<!here> @AnselmAdrian recently contributed a new runner workbench component that lets you edit parameters, launch runs, monitor jobs with live logs/history, and give stakeholders visibility into experiments - all directly in Kedro-Viz. Below is GIF of the solution (sorry abt the res, you can see original video on the PR) and the PR where you can explore it in detail (see also [issue #2483] for context). We’d love your thoughts: how useful would this be in your workflows? Do you already solve this today with another tool? Please share feedback either on the PR/issue or here on Slack 🙌

This seems quite nice and handy! I would definitely use this. I think some additional form of integration with mlflow would be great as well so we can access metrics and artifacts logged during the kedro run easily.

@Guillaume Tauzin This is interesting. So is your preference an interface for MLflow on Kedro-Viz?

@Yetunde I was thinking of an extension of what is presented here. I think it's nice to be able to define and launch runs like this, but once the run is done, I would love to be able to access the corresponding mlflow run on the mlflow UI to check that whatever I wanted kedro-mlflow to log has been properly logged. As we have the mlflow run id and tracking uri at runtime, I think it should be possible to have such a link. Maybe an actual interface for mlflow on kedro-viz would be great as well. Somewhere we can access all previously ran jobs, see which nodes were ran and even run them again.

This is great @Guillaume Tauzin! Just to check my understanding, you’re looking for a way to launch new Kedro pipeline runs, log them as experiments in MLflow, and then trigger and review those runs within Kedro-Viz. Did I capture that correctly?

@Yetunde Exactly, review those runs and always have the possibility to review the associated mlflow run by clicking on a link in kedro-viz.

This makes sense. And what changes between your experiments? Is it just parameters, do you change datasets too?

So I personally would not run actual experiments like this. I use an orchestrator when I need to. I would use this whenever I need to design, develop or debug pipelines using a limited set of data. Parameters would change indeed, datasets would get modified/added as I extend the pipelines with more and more nodes. I would typically start setting up the logging of different metrics or artifacts as well until everything works smoothly. At that point I would move to a development execution environment that is orchestrated by a dedicated orchestrator and has access to all my data and compute.