如何中断 Tornado 协程

Question

Suppose I have two functions that work like this:

@tornado.gen.coroutine
def f():
    for i in range(4):
        print("f", i)
        yield tornado.gen.sleep(0.5)

@tornado.gen.coroutine
def g():
    yield tornado.gen.sleep(1)
    print("Let's raise RuntimeError")
    raise RuntimeError

In general, function f might contain endless loop and never return (eg it can process some queue).

What I want to do is to be able to interrupt it, at any time it yields.

The most obvious way doesn't work. Exception is only raised after function f exits (if it's endless, it obviously never happens).

@tornado.gen.coroutine
def main():
    try:
        yield [f(), g()]
    except Exception as e:
        print("Caught", repr(e))

    while True:
        yield tornado.gen.sleep(10)

if __name__ == "__main__":
    tornado.ioloop.IOLoop.instance().run_sync(main)

Output:

f 0
f 1
Let's raise RuntimeError
f 2
f 3
Traceback (most recent call last):
  File "/tmp/test/lib/python3.4/site-packages/tornado/gen.py", line 812, in run
    yielded = self.gen.send(value)
StopIteration

During handling of the above exception, another exception occurred:

Traceback (most recent call last):
  <...>
  File "test.py", line 16, in g
    raise RuntimeError
RuntimeError

That is, exception is only raised when both of the coroutines return (both futures resolve).

This's partially solved by tornado.gen.WaitIterator , but it's buggy ( unless I'm mistaken ). But that's not the point.

It still doesn't solve the problem of interrupting existing coroutines. Coroutine continues to run even though the function that started it exits.

EDIT : it seems like coroutine cancellation is something not really supported in Tornado, unlike in Python's asyncio, where you can easily throw CancelledError at every yield point.

Answer 1

If you use WaitIterator according to the instructions , and use a toro.Event to signal between coroutines, it works as expected:

from datetime import timedelta
import tornado.gen
import tornado.ioloop

import toro

stop = toro.Event()


@tornado.gen.coroutine
def f():
    for i in range(4):
        print("f", i)

        # wait raises Timeout if not set before the deadline.
        try:
            yield stop.wait(timedelta(seconds=0.5))
            print("f done")
            return
        except toro.Timeout:
            print("f continuing")


@tornado.gen.coroutine
def g():
    yield tornado.gen.sleep(1)
    print("Let's raise RuntimeError")
    raise RuntimeError


@tornado.gen.coroutine
def main():
    wait_iterator = tornado.gen.WaitIterator(f(), g())
    while not wait_iterator.done():
        try:
            result = yield wait_iterator.next()
        except Exception as e:
            print("Error {} from {}".format(e, wait_iterator.current_future))
            stop.set()
        else:
            print("Result {} received from {} at {}".format(
                result, wait_iterator.current_future,
                wait_iterator.current_index))


if __name__ == "__main__":
    tornado.ioloop.IOLoop.instance().run_sync(main)

For now, pip install toro to get the Event class. Tornado 4.2 will include Event, see the changelog .

Answer 2

Since version 5, Tornado runs on asyncio event loop .

On Python 3, the IOLoop is always a wrapper around the asyncio event loop, and asyncio.Future and asyncio.Task are used instead of their Tornado counterparts.

Hence you can use asyncio Task cancellation, ie asyncio.Task.cancel .

Your example with a queue reading while-true loop, might look like this.

import logging
from asyncio import CancelledError

from tornado import ioloop, gen


async def read_off_a_queue():
    while True:
        try:
            await gen.sleep(1)
        except CancelledError:
            logging.debug('Reader cancelled')
            break
        else:
            logging.debug('Pretend a task is consumed')

async def do_some_work():
    await gen.sleep(5)
    logging.debug('do_some_work is raising')
    raise RuntimeError                     

async def main():
    logging.debug('Starting queue reader in background')
    reader_task = gen.convert_yielded(read_off_a_queue())    
    try:
        await do_some_work()
    except RuntimeError:
        logging.debug('do_some_work failed, cancelling reader')
        reader_task.cancel()
        # give the task a chance to clean up, in case it
        # catches CancelledError and awaits something
        try:
            await reader_task            
        except CancelledError:
            pass


if __name__ == '__main__':
    logging.basicConfig(level='DEBUG')        
    ioloop.IOLoop.instance().run_sync(main)

If you run it, you should see:

DEBUG:asyncio:Using selector: EpollSelector
DEBUG:root:Starting queue reader in background
DEBUG:root:Pretend a task is consumed
DEBUG:root:Pretend a task is consumed
DEBUG:root:Pretend a task is consumed
DEBUG:root:Pretend a task is consumed
DEBUG:root:do_some_work is raising
DEBUG:root:do_some_work failed, cancelling reader
DEBUG:root:Reader cancelled

Answer 3

Warning : This is not a working solution. Look at the commentary. Still if you're new (as myself), this example can show the logical flow. Thanks @nathaniel-j-smith and @wgh

What is the difference using something more primitive, like global variable for instance?

import asyncio


event = asyncio.Event()
aflag = False


async def short():
    while not aflag:
        print('short repeat')
        await asyncio.sleep(1)
    print('short end')


async def long():
    global aflag

    print('LONG START')
    await asyncio.sleep(3)

    aflag = True
    print('LONG END')


async def main():
    await asyncio.gather(long(), short())

if __name__ == '__main__':
    asyncio.run(main())

It is for asyncio , but I guess the idea stays the same. This is a semi-question (why Event would be better?). Yet solution yields exact result author needs:

LONG START
short repeat
short repeat
short repeat
LONG END
short end

UPDATE: this slides may be really helpful in understanding core of a problem.