Fixes Action.*_async futures never complete

[jmblixt3]

Replaces #1125 since original contributor went inactive

Closes #1123

If two separate client server actions are running in separate executors the future given to the ActionClient will never complete due to a race condition on the rcl handles

Tested using this from @apockill which was adapted to match rolling then test with and without locks

To reproduce initial issue remove the lock/sleep in the RaceyAction client

Adapted example Code here

client.py

import rclpy
from rclpy import Future
from rclpy.action import ActionClient
from rclpy.action.client import ClientGoalHandle
from rclpy.callback_groups import MutuallyExclusiveCallbackGroup
from rclpy.executors import MultiThreadedExecutor
from rclpy.node import Node

from test_msgs.action import Fibonacci as SomeMsg
from time import sleep

class RaceyActionClient(ActionClient):
    def _get_result_async(self, goal_handle):
        """
        Request the result for an active goal asynchronously.

        :param goal_handle: Handle to the goal to cancel.
        :type goal_handle: :class:`ClientGoalHandle`
        :return: a Future instance that completes when the get result request has been processed.
        :rtype: :class:`rclpy.task.Future` instance
        """
        if not isinstance(goal_handle, ClientGoalHandle):
            raise TypeError(
                'Expected type ClientGoalHandle but received {}'.format(type(goal_handle)))

        result_request = self._action_type.Impl.GetResultService.Request()
        result_request.goal_id = goal_handle.goal_id
        future = Future()
        with self._lock:
            sleep(1)
            sequence_number = self._client_handle.send_result_request(result_request)
            if sequence_number in self._pending_result_requests:
                raise RuntimeError(
                    'Sequence ({}) conflicts with pending result request'.format(sequence_number))
            sleep(1)
            self._pending_result_requests[sequence_number] = future
            sleep(1)
            self._result_sequence_number_to_goal_id[sequence_number] = result_request.goal_id
            sleep(1)
            future.add_done_callback(self._remove_pending_result_request)
            sleep(1)
            # Add future so executor is aware
            self.add_future(future)

        return future

class SomeActionClient(Node):
    def __init__(self):
        super().__init__("action_client")
        self._action_client = RaceyActionClient(
            self,
            SomeMsg,
            "some_action",
            callback_group=MutuallyExclusiveCallbackGroup(),
        )
        self.call_action_timer = self.create_timer(
            timer_period_sec=10,
            callback=self.send_goal,
            callback_group=MutuallyExclusiveCallbackGroup(),
        )

    def send_goal(self):
        goal_msg = SomeMsg.Goal()
        # self._action_client.wait_for_server()
        result = self._action_client.send_goal(goal_msg)
        print(result)


def main(args=None):
    rclpy.init(args=args)

    action_client = SomeActionClient()

    executor = MultiThreadedExecutor()
    rclpy.spin(action_client, executor)

if __name__ == "__main__":
    main()

server.py

import rclpy
from rclpy.action import ActionServer
from rclpy.node import Node

from test_msgs.action import Fibonacci as SomeMsg


class SomeActionServer(Node):
    def __init__(self):
        super().__init__("fibonacci_action_server")
        self._action_server = ActionServer(
            self, SomeMsg, "some_action", self.execute_callback
        )

    def execute_callback(self, goal_handle):
        self.get_logger().info("Executing goal...")
        result = SomeMsg.Result()
        goal_handle.succeed()
        return result


def main(args=None):
    rclpy.init(args=args)

    action_server = SomeActionServer()
    rclpy.spin(action_server)


if __name__ == "__main__":
    main()

