python - How to use asyncio with existing blocking library?

Question

I have few blocking functions foo, bar and I can't change those (Some internal library I don't control. Talks to one or more network services). How do I use it as async?. E.g. I wan't to do the following.

results = []
for inp in inps:
    val = foo(inp)
    result = bar(val)
    results.append(result)

This will be inefficient as I can call foo for the second input while I am waiting for the first and same for bar. How do I wrap them such that they are usable with asyncio (i.e new async, await syntax)?

Lets assume the functions are re-entrant. i.e it is fine to call foo again when already a previous foo is processing.

---

Update

Extending answer with reusable decorator. Click here for example.

def run_in_executor(f):
    @functools.wraps(f)
    def inner(*args, **kwargs):
        loop = asyncio.get_running_loop()
        return loop.run_in_executor(None, functools.partial(f, *args, **kwargs))

    return inner

Answer 1

There are (sort of) two questions here: first, how to run blocking code asynchronously, and second, how to run async code concurrently (asyncio is single-threaded, so the GIL still applies, so it isn't truly parallel, but I digress).

Concurrent tasks can be created using asyncio.ensure_future, as documented here.

To run synchronous code, you will need to run the blocking code in an executor. Example:

import concurrent.futures
import asyncio
import time

def blocking(delay):
    time.sleep(delay)
    print('Completed.')

async def non_blocking(loop, executor):
    # Run three of the blocking tasks concurrently. asyncio.wait will
    # automatically wrap these in Tasks. If you want explicit access
    # to the tasks themselves, use asyncio.ensure_future, or add a
    # "done, pending = asyncio.wait..." assignment
    await asyncio.wait(
        fs={
            # Returns after delay=12 seconds
            loop.run_in_executor(executor, blocking, 12),
            
            # Returns after delay=14 seconds
            loop.run_in_executor(executor, blocking, 14),
            
            # Returns after delay=16 seconds
            loop.run_in_executor(executor, blocking, 16)
        },
        return_when=asyncio.ALL_COMPLETED
    )

loop = asyncio.get_event_loop()
executor = concurrent.futures.ThreadPoolExecutor(max_workers=5)
loop.run_until_complete(non_blocking(loop, executor))

If you want to schedule these tasks using a for loop (as in your example), you have several different strategies, but the underlying approach is to schedule the tasks using the for loop (or list comprehension, etc), await them with asyncio.wait, and then retrieve the results. Example:

done, pending = await asyncio.wait(
    fs=[loop.run_in_executor(executor, blocking_foo, *args) for args in inps],
    return_when=asyncio.ALL_COMPLETED
)

# Note that any errors raise during the above will be raised here; to
# handle errors you will need to call task.exception() and check if it
# is not None before calling task.result()
results = [task.result() for task in done]