On Fri, Oct 14, 2016 at 6:00 PM, Matt wrote:

The mtl technique subsumes the free monad technique. if you have a term:

getCurrentTime :: MonadClock m => m UTCTime

Then you can *use* that function as anything that satisfies the constraint. Given an `IO` instance, that can just get the current time: `getCurrentTime :: IO UTCTime`. Given a mock instance, that can be `getCurrentTime :: MockClock UTCTime`. Given an instance on Free, you'd have `getCurrentTime :: Free CurrentTimeF UTCTime`

Thanks Matt. I think that was a nice explanation. Right now I'm focusing on the composition and natural transformations of free-monads, but I still haven't checked the MTL approach.

I generally find it more pleasant to write functions in mtl style. If you're after more concrete guarantees on the DSL you're building and see yourself doing a lot of introspection and optimization, then a Free monad approach fits the bill.

I definitely like monad transformers. But I guess I'd have to explain the specific case in another thread.

Matt Parsons

On Fri, Oct 14, 2016 at 11:35 AM, Damian Nadales wrote:

...

Hi,

I was looking into free monads for designing a DSL for describing scenarios of the form: scenario = do aId <- createA b0Id <- createB id b1Id <- createB id link b0 b1

In our company we use a graph database, and currently we're setting up the test data using raw queries O.O So I wanted to come up with a better abstraction, and also enable us to do property based testing (by testing on random scenarios like the one above).

Anyway, I ran into this presentation: http://www.slideshare.net/jdegoes/mtl-versus-free http://degoes.net/articles/modern-fp-part-2

In which monad transformers and free monads are compared. Do you have any experience using any of these approaches. If so would you mind sharing? ;)

Thanks! Damian _______________________________________________ Haskell-Cafe mailing list To (un)subscribe, modify options or view archives go to: http://mail.haskell.org/cgi-bin/mailman/listinfo/haskell-cafe Only members subscribed via the mailman list are allowed to post.

My gut feeling is that an an indirect function call for every (>>=), with many calls to `lift` each time, would make a deep monad transformer stack much more expensive. A free monad approach seems to be more sensible to me. But maybe GHC is doing a better job optimizing this than I would think?

Well, my gut feeling would be that free monads are more expensive than a monad transformer stack with <= 3 layers. After all, for free/operational monads, the compiler has to allocate a closure for every second argument of `>>=` as part of the monad data structure. There are less opportunities for inlining, because the interpretation of the monad is not fixed, and only decided late at runtime. That said, the above arguments are no proof. I would be interested in performance measurements as well. Maybe there is a way to generalize the "GHC state hack" to a "free monad hack"? The basic ansatz would be the ability to mark some closures as "Only entered once, may duplicate work". Best regards, Heinrich Apfelmus -- http://apfelmus.nfshost.com Joachim Breitner wrote:

Hi,

Am Freitag, den 14.10.2016, 17:35 +0200 schrieb Damian Nadales:

...

Do you have any experience using any of these approaches. If so would you mind sharing? ;)

I don’t have an answer to contribute, but I would be very interested in hearing about experiences in terms of their relative runtime performance.

My gut feeling is that an an indirect function call for every (>>=), with many calls to `lift` each time, would make a deep monad transformer stack much more expensive. A free monad approach seems to be more sensible to me. But maybe GHC is doing a better job optimizing this than I would think?

So if you have any number-supported evidence about this, possibly from a real-world application where you tried to use one or the other, please share it with us!

Thanks, Joachim

------------------------------------------------------------------------

_______________________________________________ Haskell-Cafe mailing list To (un)subscribe, modify options or view archives go to: http://mail.haskell.org/cgi-bin/mailman/listinfo/haskell-cafe Only members subscribed via the mailman list are allowed to post.

There is a free monad benchmark: https://rawgit.com/feuerbach/freemonad-benchmark/master/results.html Not very good for the free monads. But it is done only with a single transformer and for a single state. I don´t know how the MTL performance degrades when the transformer stack grows. 2016-10-15 15:49 GMT+02:00 Joachim Breitner :

Hi,

Am Freitag, den 14.10.2016, 17:35 +0200 schrieb Damian Nadales:

...

