Eric J Ma's Website

I recently switched LlamaBot and my personal website repository to pixi, and having test-driven it for a few weeks now, I think the time is right to do so. During this test-drive, I went cold turkey: I deleted ~/anaconda from my home directory, leaving me with no choice but to use pixi 100%. In this post, I'd like to share what I've learned from my perspective as a data scientist and software developer.

What is pixi?

In my mind, pixi is a package management multi-tool. pixi plays many roles, but here are the big, salient ones:

My needs under two personas

To motivate the content in this post, we first need to understand what I see as needs from a data scientist's and software developer's perspective.

Data scientist

Reproducible Environments

I need to replicate my computational environment from machine to machine. Strict versioning will be handy but shouldn't be unwieldy -- especially with lock files. Compared to my older ways of working using environment.yml files, where I manually pin versions when something broke, I now prefer to have my environment management tool automatically produce a lock file that locks in package versions defined when solving the environment.

Containerization

Containerization is also important. At work, we ship stuff within Docker containers. Whatever environment or package management tool I use must work well with Docker containers. Additionally, as a bonus, we need it to have GPU access, too! Moreover, the built container needs to be as lightweight as possible.

Composable GPU-enabled and CPU-only environments

With a single environment.yml file, I can only define a GPU-enabled or CPU-only environment. In most cases, we would default to GPU-enabled environments, which induces a huge overhead when the code is run on CPU-only environments. Ideally, I'd like to be able to do composable environment specification: a default setting for CPU-only environments, with the ability to specify GPU-only dependencies in a fashion that composes with the CPU-only environment within a single, canonical configuration file (e.g. pyproject.toml).

Software developer

Publishing to PyPI

As a Python tool developer, I create stuff that needs to be distributed to other Pythonistas. As such, I need to be able to leverage existing Python publishing tooling (e.g., PyPI or conda-forge).

Running tests

Whatever tool I use, I also need to be able to run software tests easily. Ideally, this would be done with one command, e.g., pytest, with minimal overhead. The software testing environment should be the same as my development environment.

Both personas' needs

Simplified environment specification

As mentioned above, I currently use environment.yml and pyproject.toml to specify runtime and development dependencies. Runtime dependencies are declared in pyproject.toml, while development dependencies are declared in environment.yml. However, this way of separating concerns means that we have duplication in our dependencies specification! Ideally, we'd like to do this with a single file.

Tooling for fast environment builds

Whether we are in a Docker container or not, I'd like to see much faster container and environment build times with tooling for caching when compared to what I currently experience.

(Note to self: Product-oriented folks keep talking about "don't solution, tell me the problem", but I have to say -- I only really knew how much of a problem this was once I touched the solution I'm about to talk about!)

Automatic updates of lock files

When I add a dependency to my environment, I'd like to see the lock file automatically updated so that I don't have to remember to do that in the future.

How pixi operates

With the desiderata outlined above as contextual information, let's now see how pixi works.

Installation

What's awesome is that pixi is installed via a simple, one-bash-script install. Official instructions are available here, but as of the time of writing, the only thing we need to run is:

curl -fsSL https://pixi.sh/install.sh | bash

This sets up pixi within your home directory and adds the pixi path to your PATH environment variable.

I also recommend getting set up with autocompletion for your shell. Once again, official instructions are here, and what I used for the zsh (macOS' default shell) was:

echo 'eval "$(pixi completion --shell zsh)"' >> ~/.zshrc

Starting a new project with pixi enabled

For those who are used to conda idioms, pixi won't create a centralized conda environment but instead will create a .pixi sub-directory within the directory of your pixi configuration file (either pyproject.toml or pixi.toml). This is akin to poetry for Pythonistas, or npm for those coming from the NodeJS ecosystem and cargo for Rustaceans. In line with my desire to be able to specify sets of dependencies for runtime environments and development environments and also minimize the number of configuration files present, it makes sense to initialize with a pyproject.toml configuration rather than the default pixi.toml configuration file. This is done by executing:

pixi init --format pyproject -c conda-forge -v

This gives me the following pyproject.toml file:

