grpc/examples/python/cancellation

Cancelling RPCs

RPCs may be cancelled by both the client and the server.

The Example

In the example, we implement a silly algorithm. We search for bytestrings whose hashes are similar to a given search string. For example, say we're looking for the string "doctor". Our algorithm may JrqhZVkTDoctYrUlXDbL6pfYQHU= or RC9/7mlM3ldy4TdoctOc6WzYbO4=. This is a brute force algorithm, so the server performing the search must be conscious the resources it allows to each client and each client must be conscientious of the resources demanded of the server.

In particular, we ensure that client processes cancel the stream explicitly before terminating and we ensure the server cancels RPCs that have gone on longer than a certain number of iterations.

Cancellation on the Client Side

A client may cancel an RPC for several reasons. Perhaps the data it requested has been made irrelevant. Perhaps you, as the client, want to be a good citizen of the server and are conserving compute resources.

Cancelling a Server-Side Unary RPC from the Client

The default RPC methods on a stub will simply return the result of an RPC.

>>> stub = hash_name_pb2_grpc.HashFinderStub(channel)
>>> stub.Find(hash_name_pb2.HashNameRequest(desired_name=name))
<hash_name_pb2.HashNameResponse object at 0x7fe2eb8ce2d0>

But you may use the future() method to receive an instance of grpc.Future. This interface allows you to wait on a response with a timeout, add a callback to be executed when the RPC completes, or to cancel the RPC before it has completed.

In the example, we use this interface to cancel our in-progress RPC when the user interrupts the process with ctrl-c.

stub = hash_name_pb2_grpc.HashFinderStub(channel)
future = stub.Find.future(hash_name_pb2.HashNameRequest(desired_name=name))
def cancel_request(unused_signum, unused_frame):
    future.cancel()
signal.signal(signal.SIGINT, cancel_request)

It's also important that you not block indefinitely on the RPC. Otherwise, the signal handler will never have a chance to run.

while True:
    try:
        result = future.result(timeout=_TIMEOUT_SECONDS)
    except grpc.FutureTimeoutError:
        continue
    except grpc.FutureCancelledError:
        break
    print("Got response: \n{}".format(result))
    break

Here, we repeatedly block on a result for up to _TIMEOUT_SECONDS. Doing so gives the signal handlers a chance to run. In the case that our timeout was reached, we simply continue on in the loop. In the case that the RPC was cancelled (by our user's ctrl+c), we break out of the loop cleanly. Finally, if we received the result of the RPC, we print it out for the user and exit the loop.

Cancelling a Server-Side Streaming RPC from the Client

Cancelling a Server-side streaming RPC is even simpler from the perspective of the gRPC API. The default stub method is already an instance of grpc.Future, so the methods outlined above still apply. It is also a generator, so we may iterate over it to yield the results of our RPC.

stub = hash_name_pb2_grpc.HashFinderStub(channel)
result_generator = stub.FindRange(hash_name_pb2.HashNameRequest(desired_name=name))
def cancel_request(unused_signum, unused_frame):
    result_generator.cancel()
signal.signal(signal.SIGINT, cancel_request)

However, the streaming case is complicated by the fact that there is no way to propagate a timeout to Python generators. As a result, simply iterating over the results of the RPC can block indefinitely and the signal handler may never run. Instead, we iterate over the generator on another thread and retrieve the results on the main thread with a synchronized Queue.

result_queue = Queue()
def iterate_responses(result_generator, result_queue):
    try:
        for result in result_generator:
            result_queue.put(result)
    except grpc.RpcError as rpc_error:
        if rpc_error.code() != grpc.StatusCode.CANCELLED:
            result_queue.put(None)
            raise rpc_error
    result_queue.put(None)
    print("RPC complete")
response_thread = threading.Thread(target=iterate_responses, args=(result_generator, result_queue))
response_thread.daemon = True
response_thread.start()

While this thread iterating over the results may block indefinitely, we can structure the code running on our main thread in such a way that signal handlers are guaranteed to be run at least every _TIMEOUT_SECONDS seconds.

while result_generator.running():
    try:
        result = result_queue.get(timeout=_TIMEOUT_SECONDS)
    except QueueEmpty:
        continue
    if result is None:
        break
    print("Got result: {}".format(result))

Similarly to the unary example above, we continue in a loop waiting for results, taking care to block for intervals of _TIMEOUT_SECONDS at the longest. Finally, we use None as a sentinel value to signal the end of the stream.

Using this scheme, our process responds nicely to SIGINTs while also explicitly cancelling its RPCs.

Cancellation on the Server Side

A server is reponsible for cancellation in two ways. It must respond in some way when a client initiates a cancellation, otherwise long-running computations could continue indefinitely.

It may also decide to cancel the RPC for its own reasons. In our example, the server can be configured to cancel an RPC after a certain number of hashes has been computed in order to conserve compute resources.

Responding to Cancellations from a Servicer Thread

It's important to remember that a gRPC Python server is backed by a thread pool with a fixed size. When an RPC is cancelled, the library does not terminate your servicer thread. It is your responsibility as the application author to ensure that your servicer thread terminates soon after the RPC has been cancelled.

In this example, we use the ServicerContext.add_callback method to set a threading.Event object when the RPC is terminated. We pass this Event object down through our hashing algorithm and ensure to check that the RPC is still ongoing before each iteration.

stop_event = threading.Event()
def on_rpc_done():
    # Regain servicer thread.
    stop_event.set()
context.add_callback(on_rpc_done)
secret = _find_secret(stop_event)

Initiating a Cancellation on the Server Side

Initiating a cancellation from the server side is simpler. Just call ServicerContext.cancel().

In our example, we ensure that no single client is monopolizing the server by cancelling after a configurable number of hashes have been checked.

try:
    for candidate in secret_generator:
        yield candidate
except ResourceLimitExceededError:
    print("Cancelling RPC due to exhausted resources.")
    context.cancel()

In this type of situation, you may also consider returning a more specific error using the grpcio-status package.