I’m now a uv convert. I haven’t really updated my Python tooling and workflow for the past two years. In fact, I’ve been using the same workflow ever since I started working:

python -m venv venv
source /venv/bin/activate
# While in venv
pip install -r requirements.txt

But there’s a lot of limitations in this workflow. For example, I’m limited by the Python version installed in my machine or it’s hard to just install ad-hoc dependencies for my one-off scripts. This workflow also doesn’t manage dependency conflicts, I still remember those days when I’m checking whether I should use X version of pytorch or Y version of spaCy to get things running.

The uv tool solves all of that.

1. First, uv sync does all the dependency management under the hood, and it automatically installs everything in a virtual environment. I was surprised how fast the installation is, compared to your good old pip install. Really awesome! Adding new requirements is a breeze too via uv add: I just need to think about the name of the dependency, and uv will figure out the correct version to use.

2. What’s magical for me is that the virtual environment can have its own python version. So if a project needs 3.10 and I have 3.12, I can just set 3.10 in a .python_version file (and also in my pyproject.toml), and uv will automatically install the correct python version and dependencies.

3. The best feature for me is on adding one-off dependencies for my ad-hoc scripts. This is really cool because my standalone scripts can be wholly separated from the rest of my project. When I organize my research code, I often have a separate folder for managing all my plotting code, and sometimes these only use pandas and matplotlib. I can just specify those dependencies and run the script in isolated fashion.

Yeah, uv is cool. For the past two years I never shifted to tools like poetry or pyenv, I find them quite bulky and cumbersome to use. However, uv’s interface feels lightning fast. The folks at astral are real geniuses!