[project]
name = "pixi-cuda-environment" # defaults to my directory name.
version = "0.1.0"
description = "Add a short description here"
authors = [{name = "Eric Ma", email = "e************@gmail.com"}]
requires-python = ">= 3.11"
dependencies = []

[build-system]
requires = ["setuptools"]
build-backend = "setuptools.build_meta"

[tool.pixi.project]
channels = ["conda-forge"]
platforms = ["osx-arm64"]

[tool.pixi.pypi-dependencies]
pixi-cuda-environment = { path = ".", editable = true }

[tool.pixi.tasks]

What's important to note here is that [project] -> dependencies lists the Python project's runtime dependencies, which will be respected when pixi creates or updates the environment, as it is interpreted as being pulled from PyPI. In other words, the following two configurations are equivalent:

[project]
dependencies = [
    "docutils",
    "BazSpam == 1.1",
]

[project.optional-dependencies]
PDF = ["ReportLab>=1.2", "RXP"]

and

[tool.pixi.pypi-dependencies]
docutils = "*"
BazSpam = "== 1.1"

[tool.pixi.feature.PDF.pypi-dependencies]
ReportLab = ">=1.2"
RXP = "*"

h/t Ruben Arts, one of the core developers of pixi, who educated me on this point.

The advantage of putting runtime dependencies for distributed Python packages inside the project.dependencies section is that when you build the package to be distributed on PyPI, there's no need to double-specify the dependency chain within the tool.pixi.dependencies section, since project.dependencies is already respected by pixi.

Apart from that, other default configuration settings within the configuration file should look familiar to those who have experience with pyproject.toml configuration.

Adding dependencies

As mentioned above, runtime dependencies, which must exist when someone else installs your package to be used in their project, are distinct from development dependencies, which are usually extra things needed to develop your package properly. My goal is that if someone pip installs my package, they will automatically have the full runtime dependency set. To that end, there is a two-step process that I would use to add packages to the project. The first step is to set up your development environment using pixi add <package name>, which pulls from conda-forge, and develop the data science project. Then, once one is ready for distributing the project as a Python package, we use the pixi add --pypi <package names here> command (note the --pypi flag!) to ensure that they get added into the project --> dependencies section.

For example, if my code depends on pandas to be used, I would run:

pixi add --pypi pandas

On the other hand, if my development workflow depends on ruff and pytest, I would run:

pixi add ruff pytest

Those two commands will give us the following modified pyproject.toml file:

# FILE: pyproject.toml
[project]
name = "pixi-cuda-environment"
version = "0.1.0"
description = "Add a short description here"
authors = [{name = "Eric Ma", email = "e************@gmail.com"}]
requires-python = ">= 3.11"
dependencies = ["pandas>=2.2.2,<2.3"] # <-- NOTE: stuff added via `pixi add --pypi` goes here!

[build-system]
requires = ["setuptools"]
build-backend = "setuptools.build_meta"

[tool.pixi.project]
channels = ["conda-forge"]
platforms = ["osx-arm64", "linux-64"] # <-- NOTE: I modified this manually to include linux-64, and you can add osx-64 if you're on an Intel Mac!

[tool.pixi.pypi-dependencies]
pixi-cuda-environment = { path = ".", editable = true } # <-- NOTE: this ensures that we never have to run `pip install -e .` ourselves!

[tool.pixi.tasks]

[tool.pixi.dependencies] # <-- NOTE: stuff added via `pixi add` goes here!
ruff = ">=0.5.5,<0.6"
pytest = ">=8.3.2,<8.4"

"Activating" an environment

Because there's no centralized conda environment, if you need to run commands within the pixi environment (which has the same structure as a conda environment), you run pixi shell rather than conda activate <env_name>.

