matrix-spec/proposals/2844-global-versioning.md

MSC2844: Using a global version number for the entire specification

Currently we have 4 kinds of versions, all of which have slightly different use cases and semantics which apply:

1. The individual API spec document versions, tracked as revisions (r0.6.1, for example).
2. Individual endpoint versioning underneath an API spec document version (/v1/, /v2/, etc). Note that the client-server API currently ties the major version of its spec document version to the endpoint, thus making most endpoints under it as /r0/ (currently).
3. Room versions which define a set of behaviour and algorithms on a per-room basis. These are well defined in the spec and are not covered here: https://matrix.org/docs/spec/#room-versions
4. An overarching "Matrix" version, largely for marketing purposes. So far we've only cut Matrix 1.0 back when we finalized the initial versions of the spec documents, but have not cut another one since.

This current system is slightly confusing, and has some drawbacks for being able to compile builds of the spec documents (published on matrix.org) and generally try and communicate what supported versions an implementation might have. For example, Synapse currently supports 4 different APIs, all of which have their own versions, and all of which would need to be considered and compared when validating another implementation of Matrix such as a client or push gateway. Instead, Synapse could say it supports "Matrix 1.1", making compatibility much easier to determine - this is what this proposal aims to define.

Proposal

Instead of having per-API versions (r0.6.1, etc), we have a version that spans the entire specification. This version represents versioning for the index (which has quite a bit of unversioned specification on it currently), the APIs, room versions, and the appendices (which are also currently unversioned but contain specification). Room versions are a bit more nuanced though, and are covered later in this MSC.

The version which covers the entire specification and all its parts is called the "Matrix version", and is a promotion of the previously marketing-only version number assigned to the spec. The spec core team is expected to release Matrix 1.1.0 as the first version, leaving v1.0 as the marketing era and recorded somewhere for posterity (though has no significant meaning, unchanged by this MSC).

Doing this has the benefits previously alluded to:

• Implementations of Matrix can now easily compare their supported versions using a single identifier without having to (potentially) indicate which API they built support for.
• Publishing the specification is less likely to contain broken or outdated links due to API versions not matching up properly. This is currently an issue where if we want to release a new version of the server-server specification then we must also either rebuild or manually fix the blob of HTML known as the client-server API to account for the new version - we often forget this step, sometimes because it's just too difficult.
• Explaining to people what version Matrix or any of the documents is at becomes incredibly simplified. No longer will we have to explain most of what the introduction to this proposal covers to every new person who asks.

Full Matrix version grammar

The Matrix versioning scheme takes heavy inspiration from semantic versioning, though intentionally does not follow it for reasons described throughout this proposal. Primarily, the argument against semantic versioning is held in the alternatives section below.

Given a version number MAJOR.MINOR.PATCH, incremement the:

• MAJOR version when a substantial change is made to the core of the protocol. This is reserved for interpretation by the Spec Core Team, though is intended to be for extremely invasive changes such as switching away from JSON, introducing a number of features where a MINOR version increase just doesn't feel good enough, or changes to the signing algorithms.
• MINOR version when a feature is introduced, or a backwards incompatible change has been managed through the specification. Later on, this proposal explains what it means to manage a breaking change.
• PATCH version when correctional changes are made, such as spelling, cosmetic, or other similarly small patches are done. Implementations do not need to worry about the patch version.

When present in the protocol itself, the Matrix version will always be prefixed with v. For example, v1.1.0.

Additional information can be supplied in the version number by appending a dash (-) to the end of the version and including any relevant information. This is typically used to denote alpha, beta, unstable, or other similar off-cycle release builds. This MSC does not propose a scheme for RCs or pre-releases, though the Spec Core Team may wish to do so.

See the section on brewing Matrix versions for information on how the unstable version is decided.

From an implementation perspective, compatibility is guaranteed between PATCH versions. MINOR versions have a backwards compatibility scheme described later in this proposal. MAJOR versions are expected to have zero backwards compatibility guarantees to them.

For clarity, v1.1.0 and v1.1.8 are functionally the same. v1.2.0 will probably work with v1.1.0, though implementations should be wary if they depend on a version. As mentioned, the backwards compatibility scheme section goes into more detail on this.

