Multiple Listeners: Why not multiple axum::serve ?

[moranbw]

Summary

Mostly self explanatory from title...I was looking at the example for spawning multiple listeners here: https://github.com/tokio-rs/axum/blob/main/examples/listen-multiple-addrs/src/main.rs

This approach seems to drop down to using hyper_util. But why not just spawn multiple instances of axum::serve?

This seems like an easier to understand pattern...but perhaps there are performance tradeoffs?

axum version

0.7.7

[jplatte]

That's a very good question. The discussion at #1063 that led to the example being created doesn't really answer it 😅

Back then with hyper::Server it looked like it could be more optimal to not spawn separate accept loop tasks, now I'm very skeptical. It might even be better to have two separate axum::serve tasks.

I'd be inclined to just delete the example. @mladedav thoughts?

[mladedav]

In an ideal world, I think it would be better to have it in one serve just so you can't add one later to one serve and forget to add it for althe other one. Unless that's what you want of course.

But I don't see practical issues with multiple concurrent serve calls. Well one that I can think of is that if a user fails to set non-blocking mode on ther TCP listeners then they'll be surprised that one of the servers doesn't seem to work at all. With single listener they'll blocking but it should just work by accident.

I think it should be just fine to get rid of the example since now it's needlessly complicated and the way to do this soon-ish will probably be to implement a Listener in 0.8 to achieve the same thing.

[moranbw]

@mladedav isn't tokio::net::TcpListener always non-blocking? Would this concern only apply if you are using std::net::TcpListener?

[jplatte]

It's unfortunately not, see TcpListener::from_std. But I don't think we really need to concern ourselves with that, it's not an axum-specific problem in any way.

Really the question is then: Should we update the example once #2941 is merged, or delete it instead. Even with a custom Listener, I'm not sure it's something we should promote. Or is there a reason you would consider that better, @mladedav? (excluding buggy user code)

[moranbw]

Sorry -- don't want to sidetrack from the actual conversation developing here...but @jplatte you again seemed to reference a TcpListener that is coming from std::net::TcpListener?

What if you are just creating a tokio::net::TcpListener like this, is this blocking or no:

use tokio::net::TcpListener;

use std::io;

#[tokio::main]
async fn main() -> io::Result<()> {
    let listener = TcpListener::bind("127.0.0.1:2345").await?;

    // use the listener

    Ok(())
}

Sorry if this is a naive question.

[mladedav]

Nope, buggy user code is the only concern I have. We can delete it.