Potential solutions

So far, we have outlined why these packaging ecosystems do not always work together well and a couple strategies users have used in the past to overcome them. How exactly does the conda-pypi plugin plan on addressing them? Below are a couple of methods we’ve discussed to address these issues.

Native wheel installation from a conda channel

When a user configures the conda-pypi channel and uses the Rattler solver, conda can install pure Python wheels directly without having to convert them to .conda format first.

In this approach, wheel metadata is published directly in a conda channel’s repodata.json under a v3.whl section. Each wheel appears as a metadata record alongside regular conda packages, so the solver sees both package types as a single dependency graph. When the solver selects a wheel, conda downloads the artifact from PyPI and unpacks it into the environment at install time using conda’s package extraction hook. This means that no local conversion to the .conda format is needed.

For setup instructions for the conda-pypi channel, see the Quickstart.

On-the-fly conversion of PyPI wheels to conda packages

The inspiration for this approach initially started with the conda-pupa project. The philosophy used here is that we can simply convert a wheel from PyPI into a conda package and cache it on the host locally. In conda, it’s straightforward to configure multiple channels to be used when installing packages, and by default, a “local” channel is included. As conda-pypi is run, it will begin transforming and caching wheels from PyPI into the conda packages which are then saved in this local channel.

This is the approach we currently feel most confident with implementing.

Analyze the dependency tree of your PyPI package

In this approach, we run pip with the --dry-run option and analyze the proposed solution. Of those packages, we see which ones are already available on the configured conda channels and install them with conda proper. For the ones that are not available, we pass them to pip install --no-deps and hope for an ABI compatible setting.

This was an approach we initially tried but then abandoned in favor of the “conda-pupa” approach.