Running pixi shell at my terminal results in something that looks like this:

❯ which python
python not found

❯ pixi shell
 . "/var/folders/lb/fzbrctwd2klcrwsg63vzqj0r0000gn/T/pixi_env_37v.sh"

❯  . "/var/folders/lb/fzbrctwd2klcrwsg63vzqj0r0000gn/T/pixi_env_37v.sh"

(pixi-cuda-environment)
❯ which python
/Users/ericmjl/github/incubator/pixi-cuda-environment/.pixi/envs/default/bin/python

(pixi-cuda-environment)
❯

Notice how before running pixi shell, I couldn't execute Python natively in my terminal. After all, I went cold turkey and deleted my mambaforge installation! (There is a twist: you can use pixi to install Python globally! I address this below. Though for the sake of simplicity for now, let's assume that a global Python installation doesn't exist.) But after running pixi shell, I can -- because the /Users/ericmjl/github/incubator/pixi-cuda-environment/.pixi/envs/default/bin/python path is now set on my PATH environment variable.

Adding CUDA dependencies

At this point, NVIDIA hardware is the default hardware used by data scientists for GPU-accelerated computation. As such, I need to know how to use pixi to install GPU-enabled packages. I went to my go-to, JAX and PyTorch, to figure out how to install the GPU-enabled versions of these packages.

To make this happen, we must distinguish between both environments: CPU-only, like on my Mac, and GPU-enabled, like on my home lab that runs Ubuntu. First, we define the common and CUDA-specific dependencies on CUDA-enabled systems. To start, my pyproject.toml file changes a bit on the dependencies section:

# FILE: pyproject.toml
[tool.pixi.dependencies]
ruff = ">=0.5.5,<0.6"
pytest = ">=8.3.2,<8.4"
ipython = ">=8.26.0,<9"
# NOTE: I need JAX and PyTorch to be installed in both places.
jax = "*"
pytorch = "*"

[tool.pixi.feature.cuda]
system-requirements = {cuda = "12"} # this will support CUDA minor/patch versions!

[tool.pixi.feature.cuda.target.linux-64.dependencies]
jaxlib = { version = "*", build = "cuda12" }

# Environments
[tool.pixi.environments]
cuda = ["cuda"] # maps my "cuda" environment to the "cuda" feature

For simplicity in explanation, and to show how conda package dependencies are solved, I've opted to list jax and pytorch under the tool.pixi.dependencies section rather than under project.dependencies. I am also using a floating version for simplicity's sake. If you're curious what happens, I'd encourage you to try listing jax or pytorch under project.dependencies instead and see what happens to the resolved environments!

Here, I used pixi's feature definitions. This concept comes from the Rust world if I remember correctly my conversation with the Prefix devs about specifying a CUDA-centric specification for my Linux machine. With the above modifications to my pyproject.toml file, I now have two ways to run pixi shell:

• pixi shell -e cuda will start a shell environment with cuda enabled, while
• pixi shell will start a shell environment without cuda.

We can tell which environment we're in through the environment name upon activation. Using pixi shell -e cuda on my home lab, we get:

❯ pixi shell -e cuda
 . "/tmp/pixi_env_41w.sh"

❯  . "/tmp/pixi_env_41w.sh"

(pixi-cuda-environment:cuda)
❯ ipython
Python 3.12.4 | packaged by conda-forge | (main, Jun 17 2024, 10:23:07) [GCC 12.3.0]
Type 'copyright', 'credits' or 'license' for more information
IPython 8.26.0 -- An enhanced Interactive Python. Type '?' for help.

In [1]: import jax.numpy as np; a = np.arange(3)

In [2]: a.devices()
Out[2]: {cuda(id=0)}

In [3]: import torch
t
In [4]: a = torch.tensor([1, 2, 3.0])

In [5]: a.cuda()
Out[5]: tensor([1., 2., 3.], device='cuda:0')

Notice how I've been able to run JAX and PyTorch with CUDA enabled.

