feature: uv Workspace Support

[nickthegroot]

Feature request

https://docs.astral.sh/uv/concepts/projects/workspaces/

uv allows projects to define a workspace and reference local packages in the workspace as dependencies.
For example:

[project]
# ...snip
dependencies = ["bird-feeder", "tqdm>=4,<5"]

[tool.uv.sources]
bird-feeder = { workspace = true }

Using such workspace dependencies in a bentoml project thus excludes it from using pyproject.toml based configurations.

╰─ bentoml build .
...snip
INFO: Locking PyPI package versions.
INFO: Locking packages for x86_64-manylinux_2_36. Pass `--platform` option to specify the platform.
  × No solution found when resolving dependencies:
  ╰─▶ Because bird-feeder was not found in the package registry and you require bird-feeder, we can conclude that your requirements are unsatisfiable.

Ideally, bentoml build could capture the uv workspaces and correctly add them as dependencies. The docs seem to suggest uv is already used under the hood for locking.

Motivation

Currently the only workaround is to use alternative build systems (e.g. building into wheels, then using a bentofile.yaml). Adding this would massively simplify the build process for those who utilize uv workspaces.

Other

Happy to contribute if help is desired, although can't make any promises about schedule.

[rehan243]

Thanks for bringing this up! A few general ML production patterns:

• Feature stores: Centralizing feature computation ensures consistency between
  training and serving — this eliminated a whole class of bugs for us
• Experiment tracking: Logging everything (hyperparams, data versions, metrics)
  makes reproducing results straightforward months later
• A/B testing: We never deploy a new model without running a proper A/B test.
  Statistical significance matters — we gate on p-value < 0.05

The most impactful thing we did was treating ML development like software engineering —
version control, CI/CD, code review, and automated testing.