djspiewak
commented
3 years ago This concept has been discussed a few times over the years. I could dig up some issue references, but I believe no fewer than two others. In general, I'm unconvinced.
The idea of putting major versions into groupIds certainly isn't new. Rich Hickey most notably advocated for it strongly in a talk a few years ago. Specs2 used to actually be Specs. Notably, Specs2 now has major version 4.x, which… probably says something about how this scheme is going.
This kind of name mangling is worthwhile when the expectation is for the two versions to live in parallel for a long, long time. In the case of REST APIs, the parallel existence of /v1/ and /v2/ APIs tends to be on the order of years. I find it very unlikely that CE2 and CE3 will live in parallel as coequal options for that length of time. In part because they aren't as heavily incompatible as they seem (most downstream codebases will be able to migrate simply by swapping their versions and applying the scalafix).
The other aspect of this which needs to be considered is tooling. Both CE2 and CE3 publish their semver contracts as part of the POM metadata, which sbt now respects. In the event that someone, accidentally or intentionally, engineers a dependency diamond, the results will be a build error. We still do have the problem of the linear classpath, but at least the manner in which we construct that classpath is safe.
This also brings me back to my main premise on all of these things: name mangling will never be able to encode the expressive power that a properly layered runtime can. The JVM is now capable of this, though it hasn't really been leveraged too extensively in the ecosystem, and it requires >JDK 8. Despite the sluggish nature of uptake though, it's clear where the wind is blowing. The linear classpath is on its way out, and unchecked dependency eviction is already out, which suggests to me that name mangling is a solution in search of a problem in the future. This isn't HTTP where we have absolutely no control over things other than naming. It also isn't Clojure where it's impossible to build tooling like MiMa.
Which ultimately brings me back to the original premise: names. Do names represent immutable concepts, or are they merely mutable pointers to immutable things? I very strongly believe in the latter, philosophically. This is how Git works. It's how the physical world works (think about postal addresses). The examples go on and on. This is also the semantic that software packages have converged to over the years. The added caveat is that we tack on a version qualifier to many (though not all) name references in order to give us the ability to immutably address the underlying state.
Which again comes back to tooling. Pushing the version redundantly into the name only affects the elements of the software linkage stack which are unguarded by tooling that understands versioning, which is a shrinking space that will continue to get narrower and narrower as some of these long-standing problems are addressed.
All in all, I really don't think this kind of thing carries its weight, and it ultimately saddles the ecosystem with an artificial name mangling that sticks around well beyond the point at which CE2 ceases to be relevant (as with the original Specs), not to mention the fact that the ecosystem is evolving to a point where versioning is handled in a more first-class way, and where naming tricks like this actually work against the infrastructure. It's better to take the modest pain of migration juggling in the short term rather than accepting perpetual pain of "otherness" in the unbounded long term.
Odomontois
kubukoz
Currently cats effect 3 has same namespaces and artifact prefixes as cats effect 2 despite being heavily incompatible. This creates a major pain for libraries depending on cats-effect 2, since they must prepare a major release replacing CE 2 to 3 and every user will be forced to choose between CE2 and CE3 ecosystems. CE3 also brings some innovative enhancements, some of them on a very low level, which could be dangerous for reliable systems before those changes will become mature and production-tested. This could slow CE3 adoption. To make CE3 adoption more swift and safe I propose to change namespace and artifact names for CE3 to
cats.effect3...and "org.typelevel" %% "cats-effect3-$module" correspondingly. That would also help some libraries such as tofu, that have cats effect dependency in its core, to prepare safe integration packages for the upcoming release.