I'm trying to merge a bunch of asynchronous generators in Python 3.7 while still adding new async generators on iteration. I'm currently using aiostream to merge my generators:
from asyncio import sleep, run
from aiostream.stream import merge
async def go():
yield 0
await sleep(1)
yield 50
await sleep(1)
yield 100
async def main():
tasks = merge(go(), go(), go())
async for v in tasks:
print(v)
if __name__ == '__main__':
run(main())
However, I need to be able to continue to add to the running tasks once the loop has begun. Something like.
from asyncio import sleep, run
from aiostream.stream import merge
async def go():
yield 0
await sleep(1)
yield 50
await sleep(1)
yield 100
async def main():
tasks = merge(go(), go(), go())
async for v in tasks:
if v == 50:
tasks.merge(go())
print(v)
if __name__ == '__main__':
run(main())
The closest I've got to this is using the aiostream library but maybe this can also be written fairly neatly with just the native asyncio standard library.
Here is an implementation that should work efficiently even with a large number of async iterators:
class merge:
def __init__(self, *iterables):
self._iterables = list(iterables)
self._wakeup = asyncio.Event()
def _add_iters(self, next_futs, on_done):
for it in self._iterables:
it = it.__aiter__()
nfut = asyncio.ensure_future(it.__anext__())
nfut.add_done_callback(on_done)
next_futs[nfut] = it
del self._iterables[:]
return next_futs
async def __aiter__(self):
done = {}
next_futs = {}
def on_done(nfut):
done[nfut] = next_futs.pop(nfut)
self._wakeup.set()
self._add_iters(next_futs, on_done)
try:
while next_futs:
await self._wakeup.wait()
self._wakeup.clear()
for nfut, it in done.items():
try:
ret = nfut.result()
except StopAsyncIteration:
continue
self._iterables.append(it)
yield ret
done.clear()
if self._iterables:
self._add_iters(next_futs, on_done)
finally:
# if the generator exits with an exception, or if the caller stops
# iterating, make sure our callbacks are removed
for nfut in next_futs:
nfut.remove_done_callback(on_done)
def append_iter(self, new_iter):
self._iterables.append(new_iter)
self._wakeup.set()
The only change required for your sample code is that the method is named append_iter, not merge.
This can be done using stream.flatten with an asyncio queue to store the new generators.
import asyncio
from aiostream import stream, pipe
async def main():
queue = asyncio.Queue()
await queue.put(go())
await queue.put(go())
await queue.put(go())
xs = stream.call(queue.get)
ys = stream.cycle(xs)
zs = stream.flatten(ys, task_limit=5)
async with zs.stream() as streamer:
async for item in streamer:
if item == 50:
await queue.put(go())
print(item)
Notice that you may tune the number of tasks that can run at the same time using the task_limit argument. Also note that zs can be elegantly defined using the pipe syntax:
zs = stream.call(queue.get) | pipe.cycle() | pipe.flatten(task_limit=5)
Disclaimer: I am the project maintainer.
Related
I want to achieve same effect as
# Code 1
from multiprocessing.pool import ThreadPool as Pool
from time import sleep, time
def square(a):
print('start', a)
sleep(a)
print('end', a)
return a * a
def main():
p = Pool(2)
queue = list(range(4))
start = time()
results = p.map(square, queue)
print(results)
print(time() - start)
if __name__ == "__main__":
main()
with async functions like
# Code 2
from multiprocessing.pool import ThreadPool as Pool
from time import sleep, time
import asyncio
async def square(a):
print('start', a)
sleep(a) # await asyncio.sleep same effect
print('end', a)
return a * a
async def main():
p = Pool(2)
queue = list(range(4))
start = time()
results = p.map_async(square, queue)
results = results.get()
results = [await result for result in results]
print(results)
print(time() - start)
if __name__ == "__main__":
loop = asyncio.get_event_loop()
loop.run_until_complete(main())
loop.close()
Currently Code 1 takes 4 seconds and Code 2 takes 6 seconds which means it is not running in parallel. What is the correct and cleanest way to run multiple async functions in parallel?
