A method for optimizing query routing in distributed databases to reduce latency and load

Abstract

This paper develops a query routing method for distributed databases that reduces network latency and achieves uniform load distribution across cluster nodes through a statistically grounded, autonomous adaptive tuning of scoring function weights.
A multi-factor scoring model is proposed for target-node selection, incorporating CPU load, round-trip network latency, and topological data distance. Unlike existing adaptive routing approaches that rely on heuristic or static weight assignment, the proposed method determines weights through a statistically grounded procedure based on Pearson correlation analysis between each factor and observed query response times within a sliding window, smoothed by exponential moving average (EMA). This design ensures invariance to workload type without administrator intervention.
Simulation on a five-node cluster demonstrates a 38.4% reduction in mean query latency, a 44.1% reduction in P95 latency, a 41.2% increase in throughput, and a 29.7% reduction in peak per-node CPU utilization compared to random routing. Load standard deviation across nodes decreases by a factor of 6.7.
For the first time, a weight-adaptation mechanism is proposed in which adaptation is a function of execution statistics rather than a rule set, providing theoretically grounded behavior under varying workloads. The method addresses a gap left by latency-aware, least-loaded, and geo-distributed routing, none of which jointly optimize resource, network, and topological factors adaptively.
Deployable as a middleware layer without modifying application logic. Uniform load distribution lowers peak server energy consumption, contributing to the carbon footprint reduction of data-center infrastructure.
Complexity is O(k) per routing decision; clusters exceeding 100 nodes require hierarchical adaptation. The independence assumption between factors is a known limitation addressed in future work.