`OpenEnum`: an enum to represent protobuf's enumeration field values #1079

mzabaluev · 2024-05-28T11:35:31Z

Another solution for #276, amenable to destructuring.

This changes the representation of enum fields in message structs to this generic wrapper, parameterized over the generated known enum type:

pub enum OpenEnum<T> {
    /// A known value defined in the proto
    Known(T),
    /// Unknown value as decoded from the wire
    Unknown(i32),
}

In an improvement over #1061, this allows convenient matching of enum field values as part of the message. There are also convenience methods to fallibly extract the known value in an Option or Result.

caspermeijn · 2024-05-28T14:10:35Z

Please add a description that explains your proposed change. That will make reviewing easier.

QuentinPerez · 2024-05-28T23:20:48Z

Do you think that this logic could be applied to the oneof field ?

mzabaluev · 2024-05-29T12:51:38Z

Do you think that this logic could be applied to the oneof field ?

I believe it works differently for oneofs: if an unknown field number is encountered in a message and it's not a known oneof variant or a regular field, the field is ignored. So the oneof field that would get the value if the variant field were described in the proto would get None instead.

caspermeijn · 2024-05-31T08:47:52Z

Will this break compatibility with proto2/closed enums? Personally, I don't use proto2, so I am not sure whether it is properly supported at all.

caspermeijn · 2024-05-31T08:48:10Z

I think it makes sense to provide a migration guide.

caspermeijn · 2024-05-31T08:48:26Z

What is the error message for a i32 field with #[prost(enumeration)]? Can we detect that scenario and print a nice error message?

caspermeijn · 2024-05-31T08:48:40Z

I would like to see some tests for OpenEnum.

caspermeijn · 2024-05-31T08:01:51Z

prost-derive/src/field/map.rs

+                    self.#ident.get(#take_ref key).cloned().and_then(|x| { x.known() })
                }
                #[doc=#insert_doc]
                pub fn #insert(&mut self, key: #key_ty, value: #ty) -> ::core::option::Option<#ty> {
-                    self.#ident.insert(key, value as i32).and_then(|x| {
-                        let result: ::core::result::Result<#ty, _> = ::core::convert::TryFrom::try_from(x);
-                        result.ok()
-                    })
+                    self.#ident.insert(key, value.into()).and_then(|x| { x.known() })


Is it beneficial to provide a get and insert function with the new wrapper? I feel like the map can be used directly with the helper functions provided by OpenEnum.

As can be seen in the generated code itself, these methods provide considerable convenience for the most common use cases.

Perhaps the insert method could be changed to return the possible old value as Option<OpenEnum<#ty>> so the caller can choose the fallback behavior for unknown values, but that API would be inconsistent with the get method. If it's to be changed so, the only added convenience there would be the .into() conversion call.

caspermeijn · 2024-05-31T08:06:45Z

prost-derive/src/field/scalar.rs

                        pub fn #get(&self) -> #ty {
-                            ::core::convert::TryFrom::try_from(self.#ident).unwrap_or(#default)
+                            self.#ident.unwrap_or(#default)
                        }

                        #[doc=#set_doc]
                        pub fn #set(&mut self, value: #ty) {
-                            self.#ident = value as i32;
+                            self.#ident = value.into();
                        }


Is it still beneficial to provide a get, set and push functions with the new wrapper? I feel like the field can be used directly with the helper functions provided by OpenEnum.

There is some convenience in the setter method:

msg.set_kind(Kind::Foo)

is nicer and more type-ahead friendly than

msg.kind = Kind::Foo.into()

The same applies to the push method for repeated fields.

The getter hides a bit much opinionated behavior, in my opinion, so it might be better to leave it to OpenEnum helper methods to explicitly decide on how unknown values should be treated.

Another thing to consider here is porting the existing code. If these methods get used instead of direct access to fields, which I assume happens a lot now because the fields are just i32 or Option<i32>, updating to the version that generates OpenEnum would not require changes.

I think the better API is to have some sort of set operation on OpenEnum. Maybe something like:

msg.kind.set(Kind::Foo)

Well, this change will break existing code in many ways, so I think we should directly go for the best API

caspermeijn · 2024-05-31T08:08:23Z

prost-derive/src/field/scalar.rs

@@ -555,13 +550,13 @@ impl Ty {
            Ty::Bool => quote!(bool),
            Ty::String => quote!(&str),
            Ty::Bytes(..) => quote!(&[u8]),
-            Ty::Enumeration(..) => quote!(i32),
+            Ty::Enumeration(..) => unreachable!("an enum should never be queried for its ref type"),


Can you explain me why it should never be queried by ref?

