Clarity around the MLFlow logger checkpoint issue

[Northo]

There is a long-standing bug #20664 that was later fixed #20669, and subsequently reverted, which causes an issue with the MLFlow logger, making saving checkpoints fail.

Background

There is a relatively long list of issues and PRs related to this, and there seem to be some confusion on the matter:

• MLFlowLogger fails to log artifact on Windows #20664
• fix: MLFlowLogger artifact_path #20669
• f6ef409
• MlflowException when logging checkpoints with MLFlowLogger #20932
• fix: MLFlowLogger artifact_path #21204
• Add checkpoint artifact path prefix to MLflow logger #20538
• [BUG] MLFlow logger raising invalid artifact path mlflow/mlflow#15111

The solution is currently to pin lightning==2.5.0.

Question

After some digging, I was unable to find why the issue is not fixed, though it seems the proposed fix introduced another bug.

This blocks those using MLFlow from upgrading lightning, and the relatively long list of related issues and PRs makes it confusing for those stumbling upon this issue. It would be great to get some clarity into when this will be fixed, and what is currently the hold-up.

Really love lightning ❤️ and want to keep using it for our research! Thanks for the help!

Ps. Tagging those that seem most involved:
@niander, @yxtay, @Borda, @benglewis

[benglewis]

I'd be happy to contribute/review any fix to this. I am aware that my previous fix though did (unintentionally) introduce bugs of its own, hence why I believe that @Borda reverted it and so I believe that we do need to be very careful with any fix. I believe that the core of that issue was that MLflow does not (yet) support pathlib.Path-type argument for artifact_path. I have an open PR to fix that in MLflow, but it still hasn't been merged.

[yxtay]

#21204

The above PR attempts to fix this and adds tests on the release/stable branch. Will be good to have a reviewer look at it.

Thank you.