Re: [ghc-steering-committee] #583: HasField redesign, rec: accept

12 Oct 2023


      I'm now accepting this proposal.

On Mon, 9 Oct 2023 at 10:06, Arnaud Spiwack  wrote:
...
Dear all,
Absent votes, based on my current understanding of everybody's position,
I'll mark the proposal as accepted, without modification, on Wednesday
(11th October).
On Tue, 3 Oct 2023 at 11:52, Arnaud Spiwack 
wrote:
...
Simon votes A above. I vote A too (I think I actually feel pretty
strongly about this one).
On Tue, 3 Oct 2023 at 11:52, Arnaud Spiwack 
wrote:
...
Let's do a quick vote. All, which do you prefer?
A. Introduce the breakage immediately and change the argument order of
setField (current state of the proposal)
B. Introduce a new extension -XOverloadedRecordUpdateNew which would use
a setField with a different argument order, so that it can be used during
the transition period. Eventually OverloadedRecordUpdateNew becomes
deprecated and OverloadedRecordUpdate swaps the argument order. (Moritz's
proposal)
C. Keep the originally proposed setField argument order (nobody have
suggested this yet, but I'm including it for completeness)
On Tue, 3 Oct 2023 at 10:33, Simon Peyton Jones <
simon.peytonjones@gmail.com> wrote:
...
Sure.
Just to be clear, I'm not trying to set a precedent here!
Simon
On Tue, 3 Oct 2023 at 09:26, Moritz Angermann <
moritz.angermann@gmail.com> wrote:
...
I’m not going to stay in the way of this proposal. I’hope it’s
understandable if I recuse myself.
On Tue, 3 Oct 2023 at 4:18 PM, Simon Peyton Jones <
simon.peytonjones@gmail.com> wrote:
...
We have to balance the cost of mitigation against the cost of
breakage.
- Cost of breakage.  Adam has identified only two packages that
   woud, both of which can adapt promptly.  It's true that there is more code
   out there, but the absence on Hackage is strongly indicative that there is
   unlikely to be much.
   - Cost of mitigation.  Supporting both behaviours, having extra
   language flags, that must in turn themselves be deprecated away, are real
   costs.  The suck effort away from more productive activity.
   - I completely agree that if we had a classification,
   OverloadedRecordUpdate would have been classified as Experimental, and we
   would not be discussing deprecation cycles.  And signalling what is
   experimental and what is stable is a *primary *goal of the
   stability conversation we are having.  I agree strongly with Moritz on this
   point.
   - However the user manual (Section 6.5.11
   https://ghc.gitlab.haskell.org/ghc/doc/users_guide/exts/overloaded_record_up...)
   says "*EXPERIMENTAL* *This design of this extension may well
   change in the future. It would be inadvisable to start using this extension
   for long-lived libraries just yet*".  Short of an official
   classification mechanism, which we don't have yet, it seems hard to imagine
   a clearer statement than that.
My judgement: in this case we should forge ahead without introducing
OverloadedRecordNew and supporting both behaviours. (I have not even begun
to think about how hard that would be to implement.)  Let's spend our
precious cycles elsewhere.
Simon
On Tue, 3 Oct 2023 at 08:48, Moritz Angermann <
moritz.angermann@gmail.com> wrote:
> Arnaud,
>
> That is a good argument. Without an opt-in to the new syntax, they
> can't adapt. I guess this brings us to the extension lifecycle. A solution
> could be
> 1. Deprecate OverloadedRecordUpdate in favour of
> OverloadedRecordUpdateNew. The user can adapt their code and use a
> different Extension.
> 2. Eventually Have OverloadedRecordUpdateNew be a synonym for
> changed OverloadedRecordUpdate. And Deprecate OverloadedRecordUpdateNew
>     over a longer period of time.
>
> I think this actually shows the value a `--std=experimental` could
> have. If `OverloadedRecordUpdate` was behind `--std=experimental` breaking
> it would
> be permissible by design. And we would not have this discussion, and
> Adam would not need to try to find all the potential breaking codes. This
> is part
> of the proposal to make `--std=experimental` a thing, allow fast(er)
> paced iteration on the experimental side of the compiler, while being
> extremely
> explicit about it being experimental features that can break at any
> time and without warning.
>
> The deprecation process outlined above would be fore
> non-experimental features. They can still be changed, but with a more
> rigorous change evolution.
>
> I hope this helps to clarify my position?
>
> Best,
>  Moritz
>
>
> On Tue, 3 Oct 2023 at 15:31, Arnaud Spiwack 
> wrote:
>
>> @Moritz: very clear thank you, though I'd like to understand how
>> one could adapt during the transition period to an argument swap, would you
>> say? Secondary question: if they can't adapt, isn't there a risk of making
>> the breakage more painful by delaying it, and letting more people use
>> -XOverloadedRecordUpdate with the argument order which isn't the final one?
