Stabilizing distributed OLTP under high contention via measurement-driven adaptive concurrency control

Abstract Distributed online transaction processing (OLTP) systems with high contending workloads can be extremely unstable, with negative impacts such as abort cascades, lock contention, and retry amplifications. Traditional schemes such as two-phase locking (2PL), optimistic concurrency control (OCC) and multi-version concurrency control (MVCC) rely on a fixed policy and are not flexible enough to cater for dynamic contention patterns, resulting in reduced throughput and high latency skew. In the present paper, an Adaptive Concurrency Control (ACC) framework called Measurement-Driven Adaptive Concurrency Control (MD-ACC) is proposed, where the concurrency control is interpreted as a feedback controlled process. The framework provides a thin control layer that is always updated with the metrics at runtime, including the conflict rate, abort rate, queue length and retry intensity. Real-time determination of the contention level can be achieved by using composite contention score, and stable adaptation can be achieved by using a hysteresis-based transition. MD-ACC dynamically applies coordinated policies like wait–abort regulation, adaptive retry backoff, and admission throttling based on observed conditions. The proposed MD-ACC is compared to classical baseline and adaptive/hybrid baseline (such as 2PL, OCC, MVCC, Adaptive OCC, Hybrid OCC/2PL, Dynamic Backoff OCC, CAS, and HACC). Experimental results on a distributed OLTP testbed demonstrate that the proposed approach indeed can provide up to 2.1 × higher throughput, much lower abort rates, with much better execution stability under high contention. These results provide insights into the applicability and deployability of current-state distributed transaction processing systems via measurement-based adaptation.