I'm having troubles understanding the general workflow when using metaflow. Say I need to clean and extract features from 3 different image datasets, each with their own peculiarities. The ML team expects a outputs in a standardized format, regardless of the source dataset. I'd go ahead and write my pipelines using flowspecs, then get that running on our compute. How then: • Are ML researchers supposed to access the pipeline results? Do they use the metaflow client API to collect runs of interest, and then obtain a local copy of the outputs stored on S3? • Do we trace data lineage? E.g. if image fail to be parsed because of a header issue, how can get a sense of how much data was spilled from my dataset due to the same problem? How does a ML researcher who gets a vector of nan as features trace it back to the step that caused the issue, when the whole pipeline may involve 20 different steps, some batched, some at a sample level? The UI seems limited to showing the details of compute for a run. • Do we do any form of data versioning? If I figure out a way to fix this header issue which affects 10% of our dataset, I'd like to be able to process these images again while still holding the results from the previous 90% to avoid wasting compute. Our ML researchers should then be able to verify that the version number for these bad samples has increased

good questions!