all, x/build/cmd/relui: automate go directive maintenance in golang.org/x repositories

[dmitshur]

Abstract

The value of the go directive in golang.org/x repositories is automatically maintained to be at least 1.(N-1).0, where Go 1.N is the most recent major Go release, and Go 1.(N-1) is the previous major Go release.

Background

In the beginning, there was the GOPATH mode and versions of dependencies of golang.org/x repositories weren't explicitly tracked. Go 1.11 introduced the module mode, and over time it became the default mode. All golang.org/x repositories had an initial go.mod file checked in, and that file was maintained manually. This meant that a bug fix or a new feature in one golang.org/x repository didn't benefit another golang.org/x repository until someone chose to manually update that dependency. It also meant that eventual updates sometimes jumped many versions at once to catch up. This was resolved in 2022, when an automated monthly relui workflow began to create tags and update golang.org/x dependencies across all golang.org/x repositories (issue #48523).

At this point there are upwards of 35 golang.org/x repositories. Owners of each repository update the "go" directive manually, ad-hoc, so golang.org/x repositories may receive different levels of "go" directive maintenance. For example, owners of the golang.org/x/mod module wished to use the new-to-Go-1.22 go/version package as soon as Go 1.23 came out, and so its "go" directive was recently updated to "1.22.0". On the other hand, golang.org/x/image hasn't been updated in a while, and its "go" directive is currently still at "1.18", which itself was an upgrade from "1.12" in CL 526895 as part of bringing all golang.org/x repos to use at minimum Go 1.18 language version (issue #60268).

Leaving go directive maintenance to be done entirely manually creates the possibility of some repositories staying on an older Go language version longer. When there's enough of a need to finally upgrade it to a recent Go language version, this requires a change across multiple major Go releases at once, which can be harder to review. Having continuous, smaller incremental upgrades requires creating many CLs for all of golang.org/x repositories every 6 months, which is toilsome if always done manually.

Design

Design document at go.dev/design/69095-x-repo-continuous-go.

CC @golang/release.

[gopherbot]

Change https://go.dev/cl/608955 mentions this issue: design/69095-x-repo-continuous-go.md: add design

[gabyhelp]

Related Issues and Documentation

• x/build/cmd/relui: tag golang.org/x repos continuously #48523 (closed)

_{(Emoji vote if this was helpful or unhelpful; more detailed feedback welcome in this discussion.)}

[rsc]

This proposal has been added to the active column of the proposals project
and will now be reviewed at the weekly proposal review meetings.
— rsc for the proposal review group

[findleyr]

I think we should perhaps emphasize that once the go.mod go directive is updated to Go version 1.N-1.X, is our policy is going to be that package maintainers in that module should feel free to also break the build at older Go versions. This is implied by the reference to using the go/version package, but should be highlighted.

For example, right now x/telemetry has 1.20 in its go.mod file, but still builds with 1.19 (and this "feature" will still be relied upon by v0.16.2 of gopls, which is the final version of gopls supporting building with 1.19, before #65917 takes effect).

With that said, I think this proposal is good for Go overall. Not only will it reduce the toil of go directive maintenance, but it will also help eliminate significant complexity in these repositories related to build support for older Go versions. Additionally, with the help of all the work that has been done to make it easier to use a recent Go toolchain, I think this policy may indirectly help nudge the ecosystem away from using unsupported versions of Go.

[gopherbot]

Change https://go.dev/cl/610716 mentions this issue: go.mod: for consistency with other repos, use 1.22.0 in go.mod

[rsc]

Based on the discussion above, this proposal seems like a likely accept.
— rsc for the proposal review group

The proposal is #69095 (comment).

[adonovan]

If/when this proposal is accepted, we should add links to the statement of the new policy from the Go 1 compatibility promise and from the Go release policy, to ensure it is discoverable. (And perhaps from the README or go.mod file in the root of each x/ repo.)

[ldemailly]

For libraries, I think not aggressively forcing upgrades on dependencies is a good thing, and so is staying compatible for more than 2 versions, when 1.23 is just released for instance, probably 1.20 should still be the most recent "forced" upgrade

So imo go.mod go line should just be the language features, not compiler bug fixes etc... that's handled by toolchain and build in final product and govulncheck.

An example of disruptive forced upgrade is x/tools 0.25 which suddenly jumped form go 1.19 to go 1.22(.0) for seemingly no "need"

Specifically I think a lot of even the newer libraries really only need "go 1.18", ie generics and should stay that way.

That is not to say that people shouldn't be using a recent, supported, fully patched go version (eg 1.22.7 at time of this writing), yet libraries should indicate language issue (ie that they are or are not yet using a new iterator etc) not push a cascade of updates downstream

Edit: also while here I can add my pet peeve: the go line should be language version (ie 1.23) not specific patch level (ie 1.23.0) - until 1.21 it was like that and should imo have stayed that way... semver is perfect for toolchain and specific releases but imo doesn't belong in the go line, that .0 just muddies everything (and looks off to see 1.22.0 when 1.22.7 is out)

[adonovan]

An example of disruptive forced upgrade is x/tools 0.25 which suddenly jumped form go 1.19 to go 1.22(.0) for seemingly no "need"

It might appear that there was no need because x/tools was working fine with go1.19 only a few weeks ago. But the fact that it was continuing to build with go1.19, when the only supported Go toolchains at that time were go1.21 and go1.22 (and pre-go1.23), was imposing a significant burden on those of us who work in x/tools and those who work on the toolchain (since they had to be careful not to break x/tools). It has been a significant drag on the whole team for a number of years, and was in fact part of the genesis of the forward compatibility system.

[gopherbot]

No change in consensus, so accepted. 🎉
This issue now tracks the work of implementing the proposal.
— aclements for the proposal review group

The proposal is #69095 (comment).