Can others reproduce
your conda results?
Using pixi to manage dependencies
in a Python data analysis project

James Thomas
j.thomas@bristol.ac.uk

2025-06-05

Outline

• What is pixi? (and uv?)
• Why might you want to use pixi over conda?
• How do you configure pixi?
• Example of what this means for a data science project

Note:

• I am not the developer of these tools
• You may well know more than me – tips welcome, and thank you to my colleagues Léo and Rita for the feedback 😁

What does pixi do?

Similar to conda, like uv, poetry, npm, …

• install conda packages (python or R) or PyPI packages
• manage a conda environment (.pixi) and lockfile
• multi-platform support

• Set up a new (Python) project using modern best-practices for reproducibility.
• Start using an existing project with minimal commands (less than with conda).
• Keeps environment up-to-date more easily with other developers (which could be other people, or just you in the future!) Also makes it good for bootstrapping where someone doesn’t currently have Python.
• Supports development across multiple platforms (the tool is cross-platform, but it also allows developers to be on different platforms, too).

and:

• globally-accessible tools pixi global or pixi exec
• task runner
• create multiple environments
• easy to install, fast

Can:

• Install (or just use) tools with pylint, ruff system-wide.
• Define shortcuts for frequently-run commands (called tasks) - useful for new users.
• Create separate environments for testing, documentation, so not all packages installed at once - useful if some dependencies are large.
• Becomes more useful as a project grows.

Number of pixi releases, by month

• From the developers of mamba (now used in conda - e.g. miniforge) and micromamba.
• First released in June 2023, actively developed.
• Many recent releases have been to make speed and resource-use improvements.

What does uv do?

Similar to venv + pip, like pip-tools, poetry, …

• install PyPI packages
• manage a virtual environment (.venv) and lockfile

and:

• cheap disposable environments (jupyter lab, pytest)
• inline script metadata (recreating bugs, demos)
• globally-accessible tools with uvx (ruff, visidata)
• cross-platform, easy to install, fast

• uv can be thought of as the pixi equivalent for the PyPI ecosystem (in fact, pixi uses it to install PyPI packages!)
• Similarly allows you to install different versions of Python, so again, good for bootstrapping new users.
• Advise check it out - remainder of this talk will focus on pixi, but concepts are similar.

Why properly managing dependencies
is important (a motivating example using conda)

• Promised an example (conda_demo). Will show why, when you use conda, you should be careful about keeping track of your dependencies.
• ls -A - start with a skeleton system with conda installed, a my_script.py and an environment.yml file
• cat my_script.py that performs simple hypothesis test using scipy
• cat environment.yml is hand-crafted - can see this project was created a while ago, versions are old
• conda env create
• conda activate my_env (note the env at the start of the prompt)
• which python
• python -V
• python my_script.py
• Now want to do some more analysis using Pandas (could add pandas by hand to the environment.yml, but…)
• conda install pandas
• Should also note version and update the environment.yml at the same time
• nano environment.yml
• Need to remember to do this - if working by yourself and being careful, this may be fine
• At some point, will forget
• conda env export
• But amongst desired packages there are lots of dependencies - unclear which is which
• conda env export --from-history (but won’t get the version)
• But others will probably forget to commit, or forget the arguments to the export command
• And --from-history won’t recreate using the same versions of the dependencies
• So should really export everything
• And/or at some point will likely get a full export from someone else

Someone else updates the project
(and they’re using another platform)

• cd ../conda_demo_updated_on_mac
• ls -A
• conda env update --prune
• Error is confusing - can’t even find scipy or python, but if eagle-eyed, might notice “osxarm64” (this export was from a Mac running on Apple Silicon, not Linux (WSL) running on Intel)
• Could try removing the build specifications at the end of the version numbers (after the final =) - might work but probably won’t if there are specific dependencies for one platform and not another
• Would have to look through the diff manually and the imports in the project’s Python code to try to reconstruct this list - very tedious, and this is only for 2 packages

What about the workflow with pixi?

• (pixi_demo)
• ls -A note no conda or python installed (python)
• https://pixi.sh/latest/ and curl -fsSL https://pixi.sh/install.sh | sh (care with commands like this)
• pixi
• pixi init
• ls -A
• cat pixi.toml
• pixi add python=3.10
• pixi run python -V to verify version
• pixi add numpy==1.21.6 scipy==1.7.3 (as before)
• pixi run python my_script.py (note no need to activate)
• What has happened?
• ls -A
• cat pixi.toml
• cat pixi.lock (a YAML file with all the matching versions of the packages installed, plus - scroll down - hashes)
• (version control both of these - automatically up-to-date)
• pixi list to view the actual installed packages more comfortably
• Finally, ls -A .pixi (could delete this, and it would be created again)

Making changes to the project
(adding new dependencies and new developers)

• Now see what happens when want to add new dependency or a new developer
• Want to add pandas
• pixi add pandas
• cat pixi.toml
• pixi list
• Have some new developers, on Mac and Windows - could edit pixi.toml by hand, but can use the CLI
• pixi workspace
• pixi workspace platform osx-arm64 win-64
• cat pixi.toml
• cat pixi.lock
• pixi tree

Applying this in a data science context

• (pixi_lib_demo)
• tree to show layout
• Started with pixi init --format pyproject which creates pyproject.toml, not a pixi.toml (used by other Python tools as well)
• cat pyproject.toml (similar to pixi.toml however sections start with tool.pixi. - talk through the comments, however unlikely time to go through everything)
• cat src/my_project/probability.py (now in functions taking arguments, and using typer, a CLI library)
• Look at notebook later
• pixi run
• pixi run probability (note underscore changed to a hyphen)
• pixi run probability test-coin (--help and try arguments)
• pixi run (again)
• pixi run lab
• Look at notebooks including importing probability.py and autoreloading

Other tips

• Global tools
  • pixi exec ruff (faster the next time)
  • pixi global install quarto (then just quarto)
  • useful for running tools from conda without having to install them separately

Wrapping up

• Find out more: 🔗 pixi.sh (and 🔗 docs.astral.sh/uv)

• Use for new projects

  • pixi if Conda (uv if just PyPI)
  • package in src/
  • commands pixi run analysis

• When working with someone
  new to Python

  • pixi run dashboard
    only command they need?

• Easier than conda/pip by themselves

James Thomas
j.thomas@bristol.ac.uk

jatonline.github.io/managing-
dependencies-using-uv-and-pixi