Do you have any experience using any of these approaches. If so would you mind sharing? ;)

I don’t have an answer to contribute, but I would be very interested in hearing about experiences in terms of their relative runtime performance.

My gut feeling is that an an indirect function call for every (>>=), with many calls to `lift` each time, would make a deep monad transformer stack much more expensive. A free monad approach seems to be more sensible to me. But maybe GHC is doing a better job optimizing this than I would think?

So if you have any number-supported evidence about this, possibly from a real-world application where you tried to use one or the other, please share it with us!

Thanks, Joachim

-- Joachim “nomeata” Breitner mail@joachim-breitner.de • https://www.joachim-breitner.de/ XMPP: nomeata@joachim-breitner.de • OpenPGP-Key: 0xF0FBF51F Debian Developer: nomeata@debian.org _______________________________________________ Haskell-Cafe mailing list To (un)subscribe, modify options or view archives go to: http://mail.haskell.org/cgi-bin/mailman/listinfo/haskell-cafe Only members subscribed via the mailman list are allowed to post.

-- Alberto.

I was thinking, besides the evaluation of performance, the simplicity of the approach is also important ("developer time is more expensive than CPU time" anyone?). Note that I said simple and not easy ;) I guess this aspect is a rather subjective one, but maybe there are elements that can be intuitively quantified. Right now I'm playing with free monads and MTL, to have an idea which one seems simpler to me. On Mon, Oct 17, 2016 at 8:06 PM, Julian wrote:

On 15/10/16 15:49, Joachim Breitner wrote:

...

Hi,

Am Freitag, den 14.10.2016, 17:35 +0200 schrieb Damian Nadales:

...

Do you have any experience using any of these approaches. If so would you mind sharing? ;)

I don’t have an answer to contribute, but I would be very interested in hearing about experiences in terms of their relative runtime performance.

My gut feeling is that an an indirect function call for every (>>=), with many calls to `lift` each time, would make a deep monad transformer stack much more expensive. A free monad approach seems to be more sensible to me. But maybe GHC is doing a better job optimizing this than I would think?

So if you have any number-supported evidence about this, possibly from a real-world application where you tried to use one or the other, please share it with us!

There's a paper from Oleg discussing "Freer Monads, More Extensible Effects": http://okmij.org/ftp/Haskell/extensible/more.pdf

The conclusion there seems to be that the EE approach is more "efficient". But you'll have to look at the concrete performance cases and data yourself to make a judgement.

_______________________________________________ Haskell-Cafe mailing list To (un)subscribe, modify options or view archives go to: http://mail.haskell.org/cgi-bin/mailman/listinfo/haskell-cafe Only members subscribed via the mailman list are allowed to post.

Working with stacks like "ReaderT AppConfig (StateT AppState IO)" may be hard, but using MonadIO, MonadState, MonadReader etc. is much simpler. This session explains it well: https://www.youtube.com/watch?v=GZPup5Iuaqw I haven't yet tried Oleg's approach with Freer monads and extensible effects though, it looks very interesting. Especially when creating new types of effects because writing an "interpreter" seems easier than writing something like "MyEffectT / MyEffectMonad" and giving it all the required instances. On Wed, Oct 19, 2016 at 11:57 PM David McBride wrote:

I'll say that this conversation caused me to rewrite some networking code I had written in a free monad as mtl. It is faster and simpler, right? But transformer stacks with all their typeclasses are really hard to work with and I don't think it was worth it in the end.

On Wed, Oct 19, 2016 at 8:29 AM, Damian Nadales wrote:

I was thinking, besides the evaluation of performance, the simplicity of the approach is also important ("developer time is more expensive than CPU time" anyone?). Note that I said simple and not easy ;)

I guess this aspect is a rather subjective one, but maybe there are elements that can be intuitively quantified. Right now I'm playing with free monads and MTL, to have an idea which one seems simpler to me.

On Mon, Oct 17, 2016 at 8:06 PM, Julian wrote:

...

On 15/10/16 15:49, Joachim Breitner wrote:

...

