Building against multiple Docker images

[jakirkham]

Looking at PR ( #1070 ), it seems we assume there should only be one relevant docker_image in each config. Am curious if there is a reason for that assumption? More importantly would it be reasonable to generalize this to allow more than one docker_image?

[jakirkham]

cc @isuruf @mariusvniekerk

[isuruf]

Each config should have one docker_image. If you want to build against multiple docker images, you can have multiple configs.

[jakirkham]

I don't quite follow. Was thinking docker_image could be a list in conda_build_config.yaml and it would try each one. However it appears to be grabbing the last value by sorted order. I'm guessing this will also be a problem when handing this for GPU packages generally ( conda-forge/conda-forge-pinning-feedstock#237 ). Take PR ( conda-forge/ucx-split-feedstock#6 ) for example. Is there a different way I should be doing this or is this not possible without making changes to conda-smithy?

[isuruf]

You need to make docker_image or another variable that is zipped with it be a used variable.

[jakirkham]

Why is that needed?

[isuruf]

That's how conda-smithy works. Just because there are 3 python versions in conda_build_config.yaml, that doesn't mean there are 3 jobs for recipes not using python. You need to use docker_image somewhere in the recipe (or a variable zipped with it).

[jakirkham]

I see. So you just mean it needs to be used in the meta.yaml somehow. Yeah that makes sense. Thanks. 🙂