17. Interview Prep Roadmap

The 7-Step Framework

Clarify requirements: Functional? Non-functional? Scale? Constraints? Success metrics?
Estimate scale: DAU, QPS, storage, bandwidth, training data volume.
Outline high-level architecture: Block diagram with 5—8 components.
Deep dive — data pipeline: Ingestion, preprocessing, feature store, training-serving consistency.
Deep dive — model serving: API design, caching, batch vs real-time, model versioning.
Deep dive — trade-offs & reliability: What trade-offs did you make? What breaks? How do you recover?
Monitoring, ethics & improvements: Drift detection, fairness, what would you do with more time/budget?

Common Interview Questions

Question	Key Focus Areas
Design YouTube recommendations	Two-stage retrieval + ranking, user embeddings, cold-start, watch time prediction, real-time feedback
Design a fraud detection system	Real-time inference, feature store, imbalanced data, latency budget, explainability, fallback models
Design a search ranking system	Feature engineering, click-through rate model, position bias, online vs offline evaluation, A/B testing
Design a content moderation pipeline	Multi-modal (text + image), cascading classifiers, human review queue, feedback loop, fairness
Design a personalized news feed	Candidate generation, ranking, real-time features, diversity, A/B testing, push vs pull delivery

Scaling Estimation Cheat Sheet

DAU = MAU × 0.4 (typical mobile app)
Peak QPS = DAU × requests_per_user_per_day / 86400 × 5 (peak factor)
Storage per day = QPS_avg × data_per_request × 86400
Training data size = storage_per_day × retention_days
GPU memory = model_params × bytes_per_param (FP32: 4, FP16: 2)
Throughput per GPU = batch_size / inference_time_per_sample

Must-Know Trade-offs for Interviews

Accuracy vs latency — use distillation, quantization, pruning
Batch vs real-time — use batch for training, real-time for serving
Model complexity vs interpretability — XGBoost is a good middle ground
Freshness vs stability — micro-batch features
Cost vs redundancy — multi-region active-passive
Consistency vs availability — depends on business impact of stale data

How to Present Your Design

Whiteboard: draw boxes and arrows, label technologies.
Walk through one request end-to-end: "User clicks 'refresh' → load balancer → ..."
Deep dive on the most interesting part (feature pipeline, model serving, or caching).
Discuss failure: "If Redis goes down, we fall back to per-user default features."
Summarize trade-offs and next priorities.

Exercise: Practice the 7-step framework on this question: "Design an ML system that personalizes the search results for an e-commerce platform with 100M products." Record yourself (video or audio) answering within 40 minutes. Review against the framework — did you cover all 7 steps? Did you discuss trade-offs? Did you mention failure scenarios?