[QST] What should `ExtensionDtype.type` return?

[shwina]

Greetings, Pandas devs! cuDF is building out additional dtypes such as cudf.CategoricalDtype and cudf.ListDtype based on pd.ExtensionDtype, and this is one question that came up.

The documentation states:

It’s expected ExtensionArray[item] returns an instance of ExtensionDtype.type for scalar item, assuming that value is valid (not NA). NA values do not need to be instances of type.

However, I note that pd.CategoricalDtype for instance does not adhere to this:

In [47]: import pandas as pd

In [48]: a = pd.Series(['a', 'b'], dtype='category')

In [49]: type(a[0])
Out[49]: str

In [50]: type(a.array[0])
Out[50]: str

In [51]: isinstance(a.array, pd.api.extensions.ExtensionArray)
Out[51]: True

In [52]: isinstance(a.dtype, pd.api.extensions.ExtensionDtype)
Out[52]: True

On the other hand, NumPy defines dtype.type somewhat differently:

The type object used to instantiate a scalar of this data-type.

Would love any insights as to the appropriate return value of .type.

[TomAugspurger]

The CategoricalDtype.type issue is discussed a bit at #22938 (comment). Categorical is a bit hard since it can hold anything, (including no categories, for backwards compatibility).

IIRC you're using pyarrow for ListDtype, so I would expect something like like pyarrow's ListValue or some cudf wrapper around it.

In [10]: a = pa.array([[1, 2]])

In [11]: type(a[0])
Out[11]: pyarrow.lib.ListValue

[shwina]

Thanks, this answers my question. It appears that CategoricalDtype is the exception and not the rule