MLOps 2.0: A Reference Architecture for CI/CD with Always-On Data Quality Gates

Abstract

MLOps 2.0 operationalizes machine learning by elevating data to a first-class
artifact alongside code and models. We present a reference architecture that
converges CI/CD with Continuous Data Validation (CDV), inserting data
quality gates schema, semantic, temporal, and distributional across build, test,
release, and run stages. The pipeline encodes data contracts, enforces SLOaligned
promotion criteria, and couples training/inference observability to
reduce escaped defects (silent drift, schema skew, target leakage) that CI/CD
alone cannot catch. A domain-agnostic evaluation template quantifies impacts
on lead time, rollback rate, stability, and incident frequency. Results indicate
CI/CD +CDV yields more reliable, auditable, and cost efficient ML delivery.