[ghc-steering-committee] Modification to record dot syntax propsal

Iavor Diatchki iavor.diatchki at gmail.com
Fri Mar 5 16:03:28 UTC 2021


I find that this committee is having less and less technical discussion on
concrete issues, and there is a lot of talk of process, rules, and
definitions, that I do not find useful.

It is extremely difficult to anticipate all aspects of a design without
implementing it and using it for a while, so I completely disagree that we
should not revisit accepted proposals.

While "experimental" extensions are totally fine in my book, I really do
think we should avoid intentionally introducing things that we know now are
likely to change.   The overloaded records update is such an extension: it
changes a useful, standardized, and in my experience, not uncommonly used
feature, into a less general version (in one dimension, anyway).
Furthermore, we are aware that there is already work by Adam to address
this, so the extension will likely change (or perhaps we are going to have
yet another record related extension).

So, why the hurry to add this to GHC now, when we know from experience how
painful it is to remove things, once they are out in the wild?

On Fri, Mar 5, 2021, 02:50 Spiwack, Arnaud <arnaud.spiwack at tweag.io> wrote:

> Iavor,
>
> I concur with Simon and Eric. Accepting a proposal is a promise that the
> proposal will not be debated on the design front in the future. That is,
> when the implementation comes up it is out of scope to discuss design.
> That's the deal. GHC proposals are a process. The goal of this process is
> to foster participation of the wider community to the evolution of GHC, by
> reducing stress and uncertainty, and focusing productivity. This is why
> changing the design of an accepted proposal requires another proposal,
> submitted to the same level of scrutiny as the original proposal. Even, and
> in particular, by members of the steering committee.
>
> In this particular case, the authors have agreed to the change, and this
> is something that we feel is important enough to press forward quickly. So
> in the interest of everybody's time, I suggest that we move on. But,
> generally, the point stands.
>
> /Arnaud
>
> On Thu, Mar 4, 2021 at 8:45 PM Eric Seidel <eric at seidel.io> wrote:
>
>> I agree it’s more important to get things right than to look good, but we
>> should aim to do both of course. As Simon mentioned earlier, this will
>> erode the community’s confidence if it happens too often. So I’m ok with
>> reversing our decision on this point (happily it sounds like there’s
>> already work underway on an improved design), but I think we should reflect
>> on why we didn’t have this discussion back when we were discussing the
>> original proposal.
>>
>> Sent from my iPhone
>>
>> On Mar 4, 2021, at 14:04, Iavor Diatchki <iavor.diatchki at gmail.com>
>> wrote:
>>
>> 
>> Bad look or not, it seems better to me to change our mind than accept
>> something "to save face" :)     It seems hard to imagine that I am the only
>> one here who uses type changing updates in records.  How are other members
>> of the committee reconciling this change?   Are you planning to change your
>> code to use record wild cards as Neil suggested---this seems a lot worse
>> than the status quo? Or is the plan the plan to just fork the language and
>> use the pragma to disambiguate?
>>
>> I don't mind if the extension is in GHC's code, but I think that if we
>> add it to the GHC manual, people will use it, and it will be a much bigger
>> deal to change later.
>>
>> On Thu, Mar 4, 2021 at 9:59 AM Eric Seidel <eric at seidel.io> wrote:
>>
>>> Yes that clears up the current situation. And from the current email
>>> chain it looks like the authors are aware of the plan to send record
>>> updates back for revision? I'd just want to make sure that the updated
>>> proposal reflects that split on which parts have been accepted.
>>>
>>> For what it's worth, I do recall the loss of type-changing update being
>>> part of the original proposal and thought the proposal was still good on
>>> balance. The tradeoffs were definitely discussed in the giant Github
>>> thread, I don't recall if we discussed that aspect here though..
>>> Regardless, I think it's a bad look for us to walk back a decision on
>>> account of not having read the proposal closely enough!
>>>
>>> On Thu, Mar 4, 2021, at 09:53, Simon Peyton Jones wrote:
>>> > |  That being said, I don't see anything in the revised proposal about
>>> > |  the stability of OverloadedRecordUpdate. Are you saying that as part
>>> > |  of this revision, we'll explicitly accept OverloadedRecordDot and
>>> send
>>> > |  OverloadedRecordUpdate back for revision?
>>> >
>>> > We *already* accepted both, as part of accepting the earlier
>>> > RecordDotSyntax proposal.   But in this round, Iavor has pushed back
>>> > against OverloadedRecordUpdate.  No one else has expressed a view on
>>> > this point.