Before and after log results

Before
client output

[WARN] [1720739707.125040450] [action_client.action_client]: Ignoring unexpected result response. There may be more than one action server for the action 'some_action'
^C

server output

[INFO] [1720739706.054013974] [fibonacci_action_server]: Executing goal...
^C

After
client output

[INFO] [1720739524.873096289] [fibonacci_action_server]: Executing goal...
[INFO] [1720739534.791491183] [fibonacci_action_server]: Executing goal...
[INFO] [1720739544.789914632] [fibonacci_action_server]: Executing goal...
[INFO] [1720739554.791130534] [fibonacci_action_server]: Executing goal...
[INFO] [1720739564.791638162] [fibonacci_action_server]: Executing goal...
[INFO] [1720739574.793290290] [fibonacci_action_server]: Executing goal...
[INFO] [1720739584.790164903] [fibonacci_action_server]: Executing goal...
[INFO] [1720739594.794427458] [fibonacci_action_server]: Executing goal...
^C

test_msgs.action.Fibonacci_GetResult_Response(status=4, result=test_msgs.action.Fibonacci_Result(sequence=[]))
test_msgs.action.Fibonacci_GetResult_Response(status=4, result=test_msgs.action.Fibonacci_Result(sequence=[]))
test_msgs.action.Fibonacci_GetResult_Response(status=4, result=test_msgs.action.Fibonacci_Result(sequence=[]))
test_msgs.action.Fibonacci_GetResult_Response(status=4, result=test_msgs.action.Fibonacci_Result(sequence=[]))
test_msgs.action.Fibonacci_GetResult_Response(status=4, result=test_msgs.action.Fibonacci_Result(sequence=[]))
test_msgs.action.Fibonacci_GetResult_Response(status=4, result=test_msgs.action.Fibonacci_Result(sequence=[]))
test_msgs.action.Fibonacci_GetResult_Response(status=4, result=test_msgs.action.Fibonacci_Result(sequence=[]))
test_msgs.action.Fibonacci_GetResult_Response(status=4, result=test_msgs.action.Fibonacci_Result(sequence=[]))
^C

I also initially tried to add locks to just the action client, but I was getting test failures in test_action_graph on the get_names_and_types function, so I added for action server as well.

[fujitatomoya]

i would add @aditya2592 as co-author, since this borrows the code from #1125.

[jmblixt3]

i would add @aditya2592 as co-author, since this borrows the code from #1125.

Done

[jmblixt3]

Throwing this back to draft because it still doesn't fix all conditions where this breaks. If others have ideas on how to more reliably reproduce this that would be appreciated.

[apockill]

Throwing this back to draft because it still doesn't fix all conditions where this breaks. If others have ideas on how to more reliably reproduce this that would be appreciated.

What kinds of failures are you seeing still? This is the "fix" I've been using in production, and at least with our use cases of services we haven't seen failures since:

class PatchRclpyIssue1123(RobustActionClient):

    _lock: RLock = None  # type: ignore

    @property
    def _cpp_client_handle_lock(self) -> RLock:
        if self._lock is None:
            self._lock = RLock()
        return self._lock

    async def execute(self, *args: Any, **kwargs: Any) -> None:
        # This is ugly- holding on to a lock in an async environment feels gross
        with self._cpp_client_handle_lock:
            return await super().execute(*args, **kwargs)  # type: ignore

    def send_goal_async(self, *args: Any, **kwargs: Any) -> Future:
        with self._cpp_client_handle_lock:
            return super().send_goal_async(*args, **kwargs)

    def _cancel_goal_async(self, *args: Any, **kwargs: Any) -> Future:
        with self._cpp_client_handle_lock:
            return super()._cancel_goal_async(*args, **kwargs)

    def _get_result_async(self, *args: Any, **kwargs: Any) -> Future:
        with self._cpp_client_handle_lock:
            return super()._get_result_async(*args, **kwargs)

I'm not sure if these matter, but here are the differences I can note between this above fix and this PR:

1. In my implementation I lock the async execute, since multithreaded executors are still running async tasks "concurrently", leading to possible race conditions with the client handle.
2. I use an RLock vs a Lock. Not sure if this matters in this case, I figured I'd note it

[jmblixt3]

I was defining a python action client server pair in two seperate terminals, and the custom server also contained a python client to another cpp nav2 action. After calling the custom python action server once it would work correctly the first time, but after wasn't even accepting the goal request to any subsequent calls. If I removed the nav2 action client from my custom action server it wouldn't deadlock. I suspect a lock wasn't correctly being released with the call to another action client inside the action server. I'd like to look more into try to see if the issue persists with either locking on calls to entire execute funciton as you've suggested or using Rlocks to see if the behavior persists.

