From: Tim Bates <tim@...>
Date: 2004-05-17T22:52:22+09:00
Subject: How to duck type? - the psychology of static typing in Ruby

Hi all,
Following a discussion in #ruby-lang, I have a suggestion about how to 
approach Duck Typing. Below is my dissertation on the subject. :P My 
intention is to incorporate any comments people might have into the text 
and then place it on the Wiki as an introduction to Duck Typing for the 
static typist.

For those not in on the secret, the idea is that if an object walks like 
a duck and quacks like a duck, it may as well be a duck - this being a 
metaphor for an arbitrary object that may not be exactly the same class 
your code was expecting, but still behaves the same way - see [1] if you 
don't follow.

---

Many people coming to Ruby from a statically-typed language are somewhat 
afraid of Ruby's dynamism, or "don't get it(TM)". David Black and I 
believe that this is in part because it is thought that the uncertainty 
and changeability built into Ruby are dangerous and one wants to find 
shelter from them.

Please bear with me while I describe some of the possible approaches.

1) People with a Static Typing background often have the urge to do 
something like this:

   attr_reader :date
   def date=(val)
     raise ArgumentError.new("Not a Date") if val.class != Date
   end

This is not duck typing - this is trying to get Ruby to do Static Typing.

2) Okay, you say, if that's not duck typing, let's do duck typing by 
accepting a whole bunch of different input formats and trying to turn 
them into something we know how to deal with, like this:

   def date=(val)
     class="keyword">case val
     when Date
       @date = val
     when Time
       @date = Date.new(val.year, val.month, val.day)
     when String
       if val =~ /(\d{4})\s*[-\/\\]\s*(\d{1,2})\s*[-\/\\]\s*(\d{1,2})/
         @date = Date.new($1.to_i,$2.to_i,$3.to_i)
       else
         raise ArgumentError, "Unable to parse #{val} as date"
       end
     when Array
       if val.length == 3
         @date = Date.new(val[0], val[1], val[2])
       end
     else
       raise ArgumentError, "Unable to parse #{val} as date"
     end
   end

This "normalization" approach has the advantage that the date attribute 
getter will always return a Date (producing certainty), but the setter 
can take input in a variety of formats.

2.a) Discussing this on #ruby-lang, David Black suggested the following 
optimization:

   def date=(val)
     begin
       @date = Date.new(val.year, val.month, val.day)
     rescue
       begin
         val =~ /(\d{4})\s*[-\/\\]\s*(\d{1,2})\s*[-\/\\]\s*(\d{1,2})/
         @date = Date.new($1.to_i,$2.to_i,$3.to_i)
       rescue
         begin
           @date = Date.new(val[0], val[1], val[2])
         rescue
           raise ArgumentError, "Unable to parse #{val} as date"
         end
       end
     end
   end

This has the advantage over (2) that it doesn't depend upon the class of 
val - if it acts enough like a string to use the =~ operator, then that 
clause will handle it, even if it's not descended from String - unlike 
the previous example. This makes it "more duck-typed", but still 
addresses the static-typist's fear of uncertainty and dynamism by 
providing a predictable response from #date (it will always be a Date). 
Unfortunately it's also slow.

3) Even "more duck-typed" is the approach of just testing that it 
responds to the appropriate methods, like so:

   # Accepts an object which responds to the +year+, +month+ and +day+
   # methods.
   def date=(val)
     [:year, :month, :day].each do |meth|
       raise ArgumentError unless val.responds_to?(meth)
     end
     @date = val
   end

In this case, we have removed the normalization instituted in example 
(2), but we have still ensured that the #date attribute conforms to some 
sort of interface, providing certainty. It is now the caller's 
responsibility to make sure what they pass fits the [:year, :month, 
:day] specification - but this responsibility is documented. However, 
this approach violates the Don't Repeat Yourself principle - both the 
code and the comment contain the specification, and are not therefore 
guaranteed to be in sync.

This approach is what many people believe to be embodied by "Duck 
Typing". Given an object, we're checking whether it walks and quacks 
like a duck; we're not forcing our caller to use a particular class, 
like example (1), but we are forcing our caller to put the data in a 
format we can understand, unlike (2) which attempts to deal with every 
possible representation of a date, causing volumes of maintenance work - 
imagine trying to write a normalization routine like that for every 
attribute of every class! In this way, we are moving the responsibility 
of putting the data into a reasonable format to the caller, who knows 
what format their data is in, from the receiver, who has to guess at 
every possible format the caller might send them.

