We'll need tests for this new functionality. I think thatmax_connections functionality is neat, but what's the use case again?

@1st1max_connections just says stop accepting new connections when the current active connection count isnumber. I have a proxy service that handles statsd and diamond data streams from thousands of clients, then adds some metadata and forwards to a wavefront agent. The ingress rate often exceeds the egress potential because of outside factors. Without a connection limit the server happily accepts new connections even while its egress rate plummets due to resource contention. My initial attempt to control backpressure with a semaphore on the recv side of the proxy wasn't good enough because I can't tell the event loop to stop accepting new connections. It only takes a few seconds to eat gigabytes of memory under the right conditions.

But what criteria do you use to pick up a sound number formax_connections? Is it just some arbitrary number, or there is a way to calculate an optimal one for the given system?

I'd expect it to be mostly application dependant. It would obviously need to be less than thenofile ulimit for the user running the process and then under the kernel's max open file limit and finally it needs to be a number that prevents memory exhaustion. The latter of which is going to vary depending on the memory footprint of each connection in a user's application.

In my particular case settingmax_connections to 5000 puts my memory ceiling at ~ 100MB. But I actually get better performance with a lower number. Setting it too low however, can cause degraded service to clients. There are tradeoffs everywhere, hence a default of None (no limit).

I'm writing some tests btw, but before I get too far, are you generally okay with the location of the functionality and the way the start_serving arguments are stored inServer? I frankly couldn't convince myself that I was using idiomatic conventions with that stuff. It's not clear to me when a user should interact with a Server object vs using event loop functions since most behavior is implemented in the event loop and the Server object seems to be just a placeholder for sockets and cleanup.

I'm writing some tests btw, but before I get too far, are you generally okay with the location of the functionality and the way the start_serving arguments are stored in Server?

I think that the way the patch is stuctured now is fine. But before you get too far, let's wait until@gvanrossum gives this idea a green light.

Should I be evangelizing this? I'd like to see it get accepted.@gvanrossum has publicly stated he's not focused on asyncio as much as he used to be, so perhaps there are other stakeholders we can bother?

@gvanrossum What do you think on this one? 3.7?

Definitely 3.7. Until then this could easily be implemented as a separate AbstractServer subclass.

3.7 is fine, let me know how I can help or what I should do, if anything.

Re: subclassing AbstractServer, I believe I brought this up in the forum in the context of allowing a user to bring their own Server subclass tostart_server but it was shot down. As it's architected now, I don't see how you can bring your own Server subclass without copypasta. But I may have missed something important.

Re: subclassing AbstractServer, I believe I brought this up in the forum in the context of allowing a user to bring their own Server subclass to start_server but it was shot down. As it's architected now, I don't see how you can bring your own Server subclass without copypasta. But I may have missed something important.

@gvanrossum This is something we can still do in 3.6, right? IIUC we can add aserver_factory parameter toloop.create_server?

Please reopen this PR athttps://github.com/python/cpython. This repo is going to be deleted soon.