Client: server-streaming does not collect duplicated (fault+ctx) trailers during an RpcException

[mgravell]

Scenario: the server (native in this case) deliberately faults with an RpcException including some trailers in the middle of a server-streaming operation, and that trailer has also been added to the server context, i.e.

var faultTrailers = new Metadata();
faultTrailers.Add("sum", sum);
// **and** importantly
serverContext.ResponseTrailers.Add("sum", sum);
throw new RpcException(new Status(StatusCode.Internal, "oops"), faultTrailers);

Expectation

as observed with the native Google client:

The client-side call await streamReader.MoveNext(cancellationToken) will surface an RpcException with the correct StatusCode, Detail and Trailers i.e. one of "sum", sum

Actual

as observed with Grpc-Dotnet

The client-side call await streamReader.MoveNext(cancellationToken) surfaces an RpcException with the correct StatusCode, Detail - but:

• sometimes the Trailers collection is empty
• sometimes the Trailers collection has it twice, i.e. "sum", "10,10"

I haven't figured out the rules for when it is empty and when it is duplicated; when I run just the one test, it tends to be empty; when I run a bunch of tests it tends to be duplicated. Possibly a thread race condition.

Repro

See the ServerStreaming(true) test case here - comparing NativeStreamTests to ManagedStreamTests.

A good place to put a breakpoint is in Reshape.cs, ServerStreamingAsyncImpl, on the line:

metadata?.SetTrailers(fault);

(which is where the exception is surfaced)

Note that this is the exact same (native) server in both tests, so presumably this is a client bug.

Note: I have added IsManagedClient that changes the test's expectation to reflect what I am observing (so my tests acknowledge reality for now)

[JamesNK]

@mgravell 2.30.0 is on NuGet with this fix. You can update your package reference and remove any related hacks.