Implementing gRPC server using Python

Nowadays, when people want to implement backend API, they go straight to creating application with RESTful API that communicates using JSON, without even considering other options. In the recent years though, gRPC and its protobufs started to get some traction and popularity thanks to many of their advantages. So, let’s see what’s all the buzz/hype about and implement gRPC server using Python!

TL;DR: Here is my repository with grpc branch with all the sources from this article: https://github.com/MartinHeinz/python-project-blueprint/tree/grpc

French Telco Orange Hit by Cyber-Attack

ATC Ghana supports Girls-In-ICT Program

What is gRPC, Anyway?

gRPC is Remote Procedure Call (RPC) protocol, that leverages Protocol Buffers (protobufs) as it’s message format. Using gRPC, client application can directly call method available on remote server using method stubs. It doesn’t matter in what language the server-side application is implemented as long as you have stubs (generated) for you client-side language. gRPC supports many languages, including Go, Java, Ruby, C# or our language of choice – Python.

Now, what are Protocol Buffers (protobufs)? Protobufs are alternative to formats like JSON or XML. They are smaller, simpler and more efficient way of serializing data. To use protobufs, you need to define how you want the exchanged messages to look, for example (for more specifics see this language guide):

// example.proto
syntax = "proto3";

package example;

message User {
  int32 id = 1;
  string name = 2;
}

message UserInfo {
  int32 age = 1;
  string address = 2;
  string phoneNumber = 3;
}

Aside from messages, we also need to define services and their rpc methods that will be implemented on server-side and called from client-side:

// ...

service UserService {
  rpc GetUserInfo (User) returns (UserInfo) {}
}

Why Should I Care, Anyway?

There are quite a few advantages to using gRPC over REST. First off, gRPC is much better when it comes to performance, thanks to the tight packing of protobufs, which reduces size of payloads being send. It also uses HTTP/2, instead of HTTP/1.1 like REST does. For these reasons it’s a great choice for IoT, mobile devices or other constrained/low-power environments.

Another reason to choose gRPC over REST is that REST doesn’t mandate any real structure. You might define format of requests and responses using OpenAPI, but that’s loose and optional. gRPC contracts on the other hand are stricter and clearly defined.

As already mentioned gRPC is using HTTP/2 and it’s good to mention that is’s taking full advantage of it’s features. To name a few: concurrent requests, streaming instead of request-response, smaller sensitivity to latency.

That said, there are also disadvantages and the biggest one being adoption. Not all clients (browsers) support use of HTTP/2 making it problematic for external use. Considering performance advantages though, it’s clearly great choice for internal services and communication, which you have control over.

Setting Up

To do anything with gRPC and protobufs, we need to install its compiler:

#!/bin/bash
# Download and Unzip compiler
curl -OL https://github.com/protocolbuffers/protobuf/releases/download/v3.11.4/protoc-3.11.4-linux-x86_64.zip
unzip protoc-3.11.4-linux-x86_64.zip -d protoc3

# Move the binary to directory which is PATH
sudo mv protoc3/bin/* /usr/local/bin/

sudo mv protoc3/include/* /usr/local/include/

# Change owner
sudo chown $USER /usr/local/bin/protoc
sudo chown -R $USER /usr/local/include/google

# Test if it works
protoc --version
# libprotoc 3.11.4

Considering that we are using python to build our application, we will also need grpcio and grpcio-tools libraries:

# Activate venv if you have one
source .../venv/bin/activate
pip install grpcio grpcio-tools

Let’s Build Something

With all tools ready we can now actually move onto building the application. For this example I chose simple echo server that sends you back your own messages.

First thing we should talk about though, is project layout. I chose following directory/file structure:

.
├── blueprint                 <- Source root - Name of our project
│   ├── app.py                <- Application server
│   ├── echo_client.py        <- gRPC client for testing
│   ├── generated             <- Generated gRPC Python code
│   │   ├── echo_pb2_grpc.py
│   │   └── echo_pb2.py
│   ├── grpc.py               <- Actual gRPC code
│   ├── __init__.py
│   ├── __main__.py
│   ├── proto                 <- Protobuf definitions
│   │   └── echo.proto
│   └── resources
├── tests                     <- Test suite
│   ├── conftest.py
│   ├── context.py
│   ├── __init__.py
│   └── test_grpc.py
├── Makefile
├── pytest.ini
└── requirements.txt

This layout helps us clearly separate protobuf files (.../proto), generated sources (.../generated), actual source code and our test suite.

