Introducing DjangoVer

November 18, 2024    Django, Meta, Python

Version numbering is hard, and there are lots of popular schemes out there for how to do it. Today I want to talk about a system I’ve settled on for my own Django-related packages, and which I’m calling “DjangoVer”, because it ties the version number of a Django-related package to the latest Django version that package supports.

But one quick note to start with: this is not really “introducing” the idea of DjangoVer, because I know I’ve used the name a few times already in other places. I’m also not the person who invented this, and I don’t know for certain who did — I’ve seen several packages which appear to follow some form of DjangoVer and took inspiration from them in defining my own take on it.

Django’s version scheme: an overview

The basic idea of DjangoVer is that the version number of a Django-related package should tell you which version of Django you can use it with. Which probably doesn’t help much if you don’t know how Django releases are numbered, so let’s start there. In brief:

• Django issues a “feature release” — one which introduces new features — roughly once every eight months. The current feature release series of Django is 5.1.
• Django issues “bugfix releases” — which fix bugs in one or more feature releases — roughly once each month. As I write this, the latest bugfix release for the 5.1 feature release series is 5.1.3 (along with Django 5.0.9 for the 5.0 feature release series, and Django 4.2.16 for the 4.2 feature release series).
• The version number scheme is MAJOR.FEATURE.BUGFIX, where MAJOR, FEATURE, and BUGFIX are integers.
• The FEATURE component starts at 0, then increments to 1, then to 2, then MAJOR is incremented and FEATURE goes back to 0. BUGFIX starts at 0 with each new feature release, and increments for the bugfix releases for that feature release.
• Every feature release whose FEATURE component is 2 is a long-term support (“LTS”) release.

This has been in effect since Django 2.0 was released, and the feature releases have been: 2.0, 2.1, 2.2 (LTS); 3.0, 3.1, 3.2 (LTS); 4.0, 4.1, 4.2 (LTS); 5.0, 5.1. Django 5.2 (LTS) is expected in April 2025, and then eight months later (if nothing is changed) will come Django 6.0.

I’ll talk more about SemVer in a bit, but it’s worth being crystal clear that Django does not follow Semantic Versioning, and the MAJOR number is not a signal about API compatibility. Instead, API compatibility runs LTS-to-LTS, with a simple principle: if your code runs on a Django LTS release and raises no deprecation warnings, it will run unmodified on the next LTS release. So, for example, if you have an application that runs without deprecation warnings on Django 4.2 LTS, it will run unmodified on Django 5.2 LTS (though at that point it might begin raising new deprecation warnings, and you’d need to clear them before it would be safe to upgrade any further).

DjangoVer, defined

In DjangoVer, a Django-related package has a version number of the form DJANGO_MAJOR.DJANGO_FEATURE.PACKAGE_VERSION, where DJANGO_MAJOR and DJANGO_FEATURE indicate the most recent feature release series of Django supported by the package, and PACKAGE_VERSION begins at zero and increments by one with each release of the package supporting that feature release of Django.

Since the version number only indicates the newest Django feature release supported, a package using DjangoVer should also use Python package classifiers to indicate the full range of its Django support (such as Framework :: Django :: 5.1 to indicate support for Django 5.1 — see examples on PyPI).

But while Django takes care to maintain compatibility from one LTS to the next, I do not think DjangoVer packages need to do that; they can use the simpler approach of issuing deprecation warnings for two releases, and then making the breaking change. One of the stated reasons for Django’s LTS-to-LTS compatibility policy is to help third-party packages have an easier time supporting Django releases that people are actually likely to use; otherwise, Django itself generally just follows the “deprecate for two releases, then remove it” pattern. No matter what compatibility policy is chosen, however, it should be documented clearly, since DjangoVer explicitly does not attempt to provide any information about API stability/compatibility in the version number.

That’s a bit wordy, so let’s try an example:

• If you started a new Django-related package today, you’d (hopefully) support the most recent Django feature release, which is 5.1. So the DjangoVer version of your package should be 5.1.0.
• As long as Django 5.1 is the newest Django feature release you support, you’d increment the third digit of the version number. As you add features or fix bugs you’d release 5.1.1, 5.1.2, etc.
• When Django 5.2 comes out next year, you’d (hopefully) add support for it. When you do, you’d set your package’s version number to 5.2.0. This would be followed by 5.2.1, 5.2.2, etc., and then eight months later by 6.0.0 to support Django 6.0.
• If version 5.1.0 of your package supports Django 5.1, 5.0, and 4.2 (the feature releases receiving upstream support from Django at the time of the 5.1 release), it should indicate that by including the Framework :: Django, Framework :: Django :: 4.2, Framework :: Django :: 5.0, and Framework :: Django :: 5.1 classifiers in its package metadata.

