Home

Awesome

[@@deriving protobuf]

deriving protobuf is a ppx_deriving plugin that generates Google Protocol Buffers serializers and deserializes from an OCaml type definition.

Sponsored by Evil Martians. protoc export sponsored by MaxProfitLab.

Installation

deriving protobuf can be installed via OPAM:

$ opam install ppx_deriving_protobuf

Usage

In order to use deriving protobuf, require the package ppx_deriving_protobuf.

Syntax

deriving protobuf is not a replacement for protoc and it does not attempt to generate code based on protoc definitions. Instead, it generates code based on OCaml type definitions. It can also generate input files for protoc.

deriving protobuf-generated serializers are derived from the structure of the type and several attributes: @key, @encoding, @bare and @default. Generation of the serializer is triggered by a [@@deriving Protobuf] attribute attached to the type definition.

deriving protobuf generates two functions per type:

type t = ... [@@deriving protobuf]
val t_from_protobuf : Protobuf.Decoder.t -> t
val t_to_protobuf   : t -> Protobuf.Encoder.t -> unit

In order to deserialize a value of type t from bytes message, use:

let output = Protobuf.Decoder.decode_exn t_from_protobuf message in
...

In order to serialize a value input of type t, use:

let message = Protobuf.Encoder.encode_exn t_to_protobuf input in
...

Records

A record is the most obvious counterpart for a Protobuf message. In a record, every field must have an explicitly defined key. For example, consider this protoc definition and its deriving protobuf equivalent:

message SearchRequest {
  required string query = 1;
  optional int32 page_number = 2;
  optional int32 result_per_page = 3;
}

type search_request = {
  query           : string     [@key 1];
  page_number     : int option [@key 2];
  result_per_page : int option [@key 3];
} [@@deriving protobuf]

deriving protobuf recognizes and maps option to optional fields, and list and array to repeated fields.

Optional and default fields

A [@default] attribute attached to a required field converts it to an optional field; if the field is not present, its value is assumed to be the default one, and conversely, if the value of the field is same as the default value, it is not serialized:

message Defaults {
  optional int32 results = 1 [default = 10];
}

type defaults = {
  results : int [@key 1] [@default 10];
}

Note that protoc's default behavior is to assign a type-specific default value to optional fields missing from message, i.e. 0 to integer fields, "" to string fields, and so on. With deriving protobuf, optional fields are represented with the option type; it is possible to emulate protoc's behavior by explicitly specifying int [@default 0], etc.

Integers

Unlike protoc, deriving protobuf allows a much more flexible mapping between wire representations of integral types and their counterparts in OCaml. Any combination of the known integral types (int, int32, int64, Int32.t, Int64.t, Uint32.t and Uint64.t) and wire representations (varint, zigzag, bits32 and bits64) is valid. The wire representation is specified using the @encoding attribute.

For example, consider this protoc definition and a compatible deriving protobuf one:

message Integers {
  required int32   bar = 1;
  required fixed64 baz = 2;
}

type integers = {
  bar : Uint64.t [@key 1] [@encoding `varint];
  baz : int      [@key 2] [@encoding `bits64];
}

When parsing or serializing, the values will be appropriately extended or truncated. If a value does not fit into the narrower type used for serialization or deserialization, Decoder.Error Decoder.Overflow or Encoder.Error Encoder.Overflow is raised.

The following table summarizes equivalence between integral types of protoc and encodings of deriving protobuf:

Encodingprotoc type
varintint32, int64, uint32, uint64
zigzagsint32, sint64
bits32fixed32, sfixed32
bits64fixed64, sfixed64

By default, OCaml types use the following encoding:

OCaml typeEncodingprotoc type
intvarintint32 or int64
int32 or Int32.tbits32sfixed32
Uint32.tbits32fixed32
int64 or Int64.tbits64sfixed64
Uint64.tbits64fixed64

Note that no OCaml type maps to zigzag-encoded sint32 or sint64 by default. It is necessary to use <code>[@encoding `zigzag]</code> explicitly.

