grpc_server_shutdown_and_notify reads "Send a GRPC_OP_COMPLETE event when there are no more calls being serviced."

Does this test show a violation of that guarantee in its flaky failures (it flakes around 2.5% for me)? Or is there something wrong in the test, or in our Cython code?

The gRPC Python beta server keeps track of active RPCs and which events are due; I'm trying to shed that bookkeeping and rely on semantics of "the server thread should just keep polling the server completion queue until it sees the shutdown tag, at which point it should break out of its loop and complete". That's supposed to work, right?