Nasty Python Monstrosities

Why I’m not going to distribute any Python ever again.

Wednesday, February 4, 2015 · 4 min read

Scenario: I’ve written a simple Python command-line tool. It consists of more than one file, because separation of code is important. It only needs to run on UNIX systems and shouldn’t require sudo to install. How do I distribute this code in a safe, user-friendly way? How do I teach my intro CS class how to proudly share their creations with the world?

Answer: You don’t.

Distributing Python code should be a solved problem. It’s not. Or, rather, it’s a standards problem. There are too many tools out there, and there’s no clear roadmap that explains which subset does what. I’m talking about Pip, Virtualenv, PyPi, wheels, eggs, distutils, distribute, setuptools, easy_install, and twine. Some of these have been incorporated into others, and are therefore obsolete. Some of these are endorsed by a PEP. (What’s a PEP? Do I care?)

There is certainly not one obvious way to do it. Even if you’re Dutch.

Assuming you have picked out a subset of tools to use, though, there’s no guarantee that they will cooperate. A Pip flag, for instance, might work. It might not work. It might do stuff that you didn’t expect. It might create files without telling you (correction: it’ll be noted cryptically in the Pip log). It might pass it on to setup.py, or it might not, or it might reject a valid setup.py argument because it doesn’t understand it, or it might pass it on to Python erroneously.

My Pip log contains monstrosities such as:

(Enumeration of all the things here that make me cry is left as an exercise to the reader.)

Running command /usr/bin/python -c "import setuptools, tokenize;
__file__='/path/to/setup.py'; exec(compile(getattr(tokenize, 'open',
open)(__file__).read().replace('\r\n', '\n'), __file__, 'exec'))" develop
--no-deps --user

In short, it might eat your laundry depending on local atmospheric conditions, and then it’ll flip a coin to decide whether or not to tell you.


Let’s talk about publishing. PyPi is the “standard” repository. But it makes you host your own tarballs, so essentially all it does is match package names to URLs (all other information is duplicated in the tarball itself).

Uploading to PyPi happens over an insecure connection. They recognize this in the docs, and recommend installing a separate client that lets you publish over SSL.

They also keep telling users to use OpenID. The OpenID that Google is deprecating in a couple of months. Of course, trying to log in with Google throws an internal server error.

Apparently, the whole system is so confusing that every time you upload, you are encouraged to do a dry run on the test repository just to make sure you understand it.

Now, my Python program was not a module, but a command-line tool. The way you specify a command-line tool in Python looks like this:

# ...
entry_points={
    'console_scripts': ['name_to_link = packagename.submodule:function_name'],
},
# ...

Here’s what setuptools does. It parses that monstrous string and extracts information. It then creates a new Python script in a $PATH‘d directory. That script imports your script as a module, and then calls the function you specify. It’s anyone’s guess whether important scriptey things such as environment variables or command-line arguments are preserved.

This is all in complete ignorance of the whole if __name__ == '__main__' infrastructure, for whatever reason. Of course, you can also make things runnable with python -m packagename.submodule by including a __main__ file. Yum.

Even if it did the reasonable thing and symlinked your script directly, it would actually modify your file. That’s right. It replaces shebang lines with the path to the currently-running Python interpreter. Because Explicit Is Better Than Implicit (tm).

There is no reason to do this. UNIX provides a very helpful idiom: using #!/usr/bin/env python arguments.... That’ll search the $PATH for the right interpreter and use it. No black magic needed, and you can provide flags to the python interpreter if you so desire.

Only God—make that Guido—knows what’ll happen if you’re trying to install a Python 3 script using a Python 2.7. Or want to distribute something runnable with ipython.


By the way, if you think installing is bad: you would think uninstalling should be a standardized, well-thought-out, documented process, right? You wish. There setup.py-installed programs have no fixed uninstallation procedure. If you Google around, you are led to this highly-upvoted StackOverflow answer, which suggests using an option that lists all generated files, and piping the output to xargs rm -rf.

Let’s hope they don’t have a file with a space in it, or bad things are going to happen. Keep in mind that Python is recommended for newbies, the kinds who will gladly copy-and-paste shell commands from the Internet.


Here’s the deal. Flat is not better than nested. Nested is better than flat. There’s a reason we use parsers to convert strings to abstract syntax trees. There’s a reason LISP is easier to write than x86 assembly. There’s a reason for the existence of the term “dependency tree”. They are literally nested structures. Importing from a sibling directory should not be the harrowing experience it is in Python.

And that’s why I love npm. Not because it’s written in a beautiful language or because it pretends to be easy-to-use, but because it’s simple and elegant and it doesn’t do too much or too little. Modules are nested prettily in their directories, and the only file npm touches is package.json.

Scripts are simply linked into npm’s bin folder. Your “script” could be a Bash program, for all it cares.

Installing is one command, uninstalling is one command, publishing is one command, and everything is one tool operating on one repository.

npm is transparent. Python’s plethora of packaging plakavacs are anything but.

◊ ◊ ◊