Multi Environment Support

Motivating Example

There are multiple scenarios where multiple environments are useful.

• Testing of multiple package versions, e.g. py39 and py310 or polars 0.12 and 0.13.
• Smaller single tool environments, e.g. lint or docs.
• Large developer environments, that combine all the smaller environments, e.g. dev.
• Strict supersets of environments, e.g. prod and test-prod where test-prod is a strict superset of prod.
• Multiple machines from one project, e.g. a cuda environment and a cpu environment.
• And many more. (Feel free to edit this document in our GitHub and add your use case.)

This prepares pixi for use in large projects with multiple use-cases, multiple developers and different CI needs.

Design Considerations

There are a few things we wanted to keep in mind in the design:

1. User-friendliness: Pixi is a user focussed tool that goes beyond developers. The feature should have good error reporting and helpful documentation from the start.
2. Keep it simple: Not understanding the multiple environments feature shouldn't limit a user to use pixi. The feature should be "invisible" to the non-multi env use-cases.
3. No Automatic Combinatorial: To ensure the dependency resolution process remains manageable, the solution should avoid a combinatorial explosion of dependency sets. By making the environments user defined and not automatically inferred by testing a matrix of the features.
4. Single environment Activation: The design should allow only one environment to be active at any given time, simplifying the resolution process and preventing conflicts.
5. Fixed lock files: It's crucial to preserve fixed lock files for consistency and predictability. Solutions must ensure reliability not just for authors but also for end-users, particularly at the time of lock file creation.

Feature & Environment Set Definitions

Introduce environment sets into the pixi.toml this describes environments based on feature's. Introduce features into the pixi.toml that can describe parts of environments. As an environment goes beyond just dependencies the features should be described including the following fields:

• dependencies: The conda package dependencies
• pypi-dependencies: The pypi package dependencies
• system-requirements: The system requirements of the environment
• activation: The activation information for the environment
• platforms: The platforms the environment can be run on.
• channels: The channels used to create the environment. Adding the priority field to the channels to allow concatenation of channels instead of overwriting.
• target: All the above features but also separated by targets.
• tasks: Feature specific tasks, tasks in one environment are selected as default tasks for the environment.

Default features

[dependencies] # short for [feature.default.dependencies]
python = "*"
numpy = "==2.3"

[pypi-dependencies] # short for [feature.default.pypi-dependencies]
pandas = "*"

[system-requirements] # short for [feature.default.system-requirements]
libc = "2.33"

[activation] # short for [feature.default.activation]
scripts = ["activate.sh"]

Different dependencies per feature

[feature.py39.dependencies]
python = "~=3.9.0"
[feature.py310.dependencies]
python = "~=3.10.0"
[feature.test.dependencies]
pytest = "*"

Full set of environment modification in one feature

[feature.cuda]
dependencies = {cuda = "x.y.z", cudnn = "12.0"}
pypi-dependencies = {torch = "1.9.0"}
platforms = ["linux-64", "osx-arm64"]
activation = {scripts = ["cuda_activation.sh"]}
system-requirements = {cuda = "12"}
# Channels concatenate using a priority instead of overwrite, so the default channels are still used.
# Using the priority the concatenation is controlled, default is 0, the default channels are used last.
# Highest priority comes first.
channels = ["nvidia", {channel = "pytorch", priority = -1}] # Results in:  ["nvidia", "conda-forge", "pytorch"] when the default is `conda-forge`
tasks = { warmup = "python warmup.py" }
target.osx-arm64 = {dependencies = {mlx = "x.y.z"}}

Define tasks as defaults of an environment

[feature.test.tasks]
test = "pytest"

[environments]
test = ["test"]

# `pixi run test` == `pixi run --environment test test`

The environment definition should contain the following fields:

• features: Vec<Feature>: The features that are included in the environment set, which is also the default field in the environments.
• solve-group: String: The solve group is used to group environments together at the solve stage. This is useful for environments that need to have the same dependencies but might extend them with additional dependencies. For instance when testing a production environment with additional test dependencies.

