OK so: there are always personal, professional, stylistic, and inertial reasons to use one version of Python versus another. I’ll tell you briefly about how I personally see and use the macOS system Python, the Python.org Python, the various and sundry Homebrew Pythons, and a few others like Anaconda’s Python.
Fortunately, a given version of python – that is to say, a Python installation with a version number of x.y.z – will work the same as any given Python x.y.z install, provided they are both installed and invoked correctly. At least, as far as the core language and standard library goes… what packages each Python has available to it can and will vary greatly, unless you spend a lot of time ensuring that it doesn’t.
So with that caveat, let me explain. In the beginning, when I started programming in Python around 2005, I used the macOS system Python, which at the time was called the MacOS system Python (note the capitalization difference). The version was either 2.3 or 2.5, I can’t recall. The
pip installer was not yet a thing – Python packaging was a terrific topic under which one could start a fantastic array of flamewars on listservs, or early Reddit or Digg (as StackOverflow was also not yet bethingified, and Hacker News was a small core cadre of sycophants). The executable was located at
/usr/bin/python, and packages were installed to
/Library/Python/2.5/site-packages. I didn’t know anything about Python stuff – I didn’t know why
.egg files were terrible, or how native extensions were built and linked, or what the deal was with
setuptools… or, most egregiously why you never ever want to mess around with Apple system binaries.
I used this Python because it shipped with
PyObjC is a bridge between the Python and the Objective-C runtimes. It is very old (it has been in development since Steve Jobs was at NeXT), very thorough, and very vast – it’s a constellation of interdependant Python packages that, at the time, had to be compiled from source to be properly installed. Since I had a vested interest in using
PyObjC and no idea how to install it from scratch, I took to using the system Python.
I will spare you all of the details of my subsequent learning-the-hard-way about meddling with Apple system binaries, and also forgo a lengthy digression into how I moved from trashing my MacOS installation to using package managers: first Fink, then MacPorts, and then finally Homebrew – which mercifully was the least insane and most pragmatically malleable of all of these. Between the epoch of MacPorts and the Homebrew era, I actually swore off MacOS package management for a while, and spent a brief lacuna working with the Python.org distribution, the well-known ActiveState releases, my own sub-optimal builds, and many others.
By this time, it was about 2011, and I had learned a few things: how using Homebrew’s Python for development would keep the macOS system Python clean and pristine; how
virtualenvwrapper could sequester a project further, making it independent of the Python configuration; how testing and deployment with the Vagrant virtual-machine manager would keep everything in its right place – I have OCD, so these things all appealed to my innate organizational sense as well as my technical inclinations.
And the tools had grown as I had:
pip had established itself as the de-facto Python package-management standard. A victor emerged from the packaging wars:
setuptools was now being actively developed, instead of bickered over endlessly online. GitHub and
git were more and more central every day to my development workflow, as well as that of countless others. Homebrew used
git, and while it was written in Ruby, it was being developed by people who appreciated clean and legible code – if one could read the Homebrew source, one could hack on it.
This, then, is where I have come to stand on the various and sundry versions of Python that are available:
The macOS system Python – the Python install that ships with macOS
- version: 2.7.10 (as of Mojave 10.14.5)
- verdict: DO NOT USE. This Python is an old and out-of-date version from the endangered
2.x lineage; even if you were somehow willing to deal with being locked into an old and un-upgraeable programming environment, you’d have to also take care not to disturb this Python installation, such that any macOS updates or hotfixes would be able to count on it working exactly as Apple assumes it would if it were pristine. That means you also can’t upgrade the bundled
PyObjC packages, which at the time of writing are an appalingly ancient version, 2.5.1 (by contrast, I have
PyObjC version 5.2b1 installed into my Homebrew Python’s package library). In fact, any upgrades, changes, tweaks, modifications, or ∆-minutiae you impose could screw things up later when some lazily-written Apple package script runs on the system Python and acts slightly different than it did in Cupertino. So yeah I repeat: DO NOT USE.
The official Python.org Python package distribution
- version: latest stable, latest development RCs, and archives of prior releases
- binary: $framework +
- packages: $framework +
- verdict: Are you, or would you like to be, a core Python developer? Is contributing to the Python language and its central components – the parser, the interpreter, the bytecode generator, the peephole optimizer, the standard library, the “Argument Clinic” preprocessor, the Unicode algorithms, the POSIX wrappers – if you are still reading this and are specifically not bored, this is the Python package for you to use daily. Not to develop all of those non-boring things, no – you’ll of course want to have multiple
git repositories of the Python source compiling and running in VMs orchestrated with Docker and whatnot for all that – but you, being a so-called “language lawyer,” will nonetheless want to work with the most official Python you can.
