Latency in Asynchronous Python

This week I was debugging a misbehaving Python program that makes significant use of Python’s asyncio. The program would eventually take very long periods of time to respond to network requests. My first suspicion was a CPU-heavy coroutine hogging the thread, preventing the socket coroutines from running, but an inspection with pdb showed this wasn’t the case. Instead, the program’s author had made a couple of fundamental mistakes using asyncio. Let’s discuss them using small examples.

Setting the stage: There’s a heartbeat coroutine that “beats” once per second. A real program would send out a packet as the heartbeat, but here it just prints how late it was scheduled.

async def heartbeat():
    while True:
        start = time.time()
        await asyncio.sleep(1)
        delay = time.time() - start - 1
        print(f'heartbeat delay = {delay:.3f}s')

Running this with asyncio.run(heartbeat()):

heartbeat delay = 0.001s
heartbeat delay = 0.001s
heartbeat delay = 0.001s

It’s consistently 1ms late, but good enough, especially considering what’s to come. A program that only sends a heartbeat is pretty useless, so a real program will be busy working on other things concurrently. In this example, we have little 10ms payloads of work to do, which are represented by this process() function:

JOB_DURATION = 0.01  # 10ms

async def process():
    time.sleep(JOB_DURATION) # simulate CPU time

That’s a synchronous sleep because it’s standing in for actual CPU work. Maybe it’s parsing JSON in a loop or crunching numbers in NumPy. Use your imagination. During this 10ms no other coroutines can be scheduled because this is, after all, still just a single-threaded program.

JOB_COUNT = 200

async def main():
    asyncio.create_task(heartbeat())

    await asyncio.sleep(2.5)

    print('begin processing')
    count = JOB_COUNT
    for _ in range(JOB_COUNT):
        asyncio.create_task(process())

    await asyncio.sleep(5)

This program starts the heartbeat coroutine in a task. A coroutine doesn’t make progress unless someone is waiting on it, and that something can be a task. So it will continue along independently without prodding.

The arbitrary 2.5 second sleep simulates waiting, say, for a network request. In the output we’ll see the heartbeat tick a couple of times, then it will create and process 200 jobs concurrently. In a real program we’d have some way to collect the results, but we can ignore that part for now. They’re only 10ms, so the effect on the heartbeat should be pretty small right?

heartbeat delay = 0.001s
heartbeat delay = 0.001s
begin processing
heartbeat delay = 1.534s
heartbeat delay = 0.001s
heartbeat delay = 0.001s

The heartbeat was delayed for 1.5 seconds by a mere 200 tasks each doing only 10ms of work each. What happened?

Python calls the object that schedules tasks a loop, and this is no coincidence. Everything to be scheduled gets put into a loop and is scheduled round robin, one after another. The 200 tasks got scheduled ahead of the heartbeat, and so it doesn’t get scheduled again until each of those tasks either yields (await) or completes.

It really didn’t take much to significantly hamper the heartbeat, and, with a dumb bytecode compiler, 10ms may not be much work at all. The lesson here is to avoid spawning many tasks if latency is an important consideration.

A semaphore is not the answer

My first idea at a solution: What if we used a semaphore to limit the number of “active” tasks at a time? Then perhaps the heartbeat wouldn’t have to compete with so many other tasks for time.

WORKER_COUNT = 4  # max "active" jobs at a time

async def main_with_semaphore():
    asyncio.create_task(heartbeat())

    await asyncio.sleep(2.5)

    sem = asyncio.Semaphore(WORKER_COUNT)
    async def process():
        await sem.acquire()
        time.sleep(JOB_DURATION)
        sem.release()

    print('begin processing')
    for _ in range(JOB_COUNT):
        asyncio.create_task(process())

    await asyncio.sleep(5)

When the heartbeat sleep completes, about half the jobs will be complete and the other half blocked on the semaphore. So perhaps the heartbeat gets to skip ahead of all the blocked tasks since they’re not yet ready to run?

heartbeat delay = 0.001s
heartbeat delay = 0.001s
begin processing
heartbeat delay = 1.537s
heartbeat delay = 0.001s
heartbeat delay = 0.001s

It made no difference whatsoever because the tasks each “held their place” in line in the loop! Even reducing WORKER_COUNT to 1 would have no effect. As soon as a task completes, it frees the task waiting next in line. The semaphore does practically nothing here.

Solving it with a job queue

Here’s what does work: a job queue. Create a queue to be populated with coroutines (not tasks), and have a small number of tasks run jobs from the queue. Since this is a real solution, I’ve made this example more complete.

async def main_with_queue():
    asyncio.create_task(heartbeat())

    await asyncio.sleep(2.5)

    queue = asyncio.Queue(maxsize=1)
    async def worker():
        while True:
            coro = await queue.get()
            await coro  # consider using try/except
            queue.task_done()
    workers = [asyncio.create_task(worker())
                   for _ in range(WORKER_COUNT)]

    print('begin processing')
    for _ in range(JOB_COUNT):
        await queue.put(process())
    await queue.join()
    print('end processing')

    for w in workers:
        w.cancel()

    await asyncio.sleep(2)

The task_done() and join() methods make it trivial synchronize on full job completion. I also take the time to destroy the worker tasks. It’s harmless to leave them blocked on the queue. They’ll be garbage collected so it’s not a resource leak. However, CPython complains about garbage collecting running tasks because it looks like a mistake — and it usually is.