Creating environments from features

[environments]
# implicit: default = ["default"]
default = ["py39"] # implicit: default = ["py39", "default"]
py310 = ["py310"] # implicit: py310 = ["py310", "default"]
test = ["test"] # implicit: test = ["test", "default"]
test39 = ["test", "py39"] # implicit: test39 = ["test", "py39", "default"]

Testing a production environment with additional dependencies

[environments]
# Creating a `prod` environment which is the minimal set of dependencies used for production.
prod = {features = ["py39"], solve-group = "prod"}
# Creating a `test_prod` environment which is the `prod` environment plus the `test` feature.
test_prod = {features = ["py39", "test"], solve-group = "prod"}
# Using the `solve-group` to solve the `prod` and `test_prod` environments together
# Which makes sure the tested environment has the same version of the dependencies as the production environment.

Creating environments without including the default feature

[dependencies]
python = "*"
numpy = "*"

[feature.lint.dependencies]
pre-commit = "*"

[environments]
# Create a custom environment which only has the `lint` feature (numpy isn't part of that env).
lint = {features = ["lint"], no-default-feature = true}

lock file Structure

Within the pixi.lock file, a package may now include an additional environments field, specifying the environment to which it belongs. To avoid duplication the packages environments field may contain multiple environments so the lock file is of minimal size.

- platform: linux-64
  name: pre-commit
  version: 3.3.3
  category: main
  environments:
    - dev
    - test
    - lint
  ...:
- platform: linux-64
  name: python
  version: 3.9.3
  category: main
  environments:
    - dev
    - test
    - lint
    - py39
    - default
  ...:

User Interface Environment Activation

Users can manually activate the desired environment via command line or configuration. This approach guarantees a conflict-free environment by allowing only one feature set to be active at a time. For the user the cli would look like this:

Default behavior

pixi run python
# Runs python in the `default` environment

Activating an specific environment

pixi run -e test pytest
pixi run --environment test pytest
# Runs `pytest` in the `test` environment

Activating a shell in an environment

pixi shell -e cuda
pixi shell --environment cuda
# Starts a shell in the `cuda` environment

Running any command in an environment

pixi run -e test any_command
# Runs any_command in the `test` environment which doesn't require to be predefined as a task.

Interactive selection of environments if task is in multiple environments

# In the scenario where test is a task in multiple environments, interactive selection should be used.
pixi run test
# Which env?
# 1. test
# 2. test39

Important links

• Initial writeup of the proposal: GitHub Gist by 0xbe7a
• GitHub project: #10

Real world example use cases

Polarify test setup

In polarify they want to test multiple versions combined with multiple versions of polars. This is currently done by using a matrix in GitHub actions. This can be replaced by using multiple environments.

pixi.toml

[project]
name = "polarify"
# ...
channels = ["conda-forge"]
platforms = ["linux-64", "osx-arm64", "osx-64", "win-64"]

[tasks]
postinstall = "pip install --no-build-isolation --no-deps --disable-pip-version-check -e ."

[dependencies]
python = ">=3.9"
pip = "*"
polars = ">=0.14.24,<0.21"

[feature.py39.dependencies]
python = "3.9.*"
[feature.py310.dependencies]
python = "3.10.*"
[feature.py311.dependencies]
python = "3.11.*"
[feature.py312.dependencies]
python = "3.12.*"
[feature.pl017.dependencies]
polars = "0.17.*"
[feature.pl018.dependencies]
polars = "0.18.*"
[feature.pl019.dependencies]
polars = "0.19.*"
[feature.pl020.dependencies]
polars = "0.20.*"

[feature.test.dependencies]
pytest = "*"
pytest-md = "*"
pytest-emoji = "*"
hypothesis = "*"
[feature.test.tasks]
test = "pytest"

[feature.lint.dependencies]
pre-commit = "*"
[feature.lint.tasks]
lint = "pre-commit run --all"

[environments]
pl017 = ["pl017", "py39", "test"]
pl018 = ["pl018", "py39", "test"]
pl019 = ["pl019", "py39", "test"]
pl020 = ["pl020", "py39", "test"]
py39 = ["py39", "test"]
py310 = ["py310", "test"]
py311 = ["py311", "test"]
py312 = ["py312", "test"]

.github/workflows/test.yml

jobs:
  tests-per-env:
    runs-on: ubuntu-latest
    strategy:
      matrix:
        environment: [py311, py312]
    steps:
    - uses: actions/checkout@v4
      - uses: prefix-dev/setup-pixi@v0.5.1
        with:
          environments: ${{ matrix.environment }}
      - name: Run tasks
        run: |
          pixi run --environment ${{ matrix.environment }} test
  tests-with-multiple-envs:
    runs-on: ubuntu-latest
    steps:
    - uses: actions/checkout@v4
    - uses: prefix-dev/setup-pixi@v0.5.1
      with:
       environments: pl017 pl018
    - run: |
        pixi run -e pl017 test
        pixi run -e pl018 test

