I modified test_warn_on_full_buffer() to fill the socket pair using chunks of 64 KiB instead of sending b'x' byte per byte, just to make the test faster.

Using two terminals, I ran "python -m test test_signal -m test_warn_on_full_buffer -F": I reproduced the bug in 2 minutes.

Ok. The problem is in that is that the code responsible to full the socketpair is not reliable. *Sometimes*, send() blocks, whereas if I retry *immediately*, the send() complete (the socketpair is not full).

The issue seems to be related to the test_socket() failure where send() is "asynchronous".