>>> >
>>> > But rather than debate it at length I proposed to explicitly un-accept
>>> > the OverloadedRecordUpdate part of the proposal.  It'll return as part
>>> > of a new record-update proposal that Adam is working on.
>>> >
>>> > In the meantime OverloadedRecordUpdate will be in GHC's codebase, and
>>> > (assuming that's what the majority wants) documented in the user
>>> > manual, with a prominent "may change" caveat.
>>> >
>>> > Does that make it clear?
>>> >
>>> > Simon
>>> >
>>> > |  -----Original Message-----
>>> > |  From: ghc-steering-committee <ghc-steering-committee-
>>> > |  bounces at haskell.org> On Behalf Of Eric Seidel
>>> > |  Sent: 04 March 2021 14:38
>>> > |  To: ghc-steering-committee at haskell.org
>>> > |  Subject: Re: [ghc-steering-committee] Modification to record dot
>>> > |  syntax propsal
>>> > |
>>> > |  I agree with Richard and Joachim that it should be documented in the
>>> > |  user guide. It's much better to document features with the expected
>>> > |  level of support than to let users stumble upon them and make their
>>> > |  own assumptions about stability.
>>> > |
>>> > |  That being said, I don't see anything in the revised proposal about
>>> > |  the stability of OverloadedRecordUpdate. Are you saying that as part
>>> > |  of this revision, we'll explicitly accept OverloadedRecordDot and
>>> send
>>> > |  OverloadedRecordUpdate back for revision?
>>> > |
>>> > |  On Thu, Mar 4, 2021, at 04:46, Simon Peyton Jones via ghc-steering-
>>> > |  committee wrote:
>>> > |  >
>>> > |  > Friends
>>> > |  >
>>> > |  >
>>> > |  >
>>> > |  > We've agree to accept my suggestion below.
>>> > |  >
>>> > |  >
>>> > |  >
>>> > |  > But there is one point at issue: should OverloadedRecordUpdate be
>>> > |  > documented in the user manual, albeit identified as a not-yet-
>>> > |  accepted
>>> > |  > feature?
>>> > |  >
>>> > |  >
>>> > |  >
>>> > |  >  * Richard positively wants it in the manual
>>> > |  >  * Iavor positively doesn't want it there.
>>> > |  >
>>> > |  >
>>> > |  > I don't mind either way.
>>> > |  >
>>> > |  >
>>> > |  >
>>> > |  > What do others think?  It's not a big deal, but we owe the
>>> authors a
>>> > |  > decision.  Answer by the end of the week please, and I'll make a
>>> > |  > shepherd decision based on the opinions I get.
>>> > |  >
>>> > |  >
>>> > |  >
>>> > |  > Simon
>>> > |  >
>>> > |  >
>>> > |  >
>>> > |  >
>>> > |  >
>>> > |  > *From:* Simon Peyton Jones <simonpj at microsoft.com>
>>> > |  > *Sent:* 02 March 2021 12:45
>>> > |  > *To:* ghc-steering-committee <ghc-steering-committee at haskell.org>
>>> > |  > *Cc:* Simon Peyton Jones <simonpj at microsoft.com>
>>> > |  > *Subject:* RE: [ghc-steering-committee] Modification to record dot
>>> > |  > syntax propsal
>>> > |  >
>>> > |  >
>>> > |  >
>>> > |  > Friends
>>> > |  >
>>> > |  >
>>> > |  >
>>> > |  > Having consulted the authors, I propose that we do this:
>>> > |  >
>>> > |  >
>>> > |  >
>>> > |  >  * Proceed with OverloadedRecordDot for 9.2, exactly as in the
>>> > |  > original proposal except for the extension name.
>>> > |  >  * Record update will remain exactly as now, in 9.2; that is,
>>> > |  drawing
>>> > |  > back from the original proposal.
>>> > |  >  * There may be some *code* in 9.2 that allows overloaded record
>>> > |  > update, protected by OverloadedRecordUpdate, but not in the user
>>> > |  > manual, and not treated as an accepted proposal.   I don't think
>>> we
>>> > |  > should ask the authors to remove it, and it will allow
>>> > |  experimentation.
>>> > |  >  * Adam is leading on a revised record-update proposal. This will
>>> > |  cover
>>> > |  >    * the tradeoffs between type-changing and non-type-changing
>>> > |  update
>>> > |  >    * what the current record-update syntax stands for, and/or any
>>> > |  new
>>> > |  > syntax
>>> > |  >
>>> > |  >
>>> > |  > Is that acceptable to everyone?