On the other hand, using pixi shell (without -e cuda), we have the following behaviour:

❯ pixi shell

❯  . "/tmp/pixi_env_YJX.sh"

(pixi-cuda-environment)
❯ ipython
Python 3.12.4 | packaged by conda-forge | (main, Jun 17 2024, 10:23:07) [GCC 12.3.0]
Type 'copyright', 'credits' or 'license' for more information
IPython 8.26.0 -- An enhanced Interactive Python. Type '?' for help.

In [1]: import jax.numpy as np; a = np.arange(3)
An NVIDIA GPU may be present on this machine, but a CUDA-enabled jaxlib is not installed. Falling back to cpu.

In [2]: a.devices()
Out[2]: {CpuDevice(id=0)}

In [3]: import torch

In [4]: a = torch.tensor([1, 2, 3.0])

In [5]: a.cuda()
---------------------------------------------------------------------------
AssertionError                            Traceback (most recent call last)
Cell In[5], line 1
----> 1 a.cuda()

File ~/github/incubator/pixi-cuda-environment/.pixi/envs/default/lib/python3.12/site-packages/torch/cuda/__init__.py:284, in _lazy_init()
    279     raise RuntimeError(
    280         "Cannot re-initialize CUDA in forked subprocess. To use CUDA with "
    281         "multiprocessing, you must use the 'spawn' start method"
    282     )
    283 if not hasattr(torch._C, "_cuda_getDeviceCount"):
--> 284     raise AssertionError("Torch not compiled with CUDA enabled")
    285 if _cudart is None:
    286     raise AssertionError(
    287         "libcudart functions unavailable. It looks like you have a broken build?"
    288     )

AssertionError: Torch not compiled with CUDA enabled

In [6]:

As you can tell above, our arrays are instantiated on CPU memory rather than GPU memory in the non-CUDA-enabled environment. With a relatively elegant syntax, we can create two separate environments within the same project for different hardware specifications.

Building docker containers with caching on CI/CD

Another need I have at work is the ability to build docker containers that contain our source code library and can be shipped into the cloud. Additionally, these containers should be as small as possible. Finally, when building these containers via CI/CD, we need caching enabled to ensure that build times are reasonably short when nothing changes in the environment, with the environment defined by the pyproject.toml file and generated pixi.lock file. (More on pixi.lock later.)

I went about test-driving how to make this happen. It turns out not to be too challenging! To simplify the build, thereby allowing me to iterate more, I removed pytorch from the environment (as specified above) for the following environment definitions:

# FILE: pyproject.toml
[tool.pixi.dependencies]
ruff = ">=0.5.5,<0.6"
pytest = ">=8.3.2,<8.4"
ipython = ">=8.26.0,<9"
jax = "*"

# Feature Definitions
[tool.pixi.feature.cuda]
platforms = ["linux-64"]
system-requirements = {cuda = "12"}

[tool.pixi.feature.cuda.dependencies]
jaxlib = { version = "*", build = "cuda12" }

# Environments
[tool.pixi.environments]
cuda = ["cuda"]

Then, I made a Dockerfile that looks like the following:

# FILE: Dockerfile
FROM ghcr.io/prefix-dev/pixi:latest

WORKDIR /repo

COPY pixi.lock /repo/pixi.lock
COPY pyproject.toml /repo/pyproject.toml

RUN /usr/local/bin/pixi install --manifest-path pyproject.toml --environment cuda

# Entrypoint shell script ensures that any commands we run start with `pixi shell`,
# which in turn ensures that we have the environment activated
# when running any commands.
COPY entrypoint.sh /repo/entrypoint.sh
RUN chmod 700 /repo/entrypoint.sh
ENTRYPOINT [ "/repo/entrypoint.sh" ]

Notice how there is an official pixi Docker container! Most crucially, it does not ship with anything CUDA-related, so how do we get CUDA packages into it? Well, it is now possible to use non-NVIDIA docker containers to run GPU code thanks to the many efforts of conda-forge community members who work for NVIDIA, as we can specify whether we need GPU-related packages or not entirely within our conda/pixi environments instead.

