What Data Professionals Need to Know about uv

uv is all the rage. It helps you reduce the pain of Python virtual environment management, slow package installs, and unreproducible environments.

Although uv solves a slew of headaches for Python users it's still in its early days. Although there is a lot of buzz around uv, when I polled a class of data professionals, only a few of them had tried it.

As a data professional you can use uv to save time and money. Let's see how.

The current situation

Here's the landscape of existing popular tools.

First, you've got your default Python version managers: venv, virtualenv.

Then, there's your snake-named virtual environment managers: conda, miniconda, mamba.

Next, are your Python version managers: pyenv, tox.

Plus, there are your Python package managers: pip, pipx, poetry.

Don't forget your package dependency version files: requirements.txt and pyproject.toml. And you've got tools such as pipx to help you use them.

And that's a simplified list of the options! It ain't pretty.

The worst part is that when the wires get crossed, it's a time suck. I was reminded of this not long ago. I've been a conda user for years (mamba flavor, lately). Recently, my base, default conda environment had some deep package dependency issues that caused tough-to-diagnose problems. I wanted to remove my base environment and start fresh, but you can't easily do that, at least not without risking breaking your entire setup. But reinstalling conda yet again and risking a case of "PATH issue-itis" isn't anyone's idea of a good time.

That's what caused me to make the jump to uv.

uv brings simplicity and speed to the chaos that is Python virtual environment and package management.

Goal: install uv so you can take advantage of it

First, ensure you aren't already in a Python virtual environment. If you see parentheses around a word such as (base) in your terminal prompt, you probably are in a virtual environment. If it's a conda environment, exit it with conda deactivate. If it's a venv environment, exit it with deactivate. Run those commands until you don't see any parentheses in your prompt.

Check that you don't already have uv installed with:

1uv version

I'm using uv 0.5.25 (Homebrew 2025-01-28) for this example.

Install uv on a Mac or Linux machine with homebrew:

1brew install uv

Alternatively, you can install uv with:

1curl -LsSf https://astral.sh/uv/install.sh | sh

On a Windows machine you can install uv with:

1powershell -ExecutionPolicy ByPass -c "irm https://astral.sh/uv/install.ps1 | iex"

For more assistance, see the uv installation docs.

Now that we have uv installed, let's use it!

Goal: create a Python virtual environment that will stick around for future use

The ux of uv is lovely. It borrows from the syntax of popular tools.

If you've used venv before, you'll feel right at home, just preface your usual virtual environment creation command with uv and specify the desired Python version - no need to use a tool such as pyenv to install and manage Python versions.

Here's the command to create a virtual environment with Python 3.12:

1uv venv --python 3.12

uv will find a Python 3.12 version locally if you have one installed. If you don't have a Python 3.12 version installed, uv will fetch and install the latest stable 3.12 version.

Then, follow the handy command line output instructions to activate your virtual environment:

1source .venv/bin/activate

Viola! Your terminal prompt should display the name of a virtual environment in parentheses.

Let's install some Python packages.

Goal: install packages really fast

uv is orders of magnitude faster than pip when it comes to installing packages into your virtual environment. And the syntax is familiar if you're used to using pip. For example, to install the prefect package, just throw uv in front of your usual pip command like this:

1uv pip install prefect

You upgrade a package and its dependencies, you can use the same -U flag that you'd use with pip:

1uv pip install -U prefect
2

See the versions of all your installed packages with:

1uv pip list

Now you've got a virtual environment and with the packages you need. Next, say you build a useful project and everything works. Fantastic! Let's make this reproducible.

Goal: make a reproducible environment

As with pip, you can use the -r flag with uv to install all the packages listed in a requirements.txt file with a single command.

Say this is your file:

1polars
2plotly
3prefect

Then just run the following command:

1uv pip install -r requirements.txt

In the past, you might have specified top level package versions in your requirements.txt file.

1prefect==3.11.0
2polars==0.7.3
3plotly==5.24.1

That's a step further toward reproducibility, but sometimes a dependency of one of the packages you've specified will change and break your project. We don't want that, so let's create a lock file.

uv can read a list of packages from one file and output a lock file that specifies the exact versions of all top-level packages and their dependencies - and their dependencies' dependencies, all the way down.

