AIP-84 Convert async route to sync routes

[pierrejeambrun]

As discussed in #43718 (comment), routes with blocking I/O code should be sync to not block main event loop. (db access, disk read, network call, etc...)

More information in FastAPI documentation https://fastapi.tiangolo.com/async/#path-operation-functions.

This PR converts all of the endpoints, all that can me converted to async def when we implement full async support.

(I recall Maybe one or two endpoints that are purely in memory and could stay async but I didn't bother to make an exception for them)

[bbovenzi]

I see three endpoints still using async. Do we want to change those too?

[pierrejeambrun]

Good catch,, I forgot the private API.

Updated thanks Brent.

[omkar-foss]

Hey @pierrejeambrun, I've one concern on this - if we use sync path funcs, the FastAPI requests will run in a threadpool (one request per thread), consuming more memory per request and will limit throughput on our new APIs, as there's a default limit of 40 threads on the threadpool, please see: Kludex/starlette#1724 (context: AnyIO is now used for async IO by Starlette, which in turn is used by FastAPI to handle http requests).

So, just a suggestion - instead of changing all the path funcs from async to sync, it'll be great if wrap the blocking (sync) function calls inside the path funcs to make them async, using asyncio.to_thread or asyncify or similar.

[dolfinus]

Hm, I don't get why this code:

@route.get(...)defhandler(...):
something=session.get(...)
other=session.select(...)
awaitasyncio.to_thread(handler)

does limit thoughput, but this doesn't:

@route.get(...)asyncdefhandler(...):
something=awaitasyncio.to_thread(session.get, ...)
other=awaitasyncio.to_thread(session.select, ...)

Could you please elaborate?

[pierrejeambrun]

I agree with @dolfinus, running in a separate thread manually or leveraging FastAPI to do so is more or less the same. (just less work and more code maintainability to let FastAPI handle that).

Long term we will rewrite that with full async support, in the meantime FastAPI is just sync for us.

[omkar-foss]

Hey @dolfinus, yes, the throughput in this case would be very similar for both snippets, because when calling asyncio.to_thread() (executor unspecified), the default thread pool executor will be used and be subject to same 40 thread limit. But unlike sync functions where entire request is processed in a thread, in async we'll have control over what should be processed in a separate thread.

Example: As our new APIs have a mix of CPU-bound (e.g. common params resolution, data checks, Pydantic validations, etc.) and IO-bound (e.g. DB queries, network calls) activities, I guess we could tune it something like this:

@route.get(...)asyncdefhandler(...):
# CPU boundifsome_check:
raiseHTTPException(status.HTTP_404_NOT_FOUND, "Not Found")
# IO bound, sent to it's own threadsomething=awaitasyncio.to_thread(session.get, ...)
# CPU bound againreturnSomePydanticModel(something)

We could alternatively use run_in_threadpool instead of asyncio.to_thread which is provided by FastAPI, example here. Either way we go, it'll need thorough testing to understand what's working for us in terms of performance! :)

[omkar-foss]

Long term we will rewrite that with full async support, in the meantime FastAPI is just sync for us.

That would be lovely, thank you! :)

[pierrejeambrun]

I think that would be a lot of work to maintain + code becomes hard to read + 1 mistake (someone forget to manually put into the threadpool a blocking IO call) and then the main event loop is blocked...

I think we can start like that, and if it's not enough we can go deeper into the fine tuning of what is executed in the main even loop and what is run in a separate thread. I believe CPU bound operations run in a separate thread won't bottleneck. (And if they do, most likely the main even loop would struggle too, so we would have another problem here)

[omkar-foss]

Yes sure, sounds good! Thanks @pierrejeambrun 👍🏽