Hi,

Am Freitag, den 14.10.2016, 17:35 +0200 schrieb Damian Nadales:

...

Do you have any experience using any of these approaches. If so would you mind sharing? ;)

I don’t have an answer to contribute, but I would be very interested in hearing about experiences in terms of their relative runtime performance.

My gut feeling is that an an indirect function call for every (>>=), with many calls to `lift` each time, would make a deep monad transformer stack much more expensive. A free monad approach seems to be more sensible to me. But maybe GHC is doing a better job optimizing this than I would think?

So if you have any number-supported evidence about this, possibly from a real-world application where you tried to use one or the other, please share it with us!

There's a paper from Oleg discussing "Freer Monads, More Extensible Effects": http://okmij.org/ftp/Haskell/extensible/more.pdf

The conclusion there seems to be that the EE approach is more "efficient". But you'll have to look at the concrete performance cases and data yourself to make a judgement.

_______________________________________________ Haskell-Cafe mailing list To (un)subscribe, modify options or view archives go to: http://mail.haskell.org/cgi-bin/mailman/listinfo/haskell-cafe Only members subscribed via the mailman list are allowed to post.

---

Haskell-Cafe mailing list To (un)subscribe, modify options or view archives go to: http://mail.haskell.org/cgi-bin/mailman/listinfo/haskell-cafe Only members subscribed via the mailman list are allowed to post.

_______________________________________________ Haskell-Cafe mailing list To (un)subscribe, modify options or view archives go to: http://mail.haskell.org/cgi-bin/mailman/listinfo/haskell-cafe Only members subscribed via the mailman list are allowed to post.

Can anyone comment on the use of Purescript-style effect monads as compared to MTL and Free? While I have not used them in practice, they seem to express the "intent" of monad composition a bit more directly than the approaches we use in Haskell. Cheers, Will

It seems that there are several advantages to the Purescript approach. For example, this catchException :: forall a e . (Error -> Eff e a) -> Eff (err :: EXCEPTION | e) a -> Eff e a would be unwieldy to express using typeclasses, requiring at least three constraints. I also find this style easier to read than constraints, as it requires no mental substitution. E.g. if I see (Foo m, Bar m, Bar n) => Baz -> -> m a When I get to the end of the type, I have to go back to the beginning to figure out what m is. I can't read left-to-right. This happens a lot with constraint-based monad composition. Another advantage is that the Purescipt example uses a concrete type, which is often easier to reason about than "ad-hoc" typeclass abstractions like MonadRandom. However, it looks like you still get the flexibility of ad-hoc typeclasses, because you get to pick any function that discharges the effect type in the given effect monad. Like I said, I have not used it, but these are what I've noticed from topical observation. Apologies for the formatting; copying that code example appears to have confused the iOS mail app. Cheers, Will

On Oct 19, 2016, at 12:26, Christopher Allen wrote:

It's not really more direct. It's an unordered collection of effects you can use. IME it's a less efficient mtl-style, but YMMV.

Taking an example from a PureScript tutorial:

func :: Eff (console :: CONSOLE, random :: RANDOM) Unit

Can just as easily be:

func :: (MonadConsole m, MonadGimmeRandom m) => m ()

(mangled name so it doesn't overlap with a real class)

There are other differences, but they haven't amounted to much for me yet.

Kmett's Quine has a good example of some homespun mtl-style: https://github.com/ekmett/quine

...

On Wed, Oct 19, 2016 at 12:17 PM, Will Yager wrote: Can anyone comment on the use of Purescript-style effect monads as compared to MTL and Free? While I have not used them in practice, they seem to express the "intent" of monad composition a bit more directly than the approaches we use in Haskell.

Cheers, Will _______________________________________________ Haskell-Cafe mailing list To (un)subscribe, modify options or view archives go to: http://mail.haskell.org/cgi-bin/mailman/listinfo/haskell-cafe Only members subscribed via the mailman list are allowed to post.