>>
>> @Adam: we have to recognise that Hackage is not all the Haskell
>> code. I expect Hackage to be a little more conservative in extensions than
>> end-user applications. As such we're not guaranteed an absence of breakage
>> (it's still probably quite small, but it's hard to quantify).
>>
>> On Tue, 3 Oct 2023 at 09:21, Adam Gundry 
>> wrote:
>>
>>> The backwards compatibility impact here is really extremely small,
>>> because OverloadedRecordUpdate is essentially unusable at present
>>> (as
>>> well as being explicitly documented as subject to change), so
>>> nobody
>>> uses it. While someone could implement a warning when the
>>> extension is
>>> turned on to say that it will change in the future, I'm not sure I
>>> see
>>> much point.
>>>
>>> I used
>>> https://hackage-search.serokell.io/?q=OverloadedRecordUpdate to
>>> look for occurrences of OverloadedRecordUpdate on Hackage, and I
>>> found
>>> precisely one pair of packages using it "for real":
>>>
>>> large-anon-0.3.0
>>> large-records-0.4
>>>
>>> These packages were created by my colleague Edsko in discussion
>>> with me,
>>> and will need to be amended when the proposal is implemented,
>>> because
>>> they want to take advantage of the new functionality.
>>>
>>> The other packages on Hackage containing the string
>>> OverloadedRecordUpdate are tools that reference all extensions:
>>>
>>> Cabal-syntax-3.10.1.0
>>> extensions-0.1.0.0
>>> fourmolu-0.14.0.0
>>> ghc-9.6.3
>>> ghc-boot-th-9.6.3
>>> ghc-exactprint-1.7.0.1
>>> ghc-hs-meta-0.1.2.0
>>> ghc-lib-9.6.2.20230523
>>> ghc-lib-parser-9.6.2.20230523
>>> hackport-0.8.4.0
>>> haskell-src-meta-0.8.12
>>> hindent-6.1.0
>>> hlint-3.6.1
>>> ormolu-0.7.2.0
>>>
>>> plus there are a few references in comments:
>>>
>>> lifx-lan-0.8.2
>>> optics-core-0.4.1.1
>>> pvector-0.1.1
>>> record-dot-preprocessor-0.2.16
>>> tztime-0.1.1.0
>>>
>>> While I generally agree with the "don't abruptly break existing
>>> code"
>>> position, in this case I don't think there is code out there to
>>> break.
>>>
>>> Adam
>>>
>>>
>>> On 03/10/2023 01:14, Moritz Angermann wrote:
>>> > Arnaud,
>>> >
>>> > thank you for your write up. And sorry that my view seems to be
>>> not
>>> > clear. Let me try to explain.
>>> > My position is: I'm against anything that _abruptly_ breaks
>>> existing
>>> > code. That's basically all there is to it.
>>> > Therefore I'm strongly against breaking changes without an
>>> appropriate
>>> > warning period (from the compiler).
>>> > I also strongly believe that most people do not read
>>> documentation, and
>>> > the authoritative messages are
>>> > the ones the compiler produces. As I've alluded to in a
>>> different email,
>>> > there are lots of people who work
>>> > on software as a 9-5 job, and don't spend their freetime
>>> tinkering and
>>> > playing with languages. They just
>>> > want to get the job done they are assigned. I have a lot of
>>> respect for
>>> > them. In my experience they don't
>>> > read compiler release announcements, or even go and read the
>>> compiler
>>> > documentation. They are given
>>> > the compiler version the company decided on for use in
>>> production.
>>> > That's the tool they use. And that tool
>>> > in my opinion needs to be rock solid, and not break easily
>>> across
>>> > versions. Thus yes, I would very much
>>> > like to see us not have breaking changes at all, but I can see
>>> that we
>>> > may need breaking changes
>>> > occasionally. In those cases I want it to be very visible to the
>>> > consumers that this breaking change is
>>> > coming towards them (compiler warnings). Giving them some time
>>> to adjust
>>> > (migration period), until the
>>> > breaking change happens.  Ultimately we should be able to
>>> compiler
>>> > existing code that compiles today
>>> > with at least the next compiler without it rejecting the code or
>>> needing
>>> > modifications to the code (-Werror
>>> > excluded).
>>> >
>>> > Thus what I'm arguing for is:
>>> > - Let's implement this backwards compatibility.
>>> > - Add compiler warnings about the arguments being swapped in a
>>> future
>>> > GHC version. For _a least_ one major release.
>>> > - Make the breaking change in a subsequent release.
