To Design the Adaptive Consistency via Machine-Learned Policy Control for Globally Distributed Databases

Abstract

We have presented the adaptive consistency framework for globally distributed databases that uses a machine-learned policy controller to balance latency, throughput, and correctness under dynamic workloads. This approach has treated consistency as a tunable knob, guided by real-time observability, workload characteristics, and service-level objectives (SLOs). A lightweight supervisor has collected end-to-end latency, read/write latency distribution, and data staleness metrics, and has selected a consistency level (e.g., strong, bounded staleness, or eventual) at the operation granularity or per session. The policy has learned offline from historical traces and updated online via a safe incremental learning loop that avoids destabilizing the system. The objective of this research is the formalisation of adaptive consistency as a policy-optimisation problem with stability guarantees. A learnable controller that integrates latency, staleness, and throughput signals. Practical guidelines for deployment, monitoring, and safety are also provided. We have implemented the framework on top of a representative distributed database prototype and evaluated it under synthetic and real workloads, including flash crowds, skewed key access, and partial network partitions. The results show a reduction of up to 28.6% in tail latency (p95/p99) with controlled staleness deviation, and a 75% improvement in overall throughput under bursty conditions, compared to 20% with static consistency configurations. We have considered the organisational concerns, security requirements, and opportunities for integration with the current Database-as-a-Service (DBaaS) platform.