canonical / mlflow-operator

MLFlow Operators
Apache License 2.0
13 stars 11 forks source link

Enable integrating with an existing MinIO instead of deploying your own every time #259

Open motjuste opened 3 months ago

motjuste commented 3 months ago

Context

Charmed MLFlow deploys its own MinIO to store its data, but it is often the case that MinIO is already available, for example, when integrating with Charmed Kubeflow that also deploys MinIO.

What needs to get done

TBD

Definition of Done

It should be possible to deploy Charmed MLFlow and integrate with an existing pre-deployed MinIO (e.g. done as part of deployed Charmed Kubeflow), and documentation / tutorial for doing that should be available.

syncronize-issues-to-jira[bot] commented 3 months ago

Thank you for reporting us your feedback!

The internal ticket has been created: https://warthogs.atlassian.net/browse/KF-6126.

This message was autogenerated

kimwnasptd commented 3 months ago

Thanks for the issue @motjuste!

My understanding is that the recommendation is to expect users to deploy a standalone MinIO with MLflow, even with Kubeflow. So because of this I wouldn't put this information on the tutorial but might have a dedicated how-to, but let's also hear @afgambin's opinion.

On the technical side though, aside from relating mlflow-server to Kubeflow's MinIO I expect another need for configuration (MLflow should already be related to resource-dispatcher, so a PodDefault will be created by with the credentials of Kubeflow's MinIO)