So, to build the gRPC server, we – first and foremost – need to define messages and service(s) it will use you to communicate with clients:

// echo.proto
syntax = "proto3";
package echo;

// The request message containing the user's message.
message EchoRequest {
  string message = 1;
}

// The response message containing the original message.
message EchoReply {
  string message = 1;
}

// The echo service definition.
service Echo {
  // Echo back reply.
  rpc Reply (EchoRequest) returns (EchoReply) {}
}

In this echo.proto file we can see very simple definition of message types – one for request (EchoRequest) and one for reply (EchoReply) from server. These messages are then used by Echo service, which consists of one RPC method called Reply.

To be able to use these definitions in Python code, we need to generate server and client interfaces. To do so, we can run this command:

python3 -m grpc_tools.protoc \
        --python_out=./blueprint/generated \
        --grpc_python_out=./blueprint/generated \
        ./blueprint/proto/*.proto

sed -i -E 's/^import.*_pb2/from . \0/' ./blueprint/generated/*.py

We specified quite a few arguments. First of them – -I blueprint/proto, tells grpc_tools where to look for our .proto files (it defines PATH). Next two, --python-out and --grpc_python_out specify where to output generated *_pb2.py and *_grpc_pb2.py files respectively. Last argument – ./blueprint/proto/*.proto is actual path to .proto files – this might seem like redundant as we specified PATH with -I, but you need both to make this work.

When you run this one command however, you will end up with some broken imports in these generated files. There are multiple raised issues in grpc and protobuf repositories and the easiest solution is to just fix those imports with sed.

Writing this command out would not be very enjoyable nor efficient, so I wrapped it in make target to make your (and mine) life easier. You can find complete Makefile in my repository here.

There’s quite a bit of code that gets generated using this command, so I won’t go over every little bit, but there are a few important things to look out for. Each of these *_pb2_grpc.py files has following 3 things:

Stub – First of them – Stub and in our case EchoStub – is a class used by client to connect to gRPC service
Servicer – In our case EchoServicer – is used by server to implement the gRPC service
Registration Function – Finally piece, add_EchoServicer_to_server that is needed to register servicer with gRPC server.

So, let’s go over the code to see how to use this generated code.

First thing we want to do is implement the actual service. We will do that in grpc.py file, where we will keep all gRPC specific code:

# grpc.py
from .generated import echo_pb2_grpc, echo_pb2


class Echoer(echo_pb2_grpc.EchoServicer):

    def Reply(self, request, context):
        return echo_pb2.EchoReply(message=f'You said: {request.message}')

Above, we create Echoer class which inherits from generated EchoServicer class, which contains all the methods defined in .proto file. All these methods are intended to be overridden in our implementation and here we do just that. We implement the only method Reply by returning EchoReply message that we also previously defined in the .proto file. I want to point out that the request parameter here is an instance of EchoReply – that’s why we can get the message out of it. We don’t use the context parameter here, but it contains some useful RPC-specific information, like timeout limits for example.

One neat feature I also want mention is that in case you wanted to leverage the response-streaming you could replace return with yield and return multiple responses (in for cycle).

Now that we implemented the gRPC service, we want to run it. To do so, we need a server:

# app.py
from concurrent import futures
import grpc

from .generated import echo_pb2_grpc
from .grpc import Echoer


class Server:

    @staticmethod
    def run():
        server = grpc.server(futures.ThreadPoolExecutor(max_workers=10))
        echo_pb2_grpc.add_EchoServicer_to_server(Echoer(), server)
        server.add_insecure_port('[::]:50051')
        server.start()
        server.wait_for_termination()

All this is, is just a single static method. It creates server from grpc library with a few workers – in this case 10. After that it uses previously mentioned registration function (add_EchoServicer_to_server) to bind our Echoer service to the server. Remaining 3 lines just add listening port, start the server and wait for interrupt.

