[Haskell-cafe] Improving MTL instances (was: Overlapping/Incoherent instances)
Ryan Ingram
<jgmo...@cecs.pdx.edu> wrote:
> Indeed - MTL seems to have been rewritten at some point in the past to
> prefer exhaustive enumeration to overlap.
Indeed, and I actually think this is a weakness of the current
implementation. Anyone who comes up with a new transformer that
provides different functionality than what is there needs to
explicitly provide all the relevant instances, instead of letting
MonadTrans do its thing.
Consider MonadPrompt (shameless plug, it's on hackage!) In order to
be fully interoperable with the MTL I'd need to write instances for
MonadState, MonadReader, MonadWriter, MonadError, and MonadCont for
PromptT. These are unavoidable, although for monads with a "simple
enough" interface, such as State, everything can be accomplished with
"lift".
But I also need to provide the same boilerplate instances for every
other monad transformer in the package to give them instances of
MonadPrompt. And MonadPrompt *does* have a "simple enough" interface
that it could be accomplished trivially with "lift".
And this ignores interacting with any other transformer library!
Anyone who uses MonadPrompt along with another transformer (like
DatabaseT in the PostgreSQL library) needs to write any instances they
care about themselves, which adds to the difficulty in using the
libraries together.
Of course, the point of this message isn't just to complain. The
overlap implementation was abhorrent and it *is* better now than it
was before. But perhaps there is an abstraction we are missing that
would allow for better interoperability. For example, the
type-compose library documentation at
http://haskell.org/haskellwiki/TypeCompose mentions that (f :. g) is
an applicative functor if both f and g are applicative functors, which
means there is a generic "transformer" for all applicative functors!
The presense of >>=/join for monads make this more difficult, although
there is the "product" definition:
> newtype Product m n a = Prod { runProd :: m (Either a (Product n m a)) }
which handles nesting joins by just nesting the monads recursively.
But in this case it is up to the user to figure out how to untangle
the spaghetti created, so that's no good.
So, does anyone have any good ideas for improving the interoperability of MTL?
-- ryan
_______________________________________________
Haskell-Cafe mailing list
Haskel...@haskell.org
http://www.haskell.org/mailman/listinfo/haskell-cafe
Don Stewart
> On Mon, Oct 13, 2008 at 2:04 AM, J. Garrett Morris
> <jgmo...@cecs.pdx.edu> wrote:
> > Indeed - MTL seems to have been rewritten at some point in the past to
> > prefer exhaustive enumeration to overlap.
>
> Indeed, and I actually think this is a weakness of the current
> implementation. Anyone who comes up with a new transformer that
> provides different functionality than what is there needs to
> explicitly provide all the relevant instances, instead of letting
> MonadTrans do its thing.
I just want to make one small point here encouraging people to try out
new 'mtl' libraries. There are lots of *new* monad libraries,
The new & improved mtl:
* monadLib http://hackage.haskell.org/cgi-bin/hackage-scripts/package/monadLib
As well as,
* mmtl http://hackage.haskell.org/cgi-bin/hackage-scripts/package/mmtl
* mtl-tf http://hackage.haskell.org/cgi-bin/hackage-scripts/package/mtl-tf
and the huge,
* category-extras http://hackage.haskell.org/cgi-bin/hackage-scripts/package/category-extras
'mtl' isn't magic in any way. it isn't distributed with ghc, so we can
let the (library) free market pick a winner.
Now, if the monadLib guys would advertise their work a bit more
effectively...
-- Don
J. Garrett Morris
> Of course, the point of this message isn't just to complain. The
> overlap implementation was abhorrent and it *is* better now than it
> was before.
I'm curious what you find abhorrent about the overlap implementation
that was there before - in particular, it seems like it was designed
to handle both the combinatorial explosion and the corner cases you
mentioned. Did you find writing the MonadTrans instances unpleasant?
Was it the presence of overlapping instances at all?
On Mon, Oct 13, 2008 at 12:35 AM, Don Stewart <do...@galois.com> wrote:
> I just want to make one small point here encouraging people to try out
> new 'mtl' libraries. There are lots of *new* monad libraries,
Without knowing better myself: do any of these libraries address the
issue Ryan's brought up? I know that monadLib takes the same approach
the MTL does to this, and so is likely to have the same difficulties.
/g
--
I am in here
Stephen Hicks
> On Mon, Oct 13, 2008 at 2:04 AM, J. Garrett Morris
> <jgmo...@cecs.pdx.edu> wrote:
>> Indeed - MTL seems to have been rewritten at some point in the past to
>> prefer exhaustive enumeration to overlap.
>
> Indeed, and I actually think this is a weakness of the current
> implementation. Anyone who comes up with a new transformer that
> provides different functionality than what is there needs to
> explicitly provide all the relevant instances, instead of letting
> MonadTrans do its thing.
