+1 On Fri, Apr 25, 2014 at 12:18 PM, Nathan Bouscal wrote:

+1

On Fri, Apr 25, 2014 at 12:14 PM, Niklas Haas wrote:

...

...

The previous discussion about methods on Either had some mention of adding bifunctors to base, but no one wrote up the details. So I've taken it upon myself to do so.

The following proposal is to add some modules of the bifunctors [1]

...

to base, namely:

Data.Bifunctor Data.Bifoldable Data.Bitraversable

These modules contain classes and functions for working with types similar to those identified by Functor, Foldable and Traversable, except that

On Fri, 25 Apr 2014 12:02:33 -0400, Dan Doel wrote: package there

...

are parameterized by two 'element types' so to speak.

The advantages of this change are among the following:

These are the right abstractions for many operations. For instance, Arrow is often recommended if someone wants to map over both sides (or the left side) of a pair. In fact, I'd wager that it is the single most common reason for recommending use of Arrow. But this is not really what Arrow was designed to accomplish. This is exactly what Bifunctor is for, though, and it abstracts over this kind of operation with pairs, Either, and in my experience many custom data types.

Placement in base gives a better opportunity for people to find these right abstractions. If someone goes into the documentation for Data.Either looking for a way to map both parameters, they will not, of course, be directed to the bifunctors package, even though it provides a good means of doing what they want. If Bifunctor were in base, the documentation for Either would note that it is one.

Some things to consider:

The API of the modules will shrink a bit due to Applicative becoming a superclass of Monad in 7.10. There is no reason for a separate bitraverse and bimapM and so on. Some things will likely be renamed, as well; bisequenceA => bisequence, for instance.

The 'first' and 'second' functions in Data.Bifunctor overlap with Arrow. This actually means that they are a drop-in replacement for the commonly suggested misuse of Arrow.

None of the dependencies of the bifunctors package are needed by the modules in question. They are used for other modules, or as part of an arbitrary decision of where to put an instance. For example, the tagged dependency is used to give instances for Tagged, but these could easily be moved into the tagged package if base were to adopt these classes.

+1, these are low-cost abstractions that definitely could use some more exposure and love, and with the 7.10 change in particular moving around the Applicative/Monad functions either way, it seems like a good time to refactor the Foldable/Traversable stuff and bring this in. _______________________________________________ Libraries mailing list Libraries@haskell.org http://www.haskell.org/mailman/listinfo/libraries

_______________________________________________ Libraries mailing list Libraries@haskell.org http://www.haskell.org/mailman/listinfo/libraries

Biapplicative is left out because unlike the other two, it is underdetermined. There are at least two forms of biapplicative point. There are however, no such headaches with Bifunctor/Bifoldable/Bitraversable, which get nailed down hard by the laws. A very enthusiastic +1 on the proposal from me. -Edward

On Apr 25, 2014, at 12:28 PM, Casey McCann wrote:

Was Biapplicative left out of the proposal because of dependencies? It seems at least as useful as the others.

- C.

...

On Fri, Apr 25, 2014 at 12:02 PM, Dan Doel wrote: The previous discussion about methods on Either had some mention of adding bifunctors to base, but no one wrote up the details. So I've taken it upon myself to do so.

The following proposal is to add some modules of the bifunctors [1] package to base, namely:

Data.Bifunctor Data.Bifoldable Data.Bitraversable

These modules contain classes and functions for working with types similar to those identified by Functor, Foldable and Traversable, except that there are parameterized by two 'element types' so to speak.

The advantages of this change are among the following:

These are the right abstractions for many operations. For instance, Arrow is often recommended if someone wants to map over both sides (or the left side) of a pair. In fact, I'd wager that it is the single most common reason for recommending use of Arrow. But this is not really what Arrow was designed to accomplish. This is exactly what Bifunctor is for, though, and it abstracts over this kind of operation with pairs, Either, and in my experience many custom data types.

Placement in base gives a better opportunity for people to find these right abstractions. If someone goes into the documentation for Data.Either looking for a way to map both parameters, they will not, of course, be directed to the bifunctors package, even though it provides a good means of doing what they want. If Bifunctor were in base, the documentation for Either would note that it is one.

Some things to consider:

The API of the modules will shrink a bit due to Applicative becoming a superclass of Monad in 7.10. There is no reason for a separate bitraverse and bimapM and so on. Some things will likely be renamed, as well; bisequenceA => bisequence, for instance.

