This reverts commit c24c8852.

Here's the explanation I just sent to @tenderlove:

Hey,

I've been thinking about about the transaction memory leak thing that we
were discussing.

Example code:

post = nil
Post.transaction do
  N.times { post = Post.create }
end

Post.transaction is going to create a real transaction and there will
also be a (savepoint) transaction inside each Post.create.

In an idea world, we'd like all but the last Post instance to be GC'd,
and for the last Post instance to receive its after_commit callback when
Post.transaction returns.

I can't see how this can work using your solution where the Post itself
holds a reference to the transaction it is in; when Post.transaction
returns, control does not switch to any of Post's instance methods, so
it can't trigger the callbacks itself.

What we really want is for the transaction itself to hold weak
references to the objects within the transaction. So those objects can
be GC'd, but if they are not GC'd then the transaction can iterate them
and execute their callbacks.

I've looked into WeakRef implementations that are available. On 1.9.3,
the stdlib weakref library is broken and we shouldn't use it.

There is a better implementation here:

https://github.com/bdurand/ref/blob/master/lib/ref/weak_reference/pure_ruby.rb

We could use that, either by pulling in the gem or just copying the code
in, but it still suffers from the limitation that it uses ObjectSpace
finalizers.

In my testing, this finalizers make GC quite expensive:
https://gist.github.com/3722432

Ruby 2.0 will have a native WeakRef implementation (via
ObjectSpace::WeakMap), hence won't be reliant on finalizers:
http://bugs.ruby-lang.org/issues/4168

So the ultimate solution will be for everyone to use Ruby 2.0, and for
us to just use ObjectSpace::WeakMap.

In the meantime, we have basically 3 options:

The first is to leave it as it is.

The second is to use a finalizer-based weakref implementation and take
the GC perf hit.

The final option is to store object ids rather than the actual objects.
Then use ObjectSpace._id2ref to deference the objects at the end of the
transaction, if they exist. This won't stop memory use growing within
the transaction, but it'll grow more slowly.

I benchmarked the performance of _id2ref this if the object does or does
not exist: https://gist.github.com/3722550

If it does exist it seems decent, but it's hugely more expensive if it
doesn't, probably because we have to do the rescue nil.

Probably most of the time the objects will exist. However the point of
doing this optimisation is to allow people to create a large number of
objects inside a transaction and have them be GC'd. So for that use
case, we'd be replacing one problem with another. I'm not sure which of
the two problems is worse.

My feeling is that we should just leave this for now and come back to it
when Ruby 2.0 is out.

I'm going to revert your commit because I can't see how it solves this.
Hope you don't mind... if I've misunderstood then let me know!

Jon

* There is no need to delete the primary key from cloned attributes,
  since it sets the same pk to nil afterwards.
* Check for empty? instead of any? to run initialize callbacks.

transaction.

any? will check that each item in the array is truthy, as opposed to
!empty? which will simply check that the array has length. For an empty
array, !empty? still seems to be faster than any?

deep_dup is slow. we only need to dup the values, so just do that
directly.

It's not really a good idea to have this as a global config option. We
should allow people to specify the behaviour per association.

There will now be two new values:

* :dependent => :restrict_with_exception implements the current
  behaviour of :restrict. :restrict itself is deprecated in favour of
  :restrict_with_exception.
* :dependent => :restrict_with_error implements the new behaviour - it
  adds an error to the owner if there are dependent records present

See #4727 for the original discussion of this.

This reverts commit 14fc8b34.

Reason: we need to discuss a better path from this removal.

Conflicts:
	activerecord/lib/active_record/reflection.rb
	activerecord/test/cases/base_test.rb
	activerecord/test/models/developer.rb

This feature adds a lot of complication to ActiveRecord for dubious
value. Let's talk about what it does currently:

class Customer < ActiveRecord::Base
  composed_of :balance, :class_name => "Money", :mapping => %w(balance amount)
end

Instead, you can do something like this:

    def balance
      @balance ||= Money.new(value, currency)
    end

    def balance=(balance)
      self[:value] = balance.value
      self[:currency] = balance.currency
      @balance = balance
    end

Since that's fairly easy code to write, and doesn't need anything
extra from the framework, if you use composed_of today, you'll
have to add accessors/mutators like that.

Closes #1436
Closes #2084
Closes #3807

Get rid of ActiveModel::Configuration, make better use of
ActiveSupport::Concern + class_attribute, etc.

There are two possible scenarios where the @mass_assignment_options
instance variable can become corrupted:

1. If the assign_attributes doesn't complete correctly, then
   subsequent calls to a nested attribute assignment method will use
   whatever options were passed to the previous assign_attributes call.

2. With nested assign_attributes calls, the inner call will overwrite
   the current options. This will only affect nested attributes as the
   attribute hash is sanitized before any methods are called.

To fix this we save the current options in a local variable and then
restore these options in an ensure block.

Introduced in 7ecfe3d3

Add #nodoc to initialize_dup and use :method: to document the #dup method.
Relates to issue #6235

The #relation method gets called in four places and the return value was instantly cloned in three of them. The only place that did not clone was ActiveRecord::Scoping::Default::ClassMethods#unscoped. This introduced a bug described in #5667 and should really clone the relation, too. This means all four places would clone the relation, so it doesn't make a lot of sense caching it in the first place.

The four places with calls to relations are:

activerecord/lib/active_record/scoping/default.rb:110:in `block in build_default_scope'"
activerecord/lib/active_record/scoping/default.rb:42:in `unscoped'"
activerecord/lib/active_record/scoping/named.rb:38:in `scoped'"
activerecord/lib/active_record/scoping/named.rb:52:in `scope_attributes'"

This apparently fix the warning related to @new_record variable not
being initialized in AR's test suit, when an association is built and
the object is marshalled/loaded.

See these tests in AR's base_test.rb:

test_marshalling_with_associations
test_marshalling_new_record_round_trip_with_associations

Closes #3720.