/unsync

Unsynchronize asyncio

Primary LanguagePythonMIT LicenseMIT

unsync

Unsynchronize asyncio by using an ambient event loop, or executing in separate threads or processes.

Quick Overview

Functions marked with the @unsync decorator will behave in one of the following ways:

  • async functions will run in the unsync.loop event loop executed from unsync.thread
  • Regular functions will execute in unsync.thread_executor, a ThreadPoolExecutor
    • Useful for IO bounded work that does not support asyncio
  • Regular functions marked with @unsync(cpu_bound=True) will execute in unsync.process_executor, a ProcessPoolExecutor
    • Useful for CPU bounded work

All @unsync functions will return an Unfuture object. This new future type combines the behavior of asyncio.Future and concurrent.Future with the following changes:

  • Unfuture.set_result is threadsafe unlike asyncio.Future
  • Unfuture instances can be awaited, even if made from concurrent.Future
  • Unfuture.result() is a blocking operation except in unsync.loop/unsync.thread where it behaves like asyncio.Future.result and will throw an exception if the future is not done

Examples

Simple Sleep

A simple sleeping example with asyncio:

async def sync_async():
    await asyncio.sleep(1)
    return 'I hate event loops'


async def main():
    future1 = asyncio.create_task(sync_async())
    future2 = asyncio.create_task(sync_async())

    await future1, future2

    print(future1.result() + future2.result())

asyncio.run(main())
# Takes 1 second to run

Same example with unsync:

@unsync
async def unsync_async():
    await asyncio.sleep(1)
    return 'I like decorators'

unfuture1 = unsync_async()
unfuture2 = unsync_async()
print(unfuture1.result() + unfuture2.result())
# Takes 1 second to run

Multi-threading an IO-bound function

Synchronous functions can be made to run asynchronously by executing them in a concurrent.ThreadPoolExecutor. This can be easily accomplished by marking the regular function @unsync.

@unsync
def non_async_function(seconds):
    time.sleep(seconds)
    return 'Run concurrently!'

start = time.time()
tasks = [non_async_function(0.1) for _ in range(10)]
print([task.result() for task in tasks])
print('Executed in {} seconds'.format(time.time() - start))

Which prints:

['Run concurrently!', 'Run concurrently!', ...]
Executed in 0.10807514190673828 seconds

Continuations

Using Unfuture.then chains asynchronous calls and returns an Unfuture that wraps both the source, and continuation. The continuation is invoked with the source Unfuture as the first argument. Continuations can be regular functions (which will execute synchronously), or @unsync functions.

@unsync
async def initiate(request):
    await asyncio.sleep(0.1)
    return request + 1

@unsync
async def process(task):
    await asyncio.sleep(0.1)
    return task.result() * 2

start = time.time()
print(initiate(3).then(process).result())
print('Executed in {} seconds'.format(time.time() - start))

Which prints:

8
Executed in 0.20314741134643555 seconds

Mixing methods

We'll start by converting a regular synchronous function into a threaded Unfuture which will begin our request.

@unsync
def non_async_function(num):
    time.sleep(0.1)
    return num, num + 1

We may want to refine the result in another function, so we define the following continuation.

@unsync
async def result_continuation(task):
    await asyncio.sleep(0.1)
    num, res = task.result()
    return num, res * 2

We then aggregate all the results into a single dictionary in an async function.

@unsync
async def result_processor(tasks):
    output = {}
    for task in tasks:
        num, res = await task
        output[num] = res
    return output

Executing the full chain of non_async_functionresult_continuationresult_processor would look like:

start = time.time()
print(result_processor([non_async_function(i).then(result_continuation) for i in range(10)]).result())
print('Executed in {} seconds'.format(time.time() - start))

Which prints:

{0: 2, 1: 4, 2: 6, 3: 8, 4: 10, 5: 12, 6: 14, 7: 16, 8: 18, 9: 20}
Executed in 0.22115683555603027 seconds

Preserving typing

As far as we know it is not possible to change the return type of a method or function using a decorator. Therefore, we need a workaround to properly use IntelliSense. You have three options in general:

  1. Ignore type warnings.

  2. Use a suppression statement where you reach the type warning.

    A. When defining the unsynced method by changing the return type to an Unfuture.

    B. When using the unsynced method.

  3. Wrap the function without a decorator. Example:

    def function_name(x: str) -> Unfuture[str]:
        async_method = unsync(__function_name_synced)
        return async_method(x)
    
    def __function_name_synced(x: str) -> str:
        return x + 'a'
    
    future_result = function_name('b')
    self.assertEqual('ba', future_result.result())

Custom Event Loops

In order to use custom event loops, be sure to set the event loop policy before calling any @unsync methods. For example, to use uvloop simply:

import unsync
import uvloop

@unsync
async def main():
    # Main entry-point.
    ...

uvloop.install() # Equivalent to asyncio.set_event_loop_policy(EventLoopPolicy())
main()