4) The fourth and final approach, which I believe to be the Zen of Duck 
Typing, is as follows:

   # Accepts an object which responds to the +year+, +month+ and +day+
   # methods.
   attr_accessor :date

"What?" I hear you cry. "There's no checking there at all! You could 
pass it anything!" Yes, gentle reader, but why would you? After all, the 
documentation for this method is exactly the same as the one above. If 
the programmer using this method does what the documentation says then 
the class's behaviour is exactly the same. If they hand it the wrong 
thing (accidentally, we assume) then the only difference is that it 
breaks when the setter is called, rather than some time after the getter 
is called and we try and call a non-existent method on the result.

A common response to this often contains the phrase "meaningless error 
messages", but the results of such a mistake are usually, if not always, 
far from meaningless. For the most part, they look something like this:

   NoMethodError: undefined method `year' for "notadate":String

This tells me a lot: namely, that some part of my code (whose location 
is given in the subsequent backtrace) expected "notadate" to have a 
:year method, and it didn't. From this it is fairly trivial to deduce 
that something, somewhere, has fed the wrong thing to the date= setter 
method. Chances are that if your code is well-factored, there aren't a 
whole lot of places that set the date, and the location of the error can 
be found through a little judicious testing; you've lost the certainty 
and immediacy of the inline check, but not by much, and you've gained 
the flexibility of dynamic typing, and a whole lot less code to maintain.

Now if you'd been writing and collecting unit tests as you went along, 
instead of

   NoMethodError: undefined method `year' for "notadate":String

you would be seeing

   1) Failure:
test_stuff(MyClassTest) [./test/myclasstest.rb:13]:
<false> is not true.

which makes the error even easier to find: you go to test/myclasstest.rb 
and see something like:

10:  def test_date
11:    @obj = Foo.new
12:    @obj.date = MyClass.new.notadate
13:    assert(@obj.date.respond_to?(:year))
14:  end

and now the error is trivial to trace - the moral of the story being 
that when Duck Typing, do your checking in your unit tests, rather than 
in the live code. Type errors such as this one are usually the least 
common and easiest to trace of errors; if the attribute's documentation 
specifies what it is supposed to be, as in the example above, and the 
callers of both the getter and the setter methods make no assumptions 
about any more or less than what the documentation says, then apart from 
keyboarding accidents this will never be a problem.

At [1], Dave Thomas describes Duck Typing as "a way of thinking about 
programming in Ruby." I think he means to go a step further than that - 
Duck Typing is the _best_ way of thinking about programming in Ruby, and 
possibly the _only_ way; as David Black puts it:

"I think the concept of duck typing needs to be supplemented and 
expanded on. if, as seems to be the case, Dave thinks of it as a 
component of programming style, then it doesn't address language design 
itself. As long as duck typing is viewed as a stylistic choice, rather 
than a radical language principle, the door is always open to people 
saying 'I don't do duck typing', by which they usually mean that they 
use kind_of? a lot... of course Ruby itself *does* do duck typing, 
whether a given programmer thinks they're doing it or not."

Using kind_of? (or responds_to?) a lot isn't "not doing Duck Typing", 
it's simply adding in at run time the kinds of checks that Statically 
Typed languages do at compile time, in a usually verbose and necessarily 
incomplete fashion.

Rather than trying to make Ruby do Static Typing because one is from a 
Static Typing background and that's what one is comfortable with, one 
should become comfortable with the dynamic nature of Ruby instead. I 
have found that once I stopped assuming that the callers of my method 
(who may well be me, in five minutes time, or some user of my library on 
the other side of the planet) are stupid and don't know how to read my 
documentation (you did write some, didn't you?) then writing in Ruby 
became a whole lot more natural and somewhat less verbose. The unit 
tests took care of the psychological need to check, somewhere, that the 
method was getting passed the right thing, but in reality the whole 
debacle is a non-issue; type errors are the most trivial of bugs.

And if you're still worried about that date example, an alternative 
solution is this:

   def set_date(year, month, day)
     @date = Date.new(year, month, day)
   end

which, if year, month and day are not numeric, will catch the problem 
straight away - without resorting to Static Typing or some approximation 
of it. And the way it catches it is telling:

   irb(main):027:0> Date.new(2004.0, Rational(12,2), "17")
   ArgumentError: comparison of String with 0 failed
           from /usr/lib/ruby/1.8/date.rb:560:in `<'
           from /usr/lib/ruby/1.8/date.rb:560:in `valid_civil?'
           from /usr/lib/ruby/1.8/date.rb:590:in `new'
           from (irb):27

This is not "ArgumentError: parameters must be numbers" - the error is 
discovered when the Date class attempts to compare that parameter to 
zero and can't do it, after assuming that it was valid. And it didn't 
make the mistake any harder to find, did it? Notice that it didn't balk 
at Floats or Rationals, and with no extra coding from the implementor; 
Floats and Rationals look, and quack, like numbers. That's Duck Typing 
in action.

[1] http://rubygarden.org/ruby?DuckTyping

---

Tim.

-- 
Tim Bates
tim@bates.id.au