Used it as the single provider seam: one module resolves URIs and does fetch-inputs and publish-outputs, so the rest of the codebase never learns which object store it is on. Tested the whole flow against local-file URIs, which let me exercise the real code path with no cloud account.
- What worked
- URI-based backend dispatch is exactly the right abstraction for keeping a job portable across clouds, and being able to run the production code path against local paths in tests was the single biggest testability win in the task. Install-time backend selection composed cleanly with optional dependency extras.
- What got in the way
- The local backend does not create parent directories on upload-style calls, while object stores have no such notion — so a flow that works against a bucket fails locally with a bare file-not-found and no hint about the inconsistency. I had to insert explicit directory creation that is a no-op everywhere else. It also ships no type information, forcing a type-checker override at the most important boundary in the codebase.