This is preempted by logic in prost-derive. For one example, the rust_ref_type method is used on map field's key type, which cannot be constructed as an enum.

caspermeijn · 2024-05-31T08:10:32Z

prost-derive/src/field/scalar.rs

        }
    }

    pub fn module(&self) -> Ident {
        match *self {
-            Ty::Enumeration(..) => Ident::new("int32", Span::call_site()),
+            Ty::Enumeration(..) => Ident::new("enumeration", Span::call_site()),


Does it make sense to use self.as_str() for enumerations as well?

No, that would give "enum", but we can't use that as a module name. OTOH, maybe the current behavior of Ty::as_str should be changed: "enum" is not a real type name anyway.

Changing the result of as_str to "enumeration" will change the output of Debug and Display for Ty. I don't know if this is a problem; it does not break any tests.

caspermeijn · 2024-05-31T08:31:34Z

prost/src/encoding.rs

+        for value in values {
+            encode(tag, value, buf);
+        }


Could int32::encode_repeated be reused here?

No, unless int32::encode_repeated is generalized to take an impl IntoIterator that produces i32. Which might be a good idea.

So that would need a breaking change for encode_repeated which I don't think is worth the little code duplication at this point.

caspermeijn · 2024-05-31T08:31:50Z

prost/src/encoding.rs

+        if values.is_empty() {
+            return;
+        }
+
+        encode_key(tag, WireType::LengthDelimited, buf);
+        let len: usize = values
+            .iter()
+            .map(|value| encoded_len_varint(value.to_raw() as u64))
+            .sum();
+        encode_varint(len as u64, buf);
+
+        for value in values {
+            encode_varint(value.to_raw() as u64, buf);
+        }


Could int32::encode_packed be reused here?

Same here, not worth changing the signature of int32::encode_packed for this.

caspermeijn · 2024-05-31T08:32:56Z

prost/src/encoding.rs

+        if wire_type == WireType::LengthDelimited {
+            // Packed.
+            merge_loop(values, buf, ctx, |values, buf, ctx| {
+                let mut value = Default::default();
+                merge(WireType::Varint, &mut value, buf, ctx)?;
+                values.push(value);
+                Ok(())
+            })
+        } else {
+            // Unpacked.
+            check_wire_type(WireType::Varint, wire_type)?;
+            let mut value = Default::default();
+            merge(wire_type, &mut value, buf, ctx)?;
+            values.push(value);
+            Ok(())
+        }


Could int32::merge_repeated be reused here?

No, definitely not. This function merges into a passed in &mut Vec<OpenEnum<T>>.

caspermeijn · 2024-05-31T08:33:28Z

prost/src/encoding.rs

+        key_len(tag) * values.len()
+            + values
+                .iter()
+                .map(|value| encoded_len_varint(value.to_raw() as u64))
+                .sum::<usize>()


Could int32::encoded_len_repeated be reused here?

Not without breaking changes, as explained above.

caspermeijn · 2024-05-31T08:34:07Z

prost/src/encoding.rs

+        if values.is_empty() {
+            0
+        } else {
+            let len = values
+                .iter()
+                .map(|value| encoded_len_varint(value.to_raw() as u64))
+                .sum::<usize>();
+            key_len(tag) + encoded_len_varint(len as u64) + len
+        }


Could int32::encoded_len_packed be reused here?

Not without breaking changes, as explained above.

caspermeijn · 2024-05-31T08:41:32Z

prost/src/open_enum.rs

+        let mut raw = 0;
+        <i32 as Message>::merge_field(&mut raw, tag, wire_type, buf, ctx)?;
+        *self = OpenEnum::from_raw(raw);
+        Ok(())


This is the same function as encoding::enumeration::merge, right?

No, <i32 as Message>::merge_field calls to skip_field if the tag argument is not 1.

mzabaluev · 2024-06-02T06:54:05Z

Will this break compatibility with proto2/closed enums? Personally, I don't use proto2, so I am not sure whether it is properly supported at all.

