Thanks for your interest,
I work for a synchrotron and we use the distributed control system TANGO. The main implementation is in C++, but we use a python binding called PyTango. The current server implementation (on the C++ side) does not feature an event loop but instead create a different thread for each client.
TANGO: http://www.tango-controls.org/ PyTango: http://www.esrf.eu/computing/cs/tango/tango_doc/kernel_doc/pytango/latest/in...
I wanted to add asyncio support to the library, so that we can benefit from single-threaded asynchronous programming. The problem is that client callbacks run in different threads and there is not much we can do about it until a pure python implementation is developed (and it's a lot of work). Instead, it is possible to use an asyncio event loop, run the server through run_in_executor (juste like you mentioned in your mail), and redirect all the client callbacks to the event loop. That's the part where job submission from a different thread comes in handy.
A very similar solution has been developed using gevent, but I like explicit coroutines better :p
Another use case is the communication between two event loops. From what I've seen, the current context (get/set event loop) is only related to the current thread. It makes it easy to run different event loops in different threads. Even though I'm not sure what the use case is, I suppose it's been done intentionally. Then the executor interface is useful to run things like:
executor = LoopExecutor(other_loop) result = await my_loop.run_in_executor(executor, coro_func, *args)
There is working example in the test directory: https://github.com/vxgmichel/asyncio-loopexecutor/blob/master/test/test_mult...
The coroutine(fn) cast only makes sense if a subclass of Executor is used, in order to be consistent with the Executor.submit signature. Otherwise, passing an already-called coroutine is perfectly fine. I think it is a good idea to define a simple submit function like you recommended:
def submit_to_loop(loop, coro): future = concurrent.futures.Future() callback = partial(schedule, coro, destination=future) loop.call_soon_threadsafe(callback) return future
And then use the executor interface if we realize it is actually useful. It's really not a lot of code anyway:
def __init__(self, loop=None): self.loop = loop or asyncio.get_event_loop()
def submit(self, fn, *args, **kwargs): coro = asyncio.coroutine(fn)(*args, **kwargs) return submit_to_loop(self.loop, coro)
I'll update the repository.
2015-09-27 4:52 GMT+02:00 Guido van Rossum email@example.com:
I've read your write-up with interest. You're right that it's a bit awkward to make calls from the threaded world into the asyncio world. Interestingly, there's much better support for passing work off from the asyncio event loop to a thread (run_in_executor()). Perhaps that's because the use case there was obvious from the start: some things that may block for I/O just don't have an async interface yet, so in order to use them from an asyncio task they must be off-loaded to a separate thread or else the entire event loop is blocked. (This is used for calling getaddrinfo(), for example.)
I'm curious where you have encountered the opposite use case?
I think if I had to do this myself I would go for a more minimalist interface: something like your submit() method but without the call to asyncio.coroutine(fn). Having the caller pass in the already-called coroutine object might simplify the signature even further. I'm not sure I see the advantage of trying to make this an executor -- but perhaps I'm missing something?
On Sat, Sep 26, 2015 at 7:29 AM, Vincent Michel firstname.lastname@example.org wrote:
I noticed there is currently no standard solution to submit a job from a thread to an asyncio event loop.
Here's what the asyncio documentation says about concurrency and multithreading:
To schedule a callback from a different thread, the BaseEventLoop.call_soon_threadsafe() method should be used. Example to schedule a coroutine from a different thread: loop.call_soon_threadsafe(asyncio.async, coro_func())
The issue with this method is the loss of the coroutine result.
One way to deal with this issue is to connect the asyncio.Future returned by async (or ensure_future) to a concurrent.futures.Future. It is then possible to use a subclass of concurrent.futures.Executor to submit a callback to an asyncio event loop. Such an executor can also be used to set up communication between two event loops using run_in_executor.
I posted an implementation called LoopExecutor on GitHub: https://github.com/vxgmichel/asyncio-loopexecutor The repo contains the loopexecutor module along with tests for several use cases. The README describes the whole thing (context, examples, issues, implementation).
It is interesting to note that this executor is a bit different than ThreadPoolExecutor and ProcessPoolExecutor since it can also submit a coroutine function. Example:
with LoopExecutor(loop) as executor: future = executor.submit(operator.add, 1, 2) assert future.result() == 3 future = executor.submit(asyncio.sleep, 0.1, result=3) assert future.result() == 3
This works in both cases because submit always cast the given function to a coroutine. That means it would also work with a function that returns a Future.
Here's a few topic related to the current implementation that might be interesting to discuss:
- possible drawback of casting the callback to a coroutine
- possible drawback of concurrent.future.Future using asyncio.Future._copy_state
- does LoopExecutor need to implement the shutdown method?
- removing the limitation in run_in_executor (can't submit a coroutine function)
- adding a generic Future connection function in asyncio
- reimplementing wrap_future with the generic connection
- adding LoopExecutor to asyncio (or concurrent.futures)
At the moment, the interaction between asyncio and concurrent.futures only goes one way. It would be nice to have a standard solution (LoopExecutor or something else) to make it bidirectional.
Python-ideas mailing list Pythonemail@example.com https://mail.python.org/mailman/listinfo/python-ideas Code of Conduct: http://python.org/psf/codeofconduct/
-- --Guido van Rossum (python.org/~guido)