You already know: the system design method and capacity estimation (lessons 01 and 02), data leakage (D3), and streaming pipelines (D16). Almost every case in this bank (fraud, churn, ETA, surge, ads, feeds) depends on features computed from changing data and served in milliseconds. This lesson covers the infrastructure and discipline behind them: why models silently fail when training and serving disagree, how to build point-in-time correct training data, how to choose freshness, how to size an online store, how to guarantee and monitor parity, and when a feature store is worth building at all.
How to read this lesson. No code, on purpose. Numbers marked assumed are assumptions you state out loud.
1. Where training–serving skew comes from
Training–serving skew is when the feature values a model sees in production differ from those it was trained on, for the same situation. The model raises no error; it just makes worse predictions.
The staff framing: same name, different number is the quietest bug in ML. It passes code review, unit tests and offline evaluation.