如何在python中配置gRPC HTTP/2流控制

Question

I have a gRPC server with the following proto:

syntax = "proto3";

service MyServicer {
  rpc DoSomething(stream InputBigData) returns (stream OutputBigData) {}
}
message InputBigData {
    bytes data = 1;
}
message OutputBigData {
    bytes data = 1;
}

And my server is created with the following Python code:

server = grpc.server(futures.ThreadPoolExecutor(max_workers=10),
                     options=[('grpc.max_receive_message_length', -1),
                              ('grpc.max_send_message_length', -1))])

max_receive_message_length and max_send_message_length are set to -1 to allow the transfer of big messages (typically 8Mb). The client also define the same options.

Case 1: Consider the client sends to the server InputBigData at a higher rate than the server can afford. How can I configure how many InputBigData (or bytes) can be queued in the input stream?

Case 2: Consider the client reads the response OutputBigData from the server at a lower rate than the client can afford. How can I configure how many OutputBigData (or bytes) can be queued in the output stream?

I know gRPC flow control is based on HTTP/2: https://httpwg.org/specs/rfc7540.html#FlowControl I tried to set grpc.http2.write_buffer_size at 67108864 (seems to be the max value) but nothing happened.

Here is an implementation which highlights the case 2:

# server.py
from concurrent import futures

import grpc
import myservicer_pb2_grpc, myservicer_pb2


class MyServicer(myservicer_pb2_grpc.MyServicer):

    def DoSomething(self, request_iterator, target, **kwargs):
        big_data = b'0' * 1920*1080*4
        for r in request_iterator:
            print("server received input big data")
            yield myservicer_pb2.OutputBigData(data=big_data)
            print("server sent output big data")


if __name__ == '__main__':
    server = grpc.server(futures.ThreadPoolExecutor(max_workers=10),
                         options=[('grpc.max_receive_message_length', -1),
                                  ('grpc.max_send_message_length', -1)])
    myservicer_pb2_grpc.add_MyServicerServicer_to_server(
        MyServicer(), server)
    server.add_insecure_port("[::]:50051")
    server.start()
    server.wait_for_termination()

# client.py
import time
import grpc

import myservicer_pb2_grpc
import myservicer_pb2


def big_data_generator():
    big_data = b'0' * 1920*1080*4
    for i in range(100):
        yield myservicer_pb2.InputBigData(data=big_data)


def run():
    with grpc.insecure_channel('localhost:50051',
                               options=[('grpc.max_send_message_length', -1),
                                        ('grpc.max_receive_message_length', -1)]) as channel:
        stub = myservicer_pb2_grpc.MyServicerStub(channel)
        res = stub.DoSomething(big_data_generator())

        for r in res:
            print("Client received data")
            time.sleep(10)

if __name__ == '__main__':
    run()

After 10 seconds my server output is:

server received input big data
server sent output big data
server received input big data
server sent output big data
server received input big data

And my client output is:

Client received data

My server received 3 InputBigData and sent 2 OutputBigData. It is now blocked until the client consumes the output data. In this scenario I want to increase (2 or 3 times) the output buffer size so it can continue to process more input data even if the client is late in consuming the result.

Answer 1

Thanks for the detailed question. I tried your example, but still can't tune gRPC to increase its window size freely.

gRPC Channel arguments can be found here . The flow control implementation is here There are only several might affect flow-control, which are:

• grpc.http2.bdp_probe=0 : disables automatic window increase
• grpc.http2.max_frame_size : HTTP/2 max frame size
• grpc.http2.write_buffer_size : Not really a flow-control option, it is used for GRPC_WRITE_BUFFER_HINT (write without blocking). Also, GRPC_WRITE_BUFFER_HINT is not supported yet in gRPC Python

There is no argument that could trigger a window size update. The default window size is 64KB. gRPC will increase the window size via BDP estimation. Eg, on my laptop, the client-outbound window size increased to 8380679 (~8MB). But I yet to find a way to manually intervene this process.

So, unfortunately, you might need application-level buffering. You could use coroutines in asyncio or threading with a thread-safe queue on both the client-side and server-side.