.dict[k as! Key] = value
      }
    } else {
      // We should have encoded as an array of alternating key-value pairs.

Impact on Existing Code

No direct impact, since adoption of this protocol is additive.

However, special care must be taken in adopting the protocol, since adoption on any type T which has previously been encoded as a dictionary key can introduce backwards incompatibility with archives. It is always safe to adopt CodingKeyConvertible on new types, or types newly-conforming to Codable.

Other Considerations

Conforming stdlib types to CodingKeyRepresentable

Along the above lines, we do not propose conforming any existing stdlib or Foundation type to CodingKeyRepresentable due to backwards-compatibility concerns. Should end-user code require this conversion on existing types, we recommend writing wrapper types which conform on those types' behalf (for example, a MyUUIDWrapper which contains a UUID and conforms to CodingKeyRepresentable to allow using UUIDs as dictionary keys directly).

Adding an AnyCodingKey type to the standard library

Since types that conform to CodingKeyRepresentable will need to supply a CodingKey, most likely generated dynamically from type contents, this may be a good time to introduce a general key type which can take on any String or Int value it is initialized from.

Dictionary already uses exactly such a key type internally (_DictionaryCodingKey), as do JSONEncoder / JSONDecoder with _JSONKey (and PropertyListEncoder / PropertyListDecoder with _PlistKey), so generalization could be useful. The implementation of this type could match the implementation of _AnyCodingKey provided above.

Alternatives Considered

Why not just make the type conform to CodingKey directly?

For two reasons:

  1. In the rare case in which a type already conforms to CodingKey, this runs the risk of behavior-breaking changes
  2. CodingKey requires exposure of a stringValue and intValue property, which are only relevant when encoding and decoding; forcing types to expose these properties arbitrarily seems unreasonable

Why not refine RawRepresentable, or use a RawRepresentable where RawValue == CodingKey constraint?

RawRepresentable conformance for types indicates a lossless conversion between the source type and its underlying RawValue type; this conversion is often the "canonical" conversion between a source type and its underlying representation, most commonly between enums backed by raw values, and option sets similarly backed by raw values.

In contrast, we expect conversion to and from CodingKey to be incidental , and representative only of the encoding and decoding process. We wouldn't suggest (or expect) a type's canonical underlying representation to be a CodingKey, which is what a protocol CodingKeyRepresentable: RawRepresentable where RawValue == CodingKey would require. Similarly, types which are already RawRepresentable with non- CodingKey raw values couldn't adopt conformance this way, and a big impetus for this feature is allowing Int- and String-backed enums to participate as dictionary coding keys.

Why not use an Associated Type for CodingKey

It was suggested during the pitch phase to use an associated type for the CodingKey in the CodingKeyRepresentable protocol.

The presented use case was perfectly valid - and demonstrated using the following example:

enum MyKey: Int, CodingKey {
    case a = 1
    case b = 3
    case c = 5

    var intValue: Int? { rawValue }

    var stringValue: String {
        switch self {
        case .a: return "a"
        case .b: return "b"
        case .c: return "c"
        }
    }

    init?(intValue: Int) { self.init(rawValue: intValue) }

    init?(stringValue: String) {
        guard let rawValue = RawValue(stringValue) else { return nil }
        self.init(rawValue: rawValue)
    }
}

struct MyCustomType: CodingKeyRepresentable {
    typealias CodingKey = MyKey

    var useB = false

    var codingKey: CodingKey {
        useB ? .b : .a
    }

    init?(codingKey: CodingKey) {
        switch codingKey {
        case .a: useB = false
        case .b: useB = true
        case .c: return nil // .c is unsupported
        }
    }
}

An analysis of this suggestion hints that the non-zero cost of doing type erasure for pulling out the key values at the consuming site might not carry it's weight (https://forums.swift.org/t/pitch-allow-coding-of-non-string-int-keyed-dictionary-into-a-keyedcontainer/44593/9):

Because associatedtype s have non-zero cost on the consuming side (e.g. checking for CodingKeyRepresentable conformance, using the key type), I think that the associated type definition would need to carry its weight. Despite the name, I think that the key difference between CodingKeyRepresentable and RawRepresentable is that the identity of the RawValue type is crucial to RawRepresentable , but not so in the CodingKeyRepresentable case.

