Working with threads

Practical asynchronous applications occasionally need to run network, file or computationally expensive operations. Such operations would normally block the asynchronous event loop, leading to performance issues. The solution is to run such code in worker threads. Using worker threads lets the event loop continue running other tasks while the worker thread runs the blocking call.

Caution

Do not spawn too many threads, as the context switching overhead may cause your system to slow down to a crawl. A few dozen threads should be fine, but hundreds are probably bad. Consider using AnyIO’s semaphores to limit the maximum number of threads.

Running a function in a worker thread

To run a (synchronous) callable in a worker thread:

import time

from anyio import to_thread, run

async def main():
    await to_thread.run_sync(time.sleep, 5)

run(main)

By default, tasks are shielded from cancellation while they are waiting for a worker thread to finish. You can pass the cancellable=True parameter to allow such tasks to be cancelled. Note, however, that the thread will still continue running – only its outcome will be ignored.

Calling asynchronous code from a worker thread

If you need to call a coroutine function from a worker thread, you can do this:

from anyio import from_thread, sleep, to_thread, run

def blocking_function():
    from_thread.run(sleep, 5)

async def main():
    await to_thread.run_sync(blocking_function)

run(main)

Note

The worker thread must have been spawned using run_sync() for this to work.

Calling synchronous code from a worker thread

Occasionally you may need to call synchronous code in the event loop thread from a worker thread. Common cases include setting asynchronous events or sending data to a memory object stream. Because these methods aren’t thread safe, you need to arrange them to be called inside the event loop thread using run_sync():

import time

from anyio import Event, from_thread, to_thread, run

def worker(event):
    time.sleep(1)
    from_thread.run_sync(event.set)

async def main():
    event = Event()
    await to_thread.run_sync(worker, event)
    await event.wait()

run(main)

Calling asynchronous code from an external thread

If you need to run async code from a thread that is not a worker thread spawned by the event loop, you need a blocking portal. This needs to be obtained from within the event loop thread.

One way to do this is to start a new event loop with a portal, using start_blocking_portal (which takes mostly the same arguments as run():

from anyio.from_thread import start_blocking_portal


with start_blocking_portal(backend='trio') as portal:
    portal.call(...)

If you already have an event loop running and wish to grant access to external threads, you can create a BlockingPortal directly:

from anyio import run
from anyio.from_thread import BlockingPortal


async def main():
    async with BlockingPortal() as portal:
        # ...hand off the portal to external threads...
        await portal.sleep_until_stopped()

run(main)

Spawning tasks from worker threads

When you need to spawn a task to be run in the background, you can do so using start_task_soon():

from concurrent.futures import as_completed

from anyio import sleep
from anyio.from_thread import start_blocking_portal


async def long_running_task(index):
    await sleep(1)
    print(f'Task {index} running...')
    await sleep(index)
    return f'Task {index} return value'


with start_blocking_portal() as portal:
    futures = [portal.start_task_soon(long_running_task, i) for i in range(1, 5)]
    for future in as_completed(futures):
        print(future.result())

Cancelling tasks spawned this way can be done by cancelling the returned Future.

Blocking portals also have a method similar to TaskGroup.start(): start_task() which, like its counterpart, waits for the callable to signal readiness by calling task_status.started():

from anyio import sleep, TASK_STATUS_IGNORED
from anyio.from_thread import start_blocking_portal


async def service_task(*, task_status=TASK_STATUS_IGNORED):
    task_status.started('STARTED')
    await sleep(1)
    return 'DONE'


with start_blocking_portal() as portal:
    future, start_value = portal.start_task(service_task)
    print('Task has started with value', start_value)

    return_value = future.result()
    print('Task has finished with return value', return_value)

Using asynchronous context managers from worker threads

You can use wrap_async_context_manager() to wrap an asynchronous context managers as a synchronous one:

from anyio.from_thread import start_blocking_portal


class AsyncContextManager:
    async def __aenter__(self):
        print('entering')

    async def __aexit__(self, exc_type, exc_val, exc_tb):
        print('exiting with', exc_type)


async_cm = AsyncContextManager()
with start_blocking_portal() as portal, portal.wrap_async_context_manager(async_cm):
    print('inside the context manager block')

Note

You cannot use wrapped async context managers in synchronous callbacks inside the event loop thread.

Context propagation

When running functions in worker threads, the current context is copied to the worker thread. Therefore any context variables available on the task will also be available to the code running on the thread. As always with context variables, any changes made to them will not propagate back to the calling asynchronous task.

When calling asynchronous code from worker threads, context is again copied to the task that calls the target function in the event loop thread.

Adjusting the default maximum worker thread count

The default AnyIO worker thread limiter has a value of 40, meaning that any calls to to_thread.run_sync() without an explicit limiter argument will cause a maximum of 40 threads to be spawned. You can adjust this limit like this:

from anyio import to_thread

async def foo():
    # Set the maximum number of worker threads to 60
    to_thread.current_default_thread_limiter().total_tokens = 60

Note

AnyIO’s default thread pool limiter does not affect the default thread pool executor on asyncio.