Note that this tutorial assumes basic understanding of Runhouse Functions & Modules; it is recommended that you check out our functions and modules tutorial before diving into this one.
As we’ve discussed before, once you take a Python function or module and send it to a Runhouse cluster, the cluster holds your resource in memory, and each time that function or module is called by a client, it simply accesses it in memory and calls it. Under the hood, we have a fully asynchronous server running (FastAPI), and a separate process for each environment where your Runhouse resources live. These processes all have their own async event loops, and if you run synchronous functions on Runhouse, they are ran in a separate thread to allow for many concurrent calls to the cluster. Note that if you are unfamiliar with asynchronous programming in Python, you should just continue using standard, Python sync functions and leave the performance to us.
But, what if you’re writing code that leverages Python’s powerful asynchronous functionality? Luckily, we provide rich async support in a variety of ways. First off, any function that is labeled with Python’s async keyword, when sent to a Runhouse cluster, will be executed within the environment processes’s async event loop, and not in a separate thread. This means that you should be very careful that you are not running any costly, synchronous code within an async function, to avoid blocking up your the event loop within your environment on the server. Poorly written async functions will not block the entire Runhouse daemon, but will block other functions within the same environment as the user code.
Client side, you also need to await
a call to this function the same
way you would if the function was running locally. Let’s check out an
example. First, we’ll start a local Runhouse daemon to mess with:
! runhouse server restart
Then, we’ll define a simple async
function to send to Runhouse:
async def async_test(time_to_sleep: int): import asyncio await asyncio.sleep(time_to_sleep) return time_to_sleep
We can send this to Runhouse the same way we would any other Runhouse function:
import runhouse as rh async_test_fn_remote = rh.function(async_test).to(rh.here)
INFO | 2024-04-30 18:50:35.023995 | Because this function is defined in a notebook, writing it out to a file to make it importable. Please make sure the function does not rely on any local variables, including imports (which should be moved inside the function body). Functions defined in Python files can be used normally.
INFO | 2024-04-30 18:50:35.060478 | Sending module async_test of type <class 'runhouse.resources.functions.function.Function'> to local Runhouse daemon
Then, we can call this function as we would if it were a local async
function. The network call to the remote cluster will execute
asynchronously within our local event loop (our code backed by
httpx.AsyncClient
) and the async function itself will execute within
the async event loop on the remote server.
await async_test_fn_remote(2)
2
Voila! Async functions are supported the way you’d expect them to be. There are a few other advanced cases, too:
There’s another important case that we support. Let’s say that your standard, synchronous functions are running on a remote Runhouse machine. When you call them from your local machine, there is inevitably network I/O involved in communicating with the cluster. You may want to not have your code block on this network call (for example if the function takes a long time to execute), so that you can avoid blocking your local Python code. You can choose to run this function asynchronously, locally, and this allows you to get back a coroutine from Runhouse, that you can then use to check if Note that this means your local code will have to use async primitives, even though it is calling what you defined as a sync function. Let’s check out an example of this:
def synchronous_sleep(time_to_sleep: int): import time time.sleep(time_to_sleep) return time_to_sleep sync_sleep_fn_remote = rh.function(synchronous_sleep).to(rh.here)
INFO | 2024-04-30 18:57:00.533012 | Because this function is defined in a notebook, writing it out to a file to make it importable. Please make sure the function does not rely on any local variables, including imports (which should be moved inside the function body). Functions defined in Python files can be used normally.
INFO | 2024-04-30 18:57:00.577673 | Sending module synchronous_sleep of type <class 'runhouse.resources.functions.function.Function'> to local Runhouse daemon
We can now call this function with the run_async
argument set to to
True
. This makes it not actually run locally immediately, and
instead returns a coroutine that you’d await, as if this function were
asynchronous. Note that, in your environment on your Runhouse cluster,
the functions runs in a thread, but the call to it locally is
asynchronous, and uses httpx.AsyncClient
.
await sync_sleep_fn_remote(2, run_async=True)
2
You could also use asyncio.create_task()
to not block your code on
the execution and then await
it when you want the result. When using
a function defined as async or a sync function with run_async=True
,
you always get back a coroutine, which you can do with what you please.
If I wanted, I could still call this function as a fully synchronous function:
sync_sleep_fn_remote(2)
2
The third critical case that we support is mostly applicable when you’re writing async code for the purpose of running it on the Runhouse cluster, but want to make synchronous calls to the server. The reason for you writing async code to run on the server is because our Runhouse server uses ASGI and runs everything asynchronously, so you can take advantage of the performance gains that come along with async code, but call it locally as you would a normal client calling a normal server, unaware of the backend implementation of the server. We can take the same async function I defined earlier and call it synchronously:
async_test_fn_remote(2, run_async=False)
2
That’s all there is to it! We’ve tried our hardest to make working with async code seamless from a user’s perspective. There are other edge cases we’ve put time into supporting and we’re happy to discuss architecture anytime – feel free to file an issue on Github or join us on Discord to discuss more!