[Python-Dev] ref leaks

[R. David Murray]

On Thu, 24 Apr 2014 17:17:41 +0200, Stefan Krah <stefan at bytereef.org> wrote:
> Ethan Furman <ethan at stoneleaf.us> wrote:
> > >>Any words of wisdom for tracking those leaks?
> 
> Often the easiest way is to compile --with-valgrind and run the test
> under Valgrind (obviously).
> 
> In the Valgrind output, search for "definitely lost" and ignore the rest.
> 
> If that does not turn up anything, consider a bug in regrtest.py:
> 
> It happens that certain caches are initialized incrementally in each
> repetition of the test, so the reported leaks are spurious. That is
> why some caches like small integers are "warmed up" in regrtest.py.
> 
> Perhaps there is some recently added cache that needs attention.

Note that if you do find a "real" leak, I would start with the assumption
that any leak is a leak at the python level (eg: a python cache or an
possibly-accidentally-persistent data structure) and only move to the
C level once I'd ruled out a python source for the "leak".

Unless you are dealing with a wholly or primarily C module, of course :)

--David