All that remains for the server-side is __main__.py, so that we can start it as Python module:

# __main__.py
from .app import Server

if __name__ == '__main__':
    Server.run()

With that, we are all set to start the server. You can do that with python -m blueprint or if you are using my template, then just make run.

We have server running, but we have no way of calling it… That’s where the client comes it. For demonstration purposes we will create the client in Python using the stubs generated for us, but you could write the client in completely different language.

from __future__ import print_function
import logging

import grpc

from .generated import echo_pb2
from .generated import echo_pb2_grpc


def run():
    with grpc.insecure_channel('localhost:50051') as channel:
        stub = echo_pb2_grpc.EchoStub(channel)
        response = stub.Reply(echo_pb2.EchoRequest(message='Hello World!'))
    print("Echo client received: " + response.message)


if __name__ == '__main__':
    logging.basicConfig()
    run()

For the client we need only one function, which we call run. After connecting the server, it creates stub that will allow us to call the server method, which is the next step. It calls Reply method implemented on server-side by passing in EchoRequest message with some payload. All that’s left is to just print it.

Now, let’s run the client and see if everything works as expected:

~ $ python -m blueprint.echo_client
Echo client received: You said: Hello World!

And it does work!

Testing with Pytest

As with all my little projects and articles, we are not done until there are unit tests for all the code. To write sample test for this gRPC server, I will use Pytest and its plugin pytest-grpc.

Let’s first have a look at the fixtures used to simulate request-response exchange between client and server:

# conftest.py
import pytest


@pytest.fixture(scope='module')
def grpc_add_to_server():
    from blueprint.generated.echo_pb2_grpc import add_EchoServicer_to_server

    return add_EchoServicer_to_server


@pytest.fixture(scope='module')
def grpc_servicer():
    from blueprint.grpc import Echoer

    return Echoer()


@pytest.fixture(scope='module')
def grpc_stub(grpc_channel):
    from blueprint.generated.echo_pb2_grpc import EchoStub

    return EchoStub(grpc_channel)

I think these are all pretty simple. Just make sure to use these specific names for these fixtures, as that’s what the plugin looks for. One thing to notice is the grpc_channel argument in grpc_stub. This is a fake channel supplied by pytest-grpc plugin. For more info, I recommend going directly to pytest-grpc source code, as the plugin is pretty simple. With that, let’s move on to the actual test:

# test_grpc.py

def test_reply(grpc_stub):
    value = 'test-data'
    request = blueprint.echo_pb2.EchoRequest(message=value)
    response = grpc_stub.Reply(request)

    assert response.message == f'You said: {value}'

We create this test by leveraging grpc_stub fixture which we wrote in previous step. We create EchoRequest which is passed to the grpc_stub.Reply, followed by simple assert. And, when we run the test (make run):

plugins: cov-2.8.1, metadata-1.8.0, grpc-0.7.0, html-2.0.1
collected 1 item

tests/test_grpc.py::test_reply PASSED                                                                                [100%]

----------- coverage: platform linux, python 3.7.5-final-0 -----------
Name                                   Stmts   Miss Branch BrPart  Cover
------------------------------------------------------------------------
blueprint/__init__.py                      3      0      0      0   100%
blueprint/app.py                          11      5      0      0    55%
blueprint/echo_client.py                  10     10      0      0     0%
blueprint/generated/echo_pb2.py           18      0      0      0   100%
blueprint/generated/echo_pb2_grpc.py      18      4      0      0    78%
blueprint/grpc.py                          4      0      0      0   100%
------------------------------------------------------------------------
TOTAL                                     64     19      0      0    70%
Coverage XML written to file coverage.xml


============================================================ 1 passed in 0.09s ============================================

We passed! Aaaand we are done!

Conclusion

If you take away just one thing from this article, then I think it should be the fact that we should always consider possible alternatives when deciding on solution/technology we want to use for some project. It doesn’t always need to be REST and JSON. Sometimes gRPC might fit the requirements way better.

Source: Martin Heinz

Via: martinheinz.dev

Tags: Implementing gRPC server using Python