A Formal Model for Feature Store Architecture and Governance Sivaramakrishnan Vaidyanathan

Abstract

ML systems in production have to address many challenges while ensuring consistency between the features in the training and serving phases. Feature Stores have emerged as one of the key ML infrastructure components to bridge the training and serving gaps. There are tradeoffs between different types of FS, such as latency, consistency guarantees, costs, and operational complexity. Organizations often do not have formal governance frameworks for governing Machine Learning pipelines. One example of the issues that can arise from insufficient frameworks is Training-Serving Skew, whereby feature statistics differ between environments. This leads to challenges in ensuring regulatory compliance and the ability to trace the lineage of features for model auditability and reproducibility. This presents a two-part formal model that enables mathematical optimization and structured governance. The first half frames the FS selection process as a constrained optimisation problem so that the performance of dual-database architectures can be quantitatively compared to that of unified architectures based on business priorities. The second half introduces Versioned Feature Descriptors that are canonical metadata artifacts for the permanent storage of feature definitions, complete lineage from raw data to prediction outputs, and fully machine-enforceable compliance policies. The optimization framework models serving latency, consistency gap, capital expense, and operational complexity for dual-database systems (one for online and another for offline workloads) and for unified systems (which house both workloads). The governance model prevents training-serving skew through runtime validation, ensuring that features input to a deployed model come from the desired descriptor version. Privacy and retention requirements are enforced by formal policy predicates, with the review process showing improvements in operational cost, debugging, audit, and regulatory compliance efforts. The framework formalizes Feature Store architecture evaluation, transforming decision-making from heuristic-based to a systematic architecture evaluation approach based on quantitative analysis for scalable and compliant machine learning adoption.