Except we already had godep. It worked. It was in widespread use.
> exacerbated by the relatively decent (but not perfect) backward compatibility with pre-modules code.
My experience has been that the backwards compatibility has been a nightmare. It forces itself on any downstream repository, whether or not that repository wishes to participate. Plus it does very poorly on diamond dependencies, giving error messages about as scrutable as the I Ching. If your answer is "they're doing it wrong", take it up with the Kubernetes and go-client maintainers, because I can't change what they are doing.
> There is already a global system to get hold to unique names: the domain name system. For better or worse, that's a solved problem, let's not create new naming arbiters
Versions are not addresses. Addresses are not versions. You may find a version at an address, but addresses are not versions. Docker did the same thing with image references and they too created a pig's breakfast. The difference was that they didn't have prior art to compare to and it was an oversight obvious largely in hindsight. I do not extend the same credit to Russ Cox.
> un-eyeballable, merge-review-hostile, another-arbitrary-c-hacker-mini-language, shas-upon-shas version
godep shared the same problem. We can argue that godep was good enough and the Go maintainers handled the situation poorly with this whole "ask community and then ignore the community" trick. But you're focusing on the transition; if this system had been part of Go 1.0, I would have had nothing bad to say about it.
I'm not arguing the the transition process wasn't a mess. I'm arguing that it became a mess precisely because they wanted to attempt backward compat. They could have said "this is go 2; if you want to use your library with go 2 you have to use this standard, tag your release, or upload it to some central index"
I love the mechanics of it. Most of what you describe is the result of having waited ages before implementing it, and the resulting prolonged migration period, exacerbated by the relatively decent (but not perfect) backward compatibility with pre-modules code.
Prior art isn't perfect either: I'm happy somebody tried something different from centralized naming systems or almost-but-not-quite federated artifact collections (such as the various maven repos). There is already a global system to get hold to unique names: the domain name system. For better or worse, that's a solved problem, let's not create new naming arbiters