Let's specify the top level packages we need in a requirements.in file. Then we'll generate a requirements.txt file with the versions of all specified packages and their dependencies with this command:

1uv pip compile requirements.in -o requirements.txt

Now our requirements.txt file contains the exact versions of all the top level packages and their dependencies, in tree format. When we run our requirements.txt install command, we'll get the exact versions of all the packages we want.

So what, you say? pipenv and poetry could do that, too? Fair enough, but let's talk about the need for speed.

Goal: faster package installs in Docker

Faster package installs shorten feedback loops, increase developer productivity, and speed up CI/CD pipelines, saving time, energy, and money.

uv is written in the famously fast Rust language and takes advantage of clever optimizations to reduce package install times.

I saw over a 5x speedup when using uv as opposed to pip to install 17 Python packages in a Docker image build: 7 seconds vs. 37 seconds. 🤯

Using uv to build a Docker image minor modifications. Just add:

1COPY --from=ghcr.io/astral-sh/uv:latest /uv /uvx /bin/

and change:

1RUN pip install my_package

to:

1RUN uv pip install --system my_package
2

Or instead of specifying --system every time, set the environment variable in your Dockerfile with:

1ENV UV_SYSTEM_PYTHON=1

Now let's see some almost magical uv tricks.

Goal: run a Python script in a single line in a disposable virtual environment

Use

1uv run my_script.py

Need a package such as polars?

1uv run --with polars my_script.py

You can even fire up a Jupyter lab server in a disposable virtual environment with:

1uv run --with jupyter jupyter lab

I wish this was a thing when I was learning Python.

Goal: run a file hosted on the web or in the clipboard

If your file paths are set up correctly, uv will auto-install the dependencies in a script and run them in a temporary virtual environment.

1uv run <https://example.com/my_script.py>

You can do the same thing with code in your clipboard on a Mac with this command:

1pbpaste | uv run -

Goal: easier project management

uv also solves pain points for project management. It helps you quickly start a project and provides handy commands for development. Create a project:

1uv init project1

This scaffolds a project with a pyproject.toml file, a hello.py file, and a README.md file.
It also adds.git and .gitignore files for version control. Convenient!

You can declare your Python version as well as your dependencies and project metadata in pyproject.toml. This further helps with reproducibility.

Run the script in your project with:

1uv run hello.py

Then, uv will automatically create a virtual environment with a .venv file and create a lock file, uv.lockwith the pinned versions of your dependencies and their dependencies.

Add and remove packages for your project like this:

1uv add polars
2uv remove polars

After adding some files, you may find that your pyproject.toml file, your uv.lock file, and your virtual environment aren’t in sync.

Goal: maximum strength reproducibility for your project

For maximum reproducibility, ensure that your lockfile, uv.lock is updated with your pyproject.toml.

1uv lock

Then you can ensure your environment is synced with all packages declared in your lockfile AND has no other dependencies installed:

1uv sync

Now you’ve got that environment synced and locked!

If you want to use the project you created for publishing packages, uv has your back.

Goal: publish your package

To build source distribution and wheel artifacts, make sure you have the necessary metadata in your pyproject.toml file and then run:

1uv build

Then, you'll see project1.egg-info and dist directories in your project.

Once you've set up your PyPI credentials, you can even publish your package to a PyPI-compatible index with:

1uv publish

uv provides additional handy commands for working with your package that you can see in the docs.

Conclusion

Although we've covered a lot of uv's functionality, we didn't dig into tools and the related uvx command, but we might in a future post, so keep an eye out.

At Prefect we're increasingly using uv and seeing the benefits. If you haven't yet tried uv, give it a whirl and I bet you'll see faster development and save time running your CI/CD pipelines.

Additional resources

Want to keep exploring? Check out these resources:

• Good YouTube video on uv with emphasis on its project management features and the source of the stat above on uv's popularity.
• Nice uv guide.
• uv docs.