If you read carefully you might have noticed the queue’s maximum size is set to 1: not much of a “queue”! Go developers will recognize this as being (nearly) an unbuffered channel, the default and most common kind of channel. So it’s more a synchronized rendezvous between producer (put()) and consumer (get()). The producer waits at the queue with a job until a task is free to come take it. A task waits at the queue until a producer arrives with a job for it.

heartbeat delay = 0.001s
heartbeat delay = 0.001s
begin processing
heartbeat delay = 0.014s
heartbeat delay = 0.020s
end processing
heartbeat delay = 0.002s
heartbeat delay = 0.001s

The output shows that the impact to the heartbeat was modest — about the best we could hope for from async/await — and the heartbeat continued while jobs were running. The more concurrency — the more worker tasks running on the queue — the greater the latency.

Note: Increasing the WORKER_COUNT in this toy example won’t have an impact on latency since the jobs aren’t actually concurrent. They start, run, and complete before another worker task can draw from the queue. Putting a couple awaits in process() allows for concurrency:

WORKER_COUNT = 200

async def process():
    await asyncio.sleep(0.01)
    time.sleep(JOB_DURATION)
    await asyncio.sleep(0.01)

Since there are so many worker tasks, this is back to the initial problem:

heartbeat delay = 0.001s
heartbeat delay = 0.001s
begin processing
heartbeat delay = 1.655s
end processing
heartbeat delay = 0.001s
heartbeat delay = 0.001s

As WORKER_COUNT decreases, so does heartbeat latency.

Unbounded queues

Here’s another defect from the same program. Create an unbounded queue, a producer, and a consumer. The consumer prints the queue size so we can see what’s happening:

async def producer_consumer():
    queue = asyncio.Queue()
    done = asyncio.Condition()

    async def producer():
        for i in range(100_000):
            await queue.put(i)
        await queue.join()
        async with done:
            done.notify()

    async def consumer():
        while True:
            await queue.get()
            print(f'qsize = {queue.qsize()}')
            queue.task_done()

    asyncio.create_task(producer())
    asyncio.create_task(consumer())

    async with done:
        await done.wait()

The output of this program begins:

qsize = 99999
qsize = 99998
qsize = 99997
qsize = 99996
...

So the entire queue is populated before the consumer does anything at all: tons of latency for whatever is being consumed. Since the queue is unbounded, the producer never needs to yield. You might be tempted to use asyncio.sleep(0) in the producer to yield explicitly:

    async def producer():
        for i in range(100_000):
            await queue.put(i)
            await asyncio.sleep(0)  # yield
        await queue.join()
        async with done:
            done.notify()

This even seems to work! The output looks like this:

qsize = 0
qsize = 0
qsize = 0
qsize = 0

However, this is fragile and not a real solution. If the consumer yields just two times in its own loop, its nearly back to where we started:

    async def consumer():
        while True:
            await queue.get()
            print(f'qsize = {queue.qsize()}')
            queue.task_done()
            await asyncio.sleep(0)
            await asyncio.sleep(0)

The output shows that the producer gradually creeps ahead of the consumer. On each consumer iteration, the producer iterates twice:

qsize = 0
qsize = 1
qsize = 2
qsize = 3
...

There’s a really simple solution to this: Never, ever use unbounded queues. In fact every unbounded asyncio.Queue() is a bug. It’s a serious API defect that asyncio allows unbounded queues to be created at all. The default maxsize should have been actually zero (unbuffered), not infinite. Because unbounded is the default, virtually every example of asyncio.Queue — online, offline, and even the official documentation — is broken in some way.

Important takeaways

  1. The default asyncio.Queue() is always wrong.
  2. asyncio.sleep(0) is nearly always used incorrectly.
  3. Use a maxsize=1 job queue instead of spawning many identical tasks.

Python linters should be updated to warn about 1 and 2 by default.

Update: A couple of people have pointed out an argument in the Trio documentation for unbounded queues. This argument conflates two different concepts: data structure queues and concurrent communication infrastructure queues. To distinguish, the latter is often called a channel. An unbounded queue (collections.deque) is necessary, but and unbounded channel (asyncio.Queue) is always wrong. The Trio documentation describes a web crawler, which is fundamentally a breadth-first search (read: queue-oriented) of a graph. So this is a plain old BFS queue, not a channel, which is why it’s reasonable for it to be unbounded.

Have a comment on this article? Start a discussion in my public inbox by sending an email to ~skeeto/public-inbox@lists.sr.ht [mailing list etiquette] , or see existing discussions.

null program

Chris Wellons

wellons@nullprogram.com (PGP)
~skeeto/public-inbox@lists.sr.ht (view)