In modern financial payments, every transaction must be processed exactly once and leave behind an unquestionable audit trail, yet the shift toward fully serverless, event-driven architectures on AWS introduces difficult trade-offs between raw performance and demonstrable compliance. This paper comparatively evaluates two leading AWS serverless composition patterns for high-frequency payment settlement: orchestration via AWS Step Functions Express Workflows and choreography via Amazon EventBridge combined with idempotent AWS Lambda functions. Using a purpose-built payment processing workload that simulates realistic transaction bursts, we measure end-to-end latency, throughput under varying concurrency, and the exactly-once consistency achieved through idempotency and deduplication mechanisms. Crucially, we introduce an Evidential Integrity Score that captures the ease with which a third-party auditor can reconstruct the causal chain of each settled payment from the logs, state transitions, and replay records natively produced by each pattern. Our findings reveal that the orchestration model offers straighter, more readily auditable execution histories but experiences tail-latency spikes during abrupt load surges, whereas the choreography model delivers higher sustained throughput and graceful degradation under stress, yet requires significantly greater effort to reassemble a complete, time-ordered audit narrative. We also identify specific failure-recovery scenarios where choreography’s apparent exactly-once guarantee temporarily degrades into at-least-once delivery unless carefully tuned deduplication windows are enforced. Based on these empirical insights, we provide concrete design guidelines for selecting and configuring serverless payment pipelines that simultaneously satisfy sub-100-millisecond processing targets and the evidential integrity standards mandated by financial regulators. The results help architects navigate the inherent tension between serverless agility and the fixed, provable correctness that payment infrastructures demand.
Satyanarayana Gopisetty (Mon,) studied this question.