You already know: feature engineering and target leakage (D3 Feature engineering, ML evaluation); training-serving skew (D10 bridge); partitions, incremental runs and streaming windows (ML pipelines); indexes and key-value lookups (D1 Databases, Data structures).
Why this matters for Atlas. Features are where most production ML bugs live, because the same feature is needed in two very different places: training reads millions of historical rows and must see each one exactly as it was at the time; serving reads one entity's latest values in milliseconds. Built separately, those two paths drift apart (skew), and historical joins quietly leak the future (inflated offline metrics). A feature store is the infrastructure that makes one feature definition serve both paths correctly. This lesson builds its core: point-in-time joins, offline and online stores, freshness, and consistency.
1. Leakage through time: the as-of join
The question. How can a model look great offline and disappoint online when the training code has no obvious bug?
Start from nothing. A training example says: "on date t, given what we knew then, did the student quit in the next 14 days?". Every feature must be computed as of t: using only data with timestamps ≤ t. A naive join attaches each entity's latest feature value instead, which includes data from after t, sometimes including the outcome itself (a student who quit has near-zero recent minutes). That's time travel leakage. The fix is a point-in-time (as-of) join: for each label timestamp, take the most recent feature value at or before it.