Skip to content

Latest commit

 

History

History
201 lines (153 loc) · 7.36 KB

0012-teleport-versioning.md

File metadata and controls

201 lines (153 loc) · 7.36 KB
authors state
Andrew Lytvynov ([email protected])
implemented

RFD 12 - Teleport versioning

What

Versioning scheme for Teleport releases (post-5.0).

NOTE: these are guidelines, rather than strict rules. In some cases we will ignore these guidelines based on customer needs.

Terminology

Quick note on terminology used below. I'll use naming from semver:

vX.Y.Z
 ^ ^ ^
 | | *- "patch version Z"
 | *- "minor version Y"
 *- "major version X"

Why

Teleport has ~4 big releases per year.

Up to v5.0.0, Teleport used a versioning scheme that looks like semver (but actually isn't):

  • a minor version is bumped for all regular releases (e.g. 4.3 -> 4.4)
  • a major version is bumped when a release feels particularly significant (e.g. 4.4 -> 5.0 with introduction of application access)
  • a patch version is bumped for bug patches

Our compatibility promise is:

When running multiple binaries of Teleport within a cluster (nodes, proxies, clients, etc), the following rules apply:

  • Patch versions are always compatible, for example any 4.0.1 component will work with any 4.0.3 component.
  • Other versions are always compatible with their previous release. This means you must not attempt to upgrade from 4.1 straight to 4.3. You must upgrade to 4.2 first.
  • Teleport clients tsh for users and tctl for admins may not be compatible with different versions of the teleport service.

The 2nd point is crucial: we never break compatibility with a previous release (be it a major or a minor version bump).

The downsides of this versioning scheme are:

  • the distinction between major and minor version bumps is largely driven by business/product and has no technical difference
  • a user that doesn't read our upgrade docs carefully can assume semver semantics:
    • upgrading from vX.A.B to vX.C.D is safe - it's not if C - A > 1
    • upgrading from vX.A.B to vY.C.D is going to break things - it won't, as long as these are sequential releases

Therefore, I propose to switch to a more semver-like scheme, starting with 6.0.

Details

  • Major versions are for teleport releases.
  • Minor versions are for backwards-compatible features and improvements, and non-critical bugfix batches
  • Patch versions are for quick followup regression and critical bug fixes.
  • Suffixes:
    • -dev suffix is for development builds (e.g. off of the main branch)
    • -alpha.X suffix, built from main branch, good for demos and early deploys, but not staging
    • -beta.X suffix, built from main branch, good for staging deploys, but not production
    • -rc.X suffixes are for release candidates, potentially production-ready but still in testing, built from branch/vN

The benefits are:

  • Major version bumps clearly communicate to users to exercise caution when upgrading, and read release notes
  • Minor/patch version bumps are a no-brainer and can be automated
  • Users can intuitively understand versioning semantics, due to popularity of semver

Compatibility

vN.*.* binaries are compatible with any other vN.*.* or vN-1.*.* binaries (both client and server).

Users are free to use any vN.*.* versions throughout their deployment, after upgrading everything from vN-1.*.*.

Support window

Teleport officially supports the latest major version and two previous major versions. For example:

  • v8.*.* (latest)
  • v7.*.*
  • v6.*.*

Git branches

Teleport has several branches related to releases:

  • main - main development branch
  • branch/vX - for upcoming minor releases
    • created when vX.0.0-rc.1 is cut
  • branch/vX.Y - for patch releases
    • created when vX.Y.0 is cut

Here are how different kinds of changes made map to those branches:

  • work towards the next major release
    • merged into main
  • work towards the next minor release
    • merged into main
    • backported into branch/vX
  • bugfixes that need to be backported
    • merged into main
    • backported into branch/vX
    • backported into the latest branch/vX.Y
    • repeat for all vX that need the backport
  • bugfixes that do not need to be backported
    • merged into main

Example: new release

The current release of teleport is v5 and the next will be v6.

  • initially, builds from main will be v6.0.0-dev
  • when most big changes are merged, we cut v6.0.0-alpha.1
  • when only small bugfix changes are pending, we cut v6.0.0-beta.1
  • all planned changes are merged, we make branch/v6 and cut v6.0.0-rc.1
  • during release testing, we fix bugs and cut v6.0.0-rc.2, v6.0.0-rc.3, etc
  • assuming v6.0.0-rc.3 passes the tests, create branch branch/v6.0 and tag it as v6.0.0
  • we discover a serious bug in v6.0.0, fix it in main, branch/v6, branch/v6.0 and cut v6.0.1 from branch/v6.0
  • over the next few days/weeks, we gather user feedback
  • based on feedback, we fix a number of bugs in main, backport them to branch/v6, create branch branch/v6.1 and cut v6.1.0 from it

Example: backports

The latest 3 released versions are v8.0.1, v7.1.2, v6.3.2.

  • we discover a bug affecting all versions
  • we fix it in main
  • if the bug is urgent, we backport to branch/v8, branch/v8.0, branch/v7, branch/v7.1, branch/v6, branch/v6.3 and release v8.0.2, v7.1.3, v6.3.3
  • if the bug is not urgent and no customer has requested backports to older versions, we backport to branch/v8, create branch/v8.1 and release v8.1.0
    • optionally, we wait for more bug fixes to accumulate in branch/v8 before releasing v8.1.0

Example: security vulnerability

The latest 3 released versions are v8.0.1, v7.1.2, v6.3.2.

  • we discover a security bug affecting all versions
  • we fix it in main and backport to branch/v8, branch/v8.0, branch/v7, branch/v7.1, branch/v6, branch/v6.3
  • after backports merge, we release v8.0.2, v7.1.3, v6.3.3

Why not use minor versions, like we do now?

An experienced reader might notice that this is not exactly semver. In semver, major versions are for strictly breaking changes with no compatibility.

So why not keep using minor version bumps like we do now? And reserve major version for when we really need to make a breaking change?

Because we want to keep our compatibility guarantee and avoid the Python 2/3 story. We never want a vN+1 that has no migration path other than "rebuild the cluster from scratch".

So we don't want to freeze the major version forever this way.

Docker image tags

Currently, we have 2 kinds of docker images:

  • teleport:X.Y.Z - immutable tag for a specific patch release (for example 4.4.1, 4.4.2)
  • teleport:X.Y - mutable tag pointing at the latest patch release for X.Y (for example 4.4 pointing at 4.4.2)

For new versioning, the following tags are created:

  • teleport:X.Y.Z - immutable tag for a specific patch release (for example 5.1.0, 5.1.1)
  • teleport:X - mutable tag pointing at the latest patch release for major version X (for example 5 pointing at 5.1.1)

In the above example, if we cut v5.2.0, the following tag changes happen:

  • new tag 5.2.0
  • existing tag 5 updated to point at 5.2.0

The old teleport:X.Y tags will no longer be generated, since they are an inferior version of teleport:X and shouldn't be used. This may cause customer confusion and some support load during the transition, but it's less work than maintaining the teleport:X.Y tag.