Decide how public-facing instance traits should be

[travisbrown]

Right now we have AllInstances and AllSyntax traits that are in the public API and are sometimes used outside of Cats itself—for example in Circe we have a CirceSuite helper class that extends these traits so that we don't have to import cats.implicits._ in every test file.

The problem with this kind of usage is that AllInstances and AllSyntax don't actually include all of anything, ever since we started stacking on BinCompatN layers to maintain binary compatibility on Scala 2.11. For example, want a Defer for Function0? You won't get it from AllInstances—you also need AllInstancesBinCompat0.

This means anyone using AllInstances or AllSyntax has to be careful to add the latest AllSyntaxBinCompat37 or whatever traits included in each new Cats release if they actually want everything.

This is something we haven't managed to stay on track on even within Cats itself—e.g. see #3007, which is due to the fact that cats.implicits is lagging two AllSyntaxBinCompats behind, or SyntaxSuite, which is four AllInstanceBinCompats out of date (but happens to still be working), etc.

To make things even more confusing, there's a AllSyntaxBinCompat (note the missing number) abstract class that bundles up all the syntax bin-compat traits, but no equivalent AllInstancesBinCompat.

I've also recently proposed (and merged) making the bin-compat traits package-private (#3003), which on second thoughts I think we'll have to undo, at least until 2.1.

So there's a lot of miscellaneous little brokenness and inconsistency and mildly bad UX here, some of which I'm about to open a PR to address, but there's also a bigger question of whether we actually intend the AllInstances and AllSyntax traits to be used outside of the implementation of Cats itself.

[Daenyth]

Just as a data point, I've used those in third-party code so that I could make one combined import for cats implicits + my library implicits + another library implicits, to help reduce brain effort on imports

[djspiewak]

As another random data point… I originally structured djspiewak/shims to hide the "All" traits, so the only way to actually get at anything was via the shims._ import. The problem with this is it made it impossible for any third parties to participate in the implicit prioritization, which created unavoidable conflicts. Given Cats' nature, I would suspect similar problems if we privatized the traits.

[joroKr21]

But this is only relevant for syntax now, isn't it?

[djspiewak]

But this is only relevant for syntax now, isn't it?

Yes, which significantly decreases (though does not entirely remove) the relevancy of interaction with prioritization.