All logic regarding the decision which dependencies can be installed from which channel is done by the instruction we give the solver.

The actual code regarding this is in the rattler_solve crate. This might however be hard to read. Therefore, this document will continue with simplified flow charts.

Channel specific dependencies

When a user defines a channel per dependency, the solver needs to know the other channels are unusable for this dependency.

[project]
channels = ["conda-forge", "my-channel"]

[dependencies]
packgex = { version = "*", channel = "my-channel" }

In the packagex example, the solver will understand that the package is only available in my-channel and will not look for it in conda-forge.

The flowchart of the logic that excludes all other channels:

flowchart TD
    A[Start] --> B[Given a Dependency]
    B --> C{Channel Specific Dependency?}
    C -->|Yes| D[Exclude All Other Channels for This Package]
    C -->|No| E{Any Other Dependencies?}
    E -->|Yes| B
    E -->|No| F[End]
    D --> E

Channel priority

Channel priority is dictated by the order in the project.channels array, where the first channel is the highest priority. For instance:

[project]
channels = ["conda-forge", "my-channel", "your-channel"]

If the package is found in conda-forge the solver will not look for it in my-channel and your-channel, because it tells the solver they are excluded. If the package is not found in conda-forge the solver will look for it in my-channel and if it is found there it will tell the solver to exclude your-channel for this package. This diagram explains the logic:

flowchart TD
    A[Start] --> B[Given a Dependency]
    B --> C{Loop Over Channels}
    C --> D{Package in This Channel?}
    D -->|No| C
    D -->|Yes| E{"This the first channel
     for this package?"}
    E -->|Yes| F[Include Package in Candidates]
    E -->|No| G[Exclude Package from Candidates]
    F --> H{Any Other Channels?}
    G --> H
    H -->|Yes| C
    H -->|No| I{Any Other Dependencies?}
    I -->|No| J[End]
    I -->|Yes| B

This method ensures the solver only adds a package to the candidates if it's found in the highest priority channel available. If you have 10 channels and the package is found in the 5th channel it will exclude the next 5 channels from the candidates if they also contain the package.

Use case: pytorch and nvidia with conda-forge

A common use case is to use pytorch with nvidia drivers, while also needing the conda-forge channel for the main dependencies.

[project]
channels = ["nvidia/label/cuda-11.8.0", "nvidia", "conda-forge", "pytorch"]
platforms = ["linux-64"]

[dependencies]
cuda = {version = "*", channel="nvidia/label/cuda-11.8.0"}
pytorch = {version = "2.0.1.*", channel="pytorch"}
torchvision = {version = "0.15.2.*", channel="pytorch"}
pytorch-cuda = {version = "11.8.*", channel="pytorch"}
python = "3.10.*"

What this will do is get as much as possible from the nvidia/label/cuda-11.8.0 channel, which is actually only the cuda package.

Then it will get all packages from the nvidia channel, which is a little more and some packages overlap the nvidia and conda-forge channel. Like the cuda-cudart package, which will now only be retrieved from the nvidia channel because of the priority logic.

Then it will get the packages from the conda-forge channel, which is the main channel for the dependencies.

But the user only wants the pytorch packages from the pytorch channel, which is why pytorch is added last and the dependencies are added as channel specific dependencies.

We don't define the pytorch channel before conda-forge because we want to get as much as possible from the conda-forge as the pytorch channel is not always shipping the best versions of all packages.

For example, it also ships the ffmpeg package, but only an old version which doesn't work with the newer pytorch versions. Thus breaking the installation if we would skip the conda-forge channel for ffmpeg with the priority logic.