pola-rs/polars · error · TypeError

cannot treat NumPy array of type {arr.dtype} as indices

Error message

cannot treat NumPy array of type {arr.dtype} as indices

What it means

In _convert_np_ndarray_to_indices (getitem.py:425), numpy index arrays must have integer dtype (kind 'i' or 'u'); boolean arrays route to the boolean-mask error, and every other dtype — float, str, datetime64, object — raises this TypeError naming arr.dtype.

Source

Thrown at py-polars/src/polars/_utils/getitem.py:425

    #     to absolute indexes.
    if arr.ndim == 0:
        arr = np.atleast_1d(arr)
    if arr.ndim != 1:
        msg = "only 1D NumPy arrays can be treated as indices"
        raise TypeError(msg)

    idx_type = get_index_type()

    if len(arr) == 0:
        return pl.Series("", [], dtype=idx_type)

    # Numpy array with signed or unsigned integers.
    if arr.dtype.kind not in ("i", "u"):
        if arr.dtype.kind == "b":
            _raise_on_boolean_mask()
        else:
            msg = f"cannot treat NumPy array of type {arr.dtype} as indices"
            raise TypeError(msg)

    if idx_type == UInt32:
        if arr.dtype in {np.int64, np.uint64} and arr.max() >= U32_MAX:
            msg = "index positions should be smaller than 2^32"
            raise ValueError(msg)
        if arr.dtype == np.int64 and arr.min() < -U32_MAX:
            msg = "index positions should be greater than or equal to -2^32"
            raise ValueError(msg)

    if arr.dtype.kind == "i" and arr.min() < 0:
        if idx_type == UInt32:
            if arr.dtype in (np.int8, np.int16):
                arr = arr.astype(np.int32)
        else:
            if arr.dtype in (np.int8, np.int16, np.int32):
                arr = arr.astype(np.int64)

        # Update negative indexes to absolute indexes.

View on GitHub (pinned to df599052da)

Solutions

1. Cast to integer dtype: df[arr.astype(np.int64)] (floor/round first if fractional)
2. For value-based matching use df.filter(pl.col('c').is_in(arr.tolist()))
3. Ensure index-producing numpy ops stay integral (e.g. use np.flatnonzero instead of manual float math)

Example fix

# before
df[np.array([0.5, 1.5])]

# after
df[np.array([0.5, 1.5]).round().astype(np.int64)]

import numpy as np

if arr.dtype.kind not in ("i", "u"):
    if arr.dtype.kind == "b":
        arr = np.flatnonzero(arr)  # or use df.filter(mask)
    else:
        arr = arr.astype(np.int64)
df[arr]

import numpy as np

def is_integer_index_array(arr) -> bool:
    return isinstance(arr, np.ndarray) and arr.ndim == 1 and arr.dtype.kind in ("i", "u")

When it happens

Trigger: df[np.array([0.5, 1.5])] (floats from division or np.mean); df[np.array(['a', 'b'])]; df[np.array(['2024-01-01'], dtype='datetime64[ns]')]; df[np.array([1, 2], dtype=object)] from pandas .to_numpy() on mixed columns.

Common situations: numpy operations that upcast indices to float (e.g. np.isnan-affected arrays, division); converting pandas object columns to positions; string keys mistakenly used as positions.

Related errors

• only 1D NumPy arrays can be treated as indices
• cannot treat Series of type {s.dtype} as indices
• cannot select columns using key of type {qualified_type_name
• cannot select rows using key of type {qualified_type_name(ke
• index positions should be smaller than 2^32

AI-assisted analysis of pola-rs/polars@df599052da (2026-08-16). Data as JSON: /api/errors/79592b29a443e920. Report an issue: GitHub.