-- Chris Allen Currently working on http://haskellbook.com

On Wed, Oct 19, 2016 at 3:19 PM, Christopher Allen wrote:

Here catch is:

catch :: (Exception e, MonadCatch m) => m a -> (e -> m a) -> m a

In my mind, this is inferior to the Purescript example, because you have not discharged the MonadCatch instance. That is to say, the result monad continues to have a MonadCatch instance, even though we may want to explicitly express that we have *already* caught any exceptions, so the result type need not be a MonadCatch. In the Purescript example, the result monad no longer has the EXCEPTION effect, so you have demonstrated statically that the exception has already been caught! This is very useful. You can't express this in the style you're advocating (which, to be clear, is the style I use most of the time). In other words, the purescript approach has the benefits of both explicitly enumerating the entire monad transformer stack (you can statically determine exactly which effects are possible *and* which are not, and you can discharge individual layers of the stack) and of using typeclass constraints (you aren't typing out redundant information and you don't have to specify where in the monad transformer stack the effect is handled). It's the best of both worlds.

...

Another advantage is that the Purescipt example uses a concrete type

I'm not sure I understand. What are you saying is concrete here?

I should have been more clear; the type is polymorphic, but does not have any constraints outside of the type. This isn't a formal advantage so much as a psychological/syntactic advantage. As I said, it's inconvenient to have to mentally sub in constraints on the LHS of the "=>" into the type variables on the RHS of the "=>". The Purescript approach is easier to read for me, even though I've been using the typeclass approach for years. Like I said, I haven't actually used Purescript, so I'm sure there are additional use cases I'm missing. But this is what stands out. Cheers, Will

Oh fair enough, I usually call the partially concreted or lensy style "mtl-style" and then call MonadReader/MonadState "mtl library". I lean to the former and usually use the latter if it discharges something I need right then and there. I've only used Free for something I needed to aggressively simulate/mock, and even then, some production uses of Free have gotten defenestrated in favor of something more ordinary. On Wed, Oct 19, 2016 at 3:47 PM, Richard Wallace wrote:

I would call the approach in quine a lensy-mtl style. It's ok as far as it goes, but since you are using concrete environment values it isn't great if you want to do testing of things like database code without having a "real" backend hooked up. The typical approach then is to create your own type-class and instances

class MyBackend where ...

instance (MonadReader r m, HasDb r) => MyBackend m where ...

instance (MonadState s m, HasTestState s) => MyBackend m where ...

Of course, now our problem is that our module with this abstraction depends on the module with the db and the test state. Unless we create orphan instances, which I prefer to avoid. This is one area where I like the Free monad approach more because the interpreter can be built and composed with other interpreters in completely separate modules or packages because they are just values.

Rich

PS for the record, I don't strongly prefer the mtl style or the free monad style, I think they each have good qualities and bad and which one I choose tends to depend on other factors.

On Wed, Oct 19, 2016 at 10:26 AM, Christopher Allen wrote:

...

It's not really more direct. It's an unordered collection of effects you can use. IME it's a less efficient mtl-style, but YMMV.

Taking an example from a PureScript tutorial:

func :: Eff (console :: CONSOLE, random :: RANDOM) Unit

Can just as easily be:

func :: (MonadConsole m, MonadGimmeRandom m) => m ()

(mangled name so it doesn't overlap with a real class)

There are other differences, but they haven't amounted to much for me yet.

Kmett's Quine has a good example of some homespun mtl-style: https://github.com/ekmett/quine

On Wed, Oct 19, 2016 at 12:17 PM, Will Yager wrote:

...

Can anyone comment on the use of Purescript-style effect monads as compared to MTL and Free? While I have not used them in practice, they seem to express the "intent" of monad composition a bit more directly than the approaches we use in Haskell.

Cheers, Will _______________________________________________ Haskell-Cafe mailing list To (un)subscribe, modify options or view archives go to: http://mail.haskell.org/cgi-bin/mailman/listinfo/haskell-cafe Only members subscribed via the mailman list are allowed to post.

-- Chris Allen Currently working on http://haskellbook.com _______________________________________________ Haskell-Cafe mailing list To (un)subscribe, modify options or view archives go to: http://mail.haskell.org/cgi-bin/mailman/listinfo/haskell-cafe Only members subscribed via the mailman list are allowed to post.

-- Chris Allen Currently working on http://haskellbook.com