>>> > |  >
>>> > |  >
>>> > |  >
>>> > |  > Simon
>>> > |  >
>>> > |  >
>>> > |  >
>>> > |  > *From:* ghc-steering-committee
>>> > |  > <ghc-steering-committee-bounces at haskell.org> *On Behalf Of *Simon
>>> > |  > Peyton Jones via ghc-steering-committee
>>> > |  > *Sent:* 01 March 2021 17:51
>>> > |  > *To:* Iavor Diatchki <iavor.diatchki at gmail.com>; Spiwack, Arnaud
>>> > |  > <arnaud.spiwack at tweag.io>
>>> > |  > *Cc:* ghc-steering-committee <ghc-steering-committee at haskell.org>
>>> > |  > *Subject:* Re: [ghc-steering-committee] Modification to record dot
>>> > |  > syntax propsal
>>> > |  >
>>> > |  >
>>> > |  >
>>> > |  > I don't buy the argument of "this is already accepted", as I don't
>>> > |  > think many of us had noticed that part of the proposal (I
>>> certainly
>>> > |  > didn't), and I think we should be flexible enough to revisit
>>> > |  previous
>>> > |  > decisions when we notice problems with them.
>>> > |  >
>>> > |  > I agree in principle that we can revisit decisions.   But we have
>>> to
>>> > |  be
>>> > |  > aware that it is potentially very discouraging for proposal
>>> authors
>>> > |  to
>>> > |  >
>>> > |  >  * propose something,
>>> > |  >  * have it *extensively* debated (including this very point),
>>> > |  >  * have it accepted,
>>> > |  >  * implement it,
>>> > |  > and then be told that the committee has changed its mind.  That's
>>> > |  > pretty bad from their point of view.
>>> > |  >
>>> > |  >
>>> > |  >
>>> > |  > Still, Adam is working on a new SetField proposal, so perhaps
>>> that's
>>> > |  a
>>> > |  > figleaf.  I'll consult them.
>>> > |  >
>>> > |  >
>>> > |  > Simon
>>> > |  >
>>> > |  >
>>> > |  >
>>> > |  >
>>> > |  >
>>> > |  > *From:* Iavor Diatchki <iavor.diatchki at gmail.com>
>>> > |  > *Sent:* 01 March 2021 17:23
>>> > |  > *To:* Spiwack, Arnaud <arnaud.spiwack at tweag.io>
>>> > |  > *Cc:* Simon Peyton Jones <simonpj at microsoft.com>;
>>> > |  > ghc-steering-committee <ghc-steering-committee at haskell.org>
>>> > |  > *Subject:* Re: [ghc-steering-committee] Modification to record dot
>>> > |  > syntax propsal
>>> > |  >
>>> > |  >
>>> > |  >
>>> > |  > Hello,
>>> > |  >
>>> > |  >
>>> > |  >
>>> > |  > I think there is a strong motivation to *at least* split the
>>> > |  > extensions:  with the current design, enabling the special `.`
>>> > |  notation
>>> > |  > also *disables* type changing record update, which has nothing to
>>> do
>>> > |  > with the `.` notation.
>>> > |  >
>>> > |  >
>>> > |  >
>>> > |  > My preference would be to:
>>> > |  >
>>> > |  >   1. Split the original proposal into two parts: one about `.`
>>> > |  > notation, the other about record update (as suggested by this
>>> > |  proposal)
>>> > |  >
>>> > |  >   2. Treat the `.` notation part as accepted
>>> > |  >
>>> > |  >   3. Require changes on the record update part, so that you don't
>>> > |  have
>>> > |  > to choose between it and type changing record updates, which are
>>> > |  quite
>>> > |  > useful, and I don't think we should aim for a Haskell without
>>> them.
>>> > |  >
>>> > |  >
>>> > |  >
>>> > |  > I don't buy the argument of "this is already accepted", as I don't
>>> > |  > think many of us had noticed that part of the proposal (I
>>> certainly
>>> > |  > didn't), and I think we should be flexible enough to revisit
>>> > |  previous
>>> > |  > decisions when we notice problems with them.
