Study of asynchronous transaction buffer in distributed banking system
Abstract
Report of the The C3AI 2026 conference Abstract. This paper examines the behavior of a distributed asynchronoustransaction processing system integrating a high-load mobilebanking API with a legacy core banking system (CBS) with fixed throughputand periods of complete unavailability. The Distributed TransactionalBuffer (DTB) architecture is considered, including an API service,relational storage, a Transactional Outbox, an Apache Kafka broker, anda consumer executing synchronous calls to the CBS. Unlike synchronousintegration via SOAP, the proposed architecture decouples the acceptanceof a client request from the actual execution of the transaction inthe downstream system.The objective of the study is to evaluate three properties of this architecture:backlog accumulation when the incoming flow exceeds thethroughput of the ABS, the time it takes to resolve the queue after processingis resumed, and the bounded consistency lag between commandacceptance and its actual execution. A discrete queue model is used forbacklog analysis, and the experimental part is a comparative analysis ofthe baseline mode of direct synchronous integration and the DTB modeof an asynchronous buffer. In a production scenario, the transition toDTB was accompanied by an increase in the maximum API throughputfrom 45 to 1850 TPS (Transactions Per Second), a decrease in theaverage response time from 2450 to 35 ms, and a decrease in the P99(99th percentile) latency from 8500 to 72 ms. The system maintainedavailability of the request reception layer during the 20:00–08:00 window,and the accumulated backlog, which reached 42,900 messages, wascompletely resolved by 11:00 after the opening of the trading day.The results demonstrate that DTB should be considered not only as anintegration pattern, but also as a mechanism for controlled, temporaryload balancing in a distributed system. With this approach, fault toleranceand user availability are improved at the expense of bounded consistencylag,
// Source
Authors: Yegor Sychev, Eugene Alooeff