Or
format_as_title = ->{ |val| add_header[to_camel[val.strip]] }
Also, I think #+
is better.
trans (Thomas Sawyer) wrote:
Also, I think
#+
is better.
I saw facets has some similar feature that uses #*
instead, maybe because it looks a bit closer to Haskell's composition syntax. Nevertheless, I still like #<<
better, it feels your are "connecting" the blocks together.
I would vote for #*
. I think #<<
is usually changing the left argument (in place).
aprescott (Adam Prescott) wrote:
See also: http://web.archive.org/web/20101228224741/http://drmcawesome.com/FunctionCompositionInRuby
Maybe #|
could be a possibility. (Without implementing #>
or #<
).
But I find the article's proposition about the chaining order a bit missleading:
transform = add1 | sub3 | negate
For me that feels more like "piping" add1
to sub3
to negate
, from left to right, not the other way around.
If we choose to take that path I think the following code would be a plausible implementation:
class Proc def | block proc { |*args| block.to_proc.call( self.call(*args) ) } end
end class Symbol def | block self.to_proc | block end
end
What about #*
for composing traditionally (right to left) and #|
for piping (left to right)? In mathematics, depending of the area and the subject, both ways are used, and some argue that "piping" is more natural than "precomposing". However, when functions are "piped", the arguments are usually on the left: (arguments)(function1 function2).
Update: i think having the both was a bad idea, it would be redundant.
I agree, #*
is appropriate for composition.
alexeymuranov (Alexey Muranov) wrote:
Update: i think having the both was a bad idea, it would be redundant.
I was going to say the same thing. Having both #*
and #|
is redundant and also a bit confusing, since #|
doesn't really feel to be the opposite operation of #*
at any context. We should choose one or the other but not both.
I still like #|
(chaining from left to right) a bit better, but I rather have #*
than nothing.
Positive about adding function composition. But we need method name consensus before adding it?
Is #*
OK for everyone?
Matz.
Might I humbly suggest #<-
:
to_camel = :capitalize.to_proc
add_header = ->val {"Title: " + val} format_as_title = add_header <- to_camel <- :strip
Seems to have a nice symmetry with #->
I think "<-
" reads better but I'm ok with '*
' as well.
I'm with Joshua, I think #<-
reads a lot better.
I think that the meaning of #<-
would not be symmetric with the meaning of #->
.
Also, in mathematics, arrows are more like relations than operations. When used to describe functions, usually function arguments go to the arrow's tail, function values to the arrow's head, and function's name, for example, goes on top of the arrow.
(In this sense Ruby's lambda syntax would look to me more natural in the form f = (a,b)->{ a + b }
instead of f = ->(a,b){ a + b }
.)
The main drawback of #*
in my opinion is that is does not specify the direction of composition ((f*g)(x)
is f(g(x))
or g(f(x))
?), but since in Ruby function arguments are written on the right (f(g(x))
), i think it can be assumed that the inner function is on the right and the outer is on the left.
Update : Just for reference, here is how it is done in Haskell : http://www.haskell.org/haskellwiki/Function_composition
+1 for #*
The symbol used in mathematics for function composition is a circle (∘
); the arrows are for the definitions of functions (like lambdas) only, so #<-
or whatever make no sense to me.
Finally, the f ∘ g(x)
is defined as f(g(x))
, so there is no argument there either.
marcandre (Marc-Andre Lafortune) wrote:
+1 for
#*
The symbol used in mathematics for function composition is a circle (
∘
); the arrows are for the definitions of functions (like lambdas) only, so#<-
or whatever make no sense to me.
Very good point.
Finally, the
f ∘ g(x)
is defined asf(g(x))
, so there is no argument there either.
Not true. Depending on which field of mathematics you look at, either (f ∘ g)(x)
is either f(g(x))
, or it is g(f(x))
. The later is in particular true in work involving relations, see e.g. http://en.wikipedia.org/wiki/Composition_of_relations#Definition.
Speaking from a more programming-related viewpoint, f(g(x))
is what is used e.g. in Haskell, and probably in many other functional languages, and so may be familiar with many programmers.