>>> > |  >
>>> > |  >
>>> > |  >
>>> > |  > -Iavor
>>> > |  >
>>> > |  >
>>> > |  >
>>> > |  >
>>> > |  >
>>> > |  >
>>> > |  >
>>> > |  > On Mon, Mar 1, 2021 at 1:40 AM Spiwack, Arnaud
>>> > |  <arnaud.spiwack at tweag.io> wrote:
>>> > |  >
>>> > |  > > Simon, all,
>>> > |  >
>>> > |  > >
>>> > |  >
>>> > |  > > After reading more of the PR thread (in particular the fews
>>> posts
>>> > |  after Simon's recommendation), I have to admit: I am thoroughly
>>> > |  confused. I'm not sure that two people in that thread have the same
>>> > |  motivation, end goal, or design in mind.
>>> > |  >
>>> > |  > >
>>> > |  >
>>> > |  > > The motivations provided by the modified *Alternatives* section
>>> is
>>> > |  not much more helpful (at the risk of caricaturing a little, it
>>> > |  basically reads: "we made two extensions rather than one because we
>>> > |  can"). Though it makes it clear that the end goal is to fold a bunch
>>> > |  of record-related extensions into one glorious record extension
>>> (well,
>>> > |  probably not fold them, but make a meta-extension that implies all
>>> the
>>> > |  extensions that we've decided we like).
>>> > |  >
>>> > |  > >
>>> > |  >
>>> > |  > > My starting point is that:
>>> > |  >
>>> > |  > > - Additional extensions have a maintenance cost
>>> > |  >
>>> > |  > > - Additional extensions impose a cognitive burden on their use
>>> > |  >
>>> > |  > > - I expect that a new extension will break my code in the next
>>> few
>>> > |  releases.
>>> > |  >
>>> > |  > >
>>> > |  >
>>> > |  > > Based on this, I don't care how this extension or the glorious
>>> > |  record extension are named; but if we want to have two extensions we
>>> > |  should have a serious reason. Right now, the one reason that I see
>>> > |  (and Iavor raised), is that the update part of `RecordDotSyntax` is
>>> > |  not backward compatible. Is it a strong enough reason? I don't know.
>>> > |  The only data point that I can provide is that when we discussed the
>>> > |  original proposal, I brought it up several times, and it didn't seem
>>> > |  very important at the time (the conversation focused on other points
>>> > |  of the proposal).
>>> > |  >
>>> > |  > >
>>> > |  >
>>> > |  > > So, I'm still reluctant. I feel that, at the very least, the
>>> > |  motivations are not well-enough articulated in the proposal (I'll
>>> make
>>> > |  a comment to this effect on Github when I'm done composing this
>>> > |  email).
>>> > |  >
>>> > |  > >
>>> > |  >
>>> > |  > > I appreciate that I'm in the minority here. Yet, I'm still
>>> > |  unconvinced.
>>> > |  >
>>> > |  > >
>>> > |  >
>>> > |  > > Best,
>>> > |  >
>>> > |  > > Arnaud
>>> > |  >
>>> > |  > >
>>> > |  >
>>> > |  > > On Sat, Feb 27, 2021 at 12:39 AM Simon Peyton Jones
>>> > |  <simonpj at microsoft.com> wrote:
>>> > |  >
>>> > |  > >> Generally, I'm not in favour in proposals which split
>>> extensions
>>> > |  though: we already have so many extensions. Are the reasons for this
>>> > |  split strong enough? I haven't had time to dig into the details.
>>> > |  >
>>> > |  > >> Arnaud, happily, you don't have to dig very deep - just read
>>> the
>>> > |  handful of posts following my recommendation.
>>> > |  >
>>> > |  > >>
>>> > |  >
>>> > |  > >> There seem to be two motivations.
>>> > |  >
>>> > |  > >>
>>> > |  >
>>> > |  > >>  1. There really are two orthogonal extensions, one for r.x
>>> > |  notation, and one for overloaded update.  Iavor likes the first but
>>> > |  not the second.  Neil likes both.  Having separate extensions lets
>>> us
>>> > |  experiment.
>>> > |  > >>
>>> > |  >
>>> > |  > >>  1. You suggest that changing the definition of RecordDotSyntax
>>> > |  in a subsequent release, e.g. by subsequently making it imply
>>> > |  NoFieldSelectors, would be fine. But it certainly imposes pain -
>>> some
>>> > |  programs would stop compiling.  The approach offered by this
>>> proposal
>>> > |  avoids that problem.
>>> > |  > >>
>>> > |  >
>>> > |  > >> Yes, there are lots of extensions surrounding records:
>>> > |  NoFieldSelectors, DuplicateRecordFields, NamedFieldPuns,
>>> > |  DisambiguateRecordFields, RecordWildCards.  It may not be pretty to
>>> > |  divide things up so fine, but it's not new.
