Home

Awesome

Better Protobuf / gRPC Support for Python

:octocat: If you're reading this on github, please be aware that it might mention unreleased features! See the latest released README on pypi.

This project aims to provide an improved experience when using Protobuf / gRPC in a modern Python environment by making use of modern language features and generating readable, understandable, idiomatic Python code. It will not support legacy features or environments (e.g. Protobuf 2). The following are supported:

This project is heavily inspired by, and borrows functionality from:

Motivation

This project exists because I am unhappy with the state of the official Google protoc plugin for Python.

This project is a reimplementation from the ground up focused on idiomatic modern Python to help fix some of the above. While it may not be a 1:1 drop-in replacement due to changed method names and call patterns, the wire format is identical.

Installation

First, install the package. Note that the [compiler] feature flag tells it to install extra dependencies only needed by the protoc plugin:

# Install both the library and compiler
pip install "betterproto[compiler]"

# Install just the library (to use the generated code output)
pip install betterproto

Betterproto is under active development. To install the latest beta version, use pip install --pre betterproto.

Getting Started

Compiling proto files

Given you installed the compiler and have a proto file, e.g example.proto:

syntax = "proto3";

package hello;

// Greeting represents a message you can tell a user.
message Greeting {
  string message = 1;
}

You can run the following to invoke protoc directly:

mkdir lib
protoc -I . --python_betterproto_out=lib example.proto

or run the following to invoke protoc via grpcio-tools:

pip install grpcio-tools
python -m grpc_tools.protoc -I . --python_betterproto_out=lib example.proto

This will generate lib/hello/__init__.py which looks like:

# Generated by the protocol buffer compiler.  DO NOT EDIT!
# sources: example.proto
# plugin: python-betterproto
from dataclasses import dataclass

import betterproto


@dataclass
class Greeting(betterproto.Message):
    """Greeting represents a message you can tell a user."""

    message: str = betterproto.string_field(1)

Now you can use it!

>>> from lib.hello import Greeting
>>> test = Greeting()
>>> test
Greeting(message='')

>>> test.message = "Hey!"
>>> test
Greeting(message="Hey!")

>>> serialized = bytes(test)
>>> serialized
b'\n\x04Hey!'

>>> another = Greeting().parse(serialized)
>>> another
Greeting(message="Hey!")

>>> another.to_dict()
{"message": "Hey!"}
>>> another.to_json(indent=2)
'{\n  "message": "Hey!"\n}'

Async gRPC Support

The generated Protobuf Message classes are compatible with grpclib so you are free to use it if you like. That said, this project also includes support for async gRPC stub generation with better static type checking and code completion support. It is enabled by default.

Given an example service definition:

syntax = "proto3";

package echo;

message EchoRequest {
  string value = 1;
  // Number of extra times to echo
  uint32 extra_times = 2;
}

message EchoResponse {
  repeated string values = 1;
}

message EchoStreamResponse  {
  string value = 1;
}

service Echo {
  rpc Echo(EchoRequest) returns (EchoResponse);
  rpc EchoStream(EchoRequest) returns (stream EchoStreamResponse);
}

Generate echo proto file:

python -m grpc_tools.protoc -I . --python_betterproto_out=. echo.proto

A client can be implemented as follows:

import asyncio
import echo

from grpclib.client import Channel


async def main():
    channel = Channel(host="127.0.0.1", port=50051)
    service = echo.EchoStub(channel)
    response = await service.echo(echo.EchoRequest(value="hello", extra_times=1))
    print(response)

    async for response in service.echo_stream(echo.EchoRequest(value="hello", extra_times=1)):
        print(response)

    # don't forget to close the channel when done!
    channel.close()


if __name__ == "__main__":
    loop = asyncio.get_event_loop()
    loop.run_until_complete(main())

which would output

EchoResponse(values=['hello', 'hello'])
EchoStreamResponse(value='hello')
EchoStreamResponse(value='hello')