However, we should take into account that a functional language writes e.g. reverse(sort(array))
, so it makes sense to define revsort = reverse * sort
(i.e. (f ∘ g)(x)
is f(g(x))
). But in Ruby, it would be array.sort.reverse
, so revsort = sort * reverseve
may feel much more natural (i.e. (f ∘ g)(x)
is g(f(x))
).
I agree that (f ∘ g)(x)
is g(f(x))
is more intuitive from a purely
programmatic point of view. It is "natural" for the operations to be
applied left to right, exactly like method chaining.
-- Matthew Kerwin, B.Sc (CompSci) (Hons)
"You'll never find a programming language that frees
you from the burden of clarifying your ideas." - xkcd
phluid61 (Matthew Kerwin) wrote:
I agree that
(f ∘ g)(x)
isg(f(x))
is more intuitive from a purely programmatic point of view. It is "natural" for the operations to beapplied left to right, exactly like method chaining.
When functions are applied from left to right, the argument is usually (if not always) on the left. The form (x)(fg)=((x)f)g
may look awkward (though i personally used it in a math paper), so i think usually the "exponential" notation is preferred: xfg = (xf )g, where xf corresponds to f(x)
in the usual notation.
With method chaining, IMO, the "main argument" of a method is the receiver, and it is on the left. Lambda
s and Proc
s are not chained in the same way as method calls.
Update: I agree that the common syntax for calling functions (f(x)
rather then (x)f
) should not be an obstacle if Ruby decides to consistently multiply functions putting the inner on the left and the outer on the right. Another syntax for calling functions can be invented in the future, or rubists can learn to live with this inconsistency. For example, Ruby (or Matz) can decide to multiply lambdas with the inner on the left and the outer on the right, and add the following syntax:
format_as_title = :strip.to_proc * :capitalize.to_proc * lambda { |val| "Title: " + val }
title = " over here " ^ format_as_title
Update 2012-11-11: I was not clear what i meant by "multiplying from left to right". I meant to say: putting the inner function on the left and the outer on the right. I am correcting this phrase.
In Math multiplication is always associative, even for matrix. I.e: (A*B)*C == A*(B*C)
. If we use *
for ∘
(composition) it resembles multiplication. Function composition is analog to matrix multiplication which are commonly used for transformation compositions as well. In fact, function composition is also associative.
So, when representing h = f ∘ g
as h = f * g
it makes sense to me (although Math preferring a different symbol for multiplication and composition is a good indication that we should consider this as well for Ruby - more on that later on). But Math representation is procedural, not object oriented. If we try to mix both approaches to fit Ruby philosophy this could lead to great confusion.
Ruby can be also used for procedural programming:
sqrt = ->(n){ Math.sqrt n } square_sum = ->(a, b) { a*a + b*b }
hypotenuse = sqrt * square_sum
5 == hypotenuse.call 3, 4
This makes total sense to me using procedural notation. I'm not sure how would someone use this using some OO notation instead...
Now with regards to composition notation, I think a different notation could help those reading some code and trying to understand it. Suppose this method:
def bad_name(bad_argument_name, b) bad_argument_name * b end
You can't know beforehand if bad_argument_name
is an array, a number or a proc/lambda. If we read this instead:
def bad_name(bad_argument_name, b) bad_argument_name <- b
end
we would then have a clear indication that bad_argument_name
is probably a proc/lambda. I know the same argument could be used to differentiate <<
between strings and arrays among other cases. But I think that function composition is conceptually much different from those other operations (concatenation, multiplication) than concatenation (<<
) is for strings and arrays. In both cases we are concatenating but concatenation means different things for strings and arrays in non surprising ways.
But then using this arrow notation I would expect that (a <- b
) would mean "a before b" (b(a(...)))
while (a ∘ b)
means "a after b" (a(b(...)))
.
I find it a bit awful to use "hypotenuse = square_sum <- sqrt
", although it is the way OO usually work ([4, 5].square_num.sqrt
- pseudo-code of course). But we would not be using "[4, 5].hypotenuse
", but "hypotenuse.call 4, 5
", right? So, since we're using procedural notation for procs/lambdas we should be thinking of procedural programming when deciding which operator to use.
