M9-001M9: Senior ML & AI System DesignRecSys, Ads & Search ArchitectureHard
⚡Flashcard Drill➔Mastery:
Design a Billion-Scale Four-Stage Feed Recommendation Architecture
🎯 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
Balance candidate scale and compute complexity across four stages: Dual-Tower ANN recall, distilled pre-ranker, MMoE fine ranker, and DPP diversity re-ranking within 100ms P99.
📌 Key Takeaways
- •L1 SLA: 100M pool within 100ms P99 budget across feature hydration & 4 stages
- •L2 Distillation: align coarse-ranker logits with fine-ranker Top-K via pairwise margin loss
- •L2 Decoupled Scoring: encode User history once per request and broadcast across 300 Items
- •L3 Ecological Health: DPP diversity scatter, exploration quota & server-side Bloom seen dedup
📐 Mathematical Derivations
Combine multi-task targets multiplicatively and distill fine-ranker rankings into coarse-ranker margin losses to preserve funnel consistency.
🏭 Production Trade-offs
1) Split User vs Item towers to avoid recomputing user sequence attention 300x; 2) Degrade cleanly to coarse-rank cache on P99 spike.
⚠️ Common Interview Pitfalls
- ✕Omitting in-batch popularity debiasing collapses dual-tower recall
- ✕Pointwise ranking ignores listwise diversity fatigue
- ✕Client-only seen history breaks under retries
🎯 Interviewer Follow-up Inquiries
- ?How do you shard ANN beyond single-node RAM limits at 10B scale?
- ?How do you confirm short-term click weight bumps don't hurt 14-day retention?
⚡
Test muscle memory on this topic via 3-Layer Flashcards
Active Recall: Mentally formulate Prompt ➔ Math Derivation ➔ Benchmark Implementation in 3 minutes.
📚