Kelvin Versioning

Kelvin versioning (i.e. kelver) is an extremely uncommon versioning scheme, introduced by Curtis Yarvin at Urbit, that enumerates newer versions with lower natural numbers. The metaphor is between versions of a software project and degrees of Kelvin, which is the SI base unit of temperature where 0 degrees Kelvin — written 0K and often referred to as “absolute zero” — is the measure of the total absence of kinetic energy.

In this way, as a project’s version approaches 0K, the amount of further changes allowed (i.e. the kinetic energy) approaches 0. A project with version a high degree is expected to experience many further changes — it has high kinetic energy. A project with version a low degree is expected to experience few further changes — it has low kinetic energy.

Semantic Versioning

Semantic versioning (i.e. semver) is the most standard, and extremely common, versioning scheme used in production software. It encodes a project’s version with three natural numbers, written major.minor.patch:

  • major: Versions with the same major version number are backwards compatible. In this way, an major version update indicates non-backwards-compatibile changes.
  • minor: A minor version update indicates backwards-compatible changes.
  • patch: A patch version update indicates backwards-compatible changes that only fix bugs.

The Purposes of Versioning Schemes

The contrast between kelver and semver demonstrates some fundamentally different ideas about the purposes of versioning schemes. At bottom, a versioning scheme is a scheme for uniquely naming each version of a project. This description captures both kelver and semver, and also allows completely unconventional schemes such as the following:

  • randver: the version is a random version not already used by an existing version (this is how many layuser-facing version names are chosen for systems like macos and android)
  • hashver: the version is the hash of the project (this is essentially how github versions commits)
  • idver: the version is the project (in binary)
  • natver: the version starts at 0 and increments for each new version version

randver & hashver

The hamartia of randver and hashver is that the versions of projects do not indicate anything about their relationship other that whether they are the same.

Idver indicates slightly more than randver and hashver. Taking a diff of two versions entirely describes the (syntactic) differences between those versions of the project. However, users don’t care specifically about the syntax of a project; users care about the semantics of a project so they use the project’s behaviors without knowing or understanding in particular how those behaviors are implemented. So, more developed versioning schemes have adopted the further purpose to conveniently encode important semantic relationships. From now on, “change” is used as shorthand for “semantic change”.

natver

Natver is the simplest step in the direction of satisfying this further purpose. For example, version 10 is a more developed version than version 2 and versions 2 through 9 show how version 1 was incrementally developed in version 10. Additionally, there are more significant changes between versions 10 and 100 than between version 1 and 10. Using digital encodings of natural numbers is the only acceptable way to implement this, but there has been an exception made from this for Donald Knuth who, after writing some of the most influencial work on The Art of Computer Programming and developing document and font rendering programs TeX and Metafont which are still used >60 years later, was allowed to encode the natver of TeX in unary as the sequence of digits approaching π (where π is planned to be a locked version).

Even this simple step, however, introduces many more ways to misuse the versioning scheme. In randver, hashver, and idver, releasing a new project version was trivial. But in natver, the useful information in the version has to be maintained, nontrivially, by the project developer. All of the sudden, many more factors are relevant to choosing when to make a new version than before. The developer has to be conscious now that, when they release a new version, it should be based on the immediately previous version of the project. It doesn’t make sense for development to go backwards or have multiple branches because the versioning in monotone increasing and linear. So, the developer has to anticipate that the version they are releasing is strictly more advanced than the previous version. Additionally, each new version should have roughly the same amount of change. Otherwise, it’s impossible to say if there was more change from version 1 to 10 or from version 10 to 100.

semver

Beyond natver, there are many other kinds of historical semantic relationships that could be encoded in the version. Semver has been adopted as the standard way to simply encode what are widely agreed upon as the most important of these historical semantic relationships:

  • a major version corresponds to a behavioral change that is not compatible with previous major versions
  • a minor version corresponds to an additional behavior (i.e. feature) that is compatible with previous minor versions with the same major version
  • a patch version corresponds to a change in implementation that does brings the actual behavior more in line with the intended/expected/specified behavior.

Of course, these intricate ideas multiplies the potential for misuse and burdens upkeep. For example, the project developer can make a breaking change in a minor version update, add a new feature in a fix, make a breaking change in a fix, etc. The developer simply has to be trusted to use semver correctly, although “misuses” are often just bugs. More sophisticated versioning schemes like semver have huge advantages when used well and encourage good organization of semantic relationships between versions, but are also more difficult to use and practically force users to just trust the project developer. Semvar has been widely settled upon as the best way to select and encode a minimal set of semantic relationships that’s worth the continuous effort.

kelver

Historical semantic relationships are, however, only backwards-facing. Future-facing semantic relationships are also important and are most commonly captured in the concept of stability. It is seen as good practice on GitHub for a project to include a stability badge, so clearly stability is considered an important metric, but it is also very difficult to measure if not entirely ephemeral for some projects. This is certainly one reason that semver, and a very general versioning scheme, doesn’t take a stance on encoding stability. Generally, where stability is relevant enough to measure, it is either measured at the top level of the project (above any particular version) or at the major vesion level. For example, a large and widely used project (e.g. a compiler, protocol, OS) might release a new major version every couples years. The cycle is generally like this:

  • A new major version is releases as an experimental build.
  • As the new major version gains adoption, patches are made as bugs are inevitably discovered and reported.
  • Every so often, new minor features accumulate into new minor versions
  • As the major version matures and the rate of bug appearance starts to level off, the stability increases from experimental to unstable to stable. Once the major version is considered stable, it takes it’s place as the latest version of the project. Most users are expected to adopt this version of the project, but there will of course persist a diaspora of dependents out there that are never updated..
  • Eventually, new feature proposals and experimental implementations that have major changes accumulate into a new major version proposal, which is released as an experimental build.
  • As development slows on the latest version, it eventually loses its status as the latest, and the developers more or less agree that the buts worth fixing have laready been fixed. Finally, it becomes frozen, which indicates that regular updates are no longer expected and pretty much only patches of serious new bugs are expected.
  • As more time passes, some projects that depend on a specific frozen version, age, become entrenched, and rely on that major version being availabel and a specific one of its minor/patch version to be cononical for its specified set of features and compatibilities. So, this specific version becomes locked – no more development is allowed (or at least, if there are changes, they will pretty mich just be ignored by the relevant community of dependents) on that major version.