Floats

Similarly to integers, float maps to protoc's double by default, but it is possible to specify the encoding explicitly:

message Floats {
  required float  foo = 1;
  required double bar = 2;
}

type floats = {
  foo : float [@key 1] [@encoding `bits32];
  bar : float [@key 2];
} [@@deriving protobuf]

Booleans

bool maps to protoc's bool and is encoded on wire using varint:

message Booleans {
  required bool bar = 1;
}

type booleans = {
  bar : bool [@key 1];
} [@@deriving protobuf]

Strings and bytes

All of string, String.t, bytes and Bytes.t map to protoc's string or bytes and are encoded on wire using bytes:

Note that unlike protoc, which has an additional invariant that the contents of a string must be valid UTF-8 text, deriving protobuf does not have this invariant. However, you still should observe it in your programs.

message Strings {
  required string bar = 1;
  required bytes  baz = 2;
}

type strings = {
  bar : string [@key 1];
  baz : bytes  [@key 2];
} [@@deriving protobuf]

Tuples

A tuple is treated in exactly same way as a record, except that keys are derived automatically starting at 1. The definition of search_request above could be rewritten as:

type search_request' = string * int option * int option
[@@deriving protobuf]

Additionally, a tuple can be used in any context where a scalar value is expected; in this case, it is equivalent to an anonymous inner message:

message Nested {
  message StringFloatPair {
    required string str = 1;
    required float  flo = 2;
  }
  required int32 foo = 1;
  optional StringFloatPair bar = 2;
}

type nested = {
  foo : int                     [@key 1];
  bar : (string * float) option [@key 2];
} [@@deriving protobuf]

Variants

An OCaml variant types is normally mapped to an entire Protobuf message by deriving protobuf, as opposed to protoc, which maps an enum to a simple varint. This is done because OCaml constructors can have arguments, but protoc's enums can not.

Note that even if a type doesn't have any constructor with arguments, it is still mapped to a message, because it would not be possible to extend the type later with a constructor with arguments otherwise.

Every constructor must have an explicitly specified key; if the constructor has one argument, it is mapped to an optional field with the key corresponding to the key of the constructor plus one. If there is more than one argument, they're treated like a tuple.

Consider this example:

message Variant {
  enum T {
    A = 1;
    B = 2;
    C = 3;
    D = 4;
  }
  message C {
    required string foo = 1;
    required string bar = 2;
  }
  message D {
    required string s1 = 1;
    required string s2 = 2;
  }
  required T t = 1;
  optional int32 b = 3; // (B = 2) + 1
  optional C c = 4; // (C = 3) + 1
  optional D d = 5; // (D = 4) + 1
}

type variant =
| A                              [@key 1]
| B of int                       [@key 2]
| C of string * string           [@key 3]
| D of {s1: string ; s2: string} [@key 4]
[@@deriving protobuf]

Note that decoder considers messages which contain more than one optional field invalid and rejects them.

In order to achieve better compatibility with protoc, it is possible to embed a variant where no constructors have arguments without wrapping it in a message:

enum BareVariant {
  A = 1;
  B = 2;
}
message Container {
  required T value = 1;
}

type bare_variant =
| A [@key 1]
| B [@key 2]
and container = {
  value : bare_variant [@key 1] [@bare];
} [@@deriving protobuf]

In practice, if a variant has no constructors with arguments, additional two functions are generated with the following signatures:

type t = A | B | ... [@@deriving protobuf]
val t_from_protobuf_bare : Protobuf.Decoder.t -> t
val t_to_protobuf_bare   : Protobuf.Encoder.t -> t -> unit

These functions do not expect additional framing; they just parse or serialize a single varint.

Polymorphic variants

Polymorphic variants are handled in exactly same way as regular variants. However, you can also embed them directly, like tuples, in which case the semantics is the same as defining an alias for the variant and then using that type.