The Anaconda Python distribution – from the
conda package manager
- version: 3.7.3 and 2.7.16, respectively, at the time of writing
- verdict: Are you a data scientist? Or is that something you say you are when introducing yourself at meetups and conventions? Either way, if you love data and hate all those icky compiler flags and configuration files and settings directory paths and all that other nonsense, you might just be in the market for Anaconda. Anaconda, through its
conda package-management CLU, is a whole ecosystem of data-science-y software – all of it easily and painlessly installable with just a few commands (like e.g.
conda install numpy,
conda install dlib)… Software originating from diverse ecosystems, with complex configurations and lengthy build-times, can all be installed easily in pre-tested, ready-to-run binary form through
conda. Anaconda’s Python distribution is one of the fundamental components of this system, and it will always be built with the most reasonable options and delivered within a minimal lag from its official release. You, the data scientist par excellence, would clearly rather spend your time (and your processor’s time) training unsupervised models instead of compiling stuff – and that is why you, like Sir Mixalot, choose Anaconda!
The Homebrew Python distribution – from the
- version: 3.7.3 and 2.7.16, respectively, at the time of writing
- Python 3:
- Python 2:
- Python 3: $py3framework +
- symlinked as
- Python 2: $py2framework +
- symlinked as
- Python 3:
- Python 2:
- verdict: Use this Python for Python 3, and likely for Python 2. Homebrew has come a long way from its humble beginnings, as an @mxcl side-hustle that compiled everything you asked for and promiscuously linked with anything on your system – from
SuperEnv isolation to
bottle binary distribution to the stats-driven winnowing of superfluous options, Homebrew is pretty robust and not at all challenging to use. I use the Homebrew Python 3 installation as both my development Python and my workhorse Python – meaning it’s what I use to get miscellaneous tasks done around my system. While some might divide these two roles amongst two Pythons, I find that the one-off tools I write for workhorse tasks can benefit from some of my library development work, and vice versa. Homebrew Python installations have a great feature that installed packages like Python.org Pythons often lack: upgrading the Homebrew Python package will not clobber the associated
site-packages directory, keeping your installed packages installed. In addition to Homebrew’s Python 3, I keep its Python 2 formula installed as well, mainly to satisfy the myriad requirements of the formulae that currently demand a Python 2 interpreter. I also occasionally fire up the Python 2 REPL to examine some quirk of the old environment, but I have to do that less and less frequently these days… I am champing at the bit for 2020, when Python 3 will be the Python!
The Homebrew PyPy distributions – from the
- version: 7.1.1 and 7.0.0 – equivalent to 2.7.13 and 3.6.1, respectively
- verdict: Have you heard of PyPy? It’s kind of a crazy thing to describe it to those who haven’t: it’s a JIT-optimizing GIL-free reimplementation of the Python interpreter, written in a subset of Python 2 called “RPython”. PyPy is extremely fast, ridiculously internally convoluted, frustratingly counterintuitive to deal with, outrageously slow to compile (or “translate” as it is known in PyPy-land) and highly piquantly interesting. Luckily, you can install the bottled binaries from Homebrew and play around with it without tying up your computer for six or so hours (really). Internally, concepts implemented by PyPy include pickleable stack frames (from Stackless Python), a method JIT and a tracing JIT (from the JVM), a fair generational (non-refcounted) garbage-collection scheme (from everyone who ever tried to patch Python to remove the GIL), coroutines that are not hacked-up generator objects (from Go and Greenlet) and the
async stuff from Python 3.4-plus-plus. Did you get all that? No? Don’t bother with PyPy, in that case – it will confuse the living daylights out of you. If you did get all that, and you have a codebase with no C-API extension dependencies, out of which you need to squeeze some performance gains before your next quarterly review so you can afford that time-share in Cleveland – then by all means, go for it, get yourself some PyPy and see what good can come of it. Otherwise DO NOT BOTHER unless you need to dispatch some serious free time and you don’t mind being flummoxed by a thing that deceptively looks like another thing with which you are familiar, but then isn’t. Yep.
Does that answer your question, @Kusine ? I do hope so, indeed!