[apockill]

After calling the custom python action server once it would work correctly the first time, but after wasn't even accepting the goal request to any subsequent calls.

Ahh understood. I would hazard a guess that an RLock would work in this situation.

[jmblixt3]

I tried to recreate an example outside of my environment, but have been unsuccessful

[jmblixt3]

I tried to recreate an example outside of my environment, but have been unsuccessful

I was able to recreate this issue outside my environment now #1313, and looks like there is an issue with rclcpp server, and a nested rclpy action client nested inside an rcply action server, so that seems unrelated, so I'll pull this back out of draft

[fujitatomoya]

lgtm with green CI

[fujitatomoya]

@apockill if you are willing to do 2nd review on this, that would be really appreciated.

[firesurfer]

Who ever wants to test the patch on iron. I cherry picked it on iron in this fork:
https://github.com/firesurfer/rclpy/tree/iron

Btw. from our tests today it seems to work fine on iron with our setup.

[apockill]

This looks like it would fix the issue. My only commentary is that this feels brittle, and future changes might break it easily.

Even now I have a hard time knowing which dicts need strict thread safety. For example, _remove_pending_request, _remove_pending_goal_request, _remove_pending_result_request all use dictionaries that are holding state that can be edited in other callbacks.

I wonder if a future refactor could make this harder to mess up. Off the cuff, we could wrap up all the state that is ties the C++ self._client and it's python stateful dictionaries into a single object, and protect that object. I'm just brainstorming here, but something a la:

class _ProtectedClientState:
    """Protect state between the C++ client and python that must be thread-safe"""
    
    def __init__(self):
        self.lock = RLock()
        
        # key: UUID in bytes, value: weak reference to ClientGoalHandle
        self.goal_handles = {}
        # key: goal request sequence_number, value: Future for goal response
        self.pending_goal_requests = {}
        # key: goal request sequence_number, value: UUID
        self.goal_sequence_number_to_goal_id = {}
        # key: cancel request sequence number, value: Future for cancel response
        self.pending_cancel_requests = {}
        # key: result request sequence number, value: Future for result response
        self.pending_result_requests = {}
        # key: result request sequence_number, value: UUID
        self.result_sequence_number_to_goal_id = {}
        # key: UUID in bytes, value: callback function
        self.feedback_callbacks = {}
        
    def __getattribute__(...):
       if not self.lock.acquired():
           raise HeyThatsBad("This state is meant to be protected while using these!"):
       return the requested attribute
        

[jmblixt3]

Just bumping this, to see if other maintainers can review/merge this

[fujitatomoya]

@jmblixt3 can you rebase this on rolling? and i will start the CI.

[jmblixt3]

@fujitatomoya Rebased

[fujitatomoya]

CI:

• Linux
• Linux-aarch64
• Linux-rhel
• Windows

[jmblixt3]

Anything else needed from me?

[fujitatomoya]

@sloretz @clalancette @adityapande-1995 can either of you take a look at this?

[GinesLopezz]

I'm experiencing the same issue on ROS 2 Humble. Could you backport this fix to humble? Thank you!

[mqcmd196]

I'm experiencing the same issue on ROS 2 Humble. Could you backport this fix to humble? Thank you!

+1

[leander2189]

I am having also this issue with Humble. Is there a timeframe on when the fix will be merged? Thanks!

[jmblixt3]

Just rebased this again, are there any issues with that still need to be resolved

[fujitatomoya]

@jmblixt3 no, we are just waiting for the 2nd review.

@clalancette @ahcorde @sloretz friendly ping.

[apockill]

Ship it? 👀🥺