The 'first' and 'second' functions in Data.Bifunctor overlap with Arrow. This actually means that they are a drop-in replacement for the commonly suggested misuse of Arrow.

None of the dependencies of the bifunctors package are needed by the modules in question. They are used for other modules, or as part of an arbitrary decision of where to put an instance. For example, the tagged dependency is used to give instances for Tagged, but these could easily be moved into the tagged package if base were to adopt these classes.

Discussion period: 2 weeks

[1] http://hackage.haskell.org/package/bifunctors

_______________________________________________ Libraries mailing list Libraries@haskell.org http://www.haskell.org/mailman/listinfo/libraries

---

Libraries mailing list Libraries@haskell.org http://www.haskell.org/mailman/listinfo/libraries

for :: (Traversablehttp://hackage.haskell.org/package/base-4.7.0.0/docs/Data-Traversable.html#t... t, Applicativehttp://hackage.haskell.org/package/base-4.7.0.0/docs/Control-Applicative.htm... f) => t a -> (a -> f b) -> f (t b) Unfortunate or not, bifor follows the existing convention from Traversablein base. Given two candidate conventions, I'd choose to err on the side of what we are already doing elsewhere in the library rather than devise and explain a new convention for little margin gain and much added confusion, while simultaneously hitting existing users with type errors / semantics changes. On Fri, Apr 25, 2014 at 1:14 PM, Andreas Abel wrote:

-----BEGIN PGP SIGNED MESSAGE----- Hash: SHA1

A general +1 from me.

To start the bikeshedding, let's have a look at the names in Data.Bitraversable:

bitraverse :: Applicative f => (a -> f c) -> (b -> f d) -> t a b -> f (t c d) bisequenceA:: Applicative f => t (f a) (f b) -> f (t a b) bimapM :: Monad m => (a -> m c) -> (b -> m d) -> t a b -> m (t c d) bisequence :: Monad m => t (m a) (m b) -> m (t a b) bifor :: (Bitraversable t, Applicative f) => t a b -> (a -> f c) -> (b - -> f d) -> f (t c d) biforM :: (Bitraversable t, Monad m) => t a b -> (a -> m c) -> (b -> m d) -> m (t c d)

I am unhappy about the name "bifor".

forM = flip mapM, so the right generalizations would be

for = flip map bifor = flip bimap

forA = flip mapA = flip traverse biforA = flip bimapA = flip bitraverse

Unfortunately, "for" was taken for "flip traverse" by Data.Traversable. Instead "forA" should have been used.

In my own code, I actually define

for = flip map

and use it in situations like

let ys = for xs $ \ x -> long function body

(I never use "for" from Data.Traversable anyway, since I have not needed yet an Applicative which is not a Monad.)

Proposal: use "biforA" instead of "bifor".

Cheers, Andreas

On 25.04.2014 18:36, Edward Kmett wrote:

...

Biapplicative is left out because unlike the other two, it is underdetermined. There are at least two forms of biapplicative point. There are however, no such headaches with Bifunctor/Bifoldable/Bitraversable, which get nailed down hard by the laws.

A very enthusiastic +1 on the proposal from me.

-Edward

...

On Apr 25, 2014, at 12:28 PM, Casey McCann wrote:

Was Biapplicative left out of the proposal because of dependencies? It seems at least as useful as the others.

- C.

...

On Fri, Apr 25, 2014 at 12:02 PM, Dan Doel wrote: The previous discussion about methods on Either had some mention of adding bifunctors to base, but no one wrote up the details. So I've taken it upon myself to do so.

The following proposal is to add some modules of the bifunctors [1] package to base, namely:

Data.Bifunctor Data.Bifoldable Data.Bitraversable

These modules contain classes and functions for working with types similar to those identified by Functor, Foldable and Traversable, except that there are parameterized by two 'element types' so to speak.

The advantages of this change are among the following:

These are the right abstractions for many operations. For instance, Arrow is often recommended if someone wants to map over both sides (or the left side) of a pair. In fact, I'd wager that it is the single most common reason for recommending use of Arrow. But this is not really what Arrow was designed to accomplish. This is exactly what Bifunctor is for, though, and it abstracts over this kind of operation with pairs, Either, and in my experience many custom data types.

