BUG: copy np.ndarray inputs to Index constructor by default

[mroeschke]

• closes BUG: Series stealing references from CategoricalIndex is invalid for read-only arrays #63306 (Replace xxxx with the GitHub issue number)

@jorisvandenbossche I went with your suggestion to

We changed Series to copy array input by default when we introduced CoW. We should maybe do the same for Index?

in #63306 (comment)

[jorisvandenbossche]

We essentially already do a shallow copy of Index inputs in the Series contructor (by keeping track of the references the Index has). But the problem here is that the Index was created without a copy.

To clarify my suggestion:

We changed Series to copy array input by default when we introduced CoW. We should maybe do the same for Index?

I meant "doing the same for the Index constructor" (and not doing the same for Index input to the Series constructor ..). I see that the above sentence could have been interpreted both ways, sorry ;)

So that pd.Index(arr) copies arr by default.

[mroeschke]

So that pd.Index(arr) copies arr by default.

Ah makes sense. No problem, I mimicked the copying we do in the Series constructor in the Index constructor. If you're OK with the approach, I can follow up in another PR with:

1. Applying the same changes to the index subclasses
2. Including ExtensionArrays

[jorisvandenbossche]

Looks good!

[jorisvandenbossche]

For datetime64 ndarray, we wouldn't get in the elif block below that now copies, but this is what you mentioned to handle in the Index subclasses in a follow-up?

[mroeschke]

My comment was referring passing an ndarray to DatetimeIndex et. al., but good find, yes I think I'll have to put this check higher up.

[jorisvandenbossche]

You can also use the updated description in the Series docstring as a starting point to update the docstring here as well (as the end goal should be to make it similar)