Proposal: use MLflow Tracking in papermill

[mabayona]

MLflow Tracking (https://www.mlflow.org/docs/latest/tracking.html) provides mature and easy to use library for tracking runs and experiments. Somehow this functionality is similar to some tracking functionality provided by Papermill.

The difference being that MLflow is broader in scope and provides additional features (e.g. a Tracking Server) that could be of interest for advancing Papermill functionality.

[betatim]

I think notebook authors can already use mlflow to track their parameters and models when they are executed with papermill. Do you think the papermill Python library should also provide access to the values tracked via mlflow? In what sense is mlflow broader? It was designed to track parameters and models for machine-learning runs where as papermill is about running arbitrary notebooks (as scripts).

As far as I know mlflow records its values on disk or via a tracking server. The disk based option is cool because you can easily access them with any kind of tool. For notebooks executed as jobs you'd want to record values in the notebook itself (as papermill.record does) or send the values to a tracking server. It feels like you can already do either/both with papermill as it stands now, because it doesn't prevent you from using mlflow -> I need some help understanding the proposal.