I believe prost has never been conformant with closed enums: the raw values have been left in the message as is.
I think we could try to support closed enums in two ways:

Always represent an enum field as OpenEnum, but in the closed enum case, decode unknown values as Known(Default::default()). This leaves the possibility of producing unknown values on encoding.
Represent closed enums with their generated Rust enum type. Decode unknown values as Default::default().

I prefer option 2, even though it requires more work. Protobuf edition 2023 has closed enums as a feature, so they need to be supported even if we ignore proto2.

mzabaluev · 2024-06-02T07:14:43Z

I think it makes sense to provide a migration guide.

What would be a good place for it? I can add a section to the README.

mzabaluev · 2024-06-02T07:32:47Z

I would like to see some tests for OpenEnum.

I'm going to add at least one good example/doctest on the type.

caspermeijn · 2024-06-07T11:55:39Z

Have thought some more about this PR: I don't want to break users in this way. At least not now.

I suggest making this an option in prost-build so that we can experiment with the API without breaking existing users. That way, interested users can opt in, and we don't have to create a perfect OpenEnum API on the first try.

Once we feel good about the new API, we can think about changing the default.

Please look at bytes for a good example of changing the generated data type.

ArjixWasTaken · 2024-06-28T19:42:19Z

Instead of introducing a new type, a simple Result<T, i32> sounds more logical to me.
enum-repr-derive follows this approach when parsing an enum from i32

mzabaluev · 2024-06-29T13:41:40Z

Instead of introducing a new type, a simple Result<T, i32> sounds more logical to me.

It's weird to have a Result as a struct field. The names of variants and methods of Result are less than intuitively applicable: it's not necessarily an error to receive an unknown enum value from the wire, so we should give the API users a speed bump to decide how to deal with them. There are convenience methods and the TryInto impl to convert the OpenEnum value to a Result if that's the chosen approach.

Another benefit of introducing a new type is for the add-on macros and code generators that derive something on structs generated by prost-build. These could deal with the OpenEnum type in some specific way, even without access to the proto descriptor data for the field. Doing the same (e.g. defining generic trait impls) for Result would feel like too much overloading.

caspermeijn · 2024-08-09T06:01:14Z

How are default values handles in this solution? Especially default values set for a specific field in the proto file.

mzabaluev · 2024-11-12T03:44:14Z

How are default values handles in this solution? Especially default values set for a specific field in the proto file.

Good question. Maybe there should be a const generic parameter for this, but it will then crop up everywhere.

Updated: explained below

mzabaluev · 2024-11-12T04:35:49Z

prost-derive/src/field/scalar.rs

@@ -799,7 +749,7 @@ impl DefaultValue {

    pub fn typed(&self) -> TokenStream {
        if let DefaultValue::Enumeration(_) = *self {
-            quote!(#self as i32)
+            quote!(::prost::OpenEnum::from(#self))


This handles the custom default values for message fields.

mzabaluev · 2024-11-13T09:03:31Z

The generated getter methods handled per-field defaults for optional fields. For non-optional fields, the default is initialized before merging field values from the wire.

I'm not sure how best to proceed here. Suppose we have this generated message:

pub struct Msg {
    #[prost(enumeration = "AnEnum", optional, default = "AnEnum::B")]
    pub field: Option<OpenEnum<AnEnum>>
}

It would be nice to have API implementing the default value behavior for proto2 and edition 2023.

The protobuf guide recommends a getter method that returns the default value, but whether out-of-range values are exposed is up to the enum_type protobuf feature. The "hazzer" method can be generated to check presence, but in the currently produced code one can just use Option::is_some.

The wrapper to represent field of enumerated types with the possibility of unknown values.

Replace i32 with the type-checked wrapper.

These are meant to be the primary ways to get and set open enumeration fields, instead of the getters and setters generated by Message derive.

Add typed_enum_fields method to prost-build configuration, which allows type-checked representation of enumerations in fields of message structs and variants of oneof enums. The argument and the invocation order works like with the boxed method. Depending on the syntax (and preparing for the future support of editions), the type-checked representation can be closed (for proto2) or open (for proto3). The former is represented by the generated enum type itself, while the latter is represented by OpenEnum wrapping the enum type. A new enum_type annotation is supported in the prost attribute inside derives, which allows to specify the type-checked representation of enum types in message fields and oneof variants. The accepted values are "open" or "closed".

mzabaluev · 2024-11-17T20:55:20Z

I suggest making this an option in prost-build so that we can experiment with the API without breaking existing users. That way, interested users can opt in, and we don't have to create a perfect OpenEnum API on the first try.