Why another version system?

Some of you probably didn’t even read this far before rushing to instantly post the XKCD “Standards” comic as a reply. Thank you in advance for letting the rest of us know we don’t need to bother listening to or engaging with you. For everyone else: here’s why I think in this case adding yet another “standard” is actually a good idea.

The elephant in the room here is Semantic Versioning (“SemVer”). Others have written about some of the problems with SemVer, but I’ll add my own two cents here: “compatibility” is far too complex and nebulous a concept to be usefully encoded in a simple value like a version number. And if you want my really cynical take, the actual point of SemVer in practice is to protect developers of software from users, by providing endless loopholes and ways to say “sure, this change broke your code, but that doesn’t count as a breaking change”. It’ll turn out that the developer had a different interpretation of the documentation than you did, or that the API contract was “underspecified” and now has been “clarified”, or they’ll just throw their hands up, yell “Hyrum’s Law” and say they can’t possibly be expected to preserve that behavior.

A lot of this is rooted in the belief that changes, and especially breaking changes, are inherently bad and shameful, and that if you introduce them you’re a bad developer who should be ashamed. Which is, frankly, bullshit. Useful software almost always evolves and changes over time, and it’s unrealistic to expect it not to. I wrote about this a few years back in the context of the Python 2/3 transition:

Though there is one thing I think gets overlooked a lot: usually, the anti-Python-3 argument is presented as the desire of a particular company, or project, or person, to stand still and buck the trend of the world to be ever-changing.

But really they’re asking for the inverse of that. Rather than being a fixed point in a constantly-changing world, what they really seem to want is to be the only ones still moving in a world that has become static around them. If only the Python team would stop fiddling with the language! If only the maintainers of popular frameworks would stop evolving their APIs! Then we could finally stop worrying about our dependencies and get on with our real work! Of course, it’s logically impossible for each one of those entities to be the sole mover in a static world, but pointing that out doesn’t always go well.

But that’s a rant for another day and another full post all its own. For now it’s enough to just say I don’t believe SemVer can ever deliver on what it promises. So where does that leave us?

Well, if the version number can’t tell you whether it’s safe to upgrade from one version to another, perhaps it can still tell you something useful. And for me, when I’m evaluating a piece of third-party software for possible use, one of the most important things I want to know is: is someone actually maintaining this? There are lots of potential signals to look for, but some version schemes — like CalVer — can encode this into the version number. Want to know if the software’s maintained? With CalVer you can guess a package’s maintenance status, with pretty good accuracy, from a glance at the version number.

Over the course of this year I’ve been transitioning all my personal non-Django packages to CalVer for precisely this reason. Compatibility, again, is something I think can’t possibly be encoded into a version number, but “someone’s keeping an eye on this” can be. Even if I’m not adding features to something, Python itself does a new version every year and I’ll push a new release to explicitly mark compatibility (as I did recently for the release of Python 3.13). That’ll bump the version number and let anyone who takes a quick glance at it know I’m still there and paying attention to the package.

For packages meant to be used with Django, though, the version number can usefully encode another piece of information: not just “is someone maintaining this”, but “can I use this with my Django installation”. And that is what DjangoVer is about: telling you at a glance the maintenance and Django compatibility status of a package.

DjangoVer in practice

All of my own personal Django-related packages are now using DjangoVer, and say so in their documentation. If I start any new Django-related projects they’ll do the same thing.

A quick scroll through PyPI turns up other packages doing something that looks similar; django-cockroachdb and django-snowflake, for example, versioned their Django 5.1 packages as “5.1”, and explicitly say in their READMEs to install a package version corresponding to the Django version you use (they also have a maintainer in common, who I suspect of having been an early inventor of what I’m now calling “DjangoVer”).

If you maintain a Django-related package, I’d encourage you to at least think about adopting some form of DjangoVer, too. I won’t say it’s the best, period, because something better could always come along, but in terms of information that can be usefully encoded into the version number, I think DjangoVer is the best option I’ve seen for Django-related packages.