To test-drive, I built the docker container locally on my home lab machine:

docker build -f Dockerfile . -t pixi-cuda

On my home lab machine, Docker build took about 3 minutes to complete. The longest build step was the Pixi install command in the Dockerfile; the second longest build step was exporting the layers.

[+] Building 194.3s (12/12) FINISHED                      docker:defaultl
 => [internal] load build definition from Dockerfile                0.0s
 => => transferring dockerfile: 533B                                0.0sj
 => [internal] load metadata for ghcr.io/prefix-dev/pixi:latest     0.4s
 => [internal] load .dockerignore                                   0.0so
 => => transferring context: 2B                                     0.0s
 => [1/7] FROM ghcr.io/prefix-dev/pixi:latest@sha256:45d86bb788aaa  0.0s
 => [internal] load build context                                   0.0s
 => => transferring context: 118.83kB                               0.0s
 => CACHED [2/7] WORKDIR /repo                                      0.0s
 => [3/7] COPY pixi.lock /repo/pixi.lock                            0.1s
 => [4/7] COPY pyproject.toml /repo/pyproject.toml                  0.1s
 => [5/7] RUN /usr/local/bin/pixi install --manifest-path pyproj  170.9s
 => [6/7] COPY entrypoint.sh /repo/entrypoint.sh                    0.1s
 => [7/7] RUN chmod 700 /repo/entrypoint.sh                         0.3s
 => exporting to image                                             22.2s
 => => exporting layers                                            22.2s
 => => writing image sha256:6f13e2a7b362c7b3736d4405f7f5566775320d  0.0s

To verify that I could indeed run CUDA-accelerated JAX, I entered into the container:

❯ docker run --gpus all -it docker.io/library/pixi-cuda /bin/bash
 . "/tmp/pixi_env_5FU.sh"

root@295a84b64679:/repo#  . "/tmp/pixi_env_5FU.sh"

(pixi-cuda-environment:cuda) root@295a84b64679:/repo# ipython
Python 3.12.4 | packaged by conda-forge | (main, Jun 17 2024, 10:23:07) [GCC 12.3.0]
Type 'copyright', 'credits' or 'license' for more information
IPython 8.26.0 -- An enhanced Interactive Python. Type '?' for help.

In [1]: import jax.numpy as np; a = np.arange(3)

In [2]: a.devices()
Out[2]: {cuda(id=0)}

In [3]:

I was indeed able to run the container with GPU acceleration! I had to wrestle with installing the NVIDIA docker container toolkit on my home lab, which was a one-time installation to work through, but that's tangential to exploring pixi + Docker. At this point, I'm pretty sure that pixi is a no-brainer to have: any sane ML Platform team at a company should be baking it into your AMIs (if you're on AWS) or equivalent.

What about that entrypoint.sh file? Well, did you notice this line in my shell output above?

(pixi-cuda-environment:cuda) root@295a84b64679:/repo# ipython

Yes, that's right: we have the cuda environment for my pixi-cuda-environment project enabled, rather than just a plain shell. That is courtesy of entrypoint.sh. Here's what it looks like:

#!/bin/bash
# FILE: entrypoint.sh
# Modified from: https://stackoverflow.com/a/44079215
# This script ensures that within a Docker container
# we have the right pixi environment activated
# before executing a command

# If `nvidia-smi` is available, then execute command in the `cuda` environment
# as defined in `pyproject.toml`.
if command -v nvidia-smi &> /dev/null; then
    pixi shell -e cuda
else
    pixi shell
fi
exec "$@"

I'll note that it's become standard to build Docker containers on CI/CD runners, This gives us the advantage of triggering a build automatically on every commit, as opposed to triggering a build manually like I did on my home lab. This will be discussed below!

