extradoc / talk / ep2008 / abstract_gc.txt

Full commit

Writing Python code for different GCs

(was: PyPy - garbage collection details.)


This talk will pick PyPy's GC as an example and discuss
how to write Python code that does not rely on the underlying GC


Almost all of the newer Python implementation depart from CPython's reference
counting as a memory management strategy and use more recent garbage
collection strategies. IronPython and Jython use the GCs of the .NET platform
and of the JVM respectively, PyPy contains several GCs, among them a moving
generational garbage collector. All these GCs have in common that they behave
slightly differently than reference counting in some places. Existing Python
code often uses patterns that only works well with reference counting.

In this talk we will explain one of the PyPy GCs in detail as an example of
the sort of GC techniques that are used. Then we will describe common mistakes
that are made and explain ways how to work around them. Specifically we will
talk about the different semantics of finalizers and the varying cost of some
operations on a moving GC (like id).

Proposed Length of talk required

30 minutes.

Brief Biography

Maciej Fijalkowski is a core PyPy developer, particularly interested
in the obscure details of various parts. He is always annoyed with existing code
relying on reference counting, as well as lack of knowledge about how
things like finalizers work.

Armin Rigo is the author of the Psyco JIT, as well as the key person behind 
many of PyPy's advanced details. He is one of the main implementors of PyPy's