This feature can be combined with the [@bare] annotation to create a useful shorthand:

message Packet {
  enum Type {
    REQUEST = 1;
    REPLY   = 2;
  }
  required Type  type  = 1;
  required int32 value = 2;
}

type packet = {
  type  : [ `Request [@key 1] | `Reply [@key 2] ] [@key 1] [@bare];
  value : int [@key 2];
} [@@deriving protobuf]

Type aliases

A type alias (statement of form type a = b) is treated by deriving protobuf as a definition of a message with one field with key 1:

message Alias {
  required int32 val = 1;
}

type alias = int [@@deriving protobuf]

Nested messages

When deriving protobuf encounters a non-scalar type, it generates a call to the serialization or deserialization function corresponding to the full path to the type.

Consider this definition:

type foo = bar * Baz.Quux.t [@@deriving protobuf]

The generated deserializer code will refer to bar_from_protobuf and Baz.Quux.t_from_protobuf; the serializer code will call bar_to_protobuf and Baz.Quux.t_to_protobuf.

Packed fields

Types which are encoded as varint, bits32 or bits64, that is, numeric fields or bare variants, can be declared as "packed" with the [@packed] attribute, in which case the serializer emits a more compact representation. Only protoc newer than 2.3.0 will recognize this representation. Note that the deserializer understands it regardless of the presence of [@packed] attribute.

message Packed {
  repeated int32 elem = 1 [packed=true];
}

type packed = int list [@key 1] [@packed] [@@deriving protobuf]

Parametric polymorphism

deriving protobuf is able to handle polymorphic type definitions. In this case, the serializing or deserializing function will accept one additional argument for every type variable; correspondingly, the value of this argument will be passed to serializer or deserializer of any nested parametric type.

Consider this example:

type 'a mylist =
| Nil                    [@key 1]
| Cons of 'a * 'a mylist [@key 2]
[@@deriving protobuf]

Here, the following functions will be generated:

val mylist_from_protobuf : (Protobuf.Decoder.t -> 'a) -> Protobuf.Decoder.t ->
                           'a mylist
val mylist_to_protobuf   : (Protobuf.Decoder.t -> 'a -> unit) -> Protobuf.Decoder.t ->
                           'a mylist -> unit

An example usage would be:

type a = int [@@deriving protobuf]

let get_ints message =
  let decoder = Protobuf.Decoder.of_bytes message in
  mylist_from_protobuf a_from_protobuf decoder

It's also possible to specify concrete types as parameters; in this case, deriving protobuf will infer the serializer/deserializer functions automatically. For example:

(* Combining two samples above *)
type b = a mylist [@@deriving protobuf]

Error handling

Both serializers and deserializers rigorously verify their input data. The only possible exception that can be raised during serialization is Protobuf.Encoder.Failure, and during deserialization is Protobuf.Decoder.Failure.

Decoder errors

The decoder attempts to annotate its failures with useful location information, but only if that wouldn't cost too much in terms of performance and complexity.

In general, as long as you're using the same protocol on both sides, deserialization or should never raise. The errors would mainly arise when interoperating with code generated by protoc that doesn't observe OCaml-specific invariants, or when handling malicious input.

It discerns these types of failure (represented with Decoder.Failure exception):

The decoder errors refer to fields via so-called "paths"; a path corresponds to the OCaml syntax for referring to a type, field or constructor, but can contain additional /<number> (e.g. /0) component for an immediate tuple.

For example, the string field will have the path Foo.r.ra/1:

(* foo.ml *)
type r = {
  ra: (int * string) option [@key 1];
} [@@deriving protobuf]

Encoder errors

The encoder discerns these types of failure (represented with Encoder.Failure exception):

The encoder errors use the same "path" convention as decoder errors.

Extending protocols

In real-world applications, implementations using multiple versions of the same protocol must coexist. Protocol Buffers offer an imperfect and sometimes complicated, but very powerful and practical solution to this problem.

The wire protocol is designed in a way that allows to safely extend it if one follows a set of constraints.

Always

Any of the following changes may be applied to either the sender or receiver of the message without breaking protocol:

Never

When communicating bidirectionally, violating any of the following constraints always results in exceptions or receiving garbage data:

The following sections list some exceptions to this rule when the communication is unidirectional.

On sender

Any of the following changes may be applied exclusively to the sender without breaking the existing receivers:

On receiver

Any of the following changes may be applied exclusively to the receiver without losing the ability to decode messages from existing senders:

Protoc export

deriving protobuf can export message types in proto2 language, the format that protoc accepts; protoc version 2.6 or later is required.

