Metric-based evaluation of adaptive consistency benefits in distributed database management systems

Abstract

The article presents a simulation-based methodology for testing and comparing data-consistency models in replicated distributed database systems. Such evaluation is important for modern cloud and edge deployments, where the selected consistency policy directly affects response time, coordination overhead, and the probability of stale reads under variable network conditions and skewed workloads. It also addresses the lack of reproducible test procedures for identifying when adaptive strategies are preferable to static consistency settings. In contrast to studies that focus primarily on the conceptual properties of strong, eventual, or adaptive consistency, the proposed paper emphasises the experimental side of the comparison: workload construction, metric selection, scenario design, and interpretation of the obtained results. While the evaluation includes strong and eventual consistency as reference regimes, the central emphasis is the practical benefit profile of adaptive consistency: under what conditions it preserves part of the correctness of strong coordination without inheriting its full performance cost. A modular discrete-event simulator implemented in Python/SimPy is used to model replicated nodes, stochastic communication delays, asynchronous propagation, and Zipfian access patterns. The evaluation protocol combines clientobserved read and write latency, consistency violation rate, completed operations per run, and the average number of synchronised replicas per write, while Monte Carlo repetitions are used to reduce the influence of a single stochastic trajectory. The reported results show that adaptive consistency delivers its clearest advantage in hotspot, read-dominant, and moderate-latency conditions, where it remains much closer to the strong regime in correctness while retaining part of the throughput and write-latency advantages of eventual replication. At the same time, the highlatency scenario exposes a boundary case in which the adaptive policy degrades toward neareventual behaviour. The obtained findings are useful both for evaluating adaptive-consistency mechanisms and for constructing future control policies.

У статті запропоновано методологію моделювання для тестування й порівняння моделей узгодженості даних у розподілених СКБД з реплікацією. Основну увагу зосереджено на експериментальній оцінці переваг адаптивної узгодженості порівняно зі сильною та кінцевою за різних навантажень і мережевих умов. Для дослідження використано модульний дискретно-подійне симулятор на основі Python та SimPy. Оцінювання охоплює затримки читання і запису, частоту порушень узгодженості, пропускну здатність і кількість синхронізованих реплік. Результати показують, що адаптивна узгодженість є найефективнішою за помірних затримок, домінування читань і hotspot-сценаріїв.