I have addressed this and most of the other comments in recent commits. There is now a Config::typed_enum_fields method to opt into type-checked fields.

The proto2 behavior (and the closed enums feature in future editions) is also implemented for enums, albeit with some current limitations: state available to the code generator does not track which syntax the enum types were defined in, that is needed for proto2 ← proto3 includes. I'll try to resolve this.

Getter methods are retained for the optional fields for convenience and to implement per-field default values. Protobuf guidelines also insist on there being getters returning non-nullable values.

mzabaluev mentioned this pull request May 28, 2024

OpenEnum: type-safe wrapper for enum field values #1061

Closed

mzabaluev marked this pull request as ready for review May 28, 2024 21:08

caspermeijn requested changes May 31, 2024

View reviewed changes

caspermeijn added the breaking change label May 31, 2024

caspermeijn mentioned this pull request May 31, 2024

getter/setter generated for optional enum fields should use Option #1027

Open

ArjixWasTaken mentioned this pull request Jun 28, 2024

A more ergonomic enumeration representation #276

Open

mzabaluev force-pushed the open-enum-as-enum branch from 9ec7c5d to f76175a Compare November 12, 2024 03:44

mzabaluev commented Nov 12, 2024

View reviewed changes

mzabaluev added 5 commits November 16, 2024 10:45

prost: add OpenEnum

a86239b

The wrapper to represent field of enumerated types with the possibility of unknown values.

prost: encoding module for OpenEnum

72cd750

Use OpenEnum in generated code

29da76b

Replace i32 with the type-checked wrapper.

prost: document OpenEnum

e4c84e7

feat: get, set methods for OpenEnum

3fee552

These are meant to be the primary ways to get and set open enumeration fields, instead of the getters and setters generated by Message derive.

mzabaluev force-pushed the open-enum-as-enum branch from f76175a to cb5e59c Compare November 17, 2024 18:02

mzabaluev force-pushed the open-enum-as-enum branch from cb5e59c to 905fbc8 Compare November 17, 2024 18:26

mzabaluev force-pushed the open-enum-as-enum branch from 905fbc8 to 9bd8272 Compare November 17, 2024 19:49

mzabaluev added 2 commits November 17, 2024 22:30

chore: fix clippy lints

d83a32b

fix(prost-derive): improve enum_type errors

0f7cbc8

This was referenced Nov 17, 2024

Add support for editions #1031

Open

Clean up CodeGenerator by moving compilation-global data and logic to Context #1190

Open

OpenEnum: an enum to represent protobuf's enumeration field values #1079

Are you sure you want to change the base?

OpenEnum: an enum to represent protobuf's enumeration field values #1079

Conversation

mzabaluev commented May 28, 2024 • edited Loading

caspermeijn commented May 28, 2024

QuentinPerez commented May 28, 2024

mzabaluev commented May 29, 2024 • edited Loading

caspermeijn commented May 31, 2024

caspermeijn commented May 31, 2024

caspermeijn commented May 31, 2024

caspermeijn commented May 31, 2024

Choose a reason for hiding this comment

Choose a reason for hiding this comment

mzabaluev Jun 2, 2024 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

mzabaluev Nov 11, 2024 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

mzabaluev commented Jun 2, 2024

mzabaluev commented Jun 2, 2024

mzabaluev commented Jun 2, 2024

caspermeijn commented Jun 7, 2024

ArjixWasTaken commented Jun 28, 2024 • edited Loading

mzabaluev commented Jun 29, 2024

caspermeijn commented Aug 9, 2024

mzabaluev commented Nov 12, 2024 • edited Loading

Choose a reason for hiding this comment

mzabaluev commented Nov 13, 2024 • edited Loading

mzabaluev commented Nov 17, 2024

`OpenEnum`: an enum to represent protobuf's enumeration field values #1079

`OpenEnum`: an enum to represent protobuf's enumeration field values #1079

mzabaluev commented May 28, 2024 •

edited

Loading

mzabaluev commented May 29, 2024 •

edited

Loading

mzabaluev Jun 2, 2024 •

edited

Loading

mzabaluev Nov 11, 2024 •

edited

Loading

ArjixWasTaken commented Jun 28, 2024 •

edited

Loading

mzabaluev commented Nov 12, 2024 •

edited

Loading

mzabaluev commented Nov 13, 2024 •

edited

Loading