>>> >
>>> > Alternatively I could also see:
>>> > - Adding compiler warnings now that the arguments will be
>>> swapped in a
>>> > future GHC version (for at least one major release).
>>> > - Implement the breaking change in a subsequent release.
>>> >
>>> > Either of those would be ok with me. Implementing a breaking
>>> change from
>>> > one version to the next, without an
>>> > appropriate deprecation/migration period (that is, the compiler
>>> will
>>> > warn loudly that changes are coming) is something
>>> > I am _very_ vehemently against.
>>> >
>>> > If the migration/deprecation warnings would provide a link to a
>>> GitHub
>>> > ticket or something where more information
>>> > can be found and maybe even a discussion could be had would
>>> probably
>>> > also be a good idea.
>>> >
>>> > I hope this helps clarify my position? If not, feel free to ask
>>> more,
>>> > I'm also happy to jump onto a call to explain my
>>> > position if needed.
>>> >
>>> > Best,
>>> >   Moritz
>>> >
>>> >
>>> > On Mon, 2 Oct 2023 at 23:31, Arnaud Spiwack <
>>> arnaud.spiwack@tweag.io
>>> > mailto:arnaud.spiwack@tweag.io> wrote:
>>> >
>>> >     Sorry I took a little bit of time to react to this, it was a
>>> lot to
>>> >     take in and I didn't have the mental space last week.
>>> >
>>> >     The only person that may have spoken against the current
>>> state of
>>> >     the proposal is Moritz. Yet I realise that I don't actually
>>> know,
>>> >     Moritz, what your position is.
>>> >
>>> >     To recap: to use -XOverloadedRecordUpdate in current GHC,
>>> you need
>>> >     to use -XRebindableSyntax and provide a setField function.
>>> In the
>>> >     new proposal you can use -XOverloadedRecordUpdate without
>>> >     -XRebindableSyntax, but when -XRebindableSyntax is on, the
>>> setField
>>> >     function that you have to provide has its argument swapped.
>>> The
>>> >     current documentation of OverloadedRecordUpdate has the
>>> following
>>> >     text at the top “*EXPERIMENTAL* /This design of this
>>> extension may
>>> >     well change in the future. It would be inadvisable to start
>>> using
>>> >     this extension for long-lived libraries just yet./”.
>>> >
>>> >     Now, I don't quite see how we could have a transition period
>>> that
>>> >     would allow a smooth transition there. There is no piece of
>>> code,
>>> >     with RebindableSyntax, that would compile before and after
>>> the
>>> >     change. So unless I'm missing something the position we can
>>> take as
>>> >     a committee can be either
>>> >     - Let's have the breakage without a transition period
>>> >     - Let's not make the breaking change ever and use the earlier
>>> >     argument order for set
>>> >
>>> >     Which one do you argue for, or am I missing another option?
>>> >
>>> >     On Sun, 24 Sept 2023 at 15:36, Eric Seidel >> >     mailto:eric@seidel.io> wrote:
>>> >
>>> >         I am in favor of this proposal.
>>> >
>>> >         On Thu, Sep 21, 2023, at 03:37, Arnaud Spiwack wrote:
>>> >          > Dear all.
>>> >          >
>>> >          > I submitted my recommendation 3 weeks ago, and only
>>> Simon has
>>> >         commented
>>> >          > yet. Please let me know your thoughts.
>>> >          >
>>> >          > On Wed, 6 Sept 2023 at 16:27, Arnaud Spiwack
>>> >         mailto:arnaud.spiwack@tweag.io>
>>> wrote:
>>> >          >> Dear all,
>>> >          >>
>>> >          >> Don't forget to opine here. To reiterate, I really
>>> don't
>>> >         expect the proposal to be controversial. The text of the
>>> >         proposal is rather long, but is made easy to read. So it
>>> >         shouldn't take too much of your time.
>>> >          >>
>>> >          >> /Arnaud
>>> >          >>
>>> >          >> On Thu, 31 Aug 2023 at 01:03, Simon Peyton Jones
>>> >         >> >         mailto:simon.peytonjones@gmail.com> wrote:
>>> >          >>> I support acceptance.
>>> >          >>>
>>> >          >>> Simon
>>> >          >>>
>>> >          >>> On Wed, 30 Aug 2023 at 16:09, Arnaud Spiwack
>>> >         mailto:arnaud.spiwack@tweag.io>
>>> wrote:
>>> >          >>>> Dear all,
>>> >          >>>>
>>> >          >>>> [ Proposal #583
>>> >         https://github.com/ghc-proposals/ghc-proposals/pull/583
>>> >         https://github.com/ghc-proposals/ghc-proposals/pull/583
>>> ]
>>> >          >>>>
>>> >          >>>> Our own Adam proposes to amend the design of the
>>> highly
>>> >         experimental OverloadedRecordUpdate extension as had been
>>> >         designed in proposal #158 [
>>> >
>>> https://github.com/ghc-proposals/ghc-proposals/blob/master/proposals/0158-re...