Better to be python 3.6 compatible. Thank you!
map_async() is not the same "async" as in async def - if it is fed with an async def method, it won't actually run it but return a coroutine instance immediately (try calling such a method without await). Then you awaited on the 4 coroutines one by one, that equals to sequential execution, and ended up with 6 seconds.
Please see following example:
from time import time
import asyncio
from asyncio.locks import Semaphore
semaphore = Semaphore(2)
async def square(a):
async with semaphore:
print('start', a)
await asyncio.sleep(a)
print('end', a)
return a * a
async def main():
start = time()
tasks = []
for a in range(4):
tasks.append(asyncio.ensure_future(square(a)))
await asyncio.wait(tasks)
print([t.result() for t in tasks])
print(time() - start)
if __name__ == "__main__":
loop = asyncio.get_event_loop()
loop.run_until_complete(main())
loop.close()
The Semaphore acts similarly like the ThreadPool - it allows only 2 concurrent coroutines entering the async with semaphore: block.
I think it's easier to understand in terms of code:
try:
async for item in timeout(something(), timeout=60):
await do_something_useful(item)
except asyncio.futures.TimeoutError:
await refresh()
I want the async for to run at most 60 seconds.
I needed to do something like this to create a websocket(also an async iterator) which times out if it doesn't get a message after a certain duration. I settled on the following:
socket_iter = socket.__aiter__()
try:
while True:
message = await asyncio.wait_for(
socket_iter.__anext__(),
timeout=10
)
except asyncio.futures.TimeoutError:
# streaming is completed
pass
AsyncTimedIterable could be the implementation of timeout() in your code:
class _AsyncTimedIterator:
__slots__ = ('_iterator', '_timeout', '_sentinel')
def __init__(self, iterable, timeout, sentinel):
self._iterator = iterable.__aiter__()
self._timeout = timeout
self._sentinel = sentinel
async def __anext__(self):
try:
return await asyncio.wait_for(self._iterator.__anext__(), self._timeout)
except asyncio.TimeoutError:
return self._sentinel
class AsyncTimedIterable:
__slots__ = ('_factory', )
def __init__(self, iterable, timeout=None, sentinel=None):
self._factory = lambda: _AsyncTimedIterator(iterable, timeout, sentinel)
def __aiter__(self):
return self._factory()
(original answer)
Or use this class to replace your timeout() function:
class AsyncTimedIterable:
def __init__(self, iterable, timeout=None, sentinel=None):
class AsyncTimedIterator:
def __init__(self):
self._iterator = iterable.__aiter__()
async def __anext__(self):
try:
return await asyncio.wait_for(self._iterator.__anext__(),
timeout)
except asyncio.TimeoutError:
return sentinel
self._factory = AsyncTimedIterator
def __aiter__(self):
return self._factory()
A simple approach is to use an asyncio.Queue, and separate the code into two coroutines:
queue = asyncio.Queue()
async for item in something():
await queue.put(item)
In another coroutine:
while True:
try:
item = await asyncio.wait_for(queue.get(), 60)
except asyncio.TimeoutError:
pass
else:
if item is None:
break # use None or whatever suits you to gracefully exit
await do_something_useful(item)
refresh()
Please note, it will make the queue grow if the handler do_something_useful() is slower than something() generates items. You may set a maxsize on the queue to limit the buffer size.
Answer to your question can be different based on nature of refresh function. If it's very short-running function it can be freely called inside coroutine. But if it's blocking function (due to network or CPU) it should be ran in executor to avoid freezing asyncio event loop.
Code below shows example for the first case, changing it to run refresh in executor is not hard.
Second thing to be clarified is a nature of asynchronous iterator. As far as I understand, you're using it to either get result from something or None if timeout occurred.
