This should fall under the broader data pipeline epic and will require grooming.

We have the concept of a dataset_id, that maps to runs of a training pipeline. We should polish our approach and keep track of

1. dataset variants?
2. deployment history