Fraud Screening Microservice
A stateless Java/Spring Boot service where duplicate requests, burst traffic and slow external dependencies had to be treated as normal operating conditions—not edge cases.
Keep the synchronous path small and bounded.
The decision path does the work needed to answer the request; observability and evidence emission stay off the critical path where possible.
The useful part is the constraint behind each mechanism.
Patterns only matter when they protect a concrete invariant.
Duplicate-safe processing
Idempotency checks prevent retries from turning one logical transaction into repeated fraud evaluation side effects.
Invariant: same request identity should not create duplicate work.Burst admission control
Token-bucket rate limiting protects the service and dependencies during short traffic spikes instead of letting queues grow without bound.
Failure mode: overload propagation.Dependency isolation
Circuit breakers, timeouts and fallbacks stop a slow external risk provider from consuming the entire latency budget.
Failure mode: external latency dominates request path.Operational feedback loop
CloudWatch, structured logging, alerting, runbooks, incident triage and RCA make production behavior part of the engineering loop.
Result: 99.8% uptime and MTTR 35 → 12 min.Reliability work is visible in response time, recovery time and incident quality.
The system is strongest when the operational evidence is attached to the architecture rather than presented as a separate “DevOps” layer.