Forum: Ruby `finalize' method?

Yeah, it's me again, your friendly neighbourhood power-suggester!

I'm wondering why there isn't a `finalize' method classes can define,
that will be called immediately before instances of that class are
garbage collected? I'm aware of ObjectSpace.define_finalizer, but it
seems un-OO, and it doesn't seem to work when being called from within
the instance methods. It also seems as if the Proc sent to
`define_finalizer' isn't called until *after* the object has been 
destroyed.

   class Klass
     def finalize
       # finalize something
     end
   end

A use case (the best I can come up with at the moment):

   class Contact
     def initialize(filename)
       @filename = filename
       parse_xml_file(@filename)
     end

     def finalize
       save_as_xml(@filename)
     end
   end


Cheers,
Daniel

dasch wrote:
> Yeah, it's me again, your friendly neighbourhood power-suggester!
> 
> I'm wondering why there isn't a `finalize' method classes can define,
> that will be called immediately before instances of that class are
> garbage collected? I'm aware of ObjectSpace.define_finalizer, but it
> seems un-OO, and it doesn't seem to work when being called from within
> the instance methods. It also seems as if the Proc sent to
> `define_finalizer' isn't called until *after* the object has been 
> destroyed.

That's correct, it prevents problems like the following code:

  class Klass
    def finalize
      $global_reference = self
    end
  end

Jim Weirich wrote:
> 
> That's correct, it prevents problems like the following code:
> 
>   class Klass
>     def finalize
>       $global_reference = self
>     end
>   end
> 

In that case I'd expect $global_reference to be nil after the `finalize'
method is done executing, since the Klass instance has already been
deemed ready for destruction.


Cheers,
Daniel

dasch wrote:
> In that case I'd expect $global_reference to be nil after the `finalize'
> method is done executing, since the Klass instance has already been
> deemed ready for destruction.

"after the 'finalize' method is done" => Tricky.

How would this be implemented?  Seems to me Ruby would have to scan all 
of the current image to determine if any new references were created 
during the execution of the finalize method.  Especially since the 
references may be created deep inside other methods invoked by finalize.

And it might be more subtle than just checking for variable bindings. 
Consider:

  class NamedObject
    attr_reader :name
    def initialize(name)
      @name = name
    end
    def finalize
      proc { name }
    end
  end

  k = NamedObject.new("Billy")
  p = k.finalize  # Here p is a proc that has a binding that includes
                  # a "finalized" object.

The point is that the current mechanism avoids all these questions by 
only calling the finalizer /after/ the object has been destroyed.

Jim Weirich wrote:
> references may be created deep inside other methods invoked by finalize.
>       proc { name }
>     end
>   end
> 
>   k = NamedObject.new("Billy")
>   p = k.finalize  # Here p is a proc that has a binding that includes
>                   # a "finalized" object.
> 
> The point is that the current mechanism avoids all these questions by 
> only calling the finalizer /after/ the object has been destroyed.
> 

I just think it's weird that PHP has destructors and Ruby doesn't ;)

   http://www.php.net/manual/en/language.oop5.decon.php

I'm not really into C, neither am I familiar with the Ruby
implementation, but doesn't all variables just reference an object? So
somewhere, an object is stored, and each variable referencing it really
only holds that object's id. I imagine the `finalize' method could be
run, and afterwards, the object could be replaced with nil, maybe
retaining the old object id. That way any variables that point to the
object will be nil.

Now, I don't know if that's how it works, but it sounds logical to me,
speaking as a guy who's never taken any programming lessons or
programmed/scripted using other languages than PHP, JavaScript and 
Ruby...


Cheers,
Daniel

On Sat, 26 Nov 2005, Daniel Schierbeck wrote:

> I just think it's weird that PHP has destructors and Ruby doesn't ;)
> 
>   http://www.php.net/manual/en/language.oop5.decon.php

I find this tricky too: the only model I can think of is that where
you need destructors, your constructor desstructs, and you get at
the object with a block, in the style of open("file"){|f| ...}; if
you see what I mean.
> 
> I'm not really into C, neither am I familiar with the Ruby implementation, but
> doesn't all variables just reference an object? So somewhere, an object is
> stored, and each variable referencing it really only holds that object's id. I
> imagine the `finalize' method could be run, and afterwards, the object could
> be replaced with nil, maybe retaining the old object id. That way any
> variables that point to the object will be nil.

But there is only one nil in the system, so it can't have many ids.
> 
> Now, I don't know if that's how it works, but it sounds logical to me,
> speaking as a guy who's never taken any programming lessons or
> programmed/scripted using other languages than PHP, JavaScript and Ruby...
> 
> 
> Cheers,
> Daniel
> 
        Hugh

Hugh Sasse wrote:
> On Sat, 26 Nov 2005, Daniel Schierbeck wrote:
> 
>> I just think it's weird that PHP has destructors and Ruby doesn't ;)
>>
>>   http://www.php.net/manual/en/language.oop5.decon.php
> 
> I find this tricky too: the only model I can think of is that where
> you need destructors, your constructor desstructs, and you get at
> the object with a block, in the style of open("file"){|f| ...}; if
> you see what I mean.

