[Python] Weird exception when slicing an array with non-trivial step

[pitrou]

Describe the bug, including details regarding any error messages, version, and platform.

I don't know if we can easily support this, but it should at least fail with a more readable error:

>>> a = pa.array([1.2, 3.5, None])
>>> a
<pyarrow.lib.DoubleArray object at 0x7d05b7f25900>
[
  1.2,
  3.5,
  null
]
>>> a[::-1]
Traceback (most recent call last):
  Cell In[4], line 1
    a[::-1]
  File pyarrow/array.pxi:1523 in pyarrow.lib.Array.__getitem__
  File pyarrow/array.pxi:587 in pyarrow.lib._normalize_slice
AttributeError: 'NoneType' object has no attribute 'arange'

Component(s)

Python

[pitrou]

cc @AlenkaF @thisisnic

[AlenkaF]

Hm, I get a different behaviour:

Python 3.12.10 (main, Apr  8 2025, 11:35:47) [Clang 16.0.0 (clang-1600.0.26.6)] on darwin
Type "help", "copyright", "credits" or "license" for more information.
>>> import pyarrow as pa
>>> a = pa.array([1.2, 3.5, None])
>>> a
<pyarrow.lib.DoubleArray object at 0x10c9747c0>
[
  1.2,
  3.5,
  null
]
>>> a[-1:]
<pyarrow.lib.DoubleArray object at 0x10c974820>
[
  null
]
>>> a[::-1]
<pyarrow.lib.DoubleArray object at 0x10c9748e0>
[
  null,
  3.5,
  1.2
]
>>> pa.__version__
'20.0.0.dev356+g7c820dcbda.d20250520'

[pitrou]

It seems the error happens when NumPy is not installed (arange is a NumPy function).

(@raulcd FYI)

[AlenkaF]

Ah yes. So the fix would be updating _normalize_slice:

arrow/python/pyarrow/array.pxi

Lines 575 to 591 in a2941dd

	def _normalize_slice(object arrow_obj, slice key):
	"""
	Slices with step not equal to 1 (or None) will produce a copy
	rather than a zero-copy view
	"""
	cdef:
	Py_ssize_t start, stop, step
	Py_ssize_t n = len(arrow_obj)
	start, stop, step = key.indices(n)
	if step != 1:
	indices = np.arange(start, stop, step)
	return arrow_obj.take(indices)
	else:
	length = max(stop - start, 0)
	return arrow_obj.slice(start, length)

so that it doesn't use numpy. We could use list(range(start, stop, step)) instead of np.arange(start, stop, step).
@thisisnic this could be a nice next issue when/if you have the capacity.

[raulcd]

Issue resolved by pull request 46732
#46732