Running software tests

Software tests are another integral part of our work workflow, so I explored what it would take to run tests with pixi in the loop. A few important contextual matters:

1. The software environment only exists at the terminal if we run pixi shell.
2. pixi does allow us to define tasks like in Makefiles and specify the environment in which they run.

Taking advantage of these two, we can continue configuring our pixi environment and demonstrate how to make software tests run.

Firstly, we use pixi to add pytest and pytest-cov to the environment:

pixi add pytest pytest-cov

This added pytest-cov to pyproject.toml. Because I already had pytest defined, it was not overwritten:

# FILE: pyproject.toml
[tool.pixi.dependencies]
ruff = ">=0.5.5,<0.6"
pytest = ">=8.3.2,<8.4"
ipython = ">=8.26.0,<9"
jax = "*"
pytest-cov = ">=5.0.0,<6" # <-- NOTE: This was newly added!

Then, I added a dummy test file, test_arrays.py, into the top-level directory:

# FILE: test_arrays.py
"""Example test for arrays."""
import jax.numpy as np


def test_array():
    """Test array creation."""
    a = np.arange(3)
    print(a.devices())

Finally, I added a test task under tool.pixi.tasks in pyproject.toml:

# FILE: pyproject.toml
[tool.pixi.tasks]
test = "pytest" # <-- NOTE: This was newly added!

Now, how do we run this test? Turns out it's doable this way:

pixi run test

When run outside of a specified pixi shell, it will run inside the default environment, giving an output like this:

❯ pixi run test
✨ Pixi task (test in default): pytest
================== test session starts ==================
platform linux -- Python 3.12.4, pytest-8.3.2, pluggy-1.5.0
rootdir: /home/ericmjl/github/incubator/pixi-cuda-environment
configfile: pyproject.toml
plugins: cov-5.0.0
collected 1 item

test_arrays.py .                                  [100%]

=================== 1 passed in 0.38s ===================

But what if I want to run it inside the cuda environment? The easiest way to do this is to specify the environment to run it in:

pixi run -e cuda test

NOTE: Don't get the order wrong! -e cuda must be specified before the task name (i.e. test), as anything passed in after the task name will be passed to the task's executable as an argument!

There are other ways to accomplish the same thing, but this is the easiest and most unambiguous. In any case, the output of that last command looks like this:

❯ pixi run -e cuda test
✨ Pixi task (test in cuda): pytest  ### NOTE: We are using the `cuda` env now!
==================== test session starts =====================
platform linux -- Python 3.12.4, pytest-8.3.2, pluggy-1.5.0
rootdir: /home/ericmjl/github/incubator/pixi-cuda-environment
configfile: pyproject.toml
plugins: cov-5.0.0
collected 1 item

test_arrays.py .                                       [100%]

===================== 1 passed in 0.69s ======================

Notice how we are running tests within the cuda environment!

Running tests with pixi on GitHub Actions

Now that we can run tests with pixi locally, and do it in two environments, the next thing I think we need to look at is running tests on CI/CD. Because I am on GitHub, GitHub Actions is my CI/CD choice. Let's see how to run tests with pixi on GitHub Actions. To start, we will need the GitHub Action definition, which I place at .github/workflows/test.yaml:

# FILE: .github/workflows/test.yaml
name: Run software tests

on:
  push:
  pull_request:
    branches:
      - main

jobs:
  test:
    name: Run tests
    runs-on: ubuntu-latest
    steps:
    - name: Checkout repository
      uses: actions/checkout@v4

    - uses: prefix-dev/setup-pixi@v0.8.1
      with:
        pixi-version: v0.25.0
        cache: true

    - run: pixi run test

I intentionally omitted the cuda environment because GPU runners are unavailable on GitHub actions' native runners. However, if we use self-hosted runners, it is, in principle, possible to set up the cuda environment that we defined.

