Cats: Consider Reader-style instance for MonoidK of Kleisli

[SystemFw]

The current instance of MonoidK for Kleisli (which I will refer to as Endo instance) looks like:

implicit def instance[F[_]](implicit M: Monad[F]): MonoidK[λ[α => Kleisli[F, α, α]]]

as opposed to what I'd call a Reader instance :

implicit def instance[F[_], A](implicit F: SemigroupK[F]): SemigroupK[Kleisli[F, A, ?]]

which imo should be preferred.

Concrete, driving use case
In http4s, we need to model an http service as Kleisli[OptionT[F, ?], Request, Response], and the proposed instance would allow composing services with <+>, getting the correct prioritised choice behaviour.

Rationale
An instance for Kleisli[A, ?] is more general than one for λ[α => Kleisli[F, α, α]], and therefore should be _the_ instance, using a newtype for the other one. This is consistent with both Scalaz and Haskell, where A => B gets instances and A => A gets newtyped as Endo, and also internally consistent with the other instances for Kleisli, which are prevalently Reader style.

There is some discussion about this in #247, but I believe the use case in http4s is important enough to reconsider

@rossabaker

[kailuowang]

cc @edmundnoble @ceedubs @non

[SystemFw]

I can obviously open a PR as well if this gets approved

[tpolecat]

Agree. The "EndoT" instance is interesting but I don't think it should be the default one. In scalaz I believe you get there via Endomorphic[Kleisli[F, ?, ?], A].

[rossabaker]

Before http4s is used as the justification for this, @aeons is changing the HttpService type as described to ensure it works as well on the full project as it worked in my REPL. But I think this is the right thing to do regardless.

@ChristopherDavenport offered to implement it if we decide to proceed.

[LukaJCB]

I also agree that the second instance is more useful in general. I'm a bit afraid of too many breaking changes in RC1 but I think this one should be quite minor and not a big issue for most :)

[kailuowang]

The point of 1.0.0-MF is to identify breaking changes needed by the community before we go 1.0.

[johnynek]

big +1 on this.

[kailuowang]

Looks like we should proceed with it. @ChristopherDavenport still interested?

[rossabaker]

• Ticket that would render this moot: #1932
• Related, stalled ticket: #1098

[edmundnoble]

I'm worried that we can't change this instance in place without changing behavior silently for users.

[johnynek]

I think silently breaking users is possible but probably rare. Someone would have to have an endo which also has a MonoidK on the inner type, and that MonoidK would need to be in scope.

I'd be surprised if even one person will be in this boat.

[edmundnoble]

Don't Either and Option fall under this? That seems common.

[ChristopherDavenport]

I'm interested, and will get an initial version up this evening.

[peterneyens]

I can update #1098 if needed?

[SystemFw]

Yeah it seems like most of the code is already in #1098

[ChristopherDavenport]

Cool, I can back down in favor of an existing PR.

[peterneyens]

Added in #1098