Problems integrating trio_asyncio & distributed/tornado

[dhirschfeld]

Moved from python-trio/trio#552 (comment)

import asyncio
from functools import partial
import trio
import trio_asyncio
from dask.distributed import Client


async def f():
    async with trio_asyncio.open_loop() as loop:
        client = await loop.run_asyncio(partial(Client, asynchronous=True))
        future = client.submit(lambda x: x + 1, 10)
        return await loop.run_future(future)

res = trio.run(f)

results in the error:

C:\Miniconda3\lib\site-packages\trio_asyncio\loop.py:33: TrioDeprecationWarning: trio.TaskLocal.__init__ is deprecated since Trio 0.4.0; use contextvars.ContextVar instead (https://github.com/python-trio/trio/issues/420)
  _current_loop = trio.TaskLocal(loop=None, policy=None)

Exception in default exception handler
Traceback (most recent call last):
  File "C:\Miniconda3\lib\asyncio\base_events.py", line 1291, in call_exception_handler
    self.default_exception_handler(context)
  File "C:\Miniconda3\lib\site-packages\trio_asyncio\async_.py", line 50, in default_exception_handler
    raise exception
  File "C:\Miniconda3\lib\site-packages\trio_asyncio\base.py", line 529, in _reader_loop
    await _wait_readable(fd)
  File "C:\Miniconda3\lib\site-packages\trio\_core\_io_windows.py", line 378, in wait_socket_readable
    await self._wait_socket("read", sock)
  File "C:\Miniconda3\lib\site-packages\trio\_core\_io_windows.py", line 361, in _wait_socket
    raise TypeError("need a stdlib socket")
TypeError: need a stdlib socket
<snip>

[dhirschfeld]

I'm on win64 running:

python        3.6.5
trio          0.4.0
trio_asyncio  0.7.0

[njsmith]

That check was removed in: python-trio/trio#610

There might be other problems, but I suspect the immediate roadblock here is gone...

[parity3]

So @dhirschfeld were there other problems? (too lazy to try this myself)

[dhirschfeld]

Sorry, haven't circled back to this just yet. It's on my TODO list but a pretty long way down at the moment.

[dhirschfeld]

...so, I got curious if this would Just Work now, in particular after #66. Unfortunately I'm running into another error. I first needed to patch the function to pass in the loop:

async def f():
    async with trio_asyncio.open_loop() as loop:
        client = await loop.run_asyncio(partial(Client, loop=loop, asynchronous=True))
        future = client.submit(lambda x: x + 1, 10)
        return await loop.run_future(future)

...but that then resulted in the below error:

In [3]: trio.run(f)
Traceback (most recent call last):

  File "<ipython-input-3-03c750b89f65>", line 1, in <module>
    trio.run(f)

  File "C:\Users\dhirschf\envs\dev\lib\site-packages\trio\_core\_run.py", line 1804, in run
    raise runner.main_task_outcome.error

  File "<ipython-input-2-27c49da4404c>", line 3, in f
    client = await loop.run_asyncio(partial(Client, loop=loop, asynchronous=True))

  File "c:\users\dhirschf\code\github\python-trio\trio-asyncio\trio_asyncio\adapter.py", line 82, in __await__
    f = f(*self.args)

  File "C:\Users\dhirschf\envs\dev\lib\site-packages\distributed\client.py", line 726, in __init__
    self.start(timeout=timeout)

  File "C:\Users\dhirschf\envs\dev\lib\site-packages\distributed\client.py", line 891, in start
    sync(self.loop, self._start, **kwargs)

  File "C:\Users\dhirschf\envs\dev\lib\site-packages\distributed\utils.py", line 336, in sync
    loop.add_callback(f)

AttributeError: 'TrioEventLoop' object has no attribute 'add_callback'

Does that mean that dask.distributed fundamentally won't work with trio-asyncio or just that trio-asyncio is missing some functionality which would be easy/moderately hard/hard to implement?

https://github.com/dask/distributed/blob/04de4b2adc1f0c94ab86e9aa46a4c67382a7eaea/distributed/utils.py#L339

[smurfix]

There is no add_callback function in asyncio, so no wonder trio_asyncio doesn't have one either.

[dhirschfeld]

Thx @smurfix - sounds like I'll have to take it up with distributed...

[njsmith]

Maybe you need to somehow wrap the asyncio loop object in a tornado loop object, and pass that to distributed?