Nonetheless, the biggest and most important piece to know here is that the prefix devs have provided us with a setup-pixi action! This allows us to set up pixi and automatically create the pixi environments associated with it. Better yet is the ability to do caching! This massively speeds up the time taken to create the pixi env on CI/CD, which can cut down the turn-around time to fixing bugs.

To recap, running pixi install from scratch on my home lab takes about 23 seconds for the default environment, while pixi install -e cuda takes about 3 minutes or so. With caching on GitHub actions, loading the default environment from the cache takes only 8 seconds. Imagine the speedup that one would have with caching provided by pixi!

NOTE: Home lab build times and GH Actions build times without caching are similar! The point of the table above was to illustrate the power of caching on CI/CD.

Docker container build with GitHub Actions

In addition to running tests on GitHub Actions, it's also important for me to be able to build Docker containers using Actions. This way, we offload the time otherwise needed to babysit a local machine to a remote computer that is automatically triggered by a code commit.

To do this, we need an Actions workflow file:

# FILE: .github/workflows/build-docker.yaml
name: Build docker container

on: push

jobs:
  build-container:
    name: Build container
    runs-on: ubuntu-latest

    steps:
    - name: Checkout repository
      uses: actions/checkout@v2
    - name: Docker meta
      id: meta
      uses: docker/metadata-action@v5
      with:
        # list of Docker images to use as base name for tags
        images: |
          ericmjl/pixi-cuda-environment
        # generate Docker tags based on the following events/attributes
        tags: |
          type=sha
          type=raw,value=latest

    - name: Set up QEMU
      uses: docker/setup-qemu-action@v3

    - name: Set up Docker Buildx
      uses: docker/setup-buildx-action@v3

    # NOTE: Change this to your own AWS ECR login or other container repo service,
    # such as GitHub Container Registry, which doesn't require login tokens.
    - name: Login to Docker Hub
      uses: docker/login-action@v3
      with:
        username: ${{ secrets.DOCKERHUB_USERNAME }}
        password: ${{ secrets.DOCKERHUB_TOKEN }}

    - name: Build and push
      uses: docker/build-push-action@v6
      with:
        push: true
        tags: ${{ steps.meta.outputs.tags }}
        cache-from: type=gha
        cache-to: type=gha,mode=max

Nothing within this Actions YAML file is pixi-related, but because we're building the container from the Dockerfile, which is based on the official pixi image, we take advantage of building a Docker container using the official GitHub Action -- with caching involved too!

When comparing the time it takes to generate the container image without caching, it takes about 3 minutes to build the container and another 2-ish minutes to push it to Dockerhub. (Your internet connection will determine how fast pushes happen.) With image layer caching, as long as the pyproject.toml and pixi.lock files are unchanged between commits, the entire GitHub Actions build time falls to ~30 seconds, representing an approximately 10X speed-up between iterations. Summarized as a table:

If you have anything downstream dependent on a Docker image being built, this massive reduction in iteration time is a definite big treat!

Lock files

Every pixi command that depends on an environment will result in a check that the lock files are kept in sync with the configuration file (or manifest, in pixi parlance). What pixi commands do this? What I've seen includes:

• pixi run <task> or pixi run <shell command>, ± -e <env name>
• pixi shell
• pixi install

The full behaviour is documented here. This is incredibly useful for ensuring the lock file syncs with the environment configuration file!

Putting it all together

While writing this post, I implemented and test-drove the ideas in a new repo named pixi-cuda-environment. This serves as a minimal implementation of the ideas above. h/t Ben Mares and Adrian Seyboldt, who helped review the repo.

To apply everything I learned from this exercise, I decided to update pyds-cli, llamabot, and my personal website to depend solely on pixi. This turned out to be an incredible thing to do! I could move from my MacBook Air to my home lab and get up and running easily using pixi install, simply by investing some time and thought into the configuration file. On another code repository that I wasn't the core developer of, thanks to a well-configured pixi.toml file (the alternative to pyproject.toml), I could immediately run the web app associated with the repository.