If I understand logic correctly, your code can be written clearer (similar to non-async style as asyncio is created to allow) using async_timeout context manager and without using asynchronous iterator at all:
import asyncio
from async_timeout import timeout
async def main():
while True:
try:
async with timeout(60):
res = await something()
await do_something_useful(item)
except asyncio.TimeoutError:
pass
finally:
refresh()
Your question is missing a couple of details, but assuming something() is an async iterator or generator and you want item to be sentinel everytime something has not yielded a value within the timeout, here is an implementation of timeout():
import asyncio
from typing import *
T = TypeVar('T')
# async generator, needs python 3.6
async def timeout(it: AsyncIterator[T], timeo: float, sentinel: T) -> AsyncGenerator[T, None]:
try:
nxt = asyncio.ensure_future(it.__anext__())
while True:
try:
yield await asyncio.wait_for(asyncio.shield(nxt), timeo)
nxt = asyncio.ensure_future(it.__anext__())
except asyncio.TimeoutError:
yield sentinel
except StopAsyncIteration:
pass
finally:
nxt.cancel() # in case we're getting cancelled our self
test:
async def something():
yield 1
await asyncio.sleep(1.1)
yield 2
await asyncio.sleep(2.1)
yield 3
async def test():
expect = [1, None, 2, None, None, 3]
async for item in timeout(something(), 1, None):
print("Check", item)
assert item == expect.pop(0)
asyncio.get_event_loop().run_until_complete(test())
When wait_for() times out it will cancel the task. Therefore, we need to wrap it.__anext__() in a task and then shield it, to be able to resume the iterator.
I want the coroutine to execute refresh at least every 60 seconds.
If you need to execute refresh every 60 seconds regardless of what happens with do_something_useful, you can arrange that with a separate coroutine:
import time
async def my_loop():
# ensure refresh() is invoked at least once in 60 seconds
done = False
async def repeat_refresh():
last_run = time.time()
while not done:
await refresh()
now = time.time()
await asyncio.sleep(max(60 - (now - last_run), 0))
last_run = now
# start repeat_refresh "in the background"
refresh_task = asyncio.get_event_loop().create_task(repeat_refresh())
try:
async for item in something():
if item is not None:
await do_something_useful(item)
await refresh()
finally:
done = True
Is there a good way, or a well-supported library, for merging async iterators in python3?
The desired behavior is basically the same as that of merging observables in reactivex.
That is, in the normal case, if I'm merging two async iterator, I want the resulting async iterator to yield results chronologically. An error in one of the iterators should derail the merged iterator.
(Source: http://reactivex.io/documentation/operators/merge.html)
This is my best attempt, but it seems like something there might be a standard solution to:
async def drain(stream, q, sentinal=None):
try:
async for item in stream:
await q.put(item)
if sentinal:
await q.put(sentinal)
except BaseException as e:
await q.put(e)
async def merge(*streams):
q = asyncio.Queue()
sentinal = namedtuple("QueueClosed", ["truthy"])(True)
futures = {
asyncio.ensure_future(drain(stream, q, sentinal)) for stream in streams
}
remaining = len(streams)
while remaining > 0:
result = await q.get()
if result is sentinal:
remaining -= 1
continue
if isinstance(result, BaseException):
raise result
yield result
if __name__ == "__main__":
# Example: Should print:
# 1
# 2
# 3
# 4
loop = asyncio.get_event_loop()
async def gen():
yield 1
await asyncio.sleep(1.5)
yield 3
async def gen2():
await asyncio.sleep(1)
yield 2
await asyncio.sleep(1)
yield 4
async def go():
async for x in merge(gen(), gen2()):
print(x)
loop.run_until_complete(go())
You can use aiostream.stream.merge:
from aiostream import stream
async def go():
async for x in stream.merge(gen(), gen2()):
print(x)
More examples in the documentation and this answer.
