These are the highest-frequency questions.
| Trade‑off | What to Say | |-----------|--------------| | Batch vs. real‑time inference | Batch for offline reports, recommendations precomputed nightly. Real‑time for fraud, ads (sub‑50ms). | | Model complexity vs. latency | LightGBM / distilled BERT for low latency. Ensemble for accuracy (but slower). | | Online learning vs. retraining | Online (FTRL, KF) for fast changing data. Retrain daily if patterns shift weekly. | | Feature store | Centralized feature serving (Feast, Tecton) reduces training‑serving skew. | | Embedding based retrieval | ANN (Faiss, ScaNN) vs. brute‑force. Recall‑latency balance. | machine learning system design interview pdf alex xu
A signature of Alex Xu’s style is the heavy reliance on architectural diagrams. The PDF is packed with visuals that are interview-ready. These are the highest-frequency questions
❌ Jumping to a deep neural network without a baseline.
❌ Forgetting to mention data labeling cost and label source (implicit vs. explicit feedback).
❌ Ignoring training‑serving skew (features available offline but not online).
❌ Not discussing how to handle cold start (new user/item).
❌ Missing model freshness strategy (retraining schedule, online learning). A signature of Alex Xu’s style is the