Placement in base gives a better opportunity for people to find these right abstractions. If someone goes into the documentation for Data.Either looking for a way to map both parameters, they will not, of course, be directed to the bifunctors package, even though it provides a good means of doing what they want. If Bifunctor were in base, the documentation for Either would note that it is one.

Some things to consider:

The API of the modules will shrink a bit due to Applicative becoming a superclass of Monad in 7.10. There is no reason for a separate bitraverse and bimapM and so on. Some things will likely be renamed, as well; bisequenceA => bisequence, for instance.

The 'first' and 'second' functions in Data.Bifunctor overlap with Arrow. This actually means that they are a drop-in replacement for the commonly suggested misuse of Arrow.

None of the dependencies of the bifunctors package are needed by the modules in question. They are used for other modules, or as part of an arbitrary decision of where to put an instance. For example, the tagged dependency is used to give instances for Tagged, but these could easily be moved into the tagged package if base were to adopt these classes.

Discussion period: 2 weeks

[1] http://hackage.haskell.org/package/bifunctors

- -- Andreas Abel <>< Du bist der geliebte Mensch.

Department of Computer Science and Engineering Chalmers and Gothenburg University, Sweden

andreas.abel@gu.se http://www2.tcs.ifi.lmu.de/~abel/ -----BEGIN PGP SIGNATURE----- Version: GnuPG v1.4.11 (GNU/Linux) Comment: Using GnuPG with Thunderbird - http://www.enigmail.net/

iEYEARECAAYFAlNamAEACgkQPMHaDxpUpLM43gCcCZJS6y4OIYtSU+RifojGKLha C6EAoOQUhUKbJs3Dt9OZ6ryK97cGrCji =X8HU -----END PGP SIGNATURE-----

I retract my bikeshedding attempt. On 25.04.14 8:18 PM, Dan Doel wrote:

​In my code, I use (<&>) from lens when I need a flipped fmap:

let ys = xs <&> \x -> ...

and do use Applicative for.

Maybe & and <&> become standard at some point, and then I am maybe fine to use them. Only that "<&>", unlike "for", does not evoke any association to iteration in my brain. -- Andreas Abel <>< Du bist der geliebte Mensch. andreas.abel@ifi.lmu.de http://www2.tcs.ifi.lmu.de/~abel/

On Mon, Apr 28, 2014 at 6:28 AM, João Cristóvão wrote:

...

The previous discussion about methods on Either had some mention of adding bifunctors to base, but no one wrote up the details

I was implicitly leaving that to someone more qualified than me ;)

This proposal looks great to me, and I agree that it is the right abstraction for functions like mapLeft and such, instead of Arrows.

Still, some of the underlying assumptions made are not entirely clear to me:

...

If someone goes into the documentation for Data.Either looking for a way to map both parameters, they will not, of course, be directed to the bifunctors package, even though it provides a good means of doing what they want.

I'm perfectly happy to say we should go through and put in documentation in Data.Either and Data.Maybe talking about how they can be used with the respective Data.Bifunctor and Control.Arrow combinators. That seems like an obvious win with no downside. So, what you are proposing is that mapLeft and such are _not_ standalone

functions in Data.Either, but instead we just add documentation refering them to bifunctors as the correct abstraction, a merge of my "Proposal 1" with "Proposal 8". Although I'm much more confortable with this than forwarding user to arrows, I still see a point in adding this basic functions in Data.Either.

I think it is worth voting on this independently of the Data.Either additions, and would support adding the bifunctors machinery regardless of what we do with Data.Either.

It is my understanding that Data.List has a very complete API mainly for historical reasons, and that nowadays it is recomended to use Foldable and Traversable for many of those operations. The same applies to Data.Maybe. But a novice user will then look at Data.Either and Data.Tuple, and think: are these second class citizens? I guess this is an entirely different discussion, that happend many times in the past with no relevant consensus...

The fact that "lists get their own special API" is really the exception not the rule in base. In fact, many types exported from Data.List will actually be generalizing to avoid creating dozens of import conflicts with the Prelude as part of the 7.10 transition. So pointing there to say here is an example of something more monomorphic example than elsewhere is a bad idea. I'm not in a hurry to see a huge proliferation of names for one-off purposes. That leaves me personally rather -1 on the majority of the Data.Either proposal. An additional question: would the arrow operators then be defined in terms

of bifunctors?

The arrow combinators and the bifunctor combinators are not definable in terms of one another except in special cases. -Edward