Packages that `pip` can install but `uv pip` cannot install in a conda env

[edschofield]

uv_install_logs.tar.gz

Summary

uv seems to fail to recognize some macOS Universal wheels when running in a conda environment on Darwin x86_64.

The following four packages install fine using pip install but fail with compile errors using uv pip install on Darwin x86_64 with a brand-new conda environment:

• greenlet
• pyarrow
• duckdb
• tabmat

The first two packages are dependencies for important packages (e.g. for sqlalchemy and pandas with the parquet and feather extras).

Steps to reproduce:

conda create -n minimal python=3.12
conda activate minimal
uv pip install greenlet    # compile error
# but this works:
# pip install greenlet

Using pip finds and installs the wheel (e.g. greenlet-3.2.3-cp312-cp312-macosx_11_0_universal2.whl) when run from the conda environment. Using uv pip finds and installs the wheel when run from a venv, but seems to want to build the packages instead (and fails) when run from a conda environment.

This only affects these four packages out of ~225 packages specified in a large meta-package Python Charmers uses for training courses. Manually installing the above four packages with pip install ... allows the remaining packages to install successfully using uv pip.

Log outputs attached from uv pip install --verbose greenlet and uv pip install --verbose pyarrow.

Note that I get the same compile errors with uv from a conda environment using either the latest Miniconda3 or an old Anaconda installation from August 2020.

It may also be relevant that that some other packages do successfully build from source on the same platform: uv pip install bottleneck and pip install bottleneck both compile and install from source, whether run from a conda environment or a venv.

Platform

Darwin 24.5.0 x86_64: macOS Sequoia 15.5

Version

uv 0.7.15 (Homebrew 2025-06-25)

Python version

Python 3.12

uv_install_logs.tar.gz