Apptainer/Singularity equivalent to MAMBA_DOCKERFILE_ACTIVATE for automatic env activation

[nick-youngblut]

Background & motivation

When we build Docker images from mambaorg/micromamba, we can ensure that the Conda environment is active inside every RUN layer simply by adding

ARG MAMBA_DOCKERFILE_ACTIVATE=1

In Apptainer/Singularity-based workflows, the situation is different:

• apptainer run does activate the base environment because it executes the container’s entrypoint.
• apptainer exec / apptainer shell do not activate the environment unless users remember to wrap their command with /usr/local/bin/_entrypoint.sh or pass --shell /usr/local/bin/_apptainer_shell.sh. 

For multi-stage pipelines (e.g. Nextflow) that support Conda, Docker and Apptainer, it would be very convenient to have a single Dockerfile that “just works” everywhere. Right now the lack of an Apptainer analogue to MAMBA_DOCKERFILE_ACTIVATE means the image behaves differently depending on the runtime, and tooling that shells into the container (CI jobs, apptainer exec … python ..., etc.) fails because the environment isn’t on PATH. 

Feature idea / questions

• Would it be feasible for micromamba-docker to honor something like MAMBA_APPTAINER_ACTIVATE=1 that injects the same activation shim Apptainer-side during build time?
• Alternatively, could the existing _apptainer_shell.sh script be invoked automatically when apptainer exec detects the container was built from a micromamba image?
• I’m happy to help prototype or test any proposed approach.

Minimal reproduction

FROM mambaorg/micromamba:2.1.1

COPY env.yaml /tmp/env.yaml
RUN micromamba install -y -n base -f /tmp/env.yaml && \
    micromamba clean -a -y

# Works for Docker RUN layers:
ARG MAMBA_DOCKERFILE_ACTIVATE=1
RUN python -c "import uuid, sys; print('Docker:', uuid.uuid4())"

# But after converting to Apptainer:
#   apptainer exec my_image.sif python -c "print('hello')"
# fails with `python: command not found`

[wholtz]

Hi @nick-youngblut - thanks for the feedback. We are aware that Apptainer/Singularity support isn't great.

In my previous job I was doing pipeline development in an HPC environment that used Apptainer. I spent a significant amount of time investigating better ways of supporting Apptainer/Singularity, some of which you can see in old issues (#155). I'm not an expert on Apptainer internals, so there very well could be an approach I missed. I'm open to specific implementation ideas on how to better support Apptainer, but at this point someone else is going to have to provide some new insight before I can personally move this foward.