I would really prefer to have lambda syntax as "double = <-{|n| n * 2}
" and function composition as "hypotenuse = sqrt -> square_sum
" (sqrt
after square_sum
). But since I don't believe the lambda syntax won't ever change, let's try to see this over a different perspective.
Instead of reading (a <- b)
as "a before b", I'll try to think of it as being "b applied to a" (a(b(...)))
. This also make sense to me so I can easily get used to this. It would work the same way as "*
" but there would be a clear indication that this refers to function composition rather than some generic multiplication algorithm.
Having said that, I'd like to confirm that I'm ok with either *
or <-
and I'd really like to have function composition as part of Ruby.
proc composition is not commutative, so the operator should:
i.e. the operator should be visually asymmetrical with clear directionality
e.g. <<
, <<<
, <-
a << b << c = a(b(c(x)))
perhaps it also makes sense to have the other direction: c >> b >> a = a(b(c(x)))
+1 to #*
.
+1 to rosenfeld's first 2 paragraphs ( h = f ∘ g
as h = f * g
, and matrix multiplication analogy).
-1 to "<-
". Rationale: It is too easy invent a guitar with one more string. Furthermore, when it comes to operators, I consider design by jury a weak approach.
Attached patches for Proc#*
and Method#*
for Proc and Method composition including test cases. Also raised as a pull request on GitHub at https://github.com/ruby/ruby/pull/935
One thing that might be worth discussing is the necessity of the type checking: should it be possible to compose any callable object (rather than explicitly Proc
s and Method
s)? The Ruby implementation suggested in the description of this issue suggests calling to_proc
on the given argument but we could also demand that the supplied argument responds to call
, e.g. in pure Ruby:
class Proc def *(g) proc { |*args, &blk| call(g.call(*args, &blk)) } end
end
vs.
class Proc def *(g) proc { |*args, &blk| call(g.to_proc.call(*args, &blk)) } end
end
Attached patch to support composing with any object that responds to call
(rather than raising a TypeError
if the object was not a Proc
or Method
), e.g.
class Foo def call(x, y) x + y end
end f = proc { |x| x * 2 }
g = f * Foo.new g.call(1, 2)
Regarding the syntax: I also support *
as the operator where f * g = f(g(x))
(as it seems close enough to the mathematical syntax already used by other languages such as Haskell and Idris) but if that is too divisive, we could choose a method name from the mathematical definition (https://en.wikipedia.org/wiki/Function_composition) instead:
The notation g ∘ f is read as "g circle f ", or "g round f ", or "g composed with f ", "g after f ", "g following f ", or "g of f", or "g on f ".
This opens up the following options:
Proc#compose
: f.compose(g) #=> f(g(x))
Proc#after
: f.after(g) #=> f(g(x))
Proc#following
: f.following(g) #=> f(g(x))
Proc#of
: f.of(g) #=> f(g(x))
Proc#on
: f.on(g) #=> f(g(x))
It would be nice to be able to compose functions in both ways, like in F#, you can do g << f
or g >> f
, sadly this was rejected before.
I would settle to have Proc#*
for "regular" composition and Proc#|
for "piping".
Last time there was no consensus about the syntax. Hopefully we can manage to solve this before 2.3 is released.
I'm teaching Haskell in a graduate class, so I'm quite familiar with function composition and use it a lot, but the original example isn't convincing at all. For me, in Ruby, something like val.strip.capitalize reads much, much better than some artificially forced function composition. If there were a method String#prepend, it would be even more natural: val.strip.capitalize.prepend('Title: ').
If there are better examples that feel more natural in Ruby, please post them here.
Martin Dürst wrote:
I'm teaching Haskell in a graduate class, so I'm quite familiar with function composition and use it a lot, but the original example isn't convincing at all. For me, in Ruby, something like val.strip.capitalize reads much, much better than some artificially forced function composition. If there were a method String#prepend, it would be even more natural: val.strip.capitalize.prepend('Title: ').
If there are better examples that feel more natural in Ruby, please post them here.
There is String#prepend, but it mutates the receiver.
