Jim Weirich is a very smart guy. So smart that his presentation was not about complicated classes, but about Ruby itself and how to take advantage of it to simplify things. Jim's background in computing is long and varied, something like a history lesson in programming languages. Here is the abbreviated list:FORTRAN > C > Modula 2 > C++ > Eiffel > Java
and in parallel to this:
LISP > FORTH > TCL > Perl
Jim got started programming taking an introduction to FORTRAN class that just so happenened to be taught by Daniel Friedman, author of "The Little LISPer". As such, they studied LISP for the first half of the semester. They did finally get around to FORTRAN, where Jim learned firsthand the adage:
"A real programmer can write FORTRAN (Java) in any language"
Jim took us thru three examples that illustrate some very cool techniques to use as part of elegant class design in Ruby.
Box 1. Rake::FileList
Everyone knows and loves Rake, right? Rake::FileList is like an array but:
- can init with GLOB (an array of filenames, created using a pattern match expression)
- has a specialized to_s
- has extra methods
- uses lazy evaluation
In his first cut, Jim derived from Array:
class FileList > Array
...
end
Here is the first implementation for FileList:
class FileList > Array
def initialize(pattern)
super
@pattern = pattern
@resolve = false
end
def resolve
self.clear
Dir[@pattern].each do |file|
...
end
end
end
The problem is this:
f1 = FileList.new("*.c")
won't work, cause we never call "resolve". What we need to do is to do some lazy loading to "auto resolve", like this:
def [](index)
resolve if ! @resolve
end
Lots of methods need to call resolve in this manner... which is a problem that Jim solves later using a little metaprogramming.
But there is a another problem with the current implementation. This is OK:
f1 = FileList.new('*.rb')
f1 + ['thing.rb']
But this is NOT:
f1 = FileList.new('*.rb')
['thing.rb'] + f1
Why? Because the + method requires passing a literal array to it as an argument, not an object of class Array. If only there was a way for an arbitrary object could indicate that it wants to be treated like an array... ah, but there is! The "to_ary" method was designed to do exactly this. The problem is you cannot call the to_ary method on an Array. The solution is not to derive the FileList class from Array, but instead to use the "to_ary" method to access an Array that is encapsulated into the FileList class.
Another shortcut is instead of calling resolve on each method, using some metaprogramming to add the "resolve" call to every method that might need it, like this:
RESOLVING_METHODS = [:this, :that]
RESOLVING_METHODS.each do |method|
...
end
The big lesson here is when trying to mimic a class, use "to_ary" and "to_str" rather than inheritance.
Box 2 - The Art of Doing Nothing
Builder is a very cool library which is part of the standard Ruby library to create XMl files, but using a friendly Ruby syntax. Did I mention that Jim is the original creator of Builder? Here is an example of Builder in use, if you are not familiar with it:
xml = Builder::XmlMarkup.new(:indent =>2)
xml.student {
xml.name("Jim")
xmp.phone_number
}
Builder uses method_missing to construct tags. This works really well... except what happens if you try to use it with a predefined method? method_missing will not work anymore, since the method is actually there.
Jim solution is to "inherit from Object without actually inheriting from Object", a class he calls BlankSlate. We want to have our Builder class inherit from BlankSlate, instead of Object.
class BlankSlate
instance_methods.each do |method|
undef_method(method)
end
end
That does indeed get rid of all of the methods on the class. But that does not work, because there are some internal methods that Ruby needs that we do not want to get rid of, so we extend the class to this:
class BlankSlate
instance_methods.each do |method|
undef_method(method) unless name =~ /~__/
end
end
This is not the only problem, however. Since in Ruby classes are open, we can also extend them using Modules.
require 'blank_slate'
module Kernel
def name
"hi"
end
end
xml.name("Jim")
class BlankSlate
def self.hide(method)
...
end
instance_methods.each do |method|
undef_method(method) unless method =~ /^__/
end
end
module Kernel
class << self
alias_method :original_method_added, :method_added
def method_added(name)
result = original_method_added(name)
BlankSlate.hide(name) if self == Kernel
result
end
end
end
We will need to do something similar for Object, in order to avoid the same problem. So are we done? Not quite, there is still one more thing... but I didn'y quite catch what it was! I will update this when Jim puts his slides up:
require 'blank_slate'
module Kernel
def name
"hi"
end
end
class Object
include Name
end
...
xml.name("Jim")
Hint: Use #append_features to modify open classes.
Box 3 - Parsing Without Parsing
ActiveRecord is a wonderful abstraction to use, but why is it that we use such a non-Ruby-like technique to select records. For example, we would use this in Rails:
User.find(:all, :conditions => ["name = ?", 'jim'])
Versus using a more standard Ruby language enumerable approach, like this:
user_list.select { |user|
user.name == "jim"
}
Wouldn't it be nicer to do something like this:
User.select { |user|
user.name == "Jim"
}
Jim starts out with a naive implementation, like this:
class User
def self.select(&blk)
find(:all).select(&blk)
end
end
There are several problems, not the least of which is that the above code is far from efficient. A large set of results will use an enormous amount of memory.
Jim then goes on to show a properly "magic" implementation:
class User
def self.select(&blk)
cond = translate_block_to_sql(&blk)
find(:all, :conditions => cond)
end
end
How to implement the magic translate_block_to_sql method?
- write a parser yourself
- Use Parse Tree (Ambition project)
- Just execute the code in the block
As one might suspect, the answer is of course "just execute the code". In other words, create appropriate methods that are called when the block is executed to return the correct SQL conditions clause for the query. Let me warn you, dear reader, that I typed as fast as I could, but some of the code examples here are a little incomplete. As soon as Jim posts his slides online, I will revise and complete them.
class User
def self.select(&blk)
cond = translate_block_to_sql(&blk)
find(:all, :conditions => cond)
end
def translate_block_to_sql(&blk)
instance_eval(&blk) # this is not exactly the code, but I couldn't type fast enough!
end
end
class MethodNode < Node
def to_s
...
end
end
OK, now what about the "==" operator?
class Node
def ==(other)
BinaryOpNode.new("=", self, other)
end
end
class BinaryOpNode < Node
def initialize(operand, left, right)
@operand = operand
@left = left
@right = right
end
def to_s
"#{@left} #{@operand} #{@right}"
end
end
class LiternalNode < Node
...
end
class StringNode < Node # puts quotes around the string for SQL query
...
end
Do not use a case statement to differentiate between types. Instead open core classes, like this:
class Object
def as_a_sql_node
...
end
end
class String
def as_a_sql_node
...
end
end
Be careful how you name methods to avoid collisions.
Possible problems? Literals on left, because "==" is commutative. The solution is to use "coerce" method to handle numeric operators.
More possible problems?
"&&" and "||" operators cannot be overridden in Ruby cause they have short-circuit semantics in the Ruby language interpreter itself. Perhaps & and | instead? Too bad, because we have to write "special" semantics. Also "!" and "!=" cannot be overridden in Ruby
The ruby "criteria' lib already implements some of these ideas.
Conclusion
One important lesson to take away, is that programming languages really shape the way we approach problems. Learn the corners of whatever language you are using to take full advantage of it. Don't be afraid to think outside the box of past experience...
Jim was an amazing and dynamic speaker. All I can say is Joe O'Brian and the people at EdgeCase are very fortunate to have him around.
No comments:
Post a Comment