Repository navigation
[Python] Numpy constraint is present in the build but not in the runtime constraints #53954
Description
Describe the bug, including details regarding any error messages, version, and platform.
Here https://github.com/apache/arrow/blob/apache-arrow-26.0.0/python/pyproject.toml#L25 we have:
[build-system]
requires = [
"numpy>=2.0",
]For the project dependencies, there do not appear to be any mentioned that I can see. However, when I run the code, I see:
File ...
import pyarrow as pa
File ...
from pyarrow.lib import ...
File "pyarrow/lib.pyx", line 42, in init pyarrow.lib
ImportError: pyarrow requires NumPy 2.0 or newer, found 1.26.4
when I use pyarrow==26.0.0.
It seems to me that there is a missing dependency that should be present to constraint the numpy version in the runtime dependencies (alongside the build time ones).
Component(s)
Python
Activity
[-]Numpy constraint is present in the build but not in the runtime constraints[/-][+][Python] Numpy constraint is present in the build but not in the runtime constraints[/+]mohamd764 commentedon Oct 16, 2026
I can confirm the behaviour and the version boundary (Python 3.12.15, Linux x86_64,
uv pip install):import pyarrowworks,pa.array(np.arange(3)).to_numpy()worksImportError: pyarrow requires NumPy 2.0 or newer, found 1.26.4The 26.0.0 wheel's
METADATAhas noRequires-Distlines at all, so pip/uv have nothing to resolve against.Some context in case it helps: this was an intentional change for 26.0.0 in #50769 (GH-48473, "Require numpy 2.0"). The check is in
pyarrow/lib.pyx(lines 40-45 atapache-arrow-26.0.0) and only runs ifimport numpysucceeds. NumPy is still an optional runtime dependency. With numpy uninstalled, pyarrow 26.0.0 imports fine, and a Parquet write/read round trip andpc.sumwork. That's probably why there's no runtime requirement. Addingnumpy>=2to[project].dependencies(what #53971 proposed; it was closed without being merged) would make numpy mandatory for everyone, and package metadata has no way to say "numpy>=2, but only if numpy is installed".The install docs for 26.0.0 already list "NumPy 2.0 or higher" under optional dependencies, which matches this behaviour.
Practical options for now:
pip install "numpy>=2"alongside pyarrow 26, or pinpyarrow<26if you need to stay on NumPy 1.x. On the user side, a constraints file withnumpy>=2also does the job. I tried this in a fresh Python 3.12 venv with-c constraints.txt:pyarrow==26.0.0alone: numpy is not installedpyarrow==26.0.0 "numpy<2": resolution fails (pip and uv both)pyarrow==26.0.0 pandas==2.2.3: numpy 2.5.4 is picked (uv)A constraint doesn't upgrade a numpy 1.x that's already installed if you only install pyarrow, though (checked with uv). In that case you still have to upgrade numpy yourself.
jorisvandenbossche commentedon Oct 23, 2026
@oliversheridanmethven this gives confusing situations, but so the problem is that
numpyis not a required dependency of pyarrow (various parts of pyarrow can be used without having numpy installed), but as optional runtime dependency we have a minimum version of 2.0.Unfortunately, the python package metadata don't really allow to express such relation (AFAIK) about runtime constraints.
On our zulip chat I questioned if we should actually then raise an error or rather ignore numpy (treat an older numpy as "numpy not installed"), but that won't make the situation less confusing .. (#py-chat > Running pyarrow with older NumPy: raise or ignore? @ 💬)