This project also produces server-facing stubs that can be used to implement a Python gRPC server. To use them, simply subclass the base class in the generated files and override the service methods:

import asyncio
from echo import EchoBase, EchoRequest, EchoResponse, EchoStreamResponse
from grpclib.server import Server
from typing import AsyncIterator


class EchoService(EchoBase):
    async def echo(self, echo_request: "EchoRequest") -> "EchoResponse":
        return EchoResponse([echo_request.value for _ in range(echo_request.extra_times)])

    async def echo_stream(self, echo_request: "EchoRequest") -> AsyncIterator["EchoStreamResponse"]:
        for _ in range(echo_request.extra_times):
            yield EchoStreamResponse(echo_request.value)


async def main():
    server = Server([EchoService()])
    await server.start("127.0.0.1", 50051)
    await server.wait_closed()

if __name__ == '__main__':
    loop = asyncio.get_event_loop()
    loop.run_until_complete(main())

JSON

Both serializing and parsing are supported to/from JSON and Python dictionaries using the following methods:

For compatibility the default is to convert field names to camelCase. You can control this behavior by passing a casing value, e.g:

MyMessage().to_dict(casing=betterproto.Casing.SNAKE)

Determining if a message was sent

Sometimes it is useful to be able to determine whether a message has been sent on the wire. This is how the Google wrapper types work to let you know whether a value is unset, set as the default (zero value), or set as something else, for example.

Use betterproto.serialized_on_wire(message) to determine if it was sent. This is a little bit different from the official Google generated Python code, and it lives outside the generated Message class to prevent name clashes. Note that it only supports Proto 3 and thus can only be used to check if Message fields are set. You cannot check if a scalar was sent on the wire.

# Old way (official Google Protobuf package)
>>> mymessage.HasField('myfield')

# New way (this project)
>>> betterproto.serialized_on_wire(mymessage.myfield)

One-of Support

Protobuf supports grouping fields in a oneof clause. Only one of the fields in the group may be set at a given time. For example, given the proto:

syntax = "proto3";

message Test {
  oneof foo {
    bool on = 1;
    int32 count = 2;
    string name = 3;
  }
}

On Python 3.10 and later, you can use a match statement to access the provided one-of field, which supports type-checking:

test = Test()
match test:
    case Test(on=value):
        print(value)  # value: bool
    case Test(count=value):
        print(value)  # value: int
    case Test(name=value):
        print(value)  # value: str
    case _:
        print("No value provided")

You can also use betterproto.which_one_of(message, group_name) to determine which of the fields was set. It returns a tuple of the field name and value, or a blank string and None if unset.

>>> test = Test()
>>> betterproto.which_one_of(test, "foo")
["", None]

>>> test.on = True
>>> betterproto.which_one_of(test, "foo")
["on", True]

# Setting one member of the group resets the others.
>>> test.count = 57
>>> betterproto.which_one_of(test, "foo")
["count", 57]

# Default (zero) values also work.
>>> test.name = ""
>>> betterproto.which_one_of(test, "foo")
["name", ""]

Again this is a little different than the official Google code generator:

# Old way (official Google protobuf package)
>>> message.WhichOneof("group")
"foo"

# New way (this project)
>>> betterproto.which_one_of(message, "group")
["foo", "foo's value"]

Well-Known Google Types

Google provides several well-known message types like a timestamp, duration, and several wrappers used to provide optional zero value support. Each of these has a special JSON representation and is handled a little differently from normal messages. The Python mapping for these is as follows:

Google MessagePython TypeDefault
google.protobuf.durationdatetime.timedelta0
google.protobuf.timestampTimezone-aware datetime.datetime1970-01-01T00:00:00Z
google.protobuf.*ValueOptional[...]None
google.protobuf.*betterproto.lib.google.protobuf.*None

For the wrapper types, the Python type corresponds to the wrapped type, e.g. google.protobuf.BoolValue becomes Optional[bool] while google.protobuf.Int32Value becomes Optional[int]. All of the optional values default to None, so don't forget to check for that possible state. Given:

syntax = "proto3";

import "google/protobuf/duration.proto";
import "google/protobuf/timestamp.proto";
import "google/protobuf/wrappers.proto";

message Test {
  google.protobuf.BoolValue maybe = 1;
  google.protobuf.Timestamp ts = 2;
  google.protobuf.Duration duration = 3;
}

