[ghc-steering-committee] Proposal #612: Change semantics of -fhpc to not accumulate data over multiple runs, rec: accept

Fri Feb 23 00:41:20 UTC 2024

Simon,

I see where you are coming from. So let me try to outline how I usually
handle this: this silent change (even with deprecation) likely alters one
of the build products, and hopefully(!) someone will get suspicious
(because the build products are used (otherwise why bother building them
in the first place). Thus someone flags that something with hpc is off. At
which point I'd go and manually investigate what is going on, at _that_
point, I will see this, either because I'm reading the logs, or because the
compiler reports it to me while manually running it.

1. I upgrade the compiler.
2. I observe a change.
3. I try to figure out where the change comes from.

If in (3) or maybe (2), I am told pretty directly what the change is, it's
much easier for me to adjust. Even better if the change message includes
instructions on how to adjust. Even better if someone else can observe the
change, _and_ know as well as having been given actionable advice
on how to proceed. If I can cut myself out of the whole "Something broke, I
looked at it, but can't figure out what's going on, and got way too much
on my hands, please someone else dig into it" situation, it's a win. Not
only for the person experiencing the change (they don't feel helpless, nor
do they
need to ask for assistance), but also for the people supporting these kinds
of issues as they don't come up in the first place.

The difference is between it taking 10min to figure, and providing the
agency to understand and solve it alone, or needing to reach for support it
taking 60+min to figure out and correct, because one has to dig into all
the changes and details. Maybe LLM will cut this a bit because one could
ask "I'm seeing some strange behaviour regarding -fhpc, please provide a
summary of the relevant changes between X and Y".

I like SPJ's suggestion in this thread quite a bit and will ask the Author
about their opinion.

In general we seem to be in agreement around the current default not being
ideal, and somewhat hung on the deprecation notice part. With an
eye on our stability goals, I'd like to see us try light deprecation
warnings. I see that Adams' solution is much more thorough, I also agree
that it's
a lot more involved.

If anyone has qualms around accepting this with a light deprecation notice,
please speak up, otherwise I'll mark this proposal as accepted, and
merge it in a few days after David responded.

Cheers,
 Moritz

On Sat, 27 Jan 2024 at 00:46, Simon Marlow <marlowsd at gmail.com> wrote:

> Moritz - I agree with everything you say about how users will typically
> experience the change, but I don't think a deprecation warning resolves the
> concern. As you say, people might be using HPC via automated tooling and
> will likely not even see the warning, and even if they do see it will
> likely ignore it (I don't know about you, but I don't usually want to be
> distracted by some different yak that needs shaving while I'm coding).
>
> IMO we either have to break it completely - i.e. make the existing usage
> just fail, after a warning cycle - or don't change it at all. We can't
> silently change the behaviour even with a deprecation cycle.
>
> Cheers
> Simon
>
> On Thu, 25 Jan 2024 at 10:49, Moritz Angermann <moritz.angermann at gmail.com>
> wrote:
>
>> Here's my rationale for a deprecation period.
>>
>> Someone who uses hpc, might rely on the accumulating behaviour.  They
>> need a way to be informed about the upcoming change, otherwise they won't
>> know that they need to pass a new flag going forward (and they can
>> even start doing that at that point to be defensive). I also hope the
>> deprecation note would show up in automated tooling, which can then be
>> proactively adjusted.
>>
>> To me the compiler is _the_ interface to the end user.  Chances that
>> someone reads the proposals, GHC changelog are very low. Chances someone
>> reads the user guide in detail for each GHC release are low as well,
>> only after the fact, if someone broke will someone start trying to figure
>> out why something broke. At that point, I think we've lost. They are
>> frustrated already that something they knew worked just broke.
>>
>> So to me the deprecation note is basically the signage towards the end
>> user that uses hpc, that they will need to adjust their ghc invocation if
>> they rely on the current behaviour, and also a courtesy heads up that the
>> behaviour
>> they might rely on changes, _and_ how to adjust and get their old
>> behaviour (they may need/rely on) back.
>>
>> To address your two points in detail:
>> > Is a deprecation period necessary? Or even useful (how do we signal the
>> upcoming change in default, and how do we expect people would act on it)?
>> - As the flag currently does not exist, call GHC with hpc without the new
>> flag will issue a warning. (e.g. current users will be informed).
>> - The warning message will let the user know what to pass to ghc to
>> retain the old (current) behaviour, or opt in to the new (future current)
>> behaviour, depending on their needs.
>>
>> If we did neither, the user would at some point just notice an existing
>> workflow magically stopped or in this case even worse, the silent
>> aggregation of .tix data would stop, and they
>> might _assume_ the data was the aggregate but it wasn't. A silent change
>> in default behaviour would be rather painful to notice, debug, and rectify
>> for users. At least that's my expectation.
>> (And I know I personally would be furious if a compiler pulled that on
>> me).
>>
>> Best,
>>  Moritz
>>
>> On Thu, 25 Jan 2024 at 18:35, Arnaud Spiwack <arnaud.spiwack at tweag.io>
>> wrote:
>>
>>> The proposed change seems sensible. At least the argument is convincing.
>>> I don't use hpc myself, so I don't have an actual design opinion besides
>>> that. So I vote (weakly) for the proposed new default.
>>>
>>> Is a deprecation period necessary? Or even useful (how do we signal the
>>> upcoming change in default, and how do we expect people would act on it)?
>>>
>>> On Thu, 25 Jan 2024 at 07:06, Moritz Angermann <
>>> moritz.angermann at gmail.com> wrote:
>>>
>>>> Dear Committee members,
>>>>
>>>> In Proposal #612
>>>> <https://github.com/ghc-proposals/ghc-proposals/pull/612>, David
>>>> Binder suggests changing the semantics from `-fhpc` to auto-ingest
>>>> existing .tix files for accumulation, towards _not_ auto-ingesting them
>>>> and instead overwriting
>>>> the .tix file with the coverage of the latest run.
>>>>
>>>> I agree with his assessment that the current behaviour easily leads to
>>>> unexpected surprises. He suggests adding a flag --read-tix-file= to control
>>>> this behaviour and defaulting that to _no_, with a grace and
>>>> deprecation period prior informing the user that the currently accumulating
>>>> feature will change in a future GHC release.
>>>>
>>>> Adding the flag to control this behaviour is fairly uncontroversial I
>>>> hope, however I'd like you to. Weight in on the default. Should we change
>>>> the default behaviour, or only add the flag?
>>>>
>>>> I'd recommend changing the default to --read-tix-file=no after a
>>>> deprecation period.
>>>>
>>>> Best,
>>>>  Moritz
>>>> _______________________________________________
>>>> ghc-steering-committee mailing list
>>>> ghc-steering-committee at 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 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/20240223/0ad383bc/attachment-0001.html>