Using NamedTuples as Object Storage

[julian-r]

What are your thoughts on replacing all T_obj with Named Tuples with a legacy getter layer?
I think that would give a great benefit in

• type safety.
• The immutability of tuples here can also be an advantage.
• Will probably affect memory efficiency
• Might improve performance a little

I would start implementing these things if you consider merging such changes.

[jsvine]

Thank you for this intriguing proposal, @julian-r! I've thought about something similar over the years, but have worried about backwards compatibility. Your proposed legacy getter layer certainly helps with that. Sub-concerns:

1. How large would the performance reduction be for people using the legacy getters?

2. Could the legacy getter layer handle use-cases such as pandas.DataFrame(page.chars), and other similar calls where the code expects a dict but isn't explicitly calling, e.g., obj["x0"]? I get the sense that this is possible on a library-by-library basis, but is there a more general solution? I haven't looked deeply into that.

I'm certainly open to exploring your proposal. Is there a small-scale experiment that we could do as a proof-of-concept, before trying to implement the changes across the entire library?

[jsvine]

Minor update re. 2, based on a quick experiment: It seems that pandas.DataFrame handles lists of namedtuples well, which is reassuring.