Optimizing Data Pipelines for Machine Learning in Feature Stores