Feature #16264
closedReal "callable instance method" object.
Description
It is a part of thinking about the "argument-less call style" I already made several proposals about.
Preface¶
Argument-less call style is what I now call things like map(&:foo)
and each(&Notifications.:send)
approaches, and I believe that naming the concept (even if my initial name is clamsy) will help to think about it. After using it a lot on a large production codebase (not only symbols, but method references too, which seem to be less widespread technique), I have a strong opinion that it not just "helps to save the keypresses" (which is less important), but also helps to clearer separate the concepts on a micro-level of the code. E.g. if you feel that each(&Notifications.:send)
is "more right" than select { |el| Notifications.send(el, something, something) }
, it makes you think about Notifications.send
design in a way that allows to pass there exactly that combination of arguments so it would be easily callable that way, clarifying modules responsibilities.
(And I believe that "nameless block parameters", while helping to shorter the code, lack this important characteristic of clarification.)
The problem¶
One of the problems of "argument-less calling" is passing additional arguments, things like those aren't easy to shorten:
ary1.zip(ary2, ary3).map { |lines| lines.join("\n") }
# ^^^^
construct_url.then(&HTTP.:get).body.then { |text| JSON.parse(text, symbolize_names: true) }
# ^^^^^^^^^^^^^^^^^^^^^^
(BTW, here is a blog post where I show recently found technique for solving this, pretty nice and always existing in Ruby, if slightly esotheric.)
There's a lot of proposals for "partial applications" which would be more expressive than .curry
(guilty myself), but the problematic part in all of this proposals is:
The most widespread "shortening" is
&:symbol
, andSymbol
itself is NOT a functional object, and it is wrong to extend it with functional abilities.
One of consequences of the above is, for example, that you can't use 2.6's proc combination with symbols, like File.:read >> :strip >> :reverse
. You want, but you can't.
Here (while discussing aforementioned blog posts), I stumbled upon an idea of how to solve this dilemma.
The proposal¶
I propose to have a syntax for creating a functional object that when being called, sends the specified method to its first argument. Basically, what Symbol#to_proc
does, but without "hack" of "we allow our symbols to be convertible to functional objects". Proposed syntax:
[1, 2, 3].map(&.:to_s)
Justification of the syntax:
- It is like
Foo.:method
(producing functional object that callsmethod
) - Orphan
.:method
isn't allowed currently (you need to sayself.:method
to refer to "currentself
s method"), and Matz's justification was "it would be too confusable with:method
, small typo will change the result" -- which in PROPOSED case is not as bad, as:foo
and.:foo
both meaning the same thing; - It looks kinda nice, similar to (proposed and rejected)
map { .to_s }
→ with my proposal, it ismap(&.:to_s)
, implying somehow applying.to_s
to the previous values in the chain.
The behavior: .:foo
produces object of class, say, MethodOfArgument
(class name is subject to discuss) — which makes differences of "Proc created from Symbol" (existing internally, but almost invisible) obvious and hackable.
Potential gains¶
-
New object could be used in proc composition:
File.:read >> .:strip >> JSON.:parse >> .:compact
-
When both "method" and "method of argument" are proper functional objects, a new partial application syntax can be discussed, common for them both. For example (but not necessary this method name!)
paragraph_hashes.map(&.:merge.with(author: current_author))
filenames.map(&File.:read.with(mode: 'rb'))
* (I believe at this point we'll be able to finally switch from discussing "show we extend Symbol with more callable-alike functionality" to just method's name and exact behavior)
* I am not an expert, but probably some optimizations could be applied, too
* Currently, `:sym.to_proc` is internally different from other proc, but this can't be introspected:
```ruby
:read.to_proc.inspect # => "#<Proc:0x0000556216192198(&:read)>"
# ^^^^^
- Probably, exposure of this fact could lead to some new interesting metaprogrammin/optimization techniques.
Transition¶
:foo
and .:foo
could work similarly for some upcoming versions (or indefinitely), with .:foo
being more powerful alternative, allowing features like groups_of_lines.map(&.:join.partial_apply(' '))
or something.
It would be like "real" and "imitated" keyword arguments. "Last hash without braces" was good at the beginning of the language lifecycle, but then it turned out that real ones provide a lot of benefits. Same thing here: &:symbol
is super-nice, but, honestly, it is semantically questionable, so may be slow switch to a "real thing" would be gainful for everybody?..
Updated by nobu (Nobuyoshi Nakada) almost 5 years ago
- Description updated (diff)
- Status changed from Open to Feedback
zverok (Victor Shepelev) wrote:
Transition:
:foo
and.:foo
could work similarly for some upcoming versions (or indefinitely), with.:foo
being more powerful alternative, allowing features likegroups_of_lines.map(&.:join.with(' '))
or something.
Is the point .:join.with(' ')
?
Updated by zverok (Victor Shepelev) almost 5 years ago
@nobu (Nobuyoshi Nakada), sorry, I am not sure I get your question.
Are you asking about how .:join.with(' ')
is supposed to work?
It is (theoretical and maybe not the best) example of how "partial application" could look in Ruby. It is not part of my proposal, just an idea of "if we'll have proper functional object .:join
, we can discuss how idiomatic partial application could work on functional object (blocks, methods, and method-of-arg proposed)".
Updated by shevegen (Robert A. Heiler) almost 5 years ago
I hope it is ok to link to one of your blog entries (for those who may ask - I am not
affiliated with zverok nor did I ask him about it prior to that; it may just provide
some more context to the larger topic at hand):
https://zverok.github.io/blog/2019-10-18-each_with_object.html
To the topic itself here. Hmmmmmm. Personally I am a bit biased against it, largely
due to syntax considerations. I actually think that &.:method is not very pretty.