Martin Dürst wrote:
I'm teaching Haskell in a graduate class, so I'm quite familiar with function composition and use it a lot, but the original example isn't convincing at all. For me, in Ruby, something like val.strip.capitalize reads much, much better than some artificially forced function composition. If there were a method String#prepend, it would be even more natural: val.strip.capitalize.prepend('Title: ').
If there are better examples that feel more natural in Ruby, please post them here.
I don't believe you need a pure PF language to benefit from a feature like this. Many ETL projects like transproc (https://github.com/solnic/transproc) would probably find it useful too.
Pablo Herrero wrote:
I don't believe you need a pure PF language to benefit from a feature like this. Many ETL projects like transproc (https://github.com/solnic/transproc) would probably find it useful too.
Transproc is what actually inspired me to submit a patch here: my hope is that having functional composition in the Ruby language itself will enable easier data pipelining using only Proc
s, Method
s and other objects implementing call
. The presence of curry
(http://ruby-doc.org/core-2.2.2/Proc.html#method-i-curry) seems like a good precedent for adding such functional primitives to the core.
I support the proposed Proc#*
/Method#*
syntax and semantics.
The feature being added is function composition; not relation composition, not method chaining. Its target audience is most likely to read f * g
as “f
after g
”, so that’s how it should work. Perhaps some Ruby programmers will not use this feature directly (as with Proc#curry
) because they neither program nor think in a functional style, but it should be designed to be useful and familiar to those who do. The proposed implementation achieves that.
The asterisk isn’t ideal, but it’s the best choice available.
Martin Dürst wrote:
I'm teaching Haskell in a graduate class, so I'm quite familiar with function composition and use it a lot, but the original example isn't convincing at all. For me, in Ruby, something like val.strip.capitalize reads much, much better than some artificially forced function composition. If there were a method String#prepend, it would be even more natural: val.strip.capitalize.prepend('Title: ').
If there are better examples that feel more natural in Ruby, please post them here.
I fully agree with you that a good example really helps to understand the problem which lead to a better solution.
The few times I've missed function composition is ruby is always when I wanted to replace the same pattern :
res = step1(base) res = step2(res) res = step3(res) ... res
I hate having to mutate the res variable and I would hate even more creating different temp variables especially when I do not know how many steps there will be. I would like to be able to combine the steps then apply the step combination to base
There are a lot of examples for this pattern so here is one : data retrieval with ActiveRecord ( I assume most people know this library but this is only an example )
messages = all_messages messages = restrict_to_owner(messages, owner) filters.each do |filter| messages = filter.apply(messages) end messages = messages.order(created_at: :asc) messages = paginate(messages)
With the recent addition of Hash#to_proc
and performance improvements to Proc
s in 2.3.0, I have rebased my patch (attached) to add composition between Proc
s, Method
s and other objects responding to call
with *
in the hope of reviving this proposal.
I wrote a gem with a C extension of Proc#compose
: https://github.com/mooreniemi/proc_compose#usage
What motivated me was map f (map g xs) = map (f . g) xs
, and what I still don't understand (being a newbie to extending Ruby or understanding its internals) is that .map(&some_proc).map(&some_other_proc)
still behaves better than .map(&(some_proc * some_other_proc))
given my current implementation of compose
. Although I think composition has a lot of uses, the admittedly small but free performance benefit I expected to gain was top of my list.
I do think emphasis on composition suggests a somewhat different style of writing Ruby, but I think it can be a good one, actually.
Paul: what's the performance of your compose
? If I have time later I can use https://github.com/mooreniemi/graph-function to try and see.
Alexander Moore-Niemi wrote:
Paul: what's the performance of your
compose
? If I have time later I can use https://github.com/mooreniemi/graph-function to try and see.