All taken together, although there are many benefits to adopting pixi -- reproducibility and task execution, to name two -- I think its secret sauce for Python-centric adoption lies in the following two points:

• It pulls Python packages from the two canonical sources of packages (conda and pip).
• It can unify Python project configuration into a single file (pyproject.toml).

The thought and care that went into pixi's design are quite evident. I hope that it continues to get the development attention that is commensurate with its impact!

Tips for using pixi

Any change will feel slightly uncomfortable, and I experienced some of it. I needed to adjust to this change, having come from a world where ~/anaconda and mamba were what I depended on and then suddenly going cold turkey. Here's some tips that I have for using pixi, based on test-driving it at home and at work.

Lock file errors? Keep calm and pixi install!

Lock files can represent a big change, a departure from an environment.yml file without a conda-lock file. Lock files can also look intimidating! After all, they're filled with many, many lines of auto-generated text. At the same time, they are intended to be checked into source control! (That last point broke my mental model of never checking in auto-generated files.)

At the same time, lock files can also go out of sync with your manifest file (i.e. pyproject.toml or pixi.toml) if one is not careful. pixi tries to manage this complexity by (a) auto-updating the pixi.lock file on almost all commands, making it more convenient than conda-lock, and (b) loudly erroring with a friendly error message. If ever you encounter that error, whether locally or on CI, saying that the lock files are out of date, a pixi install + git commit and git push will do the trick.

CUDA specification syntax

The syntax for specifying CUDA packages from conda-forge, is also something we need to commit to memory. The patterns to remember look like:

package_name = { version = "some version", build = "*cuda*" }
package_name = { version = "some version", build = "*cuda12*" }

Default channels

The default pixi behaviour is to look for packages from conda-forge. If your Python package can't be resolved from there, try moving it to the pypi-dependencies section instead.

Global tool installation

If you went cold turkey as I did but still need to install Python tooling globally, then there's muscle memory that you'll need to un-learn very quickly. Whereas previously I could rely on a base environment through my mambaforge installation, it no longer exists. Rather, I needed to install Python and pipx globally with pixi and then use pipx (not pixi!) to install that tool using

pipx install --python $(which python) <tool name>

In this way, I could install pyds-cli and llamabot "globally" rather than within each environment.

Define pixi tasks thoughtfully

Providing a task that allows for a one-command "get something useful accomplished" can be incredibly confidence-building for the next person who clones your repo locally and tries to run something. Examples might be a lab command to run JupyterLab, a start command that runs a Python script that demos the output of your work, or a test command to help software developers gain confidence that they've got the repo installed correctly. Be sure to provide that command in the README!

Minimal analyses

What if you're not a tool developer and just want a quick environment to do some analyses without touching other environments? For this persona, there is a simplified workflow that I have used which may be helpful as a reference.

Navigate to an empty directory. Then:

# Initialize pixi files
pixi init
# Add packages that you need
pixi add jupyter ipython ipykernel pixi-kernel numpy pandas matplotlib seaborn scikit-learn
# Run Jupyter lab
pixi run jupyter lab
# Add more packages as you see fit when it's needed; pixi will keep things in sync!
pixi add statsmodels

Thanks to pixi-kernel, there will be a Jupyter kernel named Pixi - Python 3 (ipykernel) which will contain the packages contained in your environment. If you commit the notebooks + pixi configuration files to source control, someone else can download it and reproduce the environment easily. And as long as you don't rely on data living at a hard-coded path, that other person should be able to reproduce your work as well!

Acknowledgments

I'd like to thank Sean Law, Rahul Dave, and Ruben Arts for reviewing this content, as well as Juan Orduz for battle-testing the blog post in support of his own project's migration.