Say I have two async generators:
async def get_rules():
while True:
yield 'rule=1'
asyncio.sleep(2)
async def get_snapshots():
while True:
yield 'snapshot=1'
asyncio.sleep(5)
I want to merge them into a single async generator that returns 2-tuples, with the latest value from both. Sort of combineLatest.
What is the best way to do this?
You might want to have a look at aiostream, especially stream.merge and stream.accumulate:
import asyncio
from itertools import count
from aiostream import stream
async def get_rules():
for x in count():
await asyncio.sleep(2)
yield 'rule', x
async def get_snapshots():
for x in count():
await asyncio.sleep(5)
yield 'snapshot', x
async def main():
xs = stream.merge(get_rules(), get_snapshots())
ys = stream.map(xs, lambda x: {x[0]: x[1]})
zs = stream.accumulate(ys, lambda x, e: {**x, **e}, {})
async with zs.stream() as streamer:
async for z in streamer:
print(z)
loop = asyncio.get_event_loop()
loop.run_until_complete(main())
loop.close()
Output:
{}
{'rule': 0}
{'rule': 1}
{'rule': 1, 'snapshot': 0}
{'rule': 2, 'snapshot': 0}
[...]
See the project page and the documentation for further information.
Disclaimer: I am the project maintainer.
I came up with this:
async def combine(**generators):
"""Given a bunch of async generators, merges the events from
all of them. Each should have a name, i.e. `foo=gen, bar=gen`.
"""
combined = Channel()
async def listen_and_forward(name, generator):
async for value in generator:
await combined.put({name: value})
for name, generator in generators.items():
asyncio.Task(listen_and_forward(name, generator))
async for item in combined:
yield item
async def combine_latest(**generators):
"""Like "combine", but always includes the latest value from
every generator.
"""
current = {}
async for value in combine(**generators):
current.update(value)
yield current
Call it like so:
async for item in combine_latest(rules=rulesgen, snap=snapgen):
print(item)
Output looks like this:
{'rules': 'rule-1'}
{'rules': 'rule-1', 'snap': 'snapshot-1'}
{'rules': 'rule-1', 'snap': 'snapshot-1'}
....
I am using aiochannel, but a normal asyncio.Queue should be fine, too.
In my project, I have a list to tasks that I execute with.
loop.run_until_complete(tasks)
However, there is an infinite number of tasks, so at the moment, I execute them in batches. Essentially, I have this:
def get_results(tasks):
return [result for result in loop.run_until_complete(handle_tasks(tasks))]
while True:
tasks = get_tasks()
results = get_results(tasks)
I get a number of tasks, I lunch a regular function that uses a loop to perform these tasks asynchronously and returns the results.
This approach works, but I believe it can be improved.
Instead of doing batches of tasks, I would like to do some sort of task top-up.
Something like this:
while True:
if current_tasks < max_tasks:
new_tasks = get_tasks(max_tasks - current_tasks)
add_tasks(new_tasks)
current_tasks, results = stats_and_results()
I appreciate any ideas on how to approach this problem.
Thanks!
We had a similar problem and ended up writing a small "Pool" wrapper that takes jobs and run them with a predefined concurrency.
import asyncio
import sys
class Pool:
def __init__(self, concurrency):
self._sem = asyncio.BoundedSemaphore(concurrency)
self.jobs = []
async def __aenter__(self):
return self
async def __aexit__(self, *_):
if len(self.jobs) > 0:
await asyncio.wait(self.jobs)
def put(self, coro):
assert asyncio.iscoroutine(coro)
async def wrapped():
async with self._sem:
await coro
fut = asyncio.ensure_future(wrapped())
self.jobs.append(fut)
async def __aiter__(self):
return self
async def __anext__(self):
try:
coro = self.jobs.pop(0)
except IndexError:
raise StopAsyncIteration()
else:
return await coro
You can then use it this way:
async def main():
pool = Pool(10)
for task in get_tasks():
pool.put(task)
async for result in pool:
print('got', result)
This will schedule all the task, run at most 10 of them concurrently and return the results as they come to the main() coroutine