On the consuming side of CodingKeyRepresentable.codingKey (e.g. in Dictionary ), I don't believe key type identity is necessarily useful enough:

  • The main use for the .codingKey value is immediate retrieval of the underlying String / Int values. Dictionary would either pull those values out for immediate use and throw away the original key
  • In a non-generic context (or even one not predicated on CodingKeyRepresentable conformance), you can't meaningfully get at the key type. The type-erasure song and dance you have to do to get the key values won't be able to hand you a typed key (and the pain of doing that dance is that because it doesn't make sense to expose a public protocol for doing the erasure, every consumer that wants to do this needs to reinvent the wheel and add another protocol for doing it; we had to do it a few times for Optional s and it's a bit of a shame)
  • Even if it were necessary to get a meaningful key type, the majority use-case for this feature, I believe, will be to provide dynamic-value keys for non-enumerable types (e.g. struct s like UUID [though yes, we can't make it conform]); for these types, you can't necessarily define a CodingKey s enum and instead, you'd likely want to use a more generic key type like AnyCodingKey (which by definition doesn't have identity)

On the producing side (e.g. in MyCustomType ), I'm also not sure the utility is necessarily enough: in general, the majority of CodingKeyRepresentable types (I believe) will only really care about the String / Int values of the keys, since they will be initialized dynamically (again, I think of UUID initialization from a CodingKey.stringValue — you can do this from any CodingKey ).

I believe that the constrained MyKey example above will be the minority use-case, but expressed without the associatedtype constraint too:

enum MyKey: Int, CodingKey {
    case a = 1, b = 3, c = 5

    // There are several ways to express this, just an example:
    init?(codingKey: CodingKey) {
        if let key = codingKey.intValue.flatMap(Self.init(intValue:)) {
            self = key
        } else if let key = Self(stringValue: codingKey.stringValue) {
            self = key
        } else {
            return nil
        }
    }
}

struct MyCustomType: CodingKeyRepresentable {
    var useB = false

    var codingKey: CodingKey {
        useB ? MyKey.b : MyKey.a
    }

    init?(codingKey: CodingKey) {
        switch MyKey(codingKey: codingKey) {
        case .a: useB = false
        case .b: useB = true
        default: return nil
        }    
    }
}

I personally find this equally as expressive, and I think that not requiring the associated type gives more flexibility without a significant loss, especially with non- enum types in mind.

Add workarounds to each Encoder / Decoder

Following a suggestion from @itaiferber, I have previously tried to provide a solution to this issue — not in general, but instead solving it by providing a DictionaryKeyEncodingStrategy for JSONEncoder : #26257

The idea there was to be able to express an opt-in to the new behavior directly in the JSONEncoder and JSONDecoder types by venting a new encoding/decoding 'strategy' configuration. I have since changed my personal opinion about this and I believe that the problem should not just be fixed for specific Encoder / Decoder pairs, but rather for all.

The implementation of this was not very pretty, involving casts and iterations over the dictionaries to be encoded/decoded.

Await design of newtype

I have heard mentions of a newtype design, that basically tries to solve the issue that the Tagged library solves: namely creating type safe wrappers around other primitive types.

I am in no way an expert in this, and I don't know how this would be implemented, but if it were possible to tell that SomeType is a newtype of String, then this could be used to provide a new implementation in the Dictionary Codable conformance, and since this feature does not exist in older versions of Swift (providing that this is a feature that requires changes to the Swift run-time), then adding this to the Dictionary Codable conformance would not be behavior breaking.

But those are an awful lot of ifs and buts, and it only solves one of the issues that people appear to run in to (the wrapping issue) — and not for instance String based enums or Int8-based keys.

Do nothing

It is of course possible to handle this situation manually during encoding.

A rather unintrusive way of handling the situation is by using a property wrapper as suggested here: CodableKey.

This solution needs to be applied for each Dictionary and is a quite elegant workaround. But it is still a workaround for something that could be fixed in the stdlib.

A few drawbacks to the property wrapper solution were given during the pitch phase:

  • Using Int8 (or any other numeric stdlib type for that matter) as key requires it to conform to CodingKey . This conformance would have to come from the stdlib to prevent conformance collisions across e.g. Swift packages. And IMHO those types shouldn't provide a CodingKey conformance per se...
  • It's not straightforward to simply encode/decode e.g. a Dictionary<Int8, String> that is not a property of another Codable type (also mentioned in the example in the linked post).
  • It's impossible to add a Codable conformance to an object that is already defined. So if I define a struct ( MyType ) having a Dictionary<Int8, String> in one file, I can't simply put an extension MyType: Codable { /* ... */ } into another file.
4 Likes
Terms of Service

Privacy Policy

Cookie Policy