polars and numba

Arrays from a polars frame

Series.to_numpy() and DataFrame.to_numpy() leave polars typed as ndarray[Any, Any]. What is certain is pinned, nothing more: DataFrame.to_numpy() is 2-D, and Series.to_numpy() is 1-D after a visible cast to a numeric or boolean dtype — with the dtype named for a float cast (s.cast(pl.Float64).to_numpy() is (int,) float64; an integer cast stays dtype[Any], since a null comes out float64). A plain s.to_numpy() keeps polars' own answer on purpose: a column of pl.Array or pl.Struct is 2-D, and the type of a Series does not say what it holds, so rank 1 would be a lie with no clean way out. Pinned by tests/fixtures/polars_to_numpy.py (needs polars; the test skips without it).

The cast is still seen behind the Series methods that keep the dtype and one value per entry: fill_null, fill_nan, drop_nulls, drop_nans, sort, head, tail, slice, filter, reverse, clip, limit, shift, gather, rechunk, clone, alias, rename and unique. So s.cast(pl.Float64).fill_null(np.nan).to_numpy(), the usual way to read a nullable column as floats, is (int,) float64. A method that is not on the list (cum_sum, to_frame) ends it. On the array, astype(float) names the dtype even when the rank is unknown (s.to_numpy().astype(float) is ndarray[Any, dtype[float64]]), and astype(d) with d: np.dtype[T] keeps an int-sized array (tests/fixtures/polars_cast_chain.py).

Annotate the local where the rank matters (x: F64[int] = df["x"].to_numpy()). structured=True and masked=True also keep polars' answer.

numba

A call of an @njit function is checked as a call of the function itself: named sizes are solved, a formula return is put in the caller's terms, and the call-site checks apply. Without numty every such call is expected "Never" to mypy, because the function's size variables are not solved through Dispatcher.__call__ (tests/fixtures/numba_dispatch.py).