fastgrpc3

gRPC for Python, the FastAPI way — and up to 14× faster per core.

Write gRPC services as plain async functions with Pydantic models. fastgrpc3 gives you dependency injection, typed clients, rich errors, health checking, reflection and interactive docs — and protobuf encoded and decoded in Rust. Add the Rust transport, and the same app serves 62,000 calls per second on a single core.

from pydantic import BaseModel
from fastgrpc3 import GrpcApp, GrpcService, Insecure


class Hello(BaseModel):
    name: str


class Reply(BaseModel):
    text: str


service = GrpcService("Greeter", package="hello.v1")


@service.unary()
async def SayHello(req: Hello) -> Reply:
    return Reply(text=f"Hello, {req.name}!")


app = GrpcApp(tls=Insecure())
app.add_service(service)

pip install "fastgrpc3[tonic]"
fastgrpc3 run hello:app --transport tonic

Why fastgrpc3

• Fast where it counts. On one core, an app on the tonic transport answers 14× more calls than on grpc.aio, with 17× less CPU per call and a p99 latency 8× lower — without changing a line of the handlers. On grpc.aio, the framework costs about 3 % over hand-written stubs.
• Pydantic in, Pydantic out. Requests and responses are Pydantic models, validated on the way in; protobuf runs in Rust through protoruf, with no protoc and no _pb2 files.
• Two ways to declare a contract. Write Pydantic models and get the .proto, with field numbers pinned in proto.lock and breaking changes caught in CI; or start from .proto files and generate models, services and typed clients that pass mypy --strict.
• Everything a service needs. The four kinds of call, Depends(...) with yield cleanup, hooks, middlewares, exception handlers, deadlines propagated to downstream calls, standard google.rpc error details, TestClient.
• Standard gRPC. Real protobuf on the wire: clients in any language can call your service. Health checking (Kubernetes, grpc-health-probe), server reflection (grpcurl), Swagger UI with a JSON gateway to try the RPCs from the browser.
• Secure by default. TLS(...), MTLS(...) or Insecure(), always explicit; client identities from mTLS certificates.

Performance

One CPU core, unary calls, 64 concurrent clients:

The tonic transport (fastgrpc3-transport) runs HTTP/2, gRPC framing, TLS and flow control in Rust and hands the calls to asyncio by batches; wheels for Linux, macOS and Windows. Medium and large messages, latency, several cores, dependencies, codec and method: Performance.

Install

pip install fastgrpc3              # grpc.aio transport
pip install "fastgrpc3[tonic]"     # + the Rust transport (experimental)
pip install "fastgrpc3[docs]"      # + Swagger UI and the JSON gateway

Python 3.12, 3.13 and 3.14. fastgrpc3 is at version 0.1: the API may still change. A complete example, a bookstore service with its client and tests, is in examples/.

Server and client

import asyncio
from collections.abc import AsyncIterator

from pydantic import BaseModel
from fastgrpc3 import Context, Depends, GrpcApp, GrpcError, GrpcService, Insecure, StatusCode


class GetOrder(BaseModel):
    order_number: str


class Order(BaseModel):
    order_number: str
    paid: bool = False


service = GrpcService("Orders", package="orders.v1")


def get_repo() -> dict[str, Order]:
    return {"A": Order(order_number="A", paid=True)}


@service.unary()
async def Get(req: GetOrder, ctx: Context, repo: dict[str, Order] = Depends(get_repo)) -> Order:
    if req.order_number not in repo:
        raise GrpcError(StatusCode.NOT_FOUND, "no such order")
    return repo[req.order_number]