…

On Fri, Feb 14, 2020, 04:52 Dave Hirschfeld ***@***.***> wrote: Thx @smurfix <https://github.com/smurfix> - sounds like I'll have to take it up with distributed... — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#22?email_source=notifications&email_token=AAEU42CF77W7RGWK5S7N7G3RC2HX7A5CNFSM4FH2JAYKYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOELY5THQ#issuecomment-586275230>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAEU42G2X3RTBCLNVH2NQK3RC2HX7ANCNFSM4FH2JAYA> .

[dhirschfeld]

Thanks for the suggestion @njsmith - I'll give that a go!

[dhirschfeld]

Just for informational purposes (in case anyone is interested), explicitly passing the loop was unnecessary but I run into a problem with different loops being used:

async def f():
    async with trio_asyncio.open_loop() as loop:
        print(id(loop))
        print(id(IOLoop.current().asyncio_loop))
        client = await loop.run_asyncio(partial(Client, loop=IOLoop.current(), asynchronous=True))
        future = client.submit(lambda x: x + 1, 10)
        return await loop.run_future(future)

In [25]: await f()
1656317099400
1656317099400
Traceback (most recent call last):

  File "<ipython-input-25-9fe41d264da1>", line 4, in async-def-wrapper

  File "<ipython-input-23-d02282e26f84>", line 5, in f
    client = await loop.run_asyncio(partial(Client, loop=IOLoop.current(), asynchronous=True))

  File "~\code\github\python-trio\trio-asyncio\trio_asyncio\base.py", line 231, in run_aio_coroutine
    return await run_aio_future(coro)

  File "~\code\github\python-trio\trio-asyncio\trio_asyncio\util.py", line 39, in run_aio_future
    res = await trio.hazmat.wait_task_rescheduled(abort_cb)

  File "~\envs\dev\lib\site-packages\trio\_core\_traps.py", line 165, in wait_task_rescheduled
    return (await _async_yield(WaitTaskRescheduled(abort_func))).unwrap()

  File "~\envs\dev\lib\site-packages\outcome\_sync.py", line 111, in unwrap
    raise captured_error

  File "~\envs\dev\lib\asyncio\tasks.py", line 630, in _wrap_awaitable
    return (yield from awaitable.__await__())

  File "~\envs\dev\lib\site-packages\distributed\client.py", line 957, in _start
    **self._startup_kwargs

  File "~\envs\dev\lib\site-packages\distributed\deploy\spec.py", line 365, in _
    await self._start()

  File "~\envs\dev\lib\site-packages\distributed\deploy\spec.py", line 290, in _start
    await super()._start()

  File "~\envs\dev\lib\site-packages\distributed\deploy\cluster.py", line 57, in _start
    comm = await self.scheduler_comm.live_comm()

  File "~\envs\dev\lib\site-packages\distributed\core.py", line 646, in live_comm
    connection_args=self.connection_args,

  File "~\envs\dev\lib\site-packages\distributed\comm\core.py", line 214, in connect
    future, timeout=min(deadline - time(), 1)

  File "~\envs\dev\lib\asyncio\tasks.py", line 442, in wait_for
    return fut.result()

  File "~\envs\dev\lib\site-packages\distributed\comm\tcp.py", line 349, in connect
    ip, port, max_buffer_size=MAX_BUFFER_SIZE, **kwargs

  File "~\envs\dev\lib\site-packages\tornado\tcpclient.py", line 270, in connect
    addrinfo = await self.resolver.resolve(host, port, af)

RuntimeError:
Task <Task pending coro=<BaseTCPConnector.connect() running at ~\envs\dev\lib\site-packages\distributed\comm\tcp.py:349>
cb=[_release_waiter(<Future pendi...1A4D379D8>()]>)() at ~\envs\dev\lib\asyncio\tasks.py:392]>
got Future <Future pending> attached to a different loop

[dhirschfeld]

Removing IPython from the equation (by running it as a script) results in a different error:

In [3]: from dask.distributed import LocalCluster

In [5]: cluster = LocalCluster(n_workers=1, threads_per_worker=2)