>>> <
>>> https://github.com/ghc-proposals/ghc-proposals/blob/master/proposals/0158-record-set-field.rst>
>>> ] and #405 [
>>> https://github.com/ghc-proposals/ghc-proposals/pull/405 <
>>> https://github.com/ghc-proposals/ghc-proposals/pull/405> ].
>>> >          >>>>
>>> >          >>>> Specifically, Adam proposes a modification of the
>>> type
>>> >         classes that would back the extension.
>>> >          >>>>
>>> >          >>>> In the previous design the HasField class is
>>> defined as a
>>> >         lens:
>>> >          >>>>
>>> >          >>>> class HasField (n :: k) r a | r n -> a
>>> >          >>>>   hasField :: r -> (a -> r, a)
>>> >          >>>>
>>> >          >>>> The proposal is to replace it by two classes
>>> (slightly
>>> >         simplified)
>>> >          >>>>
>>> >          >>>> class HasField (n :: k) r a | r n -> a
>>> >          >>>>   hasField :: r -> a
>>> >          >>>>
>>> >          >>>> class SetField (n::k) r a | r n -> a
>>> >          >>>>   modifyField :: (a -> a) -> r -> a
>>> >          >>>>   setField :: a -> r -> a
>>> >          >>>>
>>> >          >>>> This is originally motivated by some performance
>>> >         consideration: the prototype implementation of HasField
>>> as a
>>> >         lens can be very time consuming because instances of
>>> HasFields
>>> >         are generated eagerly at record definition sites,
>>> whereas the
>>> >         simple HasField instances can simply reuse the selectors
>>> already
>>> >         generated by GHC. But a lot of thoughts have been put
>>> into the
>>> >         new design, and my summary can certainly not do it
>>> justice: the
>>> >         proposal is very well argumented.
>>> >          >>>>
>>> >          >>>> A point I'll make here is that the new design is
>>> actually
>>> >         parametric in the data representation of the field type.
>>> >         Something that wasn't possible in the original design.
>>> >          >>>>
>>> >          >>>> This proposal is not technically backward
>>> compatible,
>>> >         because the order of argument in which
>>> OverloadedRecordUpdate
>>> >         expects the argument of setField is changed. This is not
>>> >         essential to the proposal, but this is a more consistent
>>> order
>>> >         argument with the rest of Haskell. And considering that
>>> >         OverloadedRecordUpdate is very loudly advertised as
>>> >         experimental, I recommend accepting this breakage.
>>> >          >>>>
>>> >          >>>> Overall the proposal is actually more backward
>>> compatible
>>> >         with GHC 9.8 than the original design, as the HasField
>>> class is
>>> >         left unchanged.
>>> >          >>>>
>>> >          >>>> Overall, the proposal looks quite reasonable to
>>> me, and
>>> >         well-argued. I recommend acceptance.
>>>
>>>
>>>
>>> --
>>> Adam Gundry, Haskell Consultant
>>> Well-Typed LLP, https://www.well-typed.com/
>>>
>>> Registered in England & Wales, OC335890
>>> 27 Old Gloucester Street, London WC1N 3AX, England
>>> https://www.google.com/maps/search/27+Old+Gloucester+Street,+London+WC1N+3AX,+England?entry=gmail&source=g
>>>
>>> _______________________________________________
>>> ghc-steering-committee mailing list
>>> ghc-steering-committee@haskell.org
>>>
>>> https://mail.haskell.org/cgi-bin/mailman/listinfo/ghc-steering-committee
>>>
>>
>>
>> --
>> Arnaud Spiwack
>> Director, Research at https://moduscreate.com and https://tweag.io.
>> _______________________________________________
>> ghc-steering-committee mailing list
>> ghc-steering-committee@haskell.org
>>
>> https://mail.haskell.org/cgi-bin/mailman/listinfo/ghc-steering-committee
>>
> _______________________________________________
> ghc-steering-committee mailing list
> ghc-steering-committee@haskell.org
>
> https://mail.haskell.org/cgi-bin/mailman/listinfo/ghc-steering-committee
>
--
Arnaud Spiwack
Director, Research at https://moduscreate.com and https://tweag.io.
--
Arnaud Spiwack
Director, Research at https://moduscreate.com and https://tweag.io.
--
Arnaud Spiwack
Director, Research at https://moduscreate.com and https://tweag.io.
-- 
Arnaud Spiwack
Director, Research at https://moduscreate.com and https://tweag.io.