To enable protoc export, pass a protoc option to deriving protobuf:

(* foo.ml *)
type msg = ... [@@deriving protobuf { protoc }]

Compiling this file will create a file called Foo.protoc (note the capitalization) in a directory adjacent to foo.ml; if you are using ocamlbuild and foo.ml is located in directory src/, the file will be generated at _build/src/Foo.protoc. This can be customized by providing a path explicitly, e.g. [@@deriving protobuf { protoc = "Bar.protoc" }]; the path is interpreted relative to the source file.

The mapping between OCaml types and protoc messages is straightforward.

OCaml modules become protoc packages with the same name. A nested module, e.g. module Bar in our foo.ml, becomes a nested package, Foo.Bar; it will be emitted in a file Foo.Bar.protoc, placed adjacent to Foo.protoc, since protoc requires every package to reside in its own file.

OCaml records and their fields become protoc messages and fields with the same name:

type msg = {
  name:  string [@key 1];
  value: int    [@key 2];
} [@@deriving protobuf { protoc }]

message msg {
  required string name = 1;
  required int64 value = 2;
}

OCaml variants and their constructors become protoc messages and fields with the same name; additionally generated are a nested enum called _tag whose constants have the same name as constructors with _tag appended, and a field named tag with the type _tag:

type msg =
| A [@key 1]
| B of string [@key 2]
[@@deriving protobuf { protoc }]

message msg {
  enum _tag {
    A_tag = 1;
    B_tag = 2;
  }

  required _tag tag = 1;
  oneof value {
    string B = 3;
  }
}

OCaml tuples become protoc messages with the same name whose fields are called _N with N being the field index:

type msg = int * string
[@@deriving protobuf { protoc }]

message msg {
  required int64 _0 = 1;
  required string _1 = 2;
}

OCaml aliases become protoc messages with one field called _:

type msg = int
[@@deriving protobuf { protoc }]

message msg {
  required int64 _ = 1;
}

Sometimes, a single toplevel OCaml type definition has to be translated into several messages, e.g. when a field or a constructor contains a tuple or a polymorphic variant. In this case, such messages become nested messages whose name is the name of the field or constructor with _ prepended:

type msg = {
  field: int * string [@key 1]
}
[@@deriving protobuf { protoc }]

message msg {
  message _field {
    required int64 _0 = 1;
    required string _1 = 2;
  }

  required _field field = 1;
}

Normally, when a type from another module is referenced, deriving protobuf automatically generates the corresponding protoc import directive:

type imported = Other.msg
[@@deriving protobuf { protoc }]

import "Other.protoc";
message imported {
  required Other.msg _ = 1;
}

However, when a type is referenced that was defined in a module defined earlier in the same file, the produced import directive is incorrect. (deriving protobuf does not have an accurate model of OCaml's module scoping.) In this case, the protoc_import option can help:

(* foo.ml *)
module Bar = struct
  type msg = int [@@deriving protobuf { protoc }]
end

type alias = Bar.msg
[@@deriving protobuf { protoc; protoc_import = ["Foo.Bar.protoc"] }]

// Foo.protoc
package Foo;
import "Foo.Bar.protoc";
message alias {
  required Bar.msg _ = 1;
}

// Foo.Bar.protoc
package Foo.Bar;
message msg {
  required int64 _ = 1;
}

Compatibility

Protocol Buffers specification suggests that if a message contains multiple instances of a required or optional nested message, those nested messages should be merged. However, there is no concept of "merging messages" accessible to deriving protobuf, and this feature can be considered harmful anyway: it is far too forgiving of invalid input. Thus, deriving protobuf doesn't implement this merging.

deriving protobuf is more strict than protoc with numeric types; it raises Failure (Overflow fld) rather than silently truncate values. It is thought that accidentally losing 32th or 64th bit with OCaml's int type would be a common error without this countermeasure.

Everything else should be entirely compatible with protoc.

API Documentation

The documentation for internal API is available at GitHub pages.

License

MIT