Should the clients extract the container-image for remote execution?

[Fredrik Medley]

I would like to implement support for the 'container-image' platform property to Buildbarn. The idea is to, at an early stage, patch the input root with the whole content of the container image. We realised that maybe the client (Bazel) should do that instead.

Is it in the scope for the clients to compute the content of the container image or is it up to the server side to handle the 'container-image' platform property? Is either way a good or bad way to go?

In Bazel, one could specify root directory content for each platform definition and a path where the other action files should be located in. The platform root directory would be constructed by either other build rules or by expanding a container image. When running remote-execution, then all the files in the container would be listed in the input root (making the Merkle tree calculation slow unless caching subtrees).

Best regards,

Fredrik