Observability-Centric SRE Architecture for Cloud-Native Financial Platforms: A Unified Telemetry Framework Using Metrics, Logs, and Distributed Traces

Abstract

Financial transaction platforms operating at scale expose a gap that current monitoring stacks do not close well: metrics, logs, and traces are individually mature signals, but the tooling that handles them has remained siloed [1][2]. On-call engineers responding to a live incident must context-switch across Prometheus/Grafana, the ELK Stack, and Jaeger or Tempo -- each with its own query language, each holding a different piece of the puzzle [3][5]. This paper presents a production-derived architecture that unifies all three signals under a single programmatic query abstraction layer, applies data mesh ownership principles [4] to telemetry governance, and incorporates machine learning-based anomaly detection calibrated to the seasonal periodicity of payment workloads [11][12]. Unlike prior systems that offer UI-level signal correlation or require proprietary vendor lock-in [8][9], the proposed design introduces a vendor-neutral, programmatic cross-signal query interface built entirely on open-source components -- a distinction that makes it both automation-friendly and deployable within regulated financial environments. The system was evaluated over a 90-day production deployment processing 10,000 to 50,000 transactions per second. Results show a 73% reduction in mean time to detection (MTTD), 70% reduction in mean time to resolution (MTTR), false positive rate down from 84.8% to 11.2%, and a 75% decrease in total daily alert volume alongside a 48% increase in actionable alerts [11][12]. Statistical validation confirms F1 = 0.92 (95% CI: [0.90, 0.94]) for the full hybrid detection model. We discuss implementation decisions, failure modes, cardinality management, and the organizational dynamics of adopting data mesh telemetry governance in a regulated environment.