The main point is that at first, your code will be slow (as in bound to a single core) and then you can optimize it to provide better performance. This is as opposed to threads, where your code will be immediately fast, but buggy. The difference comes from the fact that 90% correct is still buggy, but 90% paralellizable is kinda good.
Is there a transaction module that provides a null/no-op implementation for compatibility with CPython?
This is very neat! Obviously it's not a magic drop-in, but still.
Relevant bit from the docs: http://pypy.readthedocs.org/en/latest/stm.html#transaction-t...