No, not really. I really can't stretch enough that I'm a noob... :)


>> I'm not really into C, neither am I familiar with the Ruby implementation, but
>> doesn't all variables just reference an object? So somewhere, an object is
>> stored, and each variable referencing it really only holds that object's id. I
>> imagine the `finalize' method could be run, and afterwards, the object could
>> be replaced with nil, maybe retaining the old object id. That way any
>> variables that point to the object will be nil.
> 
> But there is only one nil in the system, so it can't have many ids.

Could the object then by replaced by a reference to the nil object?
Again, I'm talking theoretically...


Cheers,
Daniel

On 11/25/05, Daniel Schierbeck <daniel.schierbeck@gmail.com> wrote:
> > the object with a block, in the style of open("file"){|f| ...}; if
> object's id. I
>
> Cheers,
> Daniel
>
>
Matz has also said that define_finalizer was in ObjectSpace on purpose 
to
discourage the use of finalizers. Finalizers are kind of weird, I would
recommend adding a clean-up method to any object you think "needs" a
finalizer. At least then the object still exists as in in a known state. 
Try
to not to think of them as destructors (in C++ sense) because they 
aren't.

Jim Weirich <jim-keyword-rforum.c88827@weirichhouse.org> writes:

>
> That's correct, it prevents problems like the following code:
>
>   class Klass
>     def finalize
>       $global_reference = self
>     end
>   end

What's the problem if this method ran on the first GC once and then
the object just stayed around?  I.e. have a three-pass GC, mark, run,
sweep?

On Sat, 26 Nov 2005, Daniel Schierbeck wrote:

> > you see what I mean.
> 
> No, not really. I really can't stretch enough that I'm a noob... :)

The braces, curly brackets, are a block.  The CS people will tell
you it is a closure.  What it means is: In the current scope (local
variables defined, etc) execute open("file") to open the file for
reading, and then [when inside open it effectively calls yield] you
will have a parameter which we'll call f, and we'll do something
with it in the local scope until the closing brace of the block.  On
the closing brace, open will close the file for us.   The oft' used
each method behaves in a similar way.

For an intro:
http://www.rubycentral.com/book/tut_containers.html#S2

For info about using them with constructors:
http://www.rubycentral.com/articles/insteval.html

        [...]
> > > imagine the `finalize' method could be run, and afterwards, the object
> > > could
> > > be replaced with nil, maybe retaining the old object id. That way any
> > > variables that point to the object will be nil.
> > 
> > But there is only one nil in the system, so it can't have many ids.
> 
> Could the object then by replaced by a reference to the nil object? Again, I'm
> talking theoretically...

Ruby doesn't have references, otherwise it could.
> 
> 
> Cheers,
> Daniel
> 
        Hugh

Hugh Sasse wrote:
>>> the object with a block, in the style of open("file"){|f| ...}; if
> each method behaves in a similar way.
Oh, sorry, I misunderstood you: You mean something like this:

   class Klass
     def initialize
       yield self
       finalize
     end

     def finalize
       # do something
     end
   end

Sorry, I'm kinda groggy ;)


Cheers,
Daniel

On Sat, 26 Nov 2005, Daniel Schierbeck wrote:

> Oh, sorry, I misunderstood you: You mean something like this:
> 
>   class Klass
>     def initialize
        # setup usual stuff, then:
>       yield self
>       finalize
>     end
> 
>     def finalize
>       # do something
        # like releasing resources, yes.
>     end
>   end
> 
> Sorry, I'm kinda groggy ;)

Not that groggy, that's exactly what I meant.
> 
> Cheers,
> Daniel
> 
        Hugh

chneukirchen wrote:
> What's the problem if this method ran on the first GC once and then
> the object just stayed around?  I.e. have a three-pass GC, mark, run,
> sweep?

There are two problems:

(1) Detecting that the finalizer has created new reference to a object 
scheduled for collection is problematic.  I'm thinking this would 
require a new mark pass to detect.

(2) What is the state of an object that has been finalized, but sticks 
around afterwards?  Is it valid to call methods on it?  After all, 
supposedly all of its resources have been finalized at this point.  And 
when it becomes eligible for  collection again, does the finalizer need 
to run again or not?

Most languages that allow finalizers explicitly state that creating new 
references duing finalization is not allowed and depend upon the 
programmer to follow that rule.  Ruby handles it by making it impossible 
for the programmer to break that rule (i.e. the finalizer doesn't get a 
reference to the object, making it difficult to create a reference to 
it).

