ORDERLITE: REAL-TIME CONSISTENCY IN DISTRIBUTED CACHING AND DATABASES WITH LIGHTWEIGHT ML PREDICTION

Abstract Distributed database and cache systems struggle tremendously in ensuring real-time consistency during high write throughput, particularly under asynchronous replication. Traditional approaches sacrifice eventual consistency with a 2–5 seconds delay or place a strong consistency requirement with 5–15 milliseconds latency, neither of which are sufficient for sub-millisecond demand in applications like financial trading, gaming leaderboards, and e-commerce transaction logs. This paper introduces OrderLite, a novel framework based on a lightweight decision tree regressor for predicting replication latency to enable sub-millisecond global ordering across distributed nodes. Deployed on a single MacBook Pro (8-core M1 Max, 32GB RAM) with Python, FastAPI, and RabbitMQ, OrderLite achieves an average latency of 0.58 milliseconds and 99.1% order correctness at 120,000 writes per second. Compared to baselines such as asynchronous replication and fixed-delay synchronization, it demonstrates 42% lower latency and 22% higher consistency, demonstrating a scalable and effective solution. These results, verified through extensive local simulation, position OrderLite as an innovative contribution to real-time distributed systems research and worthy of CIT recognition.