ExtensionArray construction with given dtype (sort of shallow_copy?)

[jorisvandenbossche]

Many of the extension arrays tests are skipped for Categorical because the reconstruction of the expected result does not preserve the categoricals (so kind of the "metadata" of the dtype).

For example:

pandas/pandas/tests/extension/category/test_categorical.py

Lines 70 to 72 in 78fee04

	@pytest.mark.skip(reason="Unobserved categories preseved in concat.")
	def test_align(self, data, na_value):
	pass

because in the actual test, the expected result is constructed from a list:

pandas/pandas/tests/extension/base/reshaping.py

Lines 47 to 50 in 78fee04

	r1, r2 = pd.Series(a).align(pd.Series(b, index=[1, 2, 3]))
	# Assumes that the ctor can take a list of scalars of the type
	e1 = pd.Series(type(data)(list(a) + [na_value]))

(still with type(data)(..), but replacing that with data._constructor_from_sequence(..) in #20746).

This is kind of a recurrent pattern, so that might indicate we should find a solution for this?

So do we want a canonical way in the extension array interface to construct an ExtensionArray that has a certain dtype?

Possible solution is to add a dtype keyword to _constructor_from_sequence

[TomAugspurger]

This is kind of a recurrent pattern, so that might indicate we should find a solution for this?

Agreed. This is somewhat similar to _from_factorized right, where we pass original rather than the dtype?

No thoughts on a solution yet, but a good test may be to parametrize DecimalDtype with a decimal.Context context. Then we could ensure that the correct Context is passed through on these kinds of operations, via the dtype.

[jorisvandenbossche]

The tests that I mentioned in the original post are still skipped, so we need to check if this can now be resolved with the added dtype argument.