[Python-ideas] gc callbacks

[Terry Reedy]

On 11/29/2010 12:12 AM, Kristján Valur Jónsson wrote:
> Hi there.
>
> Some months back I mentioned that I had other stuff in store for GC.
>
> Here is an idea for you.  This particular idea is just a generalization
> of a system we’ve used in EVE for years now: garbage collection callbacks.
>
> The need, originally, was to be able to quantify the time spent in
> garbage collection runs.  Since they occur out of direct control of the
> application, we needed to have python tell us about it somehow.
>
> We added gc.register_callback(), a function that added a callable to an
> internal list of functions that get called on two occations:
>
> 1)When a garbage collection run Is about to start
>
> 2)When a garbage collection run has finished.
>
> The cases are distinguished using an integer argument.  The callbacks
> are invoked from gc with gc inhibited from reentry, so that the
> callbacks cannot themselves cause another gc run to commence.
>
> What we traditionally use this for is to start and stop a performance
> timer and other stats.
>
> More recently though, we have found another very important use for
> this.  When gc finds uncollectable objects, they are put in the
> gc.garbage list.  This then needs to be handled by the application.
> However, there is no particularly good way for the application to do
> this, except to periodically check this list.
>
> With the above callback, modules that create uncollectable objects, such
> as classes with __del__ methods, can register their callback.  At the
> end of a gc run, they can then walk gc.garbage and take appropriate
> action for those objects it recognizes.
>
> So, what do you think?  This is  a very simple addition to gc,
> orthogonal to everything and easily implemented.  I also think it is
> very useful.

I presume the slowdown in the normal case consists of a very quick 'if 
callback list is not empty' check. If others agree and you open an issue 
with a ready-to-go patch (code, doc, test, what's-new), getting in 3.2b1 
might be possible.

-- 
Terry Jan Reedy