git HEAD regression: sessions lose changes

Attached is a patch to Rails that get the tests to pass again. Warning: it does so by effectively reverting this change.

The test cases isn't minimal, but is relatively small. Setup should take less than a minute:

git clone git://github.com/rubys/awdwr.git
cd awdwr
ruby makedepot.rb 6.1-7.4 save

Once that is complete, the scenario itself can be repeated at will with the following:

ruby makedepot.rb restore 8.1-8.5

In both cases, replace with the path to your checkout of Rails.

I'm traveling at the moment, so it will be a few days before I try to create a more minimal test case; but it does look like lighthouse mangled my instructions.

A more legible set of instructions can be found at the following places:

http://intertwingly.net/blog/2009/05/15/Keeping-Up-With-Rails

http://github.com/rubys/awdwr/raw/13d847d07cd0d8ff080d8163eb5e48cfb...

Applied to this ticket, the easiest way to reproduce this is to run sections 6.1-7.4 using a checked out version of Rails and save the results; and then continue from the restore point with sections 8.1-8.5 to see the error.

It doesn't appear to be a reloading issue, but clearly is a regression. My suggestion is that the change be backed out (or disabled) until it can be debugged.

I've reviewed the code in question, and here is what I found: the commit in question optimizes out writing out sessions unless a direct assignment is made to the session object.

In my case, the session contains a Cart object. The first request will cause the Cart to be added to the session. Subsequent requests won't directly assign to the session, but will update the cart object.