A command pattern that I’ve used in multiple projects is an event-loop sidecar. The sidecar runs on it’s own thread and
does asynchronous IO; thereby increasing the responsiveness of the system which could be doing IO/CPU bound work. The
application (running on the main thread) interfaces with the sidecar through a synchronous API.
There are multiple caveats to this pattern. With python’s GIL, the sidecar should strictly be doing I/O to ensure that the pattern has the desired effects. If you have a CPU bound sidecar, you’ll end up with an architecture which performs worse than a single threaded job.
Setting up such a sidecar takes some time to get used to, but is easier in python 3.7+:
def start(self): if self.sidecar_thread is None: self.sidecar_thread = threading.Thread( target=self._create_sidecar, args=(self.loop,) ) self.sidecar_thread.start() def _create_sidecar(self, loop: asyncio.BaseEventLoop): """ Initialize the eventloop and setup a listener """ # We use an event loop initialized in the main thread and use it # over here. This allows us to submit requests on a synchronous # fashion using create_task on the main thread and processing # happens in the sidecar asyncio.set_event_loop(loop) # Register a task for socket manager setup try: self.loop.run_until_complete(self._listener()) except concurrent.futures.CancelledError as e: logging.error("Loop got shutdown, due to %s", e) async def _listener(self): """ Setup the connection and process messages """ try: self._ws = await websockets.connect(self._uri) async for msg in self._ws: if self._has_error.is_set(): raise errors.WebSocketError() # self._closing is called by exit which is described later if self._closing: break # Process messages here: except self.errors as b: # _has_error is ane Event to indicate that the socket has an error self._has_error.set() logging.warning("Error %s in WS, will raise WebSocketError", b) finally: # Close the connection and signal that we are done if self._ws: await self._ws.close() self._ws_has_closed.set()
The code takes getting used to. Hopefully the comments help in exposition. To be honest, in an earlier version,
we just let the application crash on socket errors without cleaning up. This led to gnarly memory issues, especially if
you have a long running job which reintializes this pattern. This
exit protocol to close the resource properly took
multiple iterations, but now has been perfected:
def exit(self): """ exit is called on the main thread and not on the sidecar """ # Indicates the listener to stop processing any new messages self._closing = True # Busy loop and waits for the socket to close # NOTE: assumption is that listener gets triggered frequently start = time.time() while True: if self._ws_has_closed.is_set(): break # If more than one minute is taken, just break the connection # NOTE: only minute is the max latency the application can tolerate, # before it gets repead by the scheduler in our case. time_elapsed = time.time() - start logging.warning( "Busy looping to close websocket, time elapsed %s", time_elapsed ) if time_elapsed > consts.ONE_MINUTE: logging.error("Existing web socket without clearing connection") break # Busy loop time.sleep(consts.TEN_SECONDS) # Stop the loop now self.loop.stop() # Clean up all pending tasks for task in asyncio.all_tasks(self.loop): task.cancel() # Join with the sidecar if self.sidecar_thread is not None: self.sidecar_thread.join(constants.THREAD_JOIN_TIMEOUT) if self.sidecar_thread.isAlive(): logging.error("Unable to clear up sidecar")
Every article on
asyncio will advice you to close the thread and clean up pending tasks. But not article will tell you
that closing the thread will almost never work and end up with hard to debug errors. It took me a lot of head wrangling
to figure out the right protocol to correctly close a sidecar thread which uses an event loop:
- Stop processing new messages on the sidecar.
- Close the websocket connection, you will have to busy loop in order to do so.
- Close the loop and clear pending tasks.
- Join with the sidecar.
Missing any one of these steps is a path to an endless pit of wats. Especially the first two steps, which are taken granted in languages which support such idioms out of the box like Java.