You can do stuff like:

>>> t = Test().from_dict({"maybe": True, "ts": "2019-01-01T12:00:00Z", "duration": "1.200s"})
>>> t
Test(maybe=True, ts=datetime.datetime(2019, 1, 1, 12, 0, tzinfo=datetime.timezone.utc), duration=datetime.timedelta(seconds=1, microseconds=200000))

>>> t.ts - t.duration
datetime.datetime(2019, 1, 1, 11, 59, 58, 800000, tzinfo=datetime.timezone.utc)

>>> t.ts.isoformat()
'2019-01-01T12:00:00+00:00'

>>> t.maybe = None
>>> t.to_dict()
{'ts': '2019-01-01T12:00:00Z', 'duration': '1.200s'}

Generating Pydantic Models

You can use python-betterproto to generate pydantic based models, using pydantic dataclasses. This means the results of the protobuf unmarshalling will be typed checked. The usage is the same, but you need to add a custom option when calling the protobuf compiler:

protoc -I . --python_betterproto_opt=pydantic_dataclasses --python_betterproto_out=lib example.proto

With the important change being --python_betterproto_opt=pydantic_dataclasses. This will swap the dataclass implementation from the builtin python dataclass to the pydantic dataclass. You must have pydantic as a dependency in your project for this to work.

Configuration typing imports

By default typing types will be imported directly from typing. This sometimes can lead to issues in generation if types that are being generated conflict with the name. In this case you can configure the way types are imported from 3 different options:

Direct

protoc -I . --python_betterproto_opt=typing.direct --python_betterproto_out=lib example.proto

this configuration is the default, and will import types as follows:

from typing import (
  List,
  Optional,
  Union
)
...
value: List[str] = []
value2: Optional[str] = None
value3: Union[str, int] = 1

Root

protoc -I . --python_betterproto_opt=typing.root --python_betterproto_out=lib example.proto

this configuration loads the root typing module, and then access the types off of it directly:

import typing
...
value: typing.List[str] = []
value2: typing.Optional[str] = None
value3: typing.Union[str, int] = 1

310

protoc -I . --python_betterproto_opt=typing.310 --python_betterproto_out=lib example.proto

this configuration avoid loading typing all together if possible and uses the python 3.10 pattern:

...
value: list[str] = []
value2: str | None = None
value3: str | int = 1

Development

Requirements

Setup

# Get set up with the virtual env & dependencies
poetry install -E compiler

# Activate the poetry environment
poetry shell

Code style

This project enforces black python code formatting.

Before committing changes run:

poe format

To avoid merge conflicts later, non-black formatted python code will fail in CI.

Tests

There are two types of tests:

  1. Standard tests
  2. Custom tests

Standard tests

Adding a standard test case is easy.

It will be picked up automatically when you run the tests.

Custom tests

Custom tests are found in tests/test_*.py and are run with pytest.

Running

Here's how to run the tests.

# Generate assets from sample .proto files required by the tests
poe generate
# Run the tests
poe test

To run tests as they are run in CI (with tox) run:

poe full-test

(Re)compiling Google Well-known Types

Betterproto includes compiled versions for Google's well-known types at src/betterproto/lib/google. Be sure to regenerate these files when modifying the plugin output format, and validate by running the tests.

Normally, the plugin does not compile any references to google.protobuf, since they are pre-compiled. To force compilation of google.protobuf, use the option --custom_opt=INCLUDE_GOOGLE.

Assuming your google.protobuf source files (included with all releases of protoc) are located in /usr/local/include, you can regenerate them as follows:

protoc \
    --plugin=protoc-gen-custom=src/betterproto/plugin/main.py \
    --custom_opt=INCLUDE_GOOGLE \
    --custom_out=src/betterproto/lib \
    -I /usr/local/include/ \
    /usr/local/include/google/protobuf/*.proto

TODO

Community

Join us on Slack!

License

Copyright © 2019 Daniel G. Taylor

http://dgt.mit-license.org/