Upload a wheel to PyPI

[ghickman]

You can achieve this by installing the wheel package and running python setup.py register bdist_wheel upload in your natsort directory. It can also be combined with your existing source distribution uploads like so python setup.py register sdist bdist_wheel upload.

This change makes the generated wheel compatible with both python 2 and 3.

[SethMMorton]

Done (commit f3efab7)

[ghickman]

@SethMMorton – Thank you!

[SethMMorton]

@ghickman Can you explain the benefit of having a wheel distribution over just zip and tar.gz for a pure python module?

[ghickman]

@SethMMorton – they remove the need to run a setup.py on installation (which runs two subprocesses) and this has the knock on effect of improving install time.

You can find more information here and here.

[SethMMorton]

@ghickman That's all well and good, but installing via a wheel does not install entry points, as I mention in issue #9. This is a problem, and I don't want to recommend an install method to my users that won't install all natsort components.

Is the install time of natsort so significant from source that a wheel is necessary?

[ghickman]

@SethMMorton – That's odd, AFAIK entry_points shouldn't be affected by wheeling a package. I've been using wheels with one of my own apps for a while now and not seen any issues with it's entry_points (its main usage). Also pip does the same thing and rolls a wheel which correctly installs it's entry points.

re: install speed - for a single package it doesn't make a huge difference but with a reasonable sized requirements.txt it can, especially when you look at install times for projects on CI services such as Travis-CI.

[SethMMorton]

@ghickman If you install natsort from a wheel, does the entry point get installed?

About install speed, I'm sure for a large project a wheel is much better, but for natsort, where there is only 5 files and one dependency I can't imagine it makes a difference. I'm not saying the wheel format isn't good, I'm just saying that in this specific case I'm not sure it makes a difference.

[ghickman]

@SethMMorton – Yes, tested with python 2.7 and 3.4.1 and running natsort foo3 foo1 produces the output foo1 foo3. What error are you seeing with the entry point/wheel combination?

---

Yes it won't make a noticeable difference when you're installing only natsort, however when natsort is part of said large projects dependencies then having as many of those dependencies wheeled as possible makes a huge difference. This becomes especially noticeable when installing somewhere like Travis-CI which installs a projects dependencies on every run and you end up with install times of multiple minutes for a few seconds of tests.

[SethMMorton]

Are you using virtualenv? I am not, and according to what I have seen entry points are only installed when you are in a virtualenv. Can you try in your global python environment? If a user wants to install natsort for only its shell script, I don't want to force them to use virtualenv.

---

I see now. You get better performance when installing dependencies when all of them are wheels.

[ghickman]

Yes I am but my main python is homebrewed so not easy to check my system one. That's annoying to find out though! Have you raised an issue with pip? I guess for now you'll have to drop the wheels.

[SethMMorton]

I also use a homebrewed python, just not virtualenvs. When I install natsort via a wheel to my homebrewed python location, the entry point is not installed.

Either way, I'll still upload a wheel, but in the docs I advise against it if you want the shell script. I was hoping you knew of some secret sauce to get the entry points to work properly. :(

[ghickman]

@SethMMorton – I've installed with pip 1.5.6 to a homebrewed Python 2.7.8 environment without the use of a virtualenv and the entry points script is working for me by the look of it: gist

Is there anything else I could test? Happy to help find the cause of this!

[SethMMorton]

There it is! I was still using pip 1.4.1, and you were using pip 1.5.6. When I upgraded my pip version, it worked! It seems that somewhere between these two versions, they fixed this bug.

Thanks for your help!

[ghickman]

@SethMMorton Awesome! Always happy to help and thank you for putting the time into natsort.