[Haskell-cafe] Was simplified subsumption worth it for industry Haskell programmers?

[Ruben Astudillo]

I have nothing to say about the topic you propose on here. But to avoid
having a fractured discussion, let's us all agree to answer on that reddit
thread instead of this mail thread. Otherwise some good arguments will be
repeated.

-- 
Rubén. (pgp: 1E88 3AC4 89EB FA22)

On 06-05-22 15:34, pareto optimal via Haskell-Cafe wrote:
> I originally posted on Reddit and the thread contains much debate and discussion.
> 
> I'm concerned the views of the mailing list and likely ghc devs might not be as represented there or discourse, so i'm also copying it here.
> 
> Copy-paste of post:
> 
> Warning: Long post
> 
> tl;dr
> 
> - simplified subsumption seems to make common code I write in industry clunky for no good reason in lots of places
> 
> - Are others concerned with motivating what seem like "pointless lambdas" to new hires or students for simple tasks?
> 
> - Are there more real world advantages that make these frequent annoyances worth it?
> 
> - How is quicklook impredicativity useful in industry?
> 
> The biggest advantage seems to be that laziness is more predictable.
> 
> However looking at [commits fixing simplified subsumption errors on github](https://github.com/search?q=simplified+subsumption+language%3Ahaskell&type=commits) I see very common patterns in industry Haskell now need an explicit lambda for "reasons" such as:
> 
>       readFreqSumFile :: (MonadSafe m) => FilePath -> m (FreqSumHeader, Producer FreqSumEntry m ())
>     - readFreqSumFile file = readFreqSumProd $ withFile file ReadMode PB.fromHandle
>     + readFreqSumFile file = readFreqSumProd $ withFile file ReadMode (\h -> PB.fromHandle h)
> 
> and:
> 
>     - toOrders <- asks _pdfConfToOrder
>     + toOrders <- asks (\r -> _pdfConfToOrder r)
> 
> And this typical use of id is no longer valid:
> 
>     instance MonadOrvilleControl IO where
>     -    liftWithConnection = id
>     -    liftFinally = id
>     +   liftWithConnection ioWithConn = ioWithConn
>     +   liftFinally ioFinally = ioFinally
> 
> On my $work codebase that means hundreds of changes that make our code worse with seemingly no benefit.
> 
> This case is addressed in the proposal, but seems to handwave this as:
> 
> > The benefit, in terms of programming convenience, is small.
> 
> From my perspective while updating my codebase, it certainly doesn't feel that way.
> 
> From the persective of onboarding new Haskell hires, it doesn't feel simpler. I envision a future teaching session like:
> 
> > student: This code looks correct but I get an error that means nothing to me of
> 
>      error:
>          • Couldn't match type: b0 -> b0 with: forall q. q -> q
>               Expected: p -> forall q. q -> q
>               Actual: p -> b0 -> b0
>          • In the first argument of ‘g’, namely ‘f’ In the expression: g f In an equation for ‘h’: h = g f | | h = g f | ^
> 
> > me: Ah, that's because of something called simplified subsumption which we'll cover much later.
> > me: For now, just know putting it in an explicit lambda fixes it when you notice a compile error like that.
> > me: Now lets try to move past that and get back to basic file reading and writing
> > student: oookkkay? (feeling unsure, disillusioned about Haskell requiring pointless ceremony and being overly complex for no seeming benefit)
> 
> Being a fan of and proponent of Haskell I think: If this complication is being added, surely something is made possible in return that gives more value.
> 
> This led me to [the proposal](https://github.com/ghc-proposals/ghc-proposals/blob/master/proposals/0287-simplify-subsumption.rst) and I found with simplified subsumption:
> 
> - Laziness characteristics and semantics of programs will be changed less, which I infer will lead to more predictable performance
> - I assume that simplifying a compiler step like this will speed up compile times and reduce space usage
> - Quick look impredicativity seems to be the real driving reason behind simplified subsumption and somehow makes dealing with very polymorphic code easier
> 
> At this point my thought is:
> 
> > Making highly polymorphic code simpler to write that isn't as typical in industry Haskell code in ways I can't determine without great effort was valued over "small incoveniences" that I'll run into daily
> 
> But, still wanting to give the benefit of the doubt I dive face first into [The proposal for Quicklook impredicativity](https://github.com/ghc-proposals/ghc-proposals/blob/master/proposals/0274-quick-look-impredicativity.rst).
> 
> Reading the whole thing, I still cannot ground this concept in real world terms that may effect me or that I could take advantage of.
> 
> So, I go to the paper [A quick look at impredicativity](https://www.microsoft.com/en-us/research/publication/a-quick-look-at-impredicativity/) and start reading many things I don't fully understand.
> 
> Running out of energy, I start skimming and finally find some examples in section 10 APPLICATIONS.
> 
> I see an example with gZipWithM that I still don't understand. Further down I see reference to pieces of code updated in Streamly that take advantage of quick look polymorphism and wonder why the real world example wasn't included and explained.
> 
> So, i'm left frustrated with "simplified" subsumption and posting here for help answering:
> 
> - Are others in the same boat?
> - Are there advantages i'm not seeing?
> - Can we use my reflection to improve industry/academic communication?
> 
> And finally, any revant commentary surrounding this I may be oblivious to.
> 
> Thanks!
> 
> 
> _______________________________________________
> 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.