[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: complexity of mechanism

This page is part of the web mail archives of SRFI 89 from before July 7th, 2015. The new archives for SRFI 89 contain all messages, not just those from before July 7th, 2015.



On Apr 12, felix winkelmann wrote:
> 
> I must say that even though I use keyword args regularly, I don't
> think they are the be-all and end-all solution for everything.
> Keywords are one possible way to handle complex argument sets, it's
> relatively easy to implement and scales (sort of) with regard to
> maintenance.

That same description applies to plenty of other srfis that are far
from a "be-all and end-all solution".  (That's what RnRS is for...
hopefully...)


> We can slap a keyword-enabled interface on to everything, but that
> doesn't mean we should (this also applies to SRFI-76). It's a nice
> solution if you need something quickly, but tends to get ugly later
> and is less efficient than (say) wrapping parameter-sets into custom
> data structures or providing a multi-procedure interface.

Certain situations certainly do call for lots of keywords -- and since
this is a srfi, then it sounds more like it's intended for private
solutions to abstraction problems you face, and not changing the core
with things like (member x l :test ...) instead of other member
functions.

Let me give an example -- in PLT there is a library called SlideShow
for generating presentations.  It currently has these functions for
constructing slides:

  (slide/title title-string element ...)
  (slide/title/tall title-string element ...)
  (slide/title/center title-string element ...)
  (slide element ...)
  (slide/center element ...)
  (slide/title/inset title-string slide-inset element ...)
  (slide/title/tall/inset title-string slide-inset element ...)
  (slide/title/center/inset title-string slide-inset element ...)
  (slide/inset slide-inset element ...)
  (slide/center/inset slide-inset element ...)
  (slide/name title-string element ...)
  (slide/name/tall title-string element ...)
  (slide/name/center title-string element ...)
  (slide/name/inset title-string slide-inset element ...)
  (slide/name/tall/inset title-string slide-inset element ...)
  (slide/name/center/inset title-string slide-inset element ...)

You can see how all of this can be folded into a single function.  One
popular argument is: "So what?  These functions construct different
kinds of things, so it's only proper that they have different names.".

This was exactly how this library was constructed.  Nobody seem to
like keywords too much, and it took time to reach the above
ridiculously long list to face the problem.  It is currently clear
that this is not a good solution -- nobody can use this thing without
the manual open.  (Was it `slide/title/center/inset' or
`slide/center/title/inset'?)

Another popular argument is that you could use a better interface, say
something like:

  (slide options element ...)
  ;; options is some alist or whatever

This sort of works, but it has two problems.  First, I have never seen
it used in complex situations where it's really needed.  (The closest
thing I've seen is a list of symbols that serve as gui flags, like
(message-box "title" "prompt" '(gui feature symbols)).)  My assumption
is that feels heavy enough that you'll never do it in practice.  It
feels heavy because you'll be consing some structure on every call --
yes, the same happens for a rest argument, but that doesn't feel
heavy, partly because the syntax for calling a function is very
light.  (I never saw a keyword facility that requires calling function
with something like (foo `(:bar . 3) `(:baz . ,x)).)  Again, please
don't bother to tell me how stupid this argument is: I did use "feel",
and I do believe that syntax matters (otherwise I'd use a lot more
braces in my code).

A second problem with this, which I consider more important is that
the above requires me to change the function in a way that all
call-sites need to be updated.  This is a serious problem, and
sometime it's not even feasible when the code is used in enough
places.  Another good example is make-struct-type in MzScheme.  Each
time there's a new feature added, another argument is added to the
infinite list of arguments.  The result is yet another one of these
use-with-the-manual-only functions, and in this case if you happen to
use the recently added feature, you'll need to go over the whole list
and stick in the appropriate default values (some are 0, some null,
and some #f).

Keyword arguments can perform miracles in such situations: you can
customize the function in a way that nothing needs to be updated.  It
took me some time to discover how useful this is -- I first added
keywords to Swindle as yet another hack that makes it possible to
imitate CL stuff in Scheme, but didn't use it much.  Only when I got
to write web pages I found it to be so useful.  It's exactly what Paul
Graham writes about at

  http://lib.store.yahoo.net/lib/paulgraham/bbnexcerpts.txt

So purists don't like it.  I don't care.  As a semi-purist (at times
when I choose to be one) I can dislike it too, but for getting stuff
done it's damn useful.


> So I must say that I can understand Taylor's reservations regarding the
> complexity of this facility. Specifically the descriptions of the whole
> #!rest-before-#!key or #!rest-after-#!key thingy borders on the
> hilarious.

That can be confusing, which is why we chose to stick with differen
meta keyword in the argument list -- but the functionality is very
convenient, with xml as the natural analogy for using the rest after
teh keys thing.  (But I'm talking about attributes for function calls
still; please avoid redundant arguments about encoding xml.)


> Again, I use this stuff: I like quick hacks and I have to think less
> about API-design when using keyword args. But It's still ugly and
> ad-hoc, and no SRFI material for me (everything IMHO, naturally).

As Joe Marshall said at some point -- it's a complex hack, and not too
good-looking, but it's still very practical.  (But I don't know what
you're trying to say above -- you argue for the thing, then conclude
that it's not srfi material.  I can think of several srfis that are
less useful thant this one.)

-- 
          ((lambda (x) (x x)) (lambda (x) (x x)))          Eli Barzilay:
                  http://www.barzilay.org/                 Maze is Life!