(First of all, sorry for the double reply.) I'm certainly way out of
my depth here, but would something like associated classes help here?
I'm imagining something like this (I'm sure my syntax is all wrong, though):
> class TypedMonad m where
> class MonadType m
> instance (MonadTrans m, TypedMonad n) => (MonadType n) (m n)
So then you could write something like
> instance Monad m => TypedMonad (ReaderT i m) where
> class MonadType (ReaderT i m) = MonadReader
and likewise for Reader, Writer(T), State(T), IO, etc... Then, for instance
> instance MonadWriter (StateT s (ReaderT r (WriterT w IO)))
is fully-automatic... Or wouldn't this work, at least once associated
classes is implemented?
Cheers,
steve
Iavor Diatchki
The currently released version of monadLib does not use overlapping
instances, indeed.
However, in the monadLib repo (http://github.com/yav/monadlib) there
is a file called "MonadLib4.hs" which contains a version of the
library that is implemented with overlapping instances, so you can
play around with it. I'd be interested to get feedback on this
version of moandLib.
-Iavor
Luke Palmer
> Of course, the point of this message isn't just to complain. The
> overlap implementation was abhorrent and it *is* better now than it
> was before. But perhaps there is an abstraction we are missing that
> would allow for better interoperability. For example, the
> type-compose library documentation at
> http://haskell.org/haskellwiki/TypeCompose mentions that (f :. g) is
> an applicative functor if both f and g are applicative functors, which
> means there is a generic "transformer" for all applicative functors!
> The presense of >>=/join for monads make this more difficult, although
> there is the "product" definition:
>
>> newtype Product m n a = Prod { runProd :: m (Either a (Product n m a)) }
>
> which handles nesting joins by just nesting the monads recursively.
> But in this case it is up to the user to figure out how to untangle
> the spaghetti created, so that's no good.
>
> So, does anyone have any good ideas for improving the interoperability of MTL?
http://sneezy.cs.nott.ac.uk/fplunch/weblog/?p=111
This was on planet haskell a little over a month ago. It describes
how any monad whose operations look like f (m a) -> m a for some
functor f can be automatically lifted. If it's possible to phrase a
basis for operations on some transformer this way and then provide
"adapters" for ease of use, that would be one excellent way to improve
interoperability. Some caveats are mentioned in the post...
Luke
Henning Thielemann
> On Mon, Oct 13, 2008 at 2:04 AM, J. Garrett Morris
> <jgmo...@cecs.pdx.edu> wrote:
>> Indeed - MTL seems to have been rewritten at some point in the past to
>> prefer exhaustive enumeration to overlap.
>
> Indeed, and I actually think this is a weakness of the current
> implementation. Anyone who comes up with a new transformer that
> provides different functionality than what is there needs to
> explicitly provide all the relevant instances, instead of letting
> MonadTrans do its thing.
I long thought that it is unnecessary use of type system extensions to
require multi-parameter type classes for simple monads and its
transformer versions. I thought it would be enough to have atomar monads
like ST, IO and Identity, and monads like State, Reader, Writer,
Continuation can be offered exclusively in the transforming variant.
(State s a) would have to be defined as (StateT s Identity a) instead.
This way MonadState, MonadReader and the other classes become
unnecessary. However, 'lift' remains important with this design.
wren ng thornton
> I long thought that it is unnecessary use of type system extensions to
> require multi-parameter type classes for simple monads and its
> transformer versions. I thought it would be enough to have atomar monads
> like ST, IO and Identity, and monads like State, Reader, Writer,
> Continuation can be offered exclusively in the transforming variant.
> (State s a) would have to be defined as (StateT s Identity a) instead.
> This way MonadState, MonadReader and the other classes become
> unnecessary. However, 'lift' remains important with this design.
Doing it that way removes the polymorphism that MonadState, MonadReader,
etc offer to clients. For example, the backwards-state monad[1] is a
MonadState but not a StateT (without extra plumbing). There are other
examples which don't even change the semantics. It seems a shame to
force these implementations to give different names for "the same"
functions. Are MPTCs onerous? They'll be in haskell-prime afterall. Of
course, the fundeps are another matter entirely...
[1]
http://luqui.org/blog/archives/2008/08/10/mindfuck-the-reverse-state-monad/
--
Live well,
~wren
Antoine Latter
>
> Doing it that way removes the polymorphism that MonadState, MonadReader, etc
> offer to clients. For example, the backwards-state monad[1] is a MonadState
> but not a StateT (without extra plumbing). There are other examples which
> don't even change the semantics. It seems a shame to force these
> implementations to give different names for "the same" functions. Are MPTCs
> onerous? They'll be in haskell-prime afterall. Of course, the fundeps are
> another matter entirely...
>
Slightly off topic - if you do make your backwards-state monad an
instance on MonadState be careful not to use
Control.Monad.State.Class.modify - executing this falls into a black
hole for the backwards-state monad.
-Antoine