@vstinner the PR is what I meant when commented your PR for logging pipe overflow #11566

When you're done making the requested changes, leave the comment: I have made the requested changes; please review again.

warn_on_full_buffer argument was added in Python 3.7
Without it the PR is impossible.

I'll change the title and changelog text if we decide to land the change.
Sure, the PR needs a careful review, any feedback is very appreciated.

Let' me explain the motivation better:

1. Now we collapse all signals catched on single event loop iteration to handle every signal handler only once. I think it is good, Python itself handles signals asynchronously. There is no reason to call asyncio handler for particular subscribed signal twice on single iteration.
2. Storing raised signal numbers in a set allows to ignore errors if self pipe is overloaded. We need only one pipe read event to wake up and handle both pending async signals and cross-threads calls (loop.call_soon_threadsafe()).
3. Failures on writing into self pipe are signs for either high load or blocking/long-running code in callback handlers.
   The former is pretty normal, additional logging doesn't help too much; especially in debug mode (existing checks are skipped in release mode for sake of speed anyway). Most likely, users don't use debug mode for running production code. Moreover, logging the fact that the loop is under high load on every iteration makes the load even worse :)
   Blocking calls are reported by loop._run_once() in debug mode anyway; there is no need to double the message.