FastProto

Fast, Pythonic Protocol Buffers — messages are plain, readable @dataclass types, with all encoding and decoding handled by a compiled Rust core.

Google's Python protobuf generates opaque classes full of getters/setters and a reflection API you have to learn. FastProto instead generates clean dataclasses you can construct, compare, and repr() like any other — and does the wire work in Rust.

• Idiomatic messages — generated code is a @dataclass with plain annotations (str, int, list[...], dict[...], | None); autocomplete, type checkers, and repr() all just work.
• Rust wire codec — encode/decode run in Rust via PyO3, not pure Python.
• Wire-compatible — bytes interoperate both ways with Google's reference protobuf runtime.
• Standard toolchain — ships a protoc plugin; just add --fastproto_out.

Install

pip install fastproto            # runtime (Python 3.12+)
pip install "fastproto[plugin]"  # + the protoc code generator

Code generation also needs the protoc compiler itself — install it from your package manager (brew install protobuf, apt install protobuf-compiler) or the official releases.

Quick start

1. Define user.proto:

syntax = "proto3";
package example;

enum Role {
  ROLE_UNSPECIFIED = 0;
  ROLE_ADMIN = 1;
  ROLE_USER = 2;
}

message Address {
  string city = 1;
  string street = 2;
}

message User {
  int64 id = 1;
  string name = 2;
  optional string email = 3;
  Role role = 4;
  repeated string tags = 5;
  Address address = 6;
  map<string, int32> counters = 7;

  oneof contact {
    string phone = 8;
    string telegram = 9;
  }
}

2. Generate with protoc:

protoc --proto_path=. --fastproto_out=. user.proto

This writes user_pb.py — a plain, readable dataclass module:

# @generated by fastproto. DO NOT EDIT.
# source: user.proto
# pyright: reportUnknownVariableType=false
from dataclasses import dataclass, field
from enum import IntEnum

from fastproto import Message, Scalar, message


class Role(IntEnum):
    ROLE_UNSPECIFIED = 0
    ROLE_ADMIN = 1
    ROLE_USER = 2


_ADDRESS_DESCRIPTOR = bytes.fromhex("...")  # @generated (bytes elided)


@message(_ADDRESS_DESCRIPTOR)
@dataclass(slots=True)
class Address(Message):
    city: Scalar.String = ""
    street: Scalar.String = ""


_USER_DESCRIPTOR = bytes.fromhex("...")  # @generated (bytes elided)


@message(_USER_DESCRIPTOR)
@dataclass(slots=True)
class User(Message):
    id: Scalar.Int64 = 0
    name: Scalar.String = ""
    email: Scalar.String | None = None
    role: Role = Role(0)
    tags: list[Scalar.String] = field(default_factory=list)
    address: "Address | None" = None
    counters: dict[Scalar.String, Scalar.Int32] = field(default_factory=dict)
    phone: Scalar.String | None = None
    telegram: Scalar.String | None = None

3. Use it like any dataclass:

from user_pb import Address, Role, User

user = User(
    id=42,
    name="Ada",
    role=Role.ROLE_ADMIN,
    tags=["vip", "beta"],
    address=Address(city="London", street="Baker St"),
    counters={"logins": 7},
)

data = user.to_bytes()               # serialize to protobuf wire bytes
assert User.from_bytes(data) == user  # and back

No SerializeToString() / ParseFromString() ceremony and no reflection — just to_bytes() / from_bytes() on a dataclass you can build, compare, and print.

Type mapping

• Each proto scalar maps to an alias under fastproto.Scalar. An alias is just the underlying Python type (int, str, ...) tagged with Annotated[...], so it type-checks as the base type while still recording the exact wire type.

  proto	Python	proto	Python
  double	Scalar.Double	fixed32	Scalar.Fixed32
  float	Scalar.Float	fixed64	Scalar.Fixed64
  int32	Scalar.Int32	sfixed32	Scalar.SFixed32
  int64	Scalar.Int64	sfixed64	Scalar.SFixed64
  uint32	Scalar.UInt32	bool	Scalar.Bool
  uint64	Scalar.UInt64	string	Scalar.String
  sint32	Scalar.SInt32	bytes	Scalar.Bytes
  sint64	Scalar.SInt64

• Composite fields: repeated T → list[T], map<K, V> → dict[K, V], enum → IntEnum, and optional / message / oneof fields → T | None.

• Well-known types map to native objects: google.protobuf.Timestamp → datetime, Duration → timedelta; the rest (Any, Struct, wrappers, ...) are plain dataclasses in fastproto.wellknown.

• Nested message / enum definitions generate as nested classes at any depth (Outer.Inner, Outer.Color), so the Python structure mirrors the .proto.

• Multi-file schemas work — import in a .proto becomes an import between the generated modules.

Semantics

• Presence (proto3): plain scalars use their zero value and are not nullable; optional scalars, message fields, and oneof members are T | None and track explicit presence (a set empty string is distinct from unset). An all-default message encodes to b"".
• oneof: members are plain optional fields; setting more than one raises ValueError at encode time. msg.which_oneof("group") returns the name of the set member (or None), like google's WhichOneof.
• Open enums (proto3): an enum value not defined in your schema decodes to a plain int (like google's runtime) and survives re-encoding.
• Repeated message merge: not performed. If a singular message field appears more than once on the wire, the last occurrence wins (google merges the submessages field-by-field). A wire-compat caveat for concatenated streams.
• Unknown fields: fields your schema doesn't know are preserved verbatim across decode → encode (forward compatibility). They are invisible to __init__ / repr / ==. Exception: proto2 group wire types can't be skipped, so a message containing one raises ValueError on decode.
• Strict types at the boundary: field values are taken as-is — bool does not accept int, str and bytes don't interconvert, and a repeated field must be a real list (a bare str/bytes is rejected, not iterated char-by-char). An out-of-range integer raises OverflowError on encode. from_bytes takes bytes (not yet bytearray / memoryview).
• Nesting limit: messages deeper than 100 levels — or a cyclic object graph on encode — raise ValueError instead of exhausting the stack.
• Timestamps: decoded datetimes are aware UTC; naive ones encode as UTC. protobuf's sub-microsecond precision (nanos) is truncated to microseconds.
• References: sibling, self, nested, and enum references resolve lazily on the first to_bytes() / from_bytes() — nothing for you to wire up.

empty = User()
assert empty.to_bytes() == b"" and empty.email is None
assert User(phone="1").which_oneof("contact") == "phone"
User(phone="1", telegram="a").to_bytes()  # ValueError: ... oneof ...

Performance

Protobuf runtimes pay the cost of turning bytes into Python objects either at decode time or at access time, so decode-only microbenchmarks tell half the story. Google's default backend (upb) parses into C structs and converts to Python lazily — decoding looks instant, but every field access pays a C-to-Python conversion, every time. FastProto materializes plain Python values once at decode; after that a field read is an ordinary attribute load.

On a mid-size message (509 B: strings, nested messages, maps, repeated fields, enums — Apple M-series, CPython 3.14, bench/compare.py):

So: if you decode messages and barely look inside (relays, routers), upb's lazy model wins the raw numbers. The moment you use what you decoded, the conversion bill comes due on their side — and it comes due again on every re-read — while FastProto's fields are just dataclass attributes. The encode gap is the honest price of that interface: FastProto reads live attributes off a plain Python object, upb serializes C memory it already owns. That price buys messages your editor, type checker, and repr() treat as ordinary dataclasses.

Reproduce with uv run --with protobuf python bench/compare.py.

Contributing

See CONTRIBUTING.md for setup, project layout, and the release flow.

License

MIT