Successfully building and deploying a machine-learning model can be difficult to do once. Enabling other data scientists (or yourself) to reproduce your pipeline, compare the results of different versions, track what’s running where, and redeploy and rollback updated models, is much harder.