the bug is reproduced on big multiprocessing.Process output
when the size of the output gets bigger than signed int the value becomes negative, thus
```            
assert left > 0
```
in multiprocessing/connection.py:337 raises an exception like the following

```
Traceback (most recent call last):
  File "D:\GitHub\cpython\lib\threading.py", line 917, in _bootstrap_inner
    self.run()
  File "D:\GitHub\cpython\lib\threading.py", line 865, in run
    self._target(*self._args, **self._kwargs)
  File "D:\GitHub\cpython\lib\multiprocessing\pool.py", line 470, in _handle_results
    task = get()
  File "D:\GitHub\cpython\lib\multiprocessing\connection.py", line 250, in recv
    buf = self._recv_bytes()
  File "D:\GitHub\cpython\lib\multiprocessing\connection.py", line 318, in _recv_bytes
    return self._get_more_data(ov, maxsize)
  File "D:\GitHub\cpython\lib\multiprocessing\connection.py", line 337, in _get_more_data
    assert left > 0
AssertionError
```


this assert looks invalid in this case because in C code the left value is DWORD (unsigned long), which cannot be negative by definition.

By bigger than signed int i meant bigger than positive signed int

New changeset 266f4904a222a784080e29aad0916849e507515d by Victor Stinner (Alexander Buchkovsky) in branch 'master':
bpo-34563: Fix for invalid assert on big output of multiprocessing.Process (GH-9027)
https://github.com/python/cpython/commit/266f4904a222a784080e29aad0916849e507515d

New changeset e8ca8806a9f47b3bac4ae1c6b8a54c47d1aad8f3 by Victor Stinner (Miss Islington (bot)) in branch '3.7':
bpo-34563: Fix for invalid assert on big output of multiprocessing.Process (GH-9027) (GH-9064)
https://github.com/python/cpython/commit/e8ca8806a9f47b3bac4ae1c6b8a54c47d1aad8f3

New changeset 23f427a0fdb888212136cf8745a9f5f832a3f374 by Victor Stinner (Alexander Buchkovsky) in branch '3.6':
[3.6] bpo-34563: Fix for invalid assert on big output of multiprocessing.Process (GH-9027) (GH-9069)
https://github.com/python/cpython/commit/23f427a0fdb888212136cf8745a9f5f832a3f374

Thanks Alexander Buchkovsky for the fix: it has been merged into 3.6, 3.7 and master branches. Thanks Oleksandr Buchkovskyi for the bug report!

--

I looked at the Python 2.7 code and it doesn't look to be impacted by this bug.

Modules/_multiprocessing/pipe_connection.c uses ReadFile() and PeekNamedPipe(), but the result type of the two C functions using it is not a PyObject* and conn_recv_string() looks good to me. Moreover, the type of the buflength parameter of conn_recv_string() is size_t, so there is no signed issue.

thank you for all the help in merging this fix!