On the Trade-off Between Robustness and Complexity in Data Pipelines