PayPal ML System Design: Real-Time Fraud Detection Engine

Design a real-time fraud detection system that inspects millions of e-commerce or financial transactions and responds within stringent constraints. You’ll be asked to design an end-to-end streaming ML pipeline that ingests transaction events (user ID, amount, device, location, history), computes features, scores risk, and enforces decisions (allow, flag, block) with minimal user friction.

Start by describing the high-level flow: event ingestion (Kafka/Kinesis), lightweight rule-based pre-checks, stateful stream processing for feature computation (Flink/Spark Streaming), a low-latency feature cache (Redis/FAISS), model scoring via a serving layer (TF-Serving/Triton), and a decision engine with configurable thresholds and manual-review hooks. Include offline components: feature store, batch training, labeling pipelines, automated retraining, and canary deployments to rollout models without downtime.

You must show how you meet non-functional requirements: keep total processing latency under 100 ms through compact features and local caches; scale horizontally to 10k TPS with partitioning and autoscaling; maintain 99.99% availability with redundant services and graceful degradation (fall back to rule-based checks); and reduce false positives below ~1% by combining supervised models, ensemble scoring, and human-in-the-loop review for borderline cases.

Signal the skills you’ll demonstrate: streaming architectures, feature engineering for velocity and behavioral features, online/offline model lifecycle, latency/throughput trade-offs, fault tolerance and observability (SLOs, logging, audit trails), concept-drift handling, and compliance-aware logging for audits. Use concrete choices and trade-offs rather than abstract descriptions to show practical engineering judgement.