Embed Server Performance Considerations

[adeelsohailahmed]

Thank you for this project, @gi0baro. I'm new to Granian and interested in trying it out.

I currently have an existing project that is served via Uvicorn programmatically. To keep changes minimal, I've simply swapped Uvicorn with Embed Granian Server (minimal example for brevity).

import asyncio

from granian.constants import Interfaces
from granian.server.embed import Server

from src.app import get_app
from src.settings import get_settings


async def run_server() -> None:
    app = get_app()
    settings = get_settings()

    server = Server(
        target=app,
        port=settings.port,
        address=settings.host,
        interface=Interfaces.ASGI,
        log_access=True,
    )

    await server.serve()


if __name__ == "__main__":
    asyncio.run(run_server())

I wanted to ask if there are any other disadvantages to using Granian this way besides the ones you listed here in #35

I'm fine with the 1-worker limitation, since I'm already running it in a Docker container. The other two (existing & running event loop requiring and no support for reloading) are not an issue for me either.

I did notice that it takes about 10 seconds to shut down the server (initiated by docker compose down) about every single time. Is using the embed server the reason behind it? It's a not a deal breaker per se, but it is indeed a bit annoying. 😅

[gi0baro]

First things first: the embedded server is still marked as experimental. So, while 10s still seems odd, it might be because the shutdown process is not at its best in the embedded server.

More importantly: the embedded server is not designed for that. The idea behind the embedded server is.. well to be embedded. In projects that require more fine-grained control on the bootstrap and teardown phases. Or in general regarding processes management. In short: is designed for pojects like gradio, where they need to also manage a bunch of other stuff within the same process in addition to APIs serving.

So, given your specific use case, I strongly advise you to just switch to CLI – even if performance-wise should be the same.
Especially given you run it inside containers. I don't know what get_settings does, but given you only get host and ports from that, you can just switch over GRANIAN_ env variables to achieve the same result, and you don't need to write Python code for that.

Isn't granian --interface asgi --factory --access-log src.app:get_app as the entrypoint just cleaner?

[adeelsohailahmed]

Thank you for your response! I'll update my code to use the CLI.