arrow-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Wes McKinney <wesmck...@gmail.com>
Subject Re: [DISCUSS] Result vs Status
Date Sat, 19 Oct 2019 00:58:47 GMT
I'm definitely uncomfortable with the idea of deprecating Status.

We have a few kinds of functions that can fail:

1. Functions with no "out" arguments
2. Functions with one out argument
3. Functions with multiple out arguments

IMHO functions in category 2 are the best candidates for utilizing
Status. In some cases, Case 3 may be more usable Result-based, but it
can also create more work (or confusion) on the part of the developer,
either

* The T in Result<T> has to be a struct-like value that transports
multiple pieces of data
* "Out" data may be split across a Result<T> and a separate out
argument. That's not too nice

Can't say I'm thrilled about having Result<void> or similar for Case
1-type functions (if I'm understanding what would be the solution
there).

Left to my own devices I would either use only Status or use Result
when it is convenient for functions that have a single out argument

I don't know how many functions or methods we have in the codebase
returning Status but I'd guess it's getting on the order of 1000. A
proper accounting would be helpful

- Wes

On Fri, Oct 18, 2019 at 2:41 AM Micah Kornfield <emkornfield@gmail.com> wrote:
>
> Based on the call this week, I think there are a few related questions here.
>
> 1.  Should we use Result at all?
>  - IMO Result expresses APIs more naturally then Status + Single output
> parameter.  I think most would agree if we had it from the beginning we
> would be probably use it.
> - The downside to using it is the pain in incorporating it into the
> codebase, including the potential for inconsistent APIs and breaking
> consumers of the package.  It also has the potential to cause ABI problems
> with other projects due its use of a vendored "Variant" implementations.
>
> 2.  If we agree on using Result in the code-base going forward (i.e. we
> don't remove it altogether), how do we move forward when working with APIs?
>
> This can be divided into existing and new APIs.
>
> For existing APIs we can:
> 1.  Leave existing APIs in place with no plans to migrate them to using
> Result.
> 2.  Aim to maintain both Result and Status APIs (backfill Result APIs where
> it makes sense).
> 3.  Aim to migrate to Result APIs (backfill Result APIs and deprecate
> Status APIs).  I assume this process will take at least 1 calendar year
> (its one of the things I'm hoping to get to).  This is probably best done
> incrementally by submodule.  I think for a change on this scale we should
> mark old APIs as deprecated and leave them in place for at least 2 release
> cycles (at the current cadence 6 months).
>
> For new APIs:
> 1.  Attempt to always have both versions (Result and Status) of the API
> everywhere that it makes sense.
> 2.  Produce both versions of the API until we are ready to deprecate Status
> APIs in one go.
> 3.  Only produce APIs using Result.
>
> My personal preference would be to choose to use Result and proceed with
> Option 3 for existing APIs (aim for deprecation of Status) and Option 3 for
> new APIs (only use Result going forward).  My second preference would be to
> simply remove "Result".  I don't want to be supporting parallel APIs in the
> long term.
>
> Thanks,
> Micah
>
>
>
> On Sat, Oct 5, 2019 at 4:59 AM Sutou Kouhei <kou@clear-code.com> wrote:
>
> > Hi,
> >
> > In <21183bb1-813f-b35b-b391-3f52ef2e62c9@python.org>
> >   "Re: [DISCUSS] Result vs Status" on Sat, 5 Oct 2019 12:23:05 +0200,
> >   Antoine Pitrou <antoine@python.org> wrote:
> >
> > >> OK, so what  could more context be provided on:
> > >>
> > >>> From the discussion in the sync call, it seems reasonable to require
> > that:
> > >>> Public APIs which are likely to be directly wrapped in a binding
> > should not
> > >>> use Result<> to the exclusion of Status. An equivalent Status
API
> > should
> > >>> always be provided for ease of binding.
> > >
> > > I don't know, sorry :-) I wasn't on the sync call.
> >
> >
> > We don't need Status API for bindings. We already use
> > complex types such as std::shared_ptr in our API. Bindings
> > need C++ feature for complex types. So we don't need to care
> > about Result<> or Status.
> >
> >
> > Thanks,
> > --
> > kou
> >

Mime
View raw message