Test vs Production example

This is an example of a project that has a test feature and prod environment. The prod environment is a production environment that contains the run dependencies. The test feature is a set of dependencies and tasks that we want to put on top of the previously solved prod environment. This is a common use case where we want to test the production environment with additional dependencies.

pixi.toml

[project]
name = "my-app"
# ...
channels = ["conda-forge"]
platforms = ["osx-arm64", "linux-64"]

[tasks]
postinstall-e = "pip install --no-build-isolation --no-deps --disable-pip-version-check -e ."
postinstall = "pip install --no-build-isolation --no-deps --disable-pip-version-check ."
dev = "uvicorn my_app.app:main --reload"
serve = "uvicorn my_app.app:main"

[dependencies]
python = ">=3.12"
pip = "*"
pydantic = ">=2"
fastapi = ">=0.105.0"
sqlalchemy = ">=2,<3"
uvicorn = "*"
aiofiles = "*"

[feature.test.dependencies]
pytest = "*"
pytest-md = "*"
pytest-asyncio = "*"
[feature.test.tasks]
test = "pytest --md=report.md"

[environments]
# both default and prod will have exactly the same dependency versions when they share a dependency
default = {features = ["test"], solve-group = "prod-group"}
prod = {features = [], solve-group = "prod-group"}

In ci you would run the following commands:

pixi run postinstall-e && pixi run test

Locally you would run the following command:

pixi run postinstall-e && pixi run dev

Then in a Dockerfile you would run the following command:

Dockerfile

FROM ghcr.io/prefix-dev/pixi:latest # this doesn't exist yet
WORKDIR /app
COPY . .
RUN pixi run --environment prod postinstall
EXPOSE 8080
CMD ["/usr/local/bin/pixi", "run", "--environment", "prod", "serve"]

Multiple machines from one project

This is an example for an ML project that should be executable on a machine that supports cuda and mlx. It should also be executable on machines that don't support cuda or mlx, we use the cpu feature for this.

pixi.toml

[project]
name = "my-ml-project"
description = "A project that does ML stuff"
authors = ["Your Name <your.name@gmail.com>"]
channels = ["conda-forge", "pytorch"]
# All platforms that are supported by the project as the features will take the intersection of the platforms defined there.
platforms = ["win-64", "linux-64", "osx-64", "osx-arm64"]

[tasks]
train-model = "python train.py"
evaluate-model = "python test.py"

[dependencies]
python = "3.11.*"
pytorch = {version = ">=2.0.1", channel = "pytorch"}
torchvision = {version = ">=0.15", channel = "pytorch"}
polars = ">=0.20,<0.21"
matplotlib-base = ">=3.8.2,<3.9"
ipykernel = ">=6.28.0,<6.29"

[feature.cuda]
platforms = ["win-64", "linux-64"]
channels = ["nvidia", {channel = "pytorch", priority = -1}]
system-requirements = {cuda = "12.1"}

[feature.cuda.tasks]
train-model = "python train.py --cuda"
evaluate-model = "python test.py --cuda"

[feature.cuda.dependencies]
pytorch-cuda = {version = "12.1.*", channel = "pytorch"}

[feature.mlx]
platforms = ["osx-arm64"]

[feature.mlx.tasks]
train-model = "python train.py --mlx"
evaluate-model = "python test.py --mlx"

[feature.mlx.dependencies]
mlx = ">=0.5.0,<0.6.0"

[feature.cpu]
platforms = ["win-64", "linux-64", "osx-64", "osx-arm64"]

[environments]
cuda = ["cuda"]
mlx = ["mlx"]
default = ["cpu"]

Running the project on a cuda machine

pixi run train-model --environment cuda
# will execute `python train.py --cuda`
# fails if not on linux-64 or win-64 with cuda 12.1

Running the project with mlx

pixi run train-model --environment mlx
# will execute `python train.py --mlx`
# fails if not on osx-arm64

Running the project on a machine without cuda or mlx

pixi run train-model