Support `MLFlowLogger` with only the `mlflow-skinny` dependency installed

[dconathan]

Bug description

Sorry not sure to label this as a "feature" or a "bug":

mlflow-skinny is a variant of the mlflow library that installs the bare-minimum required for tracking: https://pypi.org/project/mlflow-skinny/

As far as I can tell, this meets the requirements for using the MLFlowLogger, but currently it doesn't work because:

https://github.com/Lightning-AI/lightning/blob/4a802e00a843264afd5aea023d8475d4f1b4d360/src/pytorch_lightning/loggers/mlflow.py#L40

checks that precisely the mlflow library is installed. If this line were replaced with:

_MLFLOW_AVAILABLE = RequirementCache("mlflow>=1.0.0") or RequirementCache("mlflow-skinny>=1.0.0")

It seems to work fine.

Should I make a PR for this? Would we want to add separate tests for mlflow and mlflow-skinny environments?

How to reproduce the bug

No response

Error messages and logs

No response

Environment

No response

More info

No response

cc @Borda

[carmocca]

Sounds good to me. You will probably need separate variables for each check in case some features are not supported with the skinny version

[adeandrade]

Why was this merged on 2.0 if it was broken on 1.9?

[carmocca]

@adeandrade We only include bug-fixes in the 1.9.x releases. And this wasn't considered one because it's adding support for this alternative MLFlow package. Does our Logger not work with mlflow installed?

[adeandrade]

@carmocca pytorch-lightning was working with mlflow-skinny on 1.8 and it broke on 1.9 but I guess mlflow-skinny was not considered before. It's OK. I'll stick to 1.8 until 2.0 is released.