Client-side streaming early OK fails with `StatusCode.CANCELLED` from python client

[ebraraktas]

Bug Report

Version

└── tonic v0.7.1
└── tonic-build v0.7.1

Platform

Darwin Mustafas-MacBook-Pro.local 20.5.0 Darwin Kernel Version 20.5.0: Sat May 8 05:10:33 PDT 2021; root:xnu-7195.121.3~9/RELEASE_X86_64 x86_64

Description

I implemented simple client-side streaming endpoint using tonic-rs. It may return early if some condition occurs:

async fn record_route(
    &self,
    request: Request<Streaming<Point>>,
) -> Result<Response<RouteSummary>, Status> {
    let mut stream = request.into_inner();

    let mut summary = RouteSummary::default();

    while let Some(point) = stream.next().await {
        let point = point?;
        summary.point_count += 1;

        if point.latitude == 0 {
            println!("Return early: {}", summary.point_count);
            return Ok(Response::new(summary));
        }
    }
    println!("Return normal: {}", summary.point_count);
    Ok(Response::new(summary))
}

However, when I call this server from a simple python client it fails with the exception (StatusCode.CANCELLED) below:

# python client.py 5000
port = 5000
early_return_index = 2
Sleep 500 ms
Sleep 500 ms
Sleep 500 ms
Traceback (most recent call last):
  File "client.py", line 46, in <module>
    main()
  File "client.py", line 42, in main
    call_stream(client)
  File "client.py", line 34, in call_stream
    print(client.RecordRoute(iter_points()))
  File "/usr/local/lib/python3.8/site-packages/grpc/_channel.py", line 1131, in __call__
    return _end_unary_response_blocking(state, call, False, None)
  File "/usr/local/lib/python3.8/site-packages/grpc/_channel.py", line 849, in _end_unary_response_blocking
    raise _InactiveRpcError(state)
grpc._channel._InactiveRpcError: <_InactiveRpcError of RPC that terminated with:
        status = StatusCode.CANCELLED
        details = "Received RST_STREAM with error code 8"
        debug_error_string = "{"created":"@1651470233.217719000","description":"Error received from peer ipv6:[::1]:5000","file":"src/core/lib/surface/call.cc","file_line":1075,"grpc_message":"Received RST_STREAM with error code 8","grpc_status":1}"
>

GitHub repo to reproduce issue: https://github.com/ebraraktas/tonic-client-stream-early-ok

As the Issue section of README of the linked repo states, it runs ok with rust client. I know this may be related with grpc Python, but same server implemented in Python works smoothly if it is called from same Python client.

[LucioFranco]

Yeah interesting, I believe the cancelled comes from the python code since the h2 connection is getting RST. Curious if you could reproduce this with a rust client?

[ebraraktas]

Hi @LucioFranco , as I stated in my first message I can call rust server from rust client without an issue. I am not familiar with the internals of the tonic, but can you explain why you think that the cancelled comes from python?

In addition, I implemented same server in go and I could manage to call go server from the same python client without an issue. Interestingly, go client can call rust server with no issue as well. Here is a matrix describing the issue:

client\server	rust	go	python
rust	✅	⛔*	✅
go	✅	✅	✅
python	⛔	✅	✅

*: My rust client gives error below when I try to connect to the go server:

Error: tonic::transport::Error(Transport, hyper::Error(Connect, ConnectError("tcp connect error", Os { code: 61, kind: ConnectionRefused, message: "Connection refused" })))

[LucioFranco]

So the cancelled is likely returned from the client when it encounters a GoAway/connection drop/request ended before getting the full trailers etc. The error is not passed from server -> client. As for why go works and python does not....that one I don't know I would maybe open an issue in the python implementation?

As for your last error there that is just a regular tcp error coming from the tcp connection in the underlying OS, so maybe you're not connecting on the right port etc.

[ebraraktas]

@LucioFranco I opened an issue in the official gRPC repository and they responded that they think the issue arises from gRPC implementation of the tonic-rs. The linked message has the details that I am not very familiar with, but I hope they may be helpful.

[ebraraktas]

@LucioFranco are there any update on this issue?

[LucioFranco]

Hi @ebraraktas I've started to look but have had a few other high priority items come up. I am about to head out on vacation so not sure how far I will get on this until I get back. This sounds like it will need some low level investigation to understand where the gap is.

[LucioFranco]

Small update: I dug pretty deep into the issue on Friday afternoon (weekend hit before I had time to reply here). From what I can tell the behavior from the Rust side is interesting in that we send a RST_STREAM.Cancel even though we reply with a regular Ok status. My theory has to do with the fact that we drop the RecvStream before the sender side has half-closed it. Will keep looking into this and return with a better answer.

[LucioFranco]

Once we get the next release out of h2 hyperium/h2#633 this will fix this issue I believe.

[erebe]

I tried the repro repo with the patch, and indeed it should fix the issue.

/tmp/tonic-client-stream-early-ok/proto main 16s
venv ❯ python client.py
port = 5000
early_return_index = 3
Sleep 500 ms
Sleep 500 ms
Sleep 500 ms
Sleep 500 ms
point_count: 4

/tmp/tonic-client-stream-early-ok main 1m 25s
❯ cargo run --bin routeguide-server
    Finished dev [unoptimized + debuginfo] target(s) in 0.05s
     Running `target/debug/routeguide-server`
Return early: 4

At least no traceback

[ebraraktas]

I confirm that it is fixed. Updated & archived the repro repo. Thanks for your efforts @erebe and @LucioFranco .