API: define __copy__ NDFrame

[jreback]

from SO

this allows copy.copy(df) to work properly

by definition 'always' deep

__copy__ = copy
__deepcopy__ = copy

[immerrr]

The fact that copy doesn't work as expected without providing a __copy__ method means that pickling is broken in the same way as copy uses it as a fallback. Maybe pickle should be fixed too?

Also, I don't really get why

by definition 'always' deep

[jreback]

pickling isn't broken
a class level attribute is not copied (nor should it be)

and always deep because that's the default (copy=True)

[immerrr]

pickling isn't broken
a class level attribute is not copied (nor should it be)

It's not about copying the class-level attribute, it's about preserving object-level attribute as mandated by a modified class-level attribute. If there's some public-API-level contract about that somewhere (I admit I haven't seen any yet) that attributes mentioned in _metadata must be preserved for instances of Series and pickle doesn't obey that contract, then it sure seems broken to me.

As for __copy__ performing a deep copy, this seems a direct violation of copy.copy purpose as stated by its doc which says "Return a shallow copy of x". But, I guess, there's a precedent set by numpy in this direction (e.g. here), so that's probably ok.

[jreback]

@immerrr modifying _metadata and having it preserve is not documented atm nor really exposed to the user (its basically internal). That said I don't see any harm with doing so. This is a really simple issue of __copy__ just not doing what is needed (its close, but not perfect, just need to simply define it -end of issue).

it doesn't matter what copy.copy should perform a shallow copy. pandas objects don't have this as a guarantee anywhere (and it should not be done, copies are ALWAYS deep, its the indexes tht can be shallow). I am not sure I see a problem here either. Its actually what we do it ultimately take a view of the data when its not deep so that could be a copy (but generally is not).

[immerrr]

Ok, no contract — no worries.

copies are ALWAYS deep

Isn't there a x.copy(deep=False) that shallow copies x?