>>> > |  >
>>> > |  > >>
>>> > |  >
>>> > |  > >>
>>> > |  >
>>> > |  > >> If there are alternative suggestions, let's hear them.
>>> > |  >
>>> > |  > >>
>>> > |  >
>>> > |  > >> Simon
>>> > |  >
>>> > |  > >>
>>> > |  >
>>> > |  > >> *From:* ghc-steering-committee <ghc-steering-committee-
>>> > |  bounces at haskell.org> *On Behalf Of *Spiwack, Arnaud
>>> > |  > >> *Sent:* 26 February 2021 22:33
>>> > |  > >> *To:* Iavor Diatchki <iavor.diatchki at gmail.com>
>>> > |  > >> *Cc:* ghc-steering-committee <
>>> ghc-steering-committee at haskell.org>
>>> > |  > >> *Subject:* Re: [ghc-steering-committee] Modification to record
>>> > |  dot syntax propsal
>>> > |  >
>>> > |  > >>
>>> > |  >
>>> > |  > >>
>>> > |  >
>>> > |  > >>> I do think that  reusing the record update syntax for the
>>> > |  overloaded monomorphic update is a mistake---it is not something I
>>> had
>>> > |  noticed during our original discussion.
>>> > |  >
>>> > |  > >>
>>> > |  >
>>> > |  > >> This is the one reason I can see for cutting this extension in
>>> > |  smaller pieces. But, then again, -XOverloadedRecordUpdate would be a
>>> > |  fork-like extension.
>>> > |  >
>>> > |  > >>
>>> > |  >
>>> > |  > >> Generally, I'm not in favour in proposals which split
>>> extensions
>>> > |  though: we already have so many extensions. Are the reasons for this
>>> > |  split strong enough? I haven't had time to dig into the details.
>>> > |  >
>>> > |  > >>
>>> > |  >
>>> > |  > >> I'm not sure that not having the design of the proposal quite
>>> > |  finalised is a good reason, extensions mutate in their first
>>> > |  iterations, I don't think that it's a problem.
>>> > |  >
>>> > |  > _______________________________________________
>>> > |  > ghc-steering-committee mailing list
>>> > |  > ghc-steering-committee at haskell.org
>>> > |  >
>>> > |
>>> https://nam06.safelinks.protection.outlook.com/?url=https%3A%2F%2Fmail
>>> > |  .haskell.org%2Fcgi-bin%2Fmailman%2Flistinfo%2Fghc-steering-
>>> > |  committee&data=04%7C01%7Csimonpj%40microsoft.com
>>> %7C8abc00434aa94b7
>>> > |
>>> fa98e08d8df1b5d9f%7C72f988bf86f141af91ab2d7cd011db47%7C1%7C0%7C6375046
>>> > |
>>> 55938142566%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMz
>>> > |
>>> IiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C1000&sdata=9DGVl6oJc0%2BIQekuXf
>>> > |  zwqviiaT%2FaHZIlIk3R5aMZssA%3D&reserved=0
>>> > |  >
>>> > |  _______________________________________________
>>> > |  ghc-steering-committee mailing list
>>> > |  ghc-steering-committee at haskell.org
>>> > |
>>> https://nam06.safelinks.protection.outlook.com/?url=https%3A%2F%2Fmail
>>> > |  .haskell.org%2Fcgi-bin%2Fmailman%2Flistinfo%2Fghc-steering-
>>> > |  committee&data=04%7C01%7Csimonpj%40microsoft.com
>>> %7C8abc00434aa94b7
>>> > |
>>> fa98e08d8df1b5d9f%7C72f988bf86f141af91ab2d7cd011db47%7C1%7C0%7C6375046
>>> > |
>>> 55938152562%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMz
>>> > |
>>> IiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C1000&sdata=QZlV0RrRttaiDdrQNiRB
>>> > |  vUhS6il1DydPX5cyl%2BdILbI%3D&reserved=0
>>> >
>>> _______________________________________________
>>> ghc-steering-committee mailing list
>>> ghc-steering-committee at haskell.org
>>> https://mail.haskell.org/cgi-bin/mailman/listinfo/ghc-steering-committee
>>>
>> _______________________________________________
>> ghc-steering-committee mailing list
>> ghc-steering-committee at haskell.org
>> https://mail.haskell.org/cgi-bin/mailman/listinfo/ghc-steering-committee
>>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.haskell.org/pipermail/ghc-steering-committee/attachments/20210305/335b8882/attachment-0001.html>


More information about the ghc-steering-committee mailing list