Cloud Lakehouse Architectures for Real-Time Feature Stores and Streaming Machine Learning

Production machine learning systems in real-time fraud detection, dynamic pricing, and personalized ranking depend on low-latency access to both historical batch aggregations and sub-second streaming behavioral signals. Traditional Lambda architectures maintain separate codebases and storage engines for offline training (SQL/Spark over data lakes) and online serving (Java/Flink into key-value stores), introducing severe online-offline training-serving skew, future data leakage during historical backfills, and high operational maintenance overhead. This paper presents a unified Cloud Lakehouse Real-Time Feature Store architecture that bridges streaming stateful computation in Apache Flink with transactional Apache Iceberg lakehouse tables and tiered Redis/RocksDB online serving clusters. We formalize point-in-time correctness via bi-temporal ASOF join semantics and watermarks, guaranteeing mathematically zero future target leakage across historical training dataset generation. Using declarative feature definitions compiled simultaneously into Flink DataStream DAGs and Spark/Trino Iceberg compaction jobs, our architecture enforces single-source-of-truth feature lineage while executing exactly-once Chandy-Lamport asynchronous barrier snapshots to cloud object storage. We analyze real-world production deployments at Uber (Michelangelo Palette Feature Store) and DoorDash (Gigascale Real-Time Feature Platform combining Flink, Redis Cluster, and Apache Iceberg). Empirical evaluations at 1.2 million events per second demonstrate that unified Flink-to-Iceberg streaming ingestion with Z-order compaction and Redis Hash-tag pipelining reduces online feature retrieval P99 latency from 42.80 ms to 3.85 ms, cuts feature freshness lag from 45.0 minutes to 0.42 seconds, eliminates 2.4 percentage points of AUC loss caused by training-serving skew, and lowers monthly storage and memory TCO by 48.6% through hybrid RAM-NVMe tiering. Research Domain & Specialization Track Master of Science in Information Technology (MSIT) — Data Engineering, Streaming Systems, and MLOps. Core Technical Highlights Empirical Production Case Studies: Rigorous architectural blueprints and telemetry evaluation across enterprise cloud deployments. Four-Tier Reference Architecture: Formal specification from edge ingress to distributed hardware execution pools. Mathematical Formulation: Formally derived algorithmic and optimization models with closed-loop verification. Multi-Modal Benchmarking: Comparative empirical baselines under fault-injection and synthetic workload scaling. Author Information Author: Sonu Kumar SinghRole: Senior Consultant Affiliation: Independent Researcher Publication Status: Preprint / Working Paper (MSIT AI & Cloud Systems Research Series) Reproducibility & Open Science All included architecture diagrams, telemetry schemas, and benchmark curves are open-access under the Creative Commons Attribution 4.0 International (CC BY 4.0) license.