@service.unary_stream()
async def Watch(req: GetOrder) -> AsyncIterator[Order]:
    yield Order(order_number=req.order_number)


app = GrpcApp(address="[::]:50051", tls=Insecure())  # or TLS(cert=..., key=...)
app.add_service(service)
app.enable_health()  # grpc.health.v1: Check, Watch, List (grpc-health-probe, Kubernetes)
app.enable_reflection()  # grpc.reflection v1 and v1alpha (grpcurl list / describe / call)
app.enable_docs(gateway=True)  # Swagger UI on 127.0.0.1:8080/docs (fastgrpc3[docs])

if __name__ == "__main__":
    asyncio.run(app.run())  # stops cleanly on SIGINT / SIGTERM

from fastgrpc3 import Channel, GrpcClient, Insecure


class OrdersClient(GrpcClient):
    package = "orders.v1"
    service = "Orders"
    get = GrpcClient.unary("Get", GetOrder, Order)
    watch = GrpcClient.unary_stream("Watch", GetOrder, Order)


async def main() -> None:
    async with Channel("localhost:50051", tls=Insecure()) as channel:
        orders = OrdersClient(channel)
        order = await orders.get(GetOrder(order_number="A"))
        async for update in orders.watch(GetOrder(order_number="A")):
            print(update)

Also: dependencies with sub-dependencies, per-call cache and yield cleanup; app.add_interceptor (pre-call hooks), @app.middleware (around the handler) and @app.exception_handler; rich errors with the standard google.rpc details (ErrorInfo, BadRequest…); deadline propagation to calls made from a handler; TLS / MTLS / Insecure always explicit; fastgrpc3.testing.TestClient for tests.

fastgrpc3 run myapp.main:app --address [::]:50051 --workers 4   # one process per core

Workers share the port with SO_REUSEPORT (Linux); otherwise a busy port makes the server fail to start instead of silently sharing it.

Code first: Pydantic models -> .proto

fastgrpc3 lock, check and export take an app (myapp.main:app) or a SchemaRegistry.

from pydantic import BaseModel
from fastgrpc3 import MethodDef, SchemaRegistry, ServiceDef, field
from fastgrpc3.types import UInt32


class GetOrder(BaseModel):
    order_number: str


class Order(BaseModel):
    order_number: str
    qty: UInt32 = 0  # proto uint32, bounds checked by Pydantic
    note: str | None = field(5, default=None)  # explicit number


registry = SchemaRegistry()
registry.add_service(ServiceDef("Orders", "orders.v1", (MethodDef("Get", GetOrder, Order),)))

fastgrpc3 lock myapp.contract:registry     # number new fields, write proto.lock (commit it)
fastgrpc3 check myapp.contract:registry    # CI: fail if proto.lock is out of date
fastgrpc3 export myapp.contract:registry -o proto/   # one .proto file per package

Fields without an explicit number get one the first time they are locked, wherever they are declared; a removed field's number is reserved for good. Types without a proto equivalent (Decimal, UUID, unions…) travel as a JSON string, marked // json in the .proto.

Proto first: .proto -> Pydantic models

fastgrpc3 generate proto/ -o src/myapp/gen   # -I third_party/ for extra imports
fastgrpc3 generate proto/ -o src/myapp/gen --check   # CI: fail if not up to date

orders/v1/orders.proto gives orders/v1/orders_models.py (one ProtoModel per message, one open ProtoEnum per enum) and orders/v1/orders_service.py; _descriptor.py embeds the compiled descriptor, so nothing is compiled at startup. Defaults can live in pyproject.toml:

[tool.fastgrpc3.generate]
proto_paths = ["proto"]
output = "src/myapp/gen"

Each service module also holds the contract (Orders, to serve with GrpcService.implement(Orders)) and a typed client (OrdersClient). Generated code passes mypy --strict and ruff. Well-known types become Python types (Timestamp -> datetime, Struct -> dict[str, Any], wrappers -> X | None…).

Documentation

https://edwinalkins.github.io/fastgrpc3/ — getting started, contracts (code first and proto first), server, client, testing, the type mapping, the command line and performance.

The sources live in docs/ (MkDocs). To build the site into site/, or to serve it locally with live reload:

uv run --project docs mkdocs build --strict
uv run --project docs mkdocs serve          # http://127.0.0.1:8000

Development

uv sync                     # install the package and the dev tools
uv run pytest               # tests
uv run ruff check .         # lint
uv run ruff format --check  # formatting
uv run mypy src tests       # type checking

License

MIT, see LICENSE.