DEPR: Index.ravel returning an ndarray

[jbrockmendel]

• closes PeriodIndex.ravel returns ndarray[int64] #19956
• tests added / passed
• passes black pandas
• passes git diff upstream/master -u -- "*.py" | flake8 --diff
• whatsnew entry

Make Index.ravel() behavior match every other ravel() method in existence.

[jreback]

nice!

[jorisvandenbossche]

@jbrockmendel I don't think there was consensus on changing ravel to return an Index. See the original discussion at #19956, and you actually did a similar PR before (but to return the EA, #32641) where I said the same.

• The Index method will now be inconsistent with the Series method
• What's the use case of actually needing an Index.ravel method?

[jbrockmendel]

The Index method will now be inconsistent with the Series method

I would be OK changing the deprecation so to say the future behavior is to match self._values.ravel, matching Series behavior.

What's the use case of actually needing an Index.ravel method?

The same as the use case for Series.ravel: you want to write code that is ndarray/index/series-agnostic

[jorisvandenbossche]

I would be OK changing the deprecation so to say the future behavior is to match self._values.ravel, matching Series behavior.

IMO that sounds as a better option. Having consistency between Series and Index here would be good I think (although Index is more array-like than Series ..)

I am still skeptical about the need for this method at all, though (you can also do a ndim check if you eg want to ensure the array-like is 1D or converted to 1D, see eg Tom's comment at #19956 (comment))

[simonjayhawkins]

I am still skeptical about the need for this method at all

as there is a top level numpy function for this https://numpy.org/doc/stable/reference/generated/numpy.ravel.html.

In the future, with __array_function__ implemented, could allow much simplification of the pandas API.

so I think OK to deprecate the method, in favour of np.ravel() support.

>>> pd.__version__
'1.2.0.dev0+765.g401d7a142c'
>>>
>>> idx = pd.Index([1, 2, 3])
>>>
>>> np.ravel(idx)
array([1, 2, 3], dtype=int64)
>>>
>>> type(_)
<class 'numpy.ndarray'>
>>>

so we could use np.ravel to get the current behaviour, of implement __array_function__ if we wanted to return a Index instead.

[jorisvandenbossche]

Also for np.ravel support through __array_function__, we would need to decide on the return value, though (so I think a large part of the discussion applies to that as well).

Because np.ravel(index) doesn't necessarily (technically) need to return an ndarray, it could also return an Index (as the warning in this PR proposes) or an EA.