A potential argument is that we don't need a patch version if no implementation will ever care about it, which is a valid argument to have. This MSC believes that although the patch version is effectively useless to implementations, it is valuable as evidence of progress and finality of a given version. Going back to edit already-released versions of the specification can be damaging to the integrity of the protocol, and thus it is proposed by this MSC that the Spec Core Team remain accountable by forcing them to release with a patch version increase for minor, functionally indifferent, changes.

Structure changes and changelogs

The API documents remain mostly unchanged. We'll still have a client-server API, server-server API, etc, but won't have versions associated with those particular documents. This also means they would lose their individual changelogs in favour of a more general changelog. An exception to this rule is room versions, which are covered later in this proposal.

Though the changelog format is not covered by the MSC process, this MSC proposes that the initial changelog for the Matrix versioning scheme be broken out into sections for each API that had changes. Ideally, the changelog would also indicate if no changes were made to a particular API/area to help be clearer to implementation authors. The Push Gateway API is, for example, likely going to be one of the few which will nearly always say "No relevant changes" for years.

Endpoint versioning

Under this MSC, all HTTP endpoints in the specification are to be per-endpoint versioned. This is already the case for all APIs except the client-server API, and so this section deals specifically with that API. The deprecation of endpoints is handled later in this proposal.

Under this proposal, all endpoints in the client-server API get assigned v3 as their per-endpoint version as a starting point. This is primarily done to avoid confusion with the ancient client-server API versions which had v1 and called the rN system "v2". Though many of the endpoints available today are not present in those older API editions, it is still proposed that they start at v3 to avoid confusion with long-standing implementations.

Servers which are lucky enough to exist during this versioning scheme change are expected to continue supporting the rN system. This is done by advertising the existing client-server API versions as they always would have on /versions, though appending "v1.1.0" to indicate that this MSC is supported.

As a further clarification to a solved problem, the /versions endpoint for the client-server API does not need to advertise all patch version changes - just the major/minor versions it supports. If a server does advertise a patch version, clients are expected to resolve that to the relevant major/minor version equivalent (v1.1.8 gets treated as v1.1.0, for example). This same rule applies to widgets and the supported_api_versions action, which is currently in a draft specification state. Once widgets land in the spec, the rule comes into effect.

Brewing changes for the specification

Prior to this MSC, the Spec Core Team would release a given version of an API whenever it felt like the right time to do so. There's very little planning put into a release, and often times the call to cut a release is arbitrary. Though this MSC doesn't solve this problem neccesarily, it does change the dynamic the Spec Core Team has with the community when it comes to releases.

Instead of arbitrarily deciding when to cut a release, the Spec Core Team is expected to plan ahead and choose a date for the next major/minor release. The team is not required to use a cadence to perform releases, though is expected to perform at least one release a year. Reasonable notice is expected to be given to the community to give them a chance to push their MSCs and ideas to completion. "Reasonable" is intentionally left undefined by this MSC as it might change over time, though the current suggestion is to give at least 2 months notice. Most MSC authors are currently contributing on a volunteer or spare time basis and thus might not be able to rapidly push their ideas through the stages as quickly.

Patch releases do not require such notice and can happen whenever.

The date advertised to the community is a cutoff date, not a release date. The Spec Core Team and wider community might still need time to write up the formal specification for some MSCs or improve their implementations to be more prepared for the impending official release. In the eyes of the MSC process, the cutoff date is not enough for an implementation to switch to using stable endpoints.

Considering these are only expectations and not requirements, the Spec Core Team might break them from time to time for various reasons including urgent security releases, last minute realizations that something might not work as proposed, etc. Under the Foundation, the Spec Core Team is required to act in the best interest of the protocol and therefore should be able to reasonably justify why an expectation is being broken at the time of breaking it - in no way does this MSC propose that the Spec Core Team be able to blindside the community with a release for no justifiable reason.

To recap, the process is as follows:

1. Sometime before a given release happens, the Spec Core Team announces a cutoff date for MSCs to land that is at least 2 months in the future.
2. Upon cutoff, the Spec Core Team takes responsibility for ensuring all relevant changes are written up in a timely fashion.
3. The Spec Core Team makes the release. At this point, implementations can stop using unstable prefixes for any included MSC.

