[gclist] Finalization and the insane postman bug.

Charles Fiterman cef@geodesic.com
Thu, 04 Oct 2001 14:57:45 -0500

My conclusion about finalizers is that everything about running them is
determined by the user. Priority, order, time and context. I can easily
invent contexts where any set of requirements hold. 

What if a finalizer throws an exception? You can't tell users their
finalizers shouldn't throw exceptions because an exception is almost by
definition somewhat of a surprise. But having the finalizer run in a thread
owned by the collector requires throwing the exception away. I call this
the insane postman bug. "This letter is marked emergency, special delivery,
do not delay. I can't deal with that right now so I'll burn the letter!" I
can think of few bugs worse than the insane postman bug. Whatever is right
no implementation with that bug can be right.

Adding all this to a language would be impossibly complex. Languages need
to be simple. So what is required is a death notice. The user is informed
by some mechanism that an object died. The death notice is an identifer
associated with the objct not the object's address. Now priority, order,
time and context are in the user's control. If some objects are reference
counted they give death notices too.

At no point do objects rise from the dead to alter the bits of the living.
That's a bug the collector promices objects will be deleted when they can
no longer influence the order of execution. Finalizers alter the order of