The details will be different for different kinds of projects — open source software (GHC), closed source software (Adobe Reader), frequently-updated applications (Firefox), consumer operating systems (MacOS), widely used protocols (IPv4, IPv6) — but at the macro scale they all follow this general pattern of development.

All of this is to say that stability is an integral part of organizing the development of software, in a way parallel to versioning. Yet still some sort of standard “stability versioning” (other than rough labels like “unstable” and “stable”) are very rare. We return finally to kelver.

Kelver is a sort of co-natver; natver counts up from the start while kelver counts down to the end. But in the same way that natver is only the first step towards semver in terms of encoding the most important compatibility (historical) relationships, perhaps kelver as simple as it is the first step towards encoding the most important stability (future-facing) relationships?

I summarize the appeal of kelver, as I see it and as Urbit and Urbit developers have described it, as its focus on the goal of making an finished, totally-frozen piece of software. Not all software has this goal in development (though perhaps more should), but for the software that does, kelver delivers.

As with natver, however, kelver has some rough edges that modern software engineers expect to be smoothed out by now in their organizational systems, such as versioning. Kelver does not encode information about compatibility (historical semantic relationships), but let ignore that for the moment. The large immediate tension in kelver is that it’s almost always unreasonable to expect the developer pick a reasonable starting temperature. Of course, overshooting the starting temperature is much more convenient than undershooting (since if you undershoot then you might run out), so the incentive is to have a way higher temperature than the project should have. And since the temperature of a project will start of very high, as the final stages of the project become more clear, huge drops in temperature will occur even if the number and significance of the changes made in the versions that lie along that curve of steeply-dropping temperature are not much if at all more than the changes that happened before that point. So if a user sees that the temperature has dropped recently from 100K to 10K, they don’t know yet whether that’s because there’s been a lot of changes, or the developers just decided that the project is almost done. Given this tension, kelver incentivizes developers to overshoot temperature given uncertainty, and so the temperature only has much precision when its close to zero. In other words, until the temperature is very close to zero, it probably doesn’t encode much if any useful information.

Alternatively, even worse, if the developers undershoot the temperature because they were confident in some stability that they shouldn’t have been, they’re incenticized to pack large changes into just a few versions that are still close to zero. So in the end, you have to understand the developers very well in order to interpret how stable exactly version 15K is, and there’s pretty much no way of telling how stable 1000K is.

And these are exactly the sort of problems in natver that semver addresses (see section semver). Semver introduced the semantic “units” — major changes, minor changes, and patch changes — the quantify how much semantic change a new version represents, and kelver needs some similar kind of quantification.

Here I propose an augmented kelver, staver, that preserves the spirit of the original kelver while vastly improving its usefulness, feasibility, and incentivization of good organization by encoding slightly more information.

  • staver: the version is a pair of natural numbers, the stability which counts down from a developer-chosen number, and the patch which counts up from 0, written stability.patch. A new stability version indicates a new feature or compatibility-breaking change (which should be the same, for the kind of projects that want this kind of versioning), and a new patch version indicates a bug fix that doesn’t change the specified behavior of the current stability version.

When the developer chooses a starting stability version, they may want to overshoot slightly in case they have to make more breaking changes down the road than they expected, but its much much more feasible to anticipate how many more new features you’ll add and breaking changes you’ll have to make than how many bugs you’ll have to fix. And for the projects that desire some sort of stability-focused versioning like kelver or staver, new features are essentially synonymous with breaking changes, since a new feature is outside the specification of the previous stability version. Additionally, all bugs in the same stability version can be addressed without the worry of stability limitations, via patch versions. Kelver introduced a hack called “release candidates” which is similar to patch versions but don’t count as real versions and in that way have none of the structure imposed upon them that software engineers learned to make great use of. Patch versions maintain that usefulness, which still enforcing the restrictions of kelver when it comes to breaking changes.

Since it’s so much more feasible for the developer to choose precise stability versions rather than being strongly incentivized to overshoot, the stability version can be expected to be much more precise. Of course, sometimes developers will still want to overshoot due to large uncertainties about the future of the project, but then as the future becomes more certain and the stability version falls quickly, it more sensitively captures the solidification of the future of the project than kelver where you can only really afford to start dropping the version by much once you are at the very end of a project.

Definitions of Versioning Schemes

  • randver: The version is a random version not already used by an existing version (this is how many layuser-facing version names are chosen for systems like macos and android)
  • hashver: The version is the hash of the project (this is essentially how github versions commits)
  • idver: The version is the project (in binary)
  • natver: the version starts at 0 and increments for each new version version
  • kelver: The version is a natural number counts down for each new version, and was initialized by the developers. The final version is 0K.
  • semver: the version is a triple of a major, minor, and patch version. A new major version indicates compatibility-breaking changes, a minor version indicates compatibility-presenving new features, and a patch version indicates a compatibility-preserving bug fix
  • staver: the version is a pair of a stability version and a patch version. The stability version counts down on each breaking change or new feature and was initialized by the developers, and the patch verson indicates a compatible bug fix that doesn’t add new features. The final version is 0.N, whre N can still increment for bug fixes.