Building ML-based workflows in the real world is a trial-and-error, iterative process where an ML developer builds tens to hundreds of workflows before arriving at one that meets some task-specific acceptance criteria. This iterative process of workflow building is laborious for several reasons including the large variety of available ML models, the time required to train the workflow, difficulty keeping track of workflows built during the modeling process, and the time required for debugging trained workflows. In this thesis, we are primarily interested in two problems with the repetitive modeling process: first, how to manage ML-based workflows generated over multiple iterations of the modeling process, and second, how to efficiently debug or diagnose trained ML-based workflows. In this work, we study these questions from a systems perspective and propose novel software systems and techniques to address them. Specifically, our contributions are: 1. We propose MODELDB, a system to track provenance and performance of ML-based workflows. 2. We propose MISTIQUE, a system to store ML-based workflow intermediates in order to speed up model debugging tasks, and 3. We provide examples of new diagnostic techniques that can be designed using the data in MISTIQUE.