Perfect backwards compatibility

[MarcoGorelli]

Looks like we're close to tagging a first version

Do we want to aim for some guarantee of perfect backwards compatibility?

So, if you write your code using the 2023-08 version, then it should all still work under the 2025-01 version. The latter may have some extra features, or some extra optional arguments, but anything you write with the 2023-08 version should still work with the 2025-01 one

[rgommers]

Do we want to aim for some guarantee of perfect backwards compatibility?

Hmm, I'm not so sure about that. We've moved quite quickly, so while things look pretty good I can't say I'm sure we got it all 100% right. So I'd rather say that this is for early adopters.

What we discussed before is that this is the period between the public RFC/announcement and the first stable release towards the end of 2023.

[MarcoGorelli]

Sure, we'll probably make mistakes and require breaking changes in the short-term

But longer-term, do we want to get to the point where it should all be perfectly backwards compatible (like Rust is between its editions)? I think it'd be pretty nice

[rgommers]

Yes, definitely! I think we want to give very strong guarantees here, any bc-breaking changes between stable versions should be exceptional. We have a small and carefully designed API surface, and it should be significantly more stable than that of any dataframe library out there. More a la a C/C++ language standard.

[MarcoGorelli]

addressed by #257