M9-006M9: Senior ML & AI System DesignRisk, Pricing & Large-Scale RetrievalHard
⚡Flashcard Drill➔Mastery:
Design a Sub-10ms Real-Time Payment Fraud Detection System
🎯 Active Recall Self-Check (Draft First, Verify After)
Login-free practice tracker: recall core answer, audit pitfalls, and log review status.
Step 1 · Draft your concise answer & assumptions
Step 2 · Check Pitfalls & Follow-up Inquiries
📐 Mathematical Definition
⚡ One-Sentence Core Answer
Meet <=10ms P99 SLAs by pushing multi-hop GNN risk into async Redis updates while running 1ms rules, sliding Flink windows, and in-process ONNX GBDTs synchronously.
📌 Key Takeaways
- •L1 Strict Budget: <=10ms P99; false negatives incur chargebacks, false positives block good customers
- •L2 Flink Windows: maintain real-time 1m/1h/24h velocity & impossible-travel flags in Redis
- •L2 In-Process Inference: compile ONNX GBDTs in C++/Rust to eliminate RPC hop jitter
- •L3 Stepped Friction: Approve vs 3DS Challenge vs Block, retrained via mature point-in-time chargebacks
📐 Mathematical Derivations
Keep deep multi-hop graph neural nets off the 10ms critical path; materialize neighborhood risk scores async into Redis O(1) lookups.
🏭 Production Trade-offs
Combine fast human-reviewed fraud alerts with 60-day mature confirmed non-fraud transactions to prevent label-delay drift.
⚠️ Common Interview Pitfalls
- ✕Labeling recent uncleared transactions as clean before chargebacks arrive
- ✕Failing closed when feature caches blip, blocking legitimate checkout revenue
🎯 Interviewer Follow-up Inquiries
- ?How do emergency rule hotfixes defend against overnight proxy-pool rotations cleanly?
- ?How do you prevent future graph leakage during offline backtests of ring features?
⚡
Test muscle memory on this topic via 3-Layer Flashcards
Active Recall: Mentally formulate Prompt ➔ Math Derivation ➔ Benchmark Implementation in 3 minutes.
📚