I tested on python3.4.2, windows 7

1. Only proactorEvent issued.
2. ThreadPoolExecutor has many workers (in the attached example file, worker count 20,000)
3. The loop does run_in_executor more call than worker count (in the attached example file, 40,000 calls)
4. After some random seconds, raise BlockingIOError

BlockingIOError: [WinError 10035] A non-blocking socket operation could not be completed immediately.
exception calling callback for <Future at 0x1ab89ef0 state=finished returned NoneType>
Traceback (most recent call last):
  File "c:\Python342\Lib\concurrent\futures\_base.py", line 297, in _invoke_callbacks
    callback(self)
  File "c:\Python342\Lib\asyncio\futures.py", line 410, in <lambda>
    new_future._copy_state, fut))
  File "c:\Python342\Lib\asyncio\base_events.py", line 403, in call_soon_threadsafe
    self._write_to_self()
  File "c:\Python342\Lib\asyncio\proactor_events.py", line 449, in _write_to_self
    self._csock.send(b'\0')

I guess that proactor's _write_to_self method misses exception handle.

proactor_events.py
def _write_to_self(self):
self._csock.send(b'\0')

selector_events.py
def _write_to_self(self):
# This may be called from a different thread, possibly after
# _close_self_pipe() has been called or even while it is
# running. Guard for self._csock being None or closed. When
# a socket is closed, send() raises OSError (with errno set to
# EBADF, but let's not rely on the exact error code).
csock = self._csock
if csock is not None:
try:
csock.send(b'\0')
except OSError:
if self._debug:
logger.debug("Fail to write a null byte into the "
"self-pipe socket",
exc_info=True)

Ps: It's my first publication. Hope you understand my poor comment..

Why are you using 20000 threads?

--Guido (mobile)
On Dec 22, 2015 02:03, "Joseph Gordon" <report@bugs.python.org> wrote:

Changes by Joseph Gordon <j.gordon.matthew@gmail.com>:

----------
nosy: +josephgordon

---

Python tracker <report@bugs.python.org>
<http://bugs.python.org/issue23846\>

---

Why are you using 20000 threads?

That's a good question.

In any case it looks like self-pipe sock's buffer was overflown because call_soon_threadsafe was called too many times, and loop._read_from_self couldn't empty the buffer promptly. Then, at some point, _write_to_self failed with an IOError.

It looks like this was fixed for the selector loop, but not for proactor:

selector_events.py:

    def _write_to_self(self):
        csock = self._csock
        if csock is not None:
            try:
                csock.send(b'\0')
            except OSError:
                if self._debug:
                    logger.debug("Fail to write a null byte into the "
                                 "self-pipe socket",
                                 exc_info=True)

proactor_events.py:

    def _write_to_self(self):
        self._csock.send(b'\0')

New changeset c9f872b by Victor Stinner in branch 'master':
bpo-23846: Fix ProactorEventLoop._write_to_self() (GH-11566)
c9f872b

New changeset c9f2671 by Miss Islington (bot) in branch '3.7':
bpo-23846: Fix ProactorEventLoop._write_to_self() (GH-11566)
c9f2671

In any case it looks like self-pipe sock's buffer was overflown because call_soon_threadsafe was called too many times, and loop._read_from_self couldn't empty the buffer promptly. Then, at some point, _write_to_self failed with an IOError.

I fixed the issue. Thanks for your bug report ;-)