To be fair: I think that .map(&:strip) is not very pretty either, but it's sort of
like an "established standard" since many years in ruby. I also use it myself
sometimes, e. g. especially when I do strip/chop/chomp. I am also aware of people
trying to find a way to use succinct syntax while also being able to pass in
arguments, such as the example the example given:
&.:join.with(' ')
This may be a lot up to one's individual style and preferences, but I don't quite
like it. I don't think that many are against the functionality, e. g. any succinct
representation may be useful, but there is some trade-off in regards to syntax and
semantics.
similar to (proposed and rejected) map { .to_s }
The syntax of e. g. .map { .method } was quite clean, cleaner than most other
suggestions. :) But I did not like the implication in semantics ... we have the
object.method notation in general, and suddenly we could omit the leading object
part. That would be quite strange in my opinion.
I think what nobu is specifically asking is about the proposal for the specific
syntax that would go with e. g. the part where you would pass additional
arguments to the method(s) involved.
as :foo and .:foo both meaning the same thing
I am biased because I am much preferring :foo over .:foo if only due to bad
eyesight alone :) - but that is also a reason why I dislike &.: even more.
Perhaps I am just too lazy; on the other hand, I would prefer ruby and ruby's
syntax to not become too complicated. With &.: we now would have three
characters. How many more would we want? Five? Eight? Why stop at eight? :P
It is not solely about syntax and semantics alone, though; some of it is
about simplicity versus complexity. I understand that this is up to one's
preference a lot. I just tend to prefer simplicity usually because my brain
is not easily able to master complexity.
By the way, I agree with this statement:
Symbol itself is NOT a functional object, and it is wrong to extend it
with functional abilities.
I think it is best to retain Symbol as close as possible to matz' original
concept for Symbol, even though over time Symbol was a bit extended. It
is understandable that there are general suggestions of people wanting to
extend Symbol, but I think it is better to retain Symbol to be a simpler
concept than to extend it into many ways as a "dynamic" but complex
concept.
And I believe that "nameless block parameters", while helping to shorter
the code, lack this important characteristic of clarification.
I think the use cases are a bit different, although it also depends on
the definition(s). I was not aware of an older suggestion, or the comment
by Jeremy, so the use cases may differ (or "elegance of syntax"). For
quick debugging I think not having to rely on the names is useful; I do
however had also agree with benoit to some extent in the sense that some
ruby users may happily use them all over the place, and it's not so
pretty or useful then, IMO. I still think it is different to the proposal
here, though - but if we would reason about syntax alone then oldschool
ruby syntax beats all of these suggestions from a "pretty" point of view,
including "nameless" param style. ;) (Although it was also said in the past
that while an elegant, succinct syntax is great to have, it is not necessarily
ruby's primary design point alone. I just like clean and simple syntax
whenever it is possible.)
Martin pointed out that some of the "functional" ideas are a bit disparate,
so a larger concept that could cover the individual suggestions may be
better. I don't have a useful suggestion for that, though, partially also
because the applications are a bit dissimilar to how I may tend to use
ruby (more from an "oldschool" OOP use mostly).
Updated by Dan0042 (Daniel DeLorme) almost 5 years ago
It sounds like an interesting way to improve functional programming in ruby. I can definitely see the point of having a proper functional syntax instead of trying to hack everything via symbols. But I'm not sure it's the best way. At the core, I believe this was best explained by duerst:
we should stop dealing with individual ideas of how to improve functional programming in Ruby and work out an overview of what's missing and how to address it, understanding that Ruby is first and foremost an OO language and we'll never get to the same point as Haskell or something similar.
So I think that having a shorthand .:foo
syntax as equivalent of :foo.to_proc
can definitely be part of the solution, but ideally it would be considered together with other elements of a functional grand master plan.
similar to (proposed and rejected) map { .to_s }
The syntax of e. g. .map { .method } was quite clean, cleaner than most other suggestions.
That makes my heart ache all over again T_T
Updated by zverok (Victor Shepelev) almost 5 years ago
- Description updated (diff)
Updated description significantly to clarify intentions.
Updated by osyo (manga osyo) almost 5 years ago
Is .:hoge
a :hoge.to_proc
syntax sugar?
And, is map(&self.:hoge)
and map(&.:hoge)
different in meaning?
In Ruby, self.hoge
and hoge
without self
often have the same meaning
But, it looks very strange that map(&self.:hoge)
and
map(&.:hoge)` have different meanings.
Updated by zverok (Victor Shepelev) almost 5 years ago
is
.:hoge
a:hoge.to_proc
syntax sugar?
No, and that's a core of the proposal. :hoge.to_proc
produces regular Proc
(which is different internally, but conceals this fact). .:hoge
is meant to produce new core object.
In Ruby,
self.hoge
andhoge
withoutself
often have the same meaning
But, it looks very strange thatmap(&self.:hoge)
andmap(&.:hoge)
have different meanings.
Yes, that's an obvious compromise for this proposal. The intuition of "you could drop self.
" is already broken for self.:hoge
, which is prohibited currently; I believe it would be a "gotcha", yet bearable one.
Updated by matz (Yukihiro Matsumoto) over 4 years ago
- Status changed from Feedback to Closed
We have suspended the idea of .:
operator, so this proposal is no longer meaningful. We may revisit the method reference operator in the future. We have to remember this PR as well. FYI, we have numbered parameters now, so the proposed [1, 2, 3].map(&.:to_s)
can be expressed by [1, 2, 3].map{_1.to_s}
. This could lower the value of this proposal.
Matz.
Updated by Dan0042 (Daniel DeLorme) over 4 years ago
As a note for the future:
Since unary operations like -@
and ~@
use the @
sigil to represent the receiver, I think it would make sense for the callable instance method object to have a syntax like @.hoge