set close_fds=False when starting kernels on Windows

[minrk]

Python 3.7 sets close_fds=True by default, closing the interrupt/parent handles we are trying to pass to the kernel.

We can be more explicit and use the new STARTUPINFO.lpAttributeList (the reason for the change in default behavior) in the future to explicitly inherit only these handles, but this preserves the pre-3.7 behavior on 3.7, getting kernels back to working on py37.

closes ipython/ipykernel#374

cf spyder-ide/spyder#8013

[CAM-Gerlach]

Per @ccordoba12 's request, I tried to test this on Windows but I got a huge chain of exceptions trying to install it locally, starting with AttributeError: _DistInfoDistribution__dep_map and ending with pip._vendor.pkg_resources.RequirementParseError: Parse error at "'; extra '": Expected stringEnd, which occured even after I uninstalled the jupyter_client version in the conda env I was using for testing and updated pip and setuptools to the latest versions.

[minrk]

@CAM-Gerlach that sounds like conda-forge/testpath-feedstock#7 which I believe is fixed in the latest conda package of testpath, so I think conda update testpath should fix it.

[CAM-Gerlach]

@minrk Thanks! That did indeed resolve the problem (I'd assumed it was just some machine or OS-specific issue with pip), and testing the fix this does indeed resolve the stated problem. Ctrl-C still doesn't properly interrupt execution on Windows under QtConsole (the interrupt only gets handled at the end, after the 10 seconds have elapsed) but it does register properly at least like on Python 3.6.

[minrk]

Yes, not being able to interrupt time.sleep promptly on Windows is a function of how interrupt_main works on Windows, and time.sleep is specifically not interruptible from other threads until control returns to the Python interpreter on that thread. Most realistic code returns to the Python interpreter with a reasonable frequency, so this is less often an issue in reality than when using time.sleep to test.

Thanks for confirming! We should do a release shortly.

[CAM-Gerlach]

Thanks @minrk ! I seem to recall reading something to that effect previously and I just chalked it up (like most conda commands taking 10x longer on native Windows than on my Fedora VM running inside of it) to Windows' many limitations.

[papagala]

Sorry, I have a basic question. Which file do I have to edit in order to set close_fds=False when starting kernels on Windows?

[CAM-Gerlach]

@papagala conda install jupyter_client=5.2.4 in your Spyder environment.