-- Jim Weirich

Jim Weirich <jim-keyword-rforum.c88827@weirichhouse.org> writes:

> chneukirchen wrote:
>> What's the problem if this method ran on the first GC once and then
>> the object just stayed around?  I.e. have a three-pass GC, mark, run,
>> sweep?
>
> There are two problems:
>
> (1) Detecting that the finalizer has created new reference to a object 
> scheduled for collection is problematic.  I'm thinking this would 
> require a new mark pass to detect.

No, this is not required by my scheme.  Methods which have a finalize
will require two sweeps to actually get removed (that's not that bad,
the second one will follow soon if the first one wasn't sucessful).
You need to keep track of the objects whose finalizer has been run,
yeah.  I'm sure we can use a bit left over somewhere for that...

> (2) What is the state of an object that has been finalized, but sticks 
> around afterwards?  Is it valid to call methods on it?  After all, 
> supposedly all of its resources have been finalized at this point.  And 
> when it becomes eligible for  collection again, does the finalizer need 
> to run again or not?

It is unreclaimed. Yes. Yes. No. Yes.
You can keep track of disposal using an instance variable.

> Most languages that allow finalizers explicitly state that creating new 
> references duing finalization is not allowed and depend upon the 
> programmer to follow that rule.  Ruby handles it by making it impossible 
> for the programmer to break that rule (i.e. the finalizer doesn't get a 
> reference to the object, making it difficult to create a reference to 
> it).

That is a valid option, but I think Ruby can do better at little cost.

> method is done executing, since the Klass instance has already been
> deemed ready for destruction.

This would force the interpreter to keep a back reference for all
objects to variables which keep references to them, not very efficient
in regard of cpu and memory. Java solves this problem by reviving those
objects, which seems at least for me very odd.

So generally you should avoid finalizers, because most of the time you
are only interested in running some code before the application exits.

This should help in most cases:

BEGIN { $exit_listeners = [] }

def add_exit_listener(object)
  $exit_listeners << object
end

END { $exit_listeners.each {|listener| listener.on_exit } }

Alternatively you may declare an END block inside your initialize
method.

Hugh Sasse wrote:
>>
>> Daniel
>>
>         Hugh

What would you call such an approach? Encapsulating? I use this often
enough that I think it would fit lovely in a module:

   module Encapsulatable
     def self.included(klass)
       super
       klass.module_eval do
         def self.new(*args)
           yield obj = __new__(*args)
           obj.finalize if obj.respond_to? :finalize
         end
       end
     end
   end


Cheers,
Daniel

>>>>> "D" == Daniel Schierbeck <daniel.schierbeck@gmail.com> writes:

D> What would you call such an approach? Encapsulating? I use this often
D> enough that I think it would fit lovely in a module:

 A bad approach, at least for me.

D>          def self.new(*args)
D>            yield obj = __new__(*args)
D>            obj.finalize if obj.respond_to? :finalize
D>          end

 There is a convention in ruby : the method ::new must return an object 
of
 the class. You must change the name of the method if its return nil, or
 another object.


Guy Decoux

ts wrote:
> D>          end
> 
>  There is a convention in ruby : the method ::new must return an object of
>  the class. You must change the name of the method if its return nil, or
>  another object.

Roger. That's not that hard to do:

   module Encapsulatable
     def self.included(klass)
       super
       klass.metaclass_eval do
         private :new

         def encapsulate(*args)
           yield obj = new(*args)
           obj.finalize if obj.respond_to? :finalize
         end
       end
     end
   end

Note that this doesn't prevent the user from calling methods on the
object after its `finalize' method has been called.

   class Klass
     include Encapsulatable

     def initialize
       puts "initializing..."
     end

     def foo
       puts "called method `foo'"
     end

     def finalize
       puts "finalizing..."
     end
   end

   ref = nil
   Klass.encapsulate { |obj| obj.foo; ref = obj }
   ref.foo

But then again, this ain't Java. If you want to, you could wreak havoc
in seconds. But we're all friends, right? :)


Cheers,
Daniel

Hi,

How about making self a WeakRef in the finalize method?
http://www.ruby-doc.org/stdlib/libdoc/weakref/rdoc/

Regards,

Peter

On 11/30/05, Peter C. Verhage <peter@no-nonsense.org> wrote:
> Hi,
>
> How about making self a WeakRef in the finalize method?
> http://www.ruby-doc.org/stdlib/libdoc/weakref/rdoc/

Or just say in the docs that any references to self will become
invalid after finalize is called.  I don't think doing this is any
less safe than using #object_id and ObjectSapce::_id2ref (same type of
error if you use this on an object_id already GCed).  Right now if you
are using ObjectSapce::define_finalizer, you are likely using
ObjectSpace::_id2ref somewhere (you probably have some data-structure
holding object_id's).

Maybe this method could be called #_finalize to say that it isn't safe
(like ObjectSpace::_id2ref).