I ran the following benchmark with benchmark-ips:
require 'benchmark/ips' double = proc { |x| x * 2 } quadruple = proc { |x| x * 4 } octuple = double * quadruple inline_octuple = proc { |x| x * 2 * 4 } nested_octuple = proc { |x| quadruple.call(double.call(x)) } numbers = [1, 2, 3, 4, 5] Benchmark.ips do |x| x.report('composing procs') do numbers.map(&octuple) end x.report('chaining procs') do numbers.map(&double).map(&quadruple) end x.report('single proc') do numbers.map(&inline_octuple) end x.report('nested proc') do numbers.map(&nested_octuple) end x.compare! end
And also see a performance drop with composition over chaining multiple map
s:
Warming up -------------------------------------- composing procs 27.822k i/100ms chaining procs 32.096k i/100ms single proc 49.021k i/100ms nested proc 27.337k i/100ms Calculating ------------------------------------- composing procs 341.874k (± 0.5%) i/s - 1.725M in 5.045764s chaining procs 389.031k (± 0.7%) i/s - 1.958M in 5.032912s single proc 666.544k (± 0.6%) i/s - 3.333M in 5.001266s nested proc 321.919k (± 0.8%) i/s - 1.613M in 5.010562s Comparison: single proc: 666543.8 i/s chaining procs: 389031.4 i/s - 1.71x slower composing procs: 341873.8 i/s - 1.95x slower nested proc: 321919.1 i/s - 2.07x slower
It might be interesting to look at object allocations as we effectively create a nested Proc
which might account for the slow down.
Yukihiro Matsumoto wrote:
Positive about adding function composition. But we need method name consensus before adding it?
Is#*
OK for everyone?
Aside from implementation details, is the lack of consensus on the method name (and the resulting behaviour re left vs right operand) the main blocker here?
Paul Mucur wrote:
Yukihiro Matsumoto wrote:
Positive about adding function composition. But we need method name consensus before adding it?
Is#*
OK for everyone?Aside from implementation details, is the lack of consensus on the method name (and the resulting behaviour re left vs right operand) the main blocker here?
If Matz says so, then yes, this is the main blocker.
fyi: Composing procs is way faster now. It beats all other algorithms in Paul's benchmark on ruby 2.3.3:
Warming up -------------------------------------- composing procs 135.483k i/100ms chaining procs 55.595k i/100ms single proc 84.448k i/100ms nested proc 48.095k i/100ms Calculating ------------------------------------- composing procs 2.363M (± 6.2%) i/s - 11.787M in 5.011859s chaining procs 701.936k (± 3.4%) i/s - 3.558M in 5.074972s single proc 1.207M (± 3.6%) i/s - 6.080M in 5.044891s nested proc 579.005k (± 2.9%) i/s - 2.934M in 5.071310s Comparison: composing procs: 2362658.0 i/s single proc: 1206768.2 i/s - 1.96x slower chaining procs: 701936.2 i/s - 3.37x slower nested proc: 579005.4 i/s - 4.08x slower
Tested on linux and macOS (same result).
I want to make sure if everyone agrees with "*" instead of OP's "<<".
Besides that I also wanted to mention that Koichi concerns that function composition may be far slower than method chaining.
Matz.
matz (Yukihiro Matsumoto) wrote:
I want to make sure if everyone agrees with "*" instead of OP's "<<".
Besides that I also wanted to mention that Koichi concerns that function composition may be far slower than method chaining.Matz.
+1 for #*
Initially I thought of the F# convention #<< and it's counter part #>> as intuitive. But after giving it some thought, and practice, I prefer #* and #+.
class Proc def +(other) other.to_proc * self end def *(other) -> arg { self.call other.(arg) } end
end f = -> n { n * 2 }
g = -> n { n * 4 }
h = f * g puts h.(5) puts (h + :odd?).call(3)
Instead of composition I see potential use for shovel operations #<< and #>> for piping purposes similar to Elixir's #|>.
class Object def >>(transform) transform.(self) end
end class Integer def >>(num_or_func) return self << -num_or_func unless num_or_func.respond_to? :call super end
end class Proc def <<(input) call input end
end add_header = -> val {"Title: #{val}" }
format_as_title = add_header + :capitalize + :strip puts 'Title goes here.' >> format_as_title puts format_as_title << 'Title goes there.' puts 100 >> format_as_title puts 100 >> 2
Most languages do not define function composition in built-in operators, but provide them as function or method such as compose
.
F.Y.I) https://rosettacode.org/wiki/Function_composition
In some few languages defined function composition operator as following.
haskell : .
F# : <<
and >>
Groovy : <<
and >>
I think Ruby should proivide compose
method (and and_then
as forward composition), and alias some operator (like #<<
or `#*) to them.
