I'm experimenting with a better way of cleaning up at the end of an execution run. Without implementing true GC, I can never do it 100% right, but I can implement a predictable set of rules based on practical observation that will solve most problems that are actually observed.
Here's my proposal. At the end of this message I list some potential problems with the proposal and ask for feedback. This will probably be implemented in Python 1.5.1.
Based on some comments I received and some more thinking, I've changed this a bit since my web posting on this subject. Significant Changes are indicated in the text by [italicized remarks in square brackets].
When a Python interpreter is deleted, its variables and modules are "cleared carefully" in a partially specified order. The operation "clear carefully" is defined below; it effectively deletes the module's variables in a partially specified order.
To clear a module carefully, the following steps are taken:
[New.] Steps C1-C2 will also be used when a module is deallocated. While modules are generally not involved in cycles (except when there are mutually recursive imports), a module's dictionary generally is involved in a cycle because every function and method defined in the module references its __dict__, and these functions and methods are generally reachable from that __dict__. Thus, when a module is deleted, I explicitly clear its __dict__ carefully. (This has always been done, just not "carefully".)
M1 is done because these variables are common places for user values to hide, and they would come too late in the proposed order. (In fact nearly all reported problems with destructors not being called when expected have to do with these.)
M3 is there because __main__ is conceptually the "root" of the program -- if it is not imported by other modules, it would be deleted first by step M4 anyway, but if it is imported elsewhere, deleting __main__ is a plausible way to break a tie.
M4 is an explicit garbage collection loop -- it deletes all those modules which are referenced by no other modules, only by the table of modules (sys.modules) itself. It may not delete all modules, however, when there are mutual imports; the remaining steps take care of those.
M5 is needed to take care of mutually recursive imports, which create cycles so that M4 won't delete everything.
The special treatment of __builtin__ and sys is because these are referenced by the interpreter implicitly by many operations; __builtin__ of course contains all the built-in functions and exceptions; sys contains the standard I/O files which are referenced implicitly by various I/O operations. So they are excluded in M2 and M4. __builtin__ is deleted last because it contains the most basic and fundamental values.
The special care for cleaning up a module's dictionary is needed because there's a fundamental circular reference whenever a module defines a Python function or class. A function object contains a reference to the function's 'globals' object, which is the __dict__ of the module that defines it. Since the __dict__ normally has a reference to the function there's a cycle that needs to be broken, or else the __dict__ would never be garbage collected.
Note that a reference-count based solution doesn't work within one module, since references between functions are by name, not by value -- two mutually recursive functions can still both have a reference count of one, since they do a name lookup for each other.
C1 is an attempt to provide a way for a module to define globals that are deleted before anything else in the module. Since imported module or function names generally don't begin with an underscore, this means that such objects can be guaranteed that any imported modules or functions still exist when they are deleted -- provided, of course, that the only reference to them is in the module. (This step is already implemented in 1.5 as released.)
C2 deletes the remaining objects but leaves the "internal global variable" __builtins__ alone -- this prevents the problem that the 1.5 release has where e.g. using "None" in a destructor raises a NameError!
C3 removes the reference to the module from the modules table in a way that makes a later import of the same module fail. (It is possible for user code to delete this entry and still start a completely new import -- but if they are that clever they deserve what they get.)
P1. When all uses of a module M have the form ``from M import ...'', the module M will have a reference count of 1. So it will be deleted in step M1. This renders all but the most trivial functions defined in the module (which are presumably still referenced by other modules) useless, since the imported modules and functions that they might need are all deleted from their globals. A simple remedy of course is not to use ``from M import ...'', but this sounds like it might become a FAQ... The problem is that I don't know of a better way -- because of the circular references between functions and their module's __dict__, I can't use the reference count of the __dict__ in step M1. I think it is acceptable -- this behaviour also existed in 1.4 and earlier versions. (It has been suggested to add a reference named e.g. ".module" in the importing module, to indicate the dependency and prevent this problem; while this may do the job nicely, I'm reluctant to implement it because it may confuse introspective tools.)