In [6]: cluster.scheduler.address
Out[6]: 'tcp://127.0.0.1:60590'

import asyncio
from functools import partial
from tornado.ioloop import IOLoop
import trio
import trio_asyncio
from dask.distributed import Client


address = 'tcp://127.0.0.1:60590'


async def f():
    async with trio_asyncio.open_loop() as loop:
        client = await loop.run_asyncio(partial(Client, address, asynchronous=True))
        future = client.submit(lambda x: x + 1, 10)
        return await loop.run_future(future)
    
 
res = trio_asyncio.run(f)
print(res)

(dev) ~\code\sandbox> python .\test-trio-dask.py
distributed.client - ERROR - Error in callback <function run_aio_future.<locals>.done_cb at 0x0000019721A32DC8> of <Future: finished, type: builtins.int, key: lambda-845252994edca236a111f8f502902dff>:
Traceback (most recent call last):
  File "~\envs\dev\lib\site-packages\trio\_core\_generated_run.py", line 96, in reschedule
    return  GLOBAL_RUN_CONTEXT.runner.reschedule(task, next_send)
AttributeError: '_thread._local' object has no attribute 'runner'

During handling of the above exception, another exception occurred:

Traceback (most recent call last):
  File "~\envs\dev\lib\site-packages\distributed\client.py", line 287, in execute_callback
    fn(fut)
  File "~\code\github\python-trio\trio-asyncio\trio_asyncio\util.py", line 25, in done_cb
    trio.hazmat.reschedule(task, outcome.capture(future.result))
  File "~\envs\dev\lib\site-packages\trio\_core\_generated_run.py", line 98, in reschedule
    raise RuntimeError('must be called from async context')
RuntimeError: must be called from async context

[dhirschfeld]

Unless the traceback highlights some glaringly obvious bug in trio-asyncio I'm leaning towards probably not possible (at this point in time)

[smurfix]

Ugh. I'm not running any sort of dask server, so when I thy I get this:

OSError: Timed out trying to connect to 'tcp://127.0.0.1:60590' after 10 s: in <distributed.comm.tcp.TCPConnector object at 0x7f7828df7e80>: ConnectionRefusedError: [Errno 111] Connection refused

This error message does not make sense, either the connection was refused (which implies that the server rejected us) or it timed out (which implies that the server didn't react at all). I filed a bug at dask/distributed#3487

Anyway, if I needed a dask client I would tend to fork the thing and apply a global s/asyncio/trio/g – but as I would like to understand why this doesn't work, please tell me how you started the server part.

[dhirschfeld]

Sorry - I should've been clearer above.

Dask has a LocalCluster class which spins up a scheduler and worker processes. The first code block:

from dask.distributed import LocalCluster
cluster = LocalCluster(n_workers=1, threads_per_worker=2)
print(cluster.scheduler.address)

was run by me in a separate python session to start a cluster and get the address.

I did it in a separate process to avoid any contamination of the testing code. I've seen LocalCluster cause some issues (even if only mucking with stderr/stdout) in an IPython session. It's not recommended for production but is useful for testing.

Anyway, having started the dask cluster you would then need to change the address variable in the trio test script to connect to the scheduler you spun up externally.

• I would tend to fork the thing and apply a global s/asyncio/trio/g*

Yeah, I'm at the point of wondering if I need to start a dask-trio project! There would no doubt be a lot of benefits to it but it might be a lot of work :/ Whilst dask itself is a pretty complex project I'm hopeful that the client side is comparatively straightforward with all the complexity being in the scheduler. I haven't yet looked into that

[dhirschfeld]

At least seems potentially feasible:

Client: something that users use to submit tasks to the scheduler. Would need to be rewritten but is fairly simple. Needs to know how to serialize functions, encode msgpack, and send data over a socket.

dask/distributed#586

[smurfix]

OK, I've looked at the code a bit, particularly that of Tornado.

Bottom line: I do think it's possible to teach Tornado to be compatible with trio-asyncio, but it would result in something rather fragile at best. (Its multi-loop support doesn't help. At all. Neither does its Python 2.7 compatibility …)

If you don't want to write a dask-trio client, another way forward would be to hack up a tornado-trio module, i.e. take Tornado's IOLoop and hack it until it talks to trio instead of asyncio, taking out two layers of indirection. That should be straightforward, if a bit of work, though you'd probably have to monkeypatch the static methods in tornado.ioloop.IOLoop to impersonate your new class instead.

My preferred solution would be dask-anyio.

[dhirschfeld]

Thanks for taking a look @smurfix, and for the helpful suggestions!

My preferred solution would be dask-anyio.

That is sounding more and more like the way to go...