In my opinion, +1 for <<
and >>
instead of *
.
Because there is no language define function composition as *
and +
, but F# and groovy defined as <<
and >>
. It is intutive.
By the way, It is useful if Symbol#<<
(or Symbol#+
) is shortcut method to sym.to_proc.compose(other_proc)
.
arr.map(&:upcase.to_proc.compose(:to_s.to_proc)) arr.map(&:to_s >> :upcase)
It is more visually and intuitive.
My reference implemenation of composition is following.
Please take this to next Developers Meeting Agenda?
The best possible name is discussed for 6 years already, and feature seems pretty nice to have. I believe that sometimes it is just good to make "executive decision" about name once and for all, at least better than postpone feature for years.
(I still unhappy with yield_self
name, though :))
yuroyoro (TOMOHITO Ozaki) wrote:
Most languages do not define function composition in built-in operators, but provide them as function or method such as
compose
.F.Y.I) https://rosettacode.org/wiki/Function_composition
In some few languages defined function composition operator as following.
haskell :
.
F# :<<
and>>
Groovy :<<
and>>
I think Ruby should proivide
compose
method (andand_then
as forward composition), and alias some operator (like#<<
or `#*) to them.In my opinion, +1 for
<<
and>>
instead of*
.
Because there is no language define function composition as*
and+
, but F# and groovy defined as<<
and>>
. It is intutive.By the way, It is useful if
Symbol#<<
(orSymbol#+
) is shortcut method tosym.to_proc.compose(other_proc)
.arr.map(&:upcase.to_proc.compose(:to_s.to_proc)) arr.map(&:to_s >> :upcase)It is more visually and intuitive.
My reference implemenation of composition is following.
I do like the idea of using shovel as it's already fairly common in the language, and present in others. I'd be interested in potentially making a container variant though:
arr.map(&:to_s >> :upcase) arr.map(&[:to_s, :upcase])
Though that would involve introducing the idea of Array#to_proc. I've used a similar technique in a few gems with #[]()
, so a vague implementation may look like:
class Array def to_proc self[1..].reduce(self[0].to_proc, :compose) end
end
Wherein it'd be nice if #compose
tried to coerce its argument:
def compose(sym) fn = sym.is_a?(Proc) ? sym : sym.to_proc ...
end
Currently planning on releasing a gem this weekend that will do a similar thing, it could serve as inspiration for an implementation, but my C is no where near good enough to try at the Ruby core level.
Aside: Have we ever considered implicitly to_proc
ing anything passed to a method expecting a block? e.g. arr.map(&:to_s)
becomes arr.map(:to_s)
I just realized that infix operators evaluate before to_proc
, allowing something like this:
[1,2,3].map(&:succ.to_proc >> :to_str.to_proc)
Codified it into a gem for kicks: https://github.com/baweaver/mf
Considering the combination of OOP and FP, it seems a good idea to adding both forward and reverse combination of procs. So we pick Groovy way (adding <<
and >>
methods to Proc
).
We need more discussion if we would add combination methods to the Symbol
class.
Matz.
matz (Yukihiro Matsumoto) wrote:
Considering the combination of OOP and FP, it seems a good idea to adding both forward and reverse combination of procs. So we pick Groovy way (adding
<<
and>>
methods toProc
).
What do you think about selecting only single operation? Here is Groovy example from provided link:
final times2 = { it * 2 } final plus1 = { it + 1 } final plus1_then_times2 = times2 << plus1 final times2_then_plus1 = times2 >> plus1
It looks like <<
is less intuitive and readable. I'm also really afraid of having chance to read something like a >> b << c >> d
.
pabloh (Pablo Herrero) wrote:
Is adding composition methods to the
Symbol
class still being considered?
No, I don't think that we consider a Symbol
object a method.
Applied in changeset trunk|r65911.
proc.c: Implement Proc#* for Proc composition
proc.c (proc_compose): Implement Proc#* for Proc composition, enabling
composition of Procs and Methods. [Feature #6284]
test/ruby/test_proc.rb: Add test cases for Proc composition.
From: Paul Mucur mudge@mudge.name