[swift-evolution] [Proposal] Foundation Swift Encoders

[Itai Ferber]

> On Apr 5, 2017, at 2:29 PM, Brent Royal-Gordon via swift-evolution <swift-evolution at swift.org> wrote:
> 
>> On Apr 5, 2017, at 1:44 PM, David Hart via swift-evolution <swift-evolution at swift.org <mailto:swift-evolution at swift.org>> wrote:
>> 
>>>> For the same reasons, I continue to believe that decode functions should overload on the return type. If we follow the arguments in favor of providing a type argument, then why don't we also have type arguments for encoders: encode(_ value: T?, forKey key: Key, as type: T.self)? I'm not advocating that: I'm just pushing the argument to its logical conclusion to explain why I don't understand it.
>>> 
>>> I don’t see a way for a call to encode to become ambiguous by omitting the type argument, whereas the same is not true for a return value from decode. The two seem fundamentally different.
>> 
>> When decoding to a property, there will be no ambiguity. And for other cases, Swift developers are already quite used to handling that kind of ambiguity, like for literals:
>> 
>> let x: UInt = 10
>> let y = 20 as CGFloat
> 
> But in the literal case, they *don't* have to deal with ambiguity for two reasons:
> 
> 1. The literal provides some hint of the type; integer, float, string, array, and dictionary literals are all easy to distinguish from one another.
> 
> 2. Each literal syntax has a default type. That is not true and cannot *be* true for `decode()`.
> 
> Plus there's a third reason:
> 
> 3. `Decoder` doesn't guarantee there's a safety net if you use the wrong type. If you, say, decode an `Int32` using `Int64`, a decoder for some low-level binary type would be perfectly within its rights to read part of the next field, access everything subsequent to that point in a misaligned way, and go totally off the rails (as long as it doesn't violate memory safety).
> 
> That third reason is exactly the same as why `unsafeBitCast(_:to:)`, `bindMemory(to:capacity:)`, etc. all have a type-pinning parameter. Although Swift places no restrictions on return-type inference, in practice the core team thinks unconstrained return types are dangerous and should be used with care, only permitted when an API explicitly exists to ease conversions between different types. (`numericCast(_:)` is one example; I'm not sure if there are any others.) That's just their opinion, and of course you are always free to disagree with them, but I think it's a solid and easily justified one.
Well put; this is basically what I was going to say.

> Besides, if you really want this, it's easy to add with a pair of extensions:
> 
> 	extension KeyedDecodingContainer {
> 		func decode<T: Encodable>(forKey key: Key) throws -> T {
> 			return try decode(T.self, forKey: key)
> 		}
> 	}
> 	extension UnkeyedDecodingContainer {
> 		func decode<T: Encodable>() throws -> T {
> 			return try decode(T.self)
> 		}
> 	}
> 
> (P.S. There might be a way to square this circle: If a CodingKey knew its type, the mere act of providing a CodingKey would be enough to pin the type. This would not only avoid both an explicit type-pinning parameter *and* unconstrained generic return types, it would also prevent you from accidentally specifying the wrong type during decoding. Rough example:
> 
> 	protocol CodingKey {
> 		associatedtype Value: Encodable
> 		
> 		var stringValue: String { get }
> 		var intValue: Int? { get }
> 		
> 		init(stringValue: String, intValue: Int?, as _: Value.Type)
> 	}
> 	extension CodingKey {
> 		var intValue: Int { return nil }
> 	}
> 	
> 	extension Person: Decodable {
> 		struct CodingKeys<Value>: CodingKey {
> 			let stringValue: String
> 			
> 			init(stringValue: String, intValue: Int? = nil, as _: Value.Type) {
> 				self.stringValue = stringValue
> 			}
> 			
> 			static let name = CodingKeys(stringValue: "name", as: String.self)
> 			static let age = CodingKeys(stringValue: "age", as: Int.self)
> 			static let pets = CodingKeys(stringValue: "pets", as: [Pet].self)
> 		}
> 		
> 		init(from decoder: Decoder) throws {
> 			let c = try decoder.container(keyedBy: CodingKeys.self)
> 			
> 			name = c.decode(.name)
> 			age = c.decode(.age)
> 			pets = c.decode(.pets)
> 		}
> 	}
> 
> But this doesn't work for two reasons: You can't pass an entire generic type to `container(keyedBy:)` and you can't put constants in a generic type. You also lose the ability to construct a CodingKey from a String or Int, you lose the guarantee that all possible instances are known at compile time (I could imagine a linter checking that you've encoded/decoded all CodingKeys), and you lose all the convenient magic of enum raw types. You could address the "doesn't work" issues, but only by adding even more boilerplate. Still, if there's a way to do this that *doesn't* have so many disadvantages, we should seriously consider taking it.)
FWIW, there’s another reason why a CodingKey doesn’t know about the type you’ll be trying to decode with it: data migration.
Over time as your type changes, it’s perfectly reasonable to change the types of its properties (e.g. you realize that string identifiers are not needed somewhere in the app, and integer identifiers will do — better for efficiency). If some of those old types were written into archives, you’ll want to be able to still read them in new versions of your app, and potentially migrate them forward if possible. In that case, you wouldn’t want to have to keep an old key around just to represent the other type; you could just attempt to decode with a key and one type, and if it fails, try with the same key but a different type.

> -- 
> Brent Royal-Gordon
> Architechies
> 
> _______________________________________________
> swift-evolution mailing list
> swift-evolution at swift.org
> https://lists.swift.org/mailman/listinfo/swift-evolution

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.swift.org/pipermail/swift-evolution/attachments/20170405/7ff1621b/attachment.html>