The gRPC server bidirectional streams do not receive exceptions when the client loses connectivity

[doctorpangloss]

Version

4.3.7

Context

Using request.exceptionHandler(...), bidirectional streams can respond to ungraceful stream closure, which could be killing an app, a browser crashing, a phone running out of battery, network disconnection or other failure.

This package does not correctly receive exceptions the way Netty does. It lacks the configuration that the Netty server has for keep alives. This appears to be the root issue.

Steps to reproduce

1. Generate a bidi streaming method stub using the original Netty-based Vertx package's jprotoc based code generator. You will have a method of the signature:

void bidiCall(ReadStream<X> request, WriteStream<Y> response);

2. Bind it to the Netty-based VertxServer. Configure it with timeouts:

var builder = VertxServerBuilder.forPort(vertx, port);
var nettyServerBuilder = builder.nettyBuilder();
nettyServerBuilder
.maxConnectionIdle(Long.MAX_VALUE, TimeUnit.NANOSECONDS)
.maxConnectionAge(29, TimeUnit.DAYS)
.keepAliveTime(400, TimeUnit.MILLISECONDS)
.keepAliveTimeout(8000, TimeUnit.MILLISECONDS)
.permitKeepAliveTime(100, TimeUnit.MILLISECONDS)
.permitKeepAliveWithoutCalls(true);
builder.build();

3. Proxy the server with ToxiProxy.
4. Connect to the bidi RPC endpoint.
5. Observe you are connected.
6. Disconnect the network using ToxiProxy.
7. After a short while, observe a StatusRuntimeException with status Status.CANCELLED is delivered to request.exceptionHandler.
8. Bind the same service to a GrpcServer.
9. Proxy it with ToxiProxy.
10. Connect to the bidi RPC endpoint.
11. Observe you are connected.
12. Disconnect the network using ToxiProxy.
13. Observe no exception is received nor the endHandler called in a short while.

Extra

This is reproduced in the tests of my own open source program. I can share that one. Otherwise I will need help authoring tests in this package, since Maven in my IntelliJ does not appear to generate the protos / reference them correctly and it's impossible to code. Additionally I suggest adding TestContainers as a dependency for using ToxiProxy.

[doctorpangloss]

I ended up using the KeepAliveManager from io.grpc. You will need to deal with adapting ScheduledFuture.