https://bugs.python.org/issue27906 describes a situation where accept rate of connection is too slow for use cases where spikes of incoming connection warrants fast accept. The fix for that was to accept socket connection in a tight loop until it reaches "backlog" connections.

This doesn't work very well in a web server scenario where we have many processes listening on the same socket. Each process should not accept up to "backlog" connections, for better load balancing among processes. It would be ideal if this is a separate argument for the server configuration so that the application can decide up to how many connections it is willing to accept in the loop, independent of the backlog parameter for listen() system call.

Let me know if this makes sense.

Lisa

In principle this sounds good, but I'm not sure how the API and implementation would look like. Maybe you can share a bit more details?

One rough idea would be like this: https://github.com/python/cpython/compare/master...lguo2020:fix-issue-33811?expand=1. Another option is to associate it with the loop:

loop.set_max_accept(2)

and then later

self._loop._start_serving(....., max_accept=self._loop._max_accept)