ENH: Determination of result "flavor" when based on multiple input flavors

[rhshadrach]

In various places, we may want to determine the dtype flavor* for newly created columns when based on multiple input flavors. A list of areas is at the bottom.

The implementation (a function in core.dtypes) could take an array, Series, or DataFrame and determine the proper flavor. If the input is a single flavor, then this is trivial. We'd need to decide what the rule is in case of multiple flavors.

Option 1: pyarrow -> nullable -> numpy (or any of the 5 other permutations as suboptions)

If there are any pyarrow then go with that; if no pyarrow and there are nullable then go with that, and finally only have numpy if all columns are numpy. I think we'd want to consider str dtype (whether backed by pyarrow or numpy) as being numpy for the purposes of this. This might be problematic when the data has only str columns, but I don't see a better alternative.

*flavor may not be the appropriate term here, I recall some objections to its usage in the past.

Areas

• In pivot_table ref
• In apply/agg/transform/map when a UDF is returning all scalars. Here, the input columns are to be used, which can in some cases be more than 1.

[rhshadrach]

Milestoned as 4.0 just for visibility.

[mroeschke]

I think find_common_type more or less does this? However, I think it's sensitive to the order of the passed ExtensionDtypes as it will dispatch to _get_common_dtype so the resulting "flavor" may not be communicative currently

[rhshadrach]

I don't think that's suitable, it will find the common dtype that can hold values from all dtypes which is often object. In the example below, the flavor should be pyarrow.

import datetime as dt
import pyarrow as pa
import pandas as pd
from pandas.core.dtypes.cast import find_common_type

df = pd.DataFrame(
    {
        "a": pd.array([1], dtype="int64[pyarrow]"),
        "b": pd.array([dt.datetime(2026, 1, 1)], dtype=pd.ArrowDtype(pa.timestamp("ms"))),
    }
)
find_common_type(df.dtypes.tolist())
# dtype('O')