Implement #[serde(other)] on enum variant

[roblabla]

Part of #912.

It might require more complex testing WRT non-unit other fields?

Untagged enums are not supported. I'm not quite sure the checks are correct, so if someone could make sure they make sense (they might be a bit too lax).

[dtolnay]

This should be written as else if let Some(other_idx) = other_idx to avoid the unwrap below.

[dtolnay]

What currently happens if you write serde(other) on a variant that is not unit?

[dtolnay]

Here are two cases that I would expect to work but do not. Do you think these would be possible to support?

#[derive(Deserialize, PartialEq, Debug)]
enum E {
    A,
    #[serde(other)]
    Other(String),
}

fn main() {
    assert_eq!(
        serde_json::from_str::<E>(r#" "foo" "#).unwrap(),
        E::Other("foo".to_owned()),
    );
}

#[derive(Deserialize, PartialEq, Debug)]
struct Artifact {}

#[derive(Deserialize, PartialEq, Debug)]
enum Message {
    CompilerArtifact(Artifact),
    #[serde(other)]
    Unknown,
}

fn main() {
    assert_eq!(
        serde_json::from_str::<Message>(r#" {"CompilerMessage":[]} "#).unwrap(),
        Message::Unknown,
    );
}

[roblabla]

🤔 this is a bit of a puzzle. The first one IMO doesn't make sense: it's deserializing an externally tagged enum, so we expect something in the form { "Variant": data }, but instead getting "data". I expect this to fail. If you still think it should work, I can look into it though.

The second one should work though, so I probably messed up somewhere! (I mostly tested on internally tagged enums for my use-case ^^'). I'll take a look.

[dtolnay]

I would expect the first one to work by analogy with how E::A is represented: playground link.

[roblabla]

Ah, so this is quite different from how I originally envisioned this to work. I thought the data contained in the Other variant would be the "inner" object, e.g. for an externally tagged enum { "UnknownVariant": "data" } => Enum::Other("data").

So here's a proposal:

#[derive(Deserialize)]
enum Test {
    A,
    #[serde(other)]
    B(#[serde(other_tag)] String, serde_json::Value)
}

The field equipped with #[serde(other_tag)] (bikeshed for a better name welcome) would automatically be populated with the tag name. (Obviously, it doesn't have to be a String, anything deserializable should work).

Then, we'd get { "Unknown": "data" } => Test::B("Unknown", Str("data")).

The upside is that we can keep parsing data even in the Unknown case (maybe we can still get some useful data that we know are guaranteed to always be present regardless of the enum ?) while having a consistent way to access the variant name.

Would this work?

[dtolnay]

I guess my confusion is in when you would ever want to capture the data of an unknown variant but not its variant name. I would expect the default least-effort way of using serde(other) would capture both of those or neither of those. Skimming through the use cases in #912, either the other variant is a unit variant and captures nothing, or is a newtype variant that captures everything necessary for round-tripping the input data through that Rust data structure, which includes the unrecognized tag. In your suggestion involving serde(other_tag) the default least-effort use of serde(other) (so without other_tag) captures the data but not the tag, which is not a behavior that has been requested.

[roblabla]

Alright, here's my reasoning:

As a user, I want the ability to access both the data and the tag in a simple way. Because of this, I want Other to behave like the other enum variant, because that's the intuitive thing. Here's an example: we have an API that sends requests following this schema:

{
    "action": "some_action",
    "version": "1.0.0",
    "field_of_some_action": "1",
    "field_of_some_action2": "2",
}

(this is an internally tagged enum). The API contract says version will always be present. As such, it might be useful for the Other variant to have the ability to parse it, in order to inform the user of what the latest version is. This is a totally made-up example, but it's a pattern that I've seen in the past.

So, how would I parse this API?

#[derive(Deserialize, PartialEq, Debug)]
#[serde(tag = "action")]
enum E {
    Action1, // whatever
    #[serde(other)]
    UnknownAction { version: String },
}

fn main() {
    assert_eq!(
        serde_json::from_str::<E>(r#" "{ "action": "unknown", "version": "2.0.0" }" "#).unwrap(),
        E::UnknownAction { version: "2.0.0" }),
    );
}

Now, obviously, in many/most cases, we'd still want the ability to access the tag. For this case, we need to put it somewhere in that UnknownAction enum variant. The question is: where? I can only come up with two ideas:

1. Have an attribute defining where. That's #[serde(other_tag)]
2. Enforce that the first element of the other variant has to be a String, and will be the enum.
3. "Reverse" the problem and put a tag on data fields? I'm not sure how that would even look. UnknownAction { #[serde(data_field)] version: u32, tag: String }

The first felt the most intuitive to me. With this case, in order to access the action field above, we'd do this:

#[derive(Deserialize, PartialEq, Debug)]
#[serde(tag = "action")]
enum E {
    Action1, // whatever
    #[serde(other)]
    UnknownAction { #[serde(other_tag)] action: String, version: String },
}

fn main() {
    assert_eq!(
        serde_json::from_str::<E>(r#" "{ "action": "unknown", "version": "2.0.0" }" "#).unwrap(),
        E::UnknownAction { action: "unknown", version: "2.0.0" }),
    );
}

I'm still trying to figure out if this is even implementable though.

---

Maybe I could just enforce that the Other is a unit struct for now, and then we can discuss what the proper course of action is for accessing the tag/data in a separate issue and follow-up PR ?

[dtolnay]

Thanks, this looks like the right behavior for unit variants.

[derekdreery]

Does this implement the following use-case:

Data shape:

{
    "errors": ["ErrorKind1", "ErrorKind2", ...]
}

So we get 0 or more errors.

Rust code

#[derive(Deserialize)]
struct Result {
    errors: Vec<ErrorKind>
}

enum ErrorKind {
    ErrorKind1,
    ErrorKind2,
    // ...
    #[serde(other)]
    Unrecognised(String),
}

That is, I receive a list of errors, and most of the time I can strongly type them, but on occasion I don't recognize the string, and so fall back to just storing the string value.

This seems to be different to the use-case discussed in the PR.

[roblabla]

The pr only allows the #[serde(other)] variant to be unit-style. A future pr could extend it to allow recovering the tag. From my testing, however, there are several problems that don't have am obvious solution:

• The enums can parse their discriminant as either String or usize. How do we deal with unknown numeric tags in the case of Unknown(String), and conversely, deal with string tags in the case of Unknown(num)?

• For some of my use-cases, I need the ability to parse data in the other tag. I gave some examples in my last comment above. How should we deal with that?

[derekdreery]

A conservative approach would be to just reject the parse if the discriminant type and the extra type don't match, and just support numbers and strings. Then the functionality you mention in your 2 points could be added incrementally, whilst retaining backwards compatibility.

[derekdreery]

I'll have a go at implementing this and see what it looks like.

[spease]

Guess this has gotten stalled?

[nixpulvis]

The pr only allows the #[serde(other)] variant to be unit-style.

This gets me started, thanks.

I'm using serde_json and I'm finding myself wanting to just store a generic Value inside the other variant. I'm guessing this use case is pretty common across formats.

EDIT: Using an Other(serde_json::Value) inside a #[serde(untagged)] seems to be working for this case, since it checks in order.

[maxcountryman]

EDIT: Using an Other(serde_json::Value) inside a #[serde(untagged)] seems to be working for this case, since it checks in order.

What if I have a tag I want but also want to have a blanket Other variant (specifically when there's a tag that I have intentionally not provided as an enum variant)? I'm not seeing how these two can work together.