So this inspired me to start thinking outside our current box. The very first thing about DataInfo is that it is a descriptor class that is storing the per-instance attribute information on the parent object and always dynamically finding the parent. You made that faster, which is definitely good. Overall DataInfo is a bit complicated and has all these redirects and knobs.

Is there any reasonable way we can scrap that whole paradigm and simplify a whole lot? Just make info a property pointing to an ordinary object stored in _info. The info "attributes" are properties that allow quite fast access, with setters that do validation as necessary.

Maybe we've been dancing around this idea, but I wonder if something like below could fly? It might end up making the DataInfo concept much simpler, cleaner, and faster.

master...taldcroft:table-info-play

In [2]: c = Column([1,2], name='hello')

In [3]: timeit c.name
165 ns ± 0.627 ns per loop (mean ± std. dev. of 7 runs, 10000000 loops each)

In [4]: timeit c.finfo.name
305 ns ± 2.74 ns per loop (mean ± std. dev. of 7 runs, 1000000 loops each)

In [5]: timeit c.info.name
3.89 µs ± 56.5 ns per loop (mean ± std. dev. of 7 runs, 100000 loops each)

@taldcroft - interesting idea. Part of this is what I had in mind by auto-generation - one can just generate the name properties, etc., by a metaclass from attrs_from_parent, etc. I'm not sure making info a property is going to gain much over it being a descriptor (a property is a descriptor, after all), but I think making it simpler/more permanent is good, and I like especially the much closer link to the parent that you have (though obviously we need to look into why we had that weakref in the first place! I experimented with not setting _parent on every call but got quite a few errors).

p.s. The good part is that we've got lots of tests, so one can quickly find out whether a new idea will work!

closing in favour of #8998 (which includes this commit, though it partially undoes it)