Because the release schedule is well-advertised, it should be clear to everyone what the next non-patch version number will be. By default, the assumption can be made that the MINOR version will increase by 1. For the purposes of producing built copies of the spec, the version number for unstable (unreleased) versions shall be the next expected version number followed by -unstable. For example, if v1.1.0 were the current release, the unstable specification would be built as v1.2.0-unstable. In the event that a change lands where the major version needs incremementing, v1.2.0 (in this example) would never see the light of day and instead turn into v2.0.0-unstable.

Room versions & brewing room versions

Author's note: Having many things with the root word "version" can be confusing, so for this section "room versions" are called "room editions" and the Matrix version refers to what this proposal is introducing. This MSC does not propose renaming "room versions" - that is another MSC's problem.

Room editions are a bit special in that they have their own versioning scheme as other servers need to be aware of how to process the room. As such, a room edition's versioning scheme is not altered by this proposal, however the publishing of the (in)stability of a given edition is now covered by the newly proposed Matrix version.

Whenever a room edition transitions from stable to unstable, or unstable to stable, or is introduced then it would get counted as a feature for a MINOR release of Matrix. We don't currently have a plan to remove any room editions, so they are not covered as a potential process for this MSC.

The brewing process for a room edition has typically been a last minute MSC to declare an edition number and which other MSCs are included. This MSC proposes changing that slightly to instead allow accepted MSCs to be batched up at release time into a new room edition. Each MSC will have needed testing and implementation validation prior to their acceptance, so it should be fine to merge them together. If a concern or conflict were to arise about doing so, an explicit MSC could be raised just like in today's model where the conflicts/concerns are resolved. Such concerns might be the compatibility of the features, and some conflicts might be two MSCs changing the same authorization rule (for example). The new, explicit, MSC would require implementation validation and testing to be approved and thus would then be fine.

Though somewhat implicit, it might not be clear: multiple room editions can be included in a single Matrix release, though how they get built is not really addressed by this MSC. Using the spirit of the ideas proposed, it would likely mean as many MSCs as room editions to bring them into effect. This sort of practice is believed to be extremely rare, however - it'd be more likely that the Spec Core Team uses their capability to do an emergency release of Matrix to get a room edition into the wild rather than brew two (or more) room editions at once.

The recommended default room edition is not automatically changed under this proposal either - changing the recommended default requires an MSC just like any other part of Matrix and would fall under general feature changes.

Deprecation approach

This MSC doesn't change much about how things get deprecated, though it would be codified by this MSC.

An MSC is required to transition something from stable (the default) to deprecated. Once something has been deprecated for suitably long enough, it is eligible for removal from the specification with another MSC. Today's process is the same, though not defined explicitly.

Also not mentioned in today's system is that implementations are not required to implement deprecated endpoints. This MSC doesn't change that, but does put some rules around how deprecation works for a given endpoint. Specifically, if a server wants to support a Matrix version where an endpoint is not deprecated, then it must serve it.

As an example, if /test were introduced in v1.1, deprecated in v1.2, and removed in v1.3 then an implementation can support v1.1, v1.2, and v1.3 by implementing /test as it was defined in v1.2 (minus the deprecation flag). If the implementation wanted to support just v1.2 and v1.3, then it could optionally implement /test, though is encouraged to. Further, if the implementation only wanted to support v1.3, then it should not implement /test at all because it was removed.

Generally deprecation is paired with replacement or breaking changes. For example, if /v3/sync were to be modified such that it needed to be bumped to v4, the MSC which does so would deprecate /v3/sync in favour of its proposed /v4/sync. Because endpoints are versioned on a per-endpoint basis, /v4/sync will still work with a server that supports /v3/profile (for example) - the version number doesn't mean an implementation can only use v4 endpoints.

This sort of approach could be potentially confusing and non-standard as it would mean for an amount of time in the specification there would be two versions of an endpoint: one deprecated and one not. Most specifications and protocols do not use this sort of approach and instead opt to replace the whole API or outright remove the endpoint, however the Matrix specification tends to have a longer-lived cycle associated with it and thus means we should support a larger than average backwards compatibility period.

Supported versions schedule

Currently implementations are left to fend for themselves on deciding which versions of their APIs to support. This MSC doesn't change this due to a wide variety of potential use cases for different environments and older specifications, though this MSC does bring forth a guiding principle which is applicable to most standard implementations.

Implementations are encouraged to support the latest version of Matrix as soon as feasibly possible, and the two most recent MINOR versions past that. Given a cadence of about 1 release every 6 months, this should mean that the standard implementation supports roughly 1.5 years worth of Matrix history.

Room versions are special in that they will essentially always be included in a Matrix release, even if unstable. The current specification says that implementations don't have to implement unstable room versions, and this is true under this MSC too.

As with the current day, implementations should consider what makes sense for them to support. It's entirely possible that a Matrix version happens with no client-server API changes, for instance, which may mean that clients won't bother even checking for the no-op version - a server implementation might not want to count that version towards it's "last 2 major releases" schedule to avoid breaking clients. It may instead be desirable for implementations to support Matrix releases based on time rather than by the number of releases.

For extreme clarity, the suggested schedule for supported versions would be (all examples):

• v1.1.0 - Released in January - No support needed due to age.
• v1.2.0 - Released in July - Should be supported as part of the 1.2.x series.
• v1.2.5 - Released in August - Should be supported because it's part of the 1.2.x series.
• v1.3.0 - Released in November - Should be supported.
• v1.4.0 - Released in December (current day) - The current release so should definitely be supported.

Potential issues

None appear to be relevant to be discussed on their own - they are discussed in their respective sections above when raised.

Alternatives

There are some strong opinions that we should use proper semantic versioning for the specification instead of the inspired system proposed here. So, why shouldn't we use semantic versioning?

1. It's meant for software and library compatibility, not specifications. Though it could theoretically be used as a specification version, the benefits of doing so are not immediately clear. The scheme proposed here is simple enough where rudimentary comparisons are still possible between versions, and existing semantic versioning libraries can still be made to work. Further, the specification's version number should not be relied upon by a library for its versioning scheme - libraries, applications, etc should have their own versioning scheme so they may work independently of the spec's release schedule.

2. It has potential for causing very high major version numbers. Though largely an aesthetic concern, it can be hard to market Matrix v45 (or even Matrix v4) to potential ecosystem adopters due to the apparant unstable-ness of the specification. Similarly, the major version is used for advertising purposes which could be confusing or overly noisy to say there's a major version every few releases. By instead staying in the 1.x series for a long period of time, the specification appears stable and easy to work with, attracting potential adopters and making that 2.0 release feel all that more special.

3. The semantic versioning spec is not followed in practice. Most uses of semantic versioning are actually off-spec adaptations which are largely compatible with the ideals of the system. This, however, puts Matrix in a difficult spot as it would want to say we follow semantic versioning, but can't because there's no relevant specification document to link to. Even if there was, it would appear as though we were encouraging the idea of forking a specification as a specification ourselves, which may be confusing if not sending the wrong message entirely. Though the system proposed here is a reinvention of semantic versioning to a degree, this proposed system is different from how semantic versioning works in so many ways it is not entirely comparable.

4. The benefit of saying we use a well-popularized versioning system is not a strong enough argument to be considered here.

This MSC is also inherently incompatible with semantic versioning due to its approach to deprecation. Instead of encouraging breaking changes (removal of endpoints) be major version changes, this MSC says that happens at the minor version change level. As mentioned in the relevant section, this is not foreseen to be an issue for Matrix given its a system already used by the protocol and is common enough to at least be moderately familiar - the arguments for using semantic versioning in this respect do not hold up, per above.

Security considerations

None relevant - if we need to make a security release for Matrix then we simply make a release and advertise accordingly.

Unstable prefix

The author does not recommend that this MSC be implemented prior to it landing due to the complexity involved as well as the behavioural changes not being possible to implement. However, if an implementation wishes to try anyways, it should use org.matrix.msc2844 in the unstable_features of /versions and use /_matrix/client/unstable/org.matrix.msc2844 in place of /_matrix/client/r0.

This MSC is largely proven as possible through an in-development build of the specification which uses an alternative toolchain for rendering the specification: https://adoring-einstein-5ea514.netlify.app/ (see the 'releases' dropdown in the top right; link may not be available in the distant future - sorry).