Real-Time Features & Serving — Week 2 CheckWeek 2 of Data Engineering for ML & AI: online beats batch when milliseconds are money — serving architectures, caching, freshness.

Whatnot found online inference 6× faster/better than batch pre-computation for feed ranking. When does ONLINE win structurally?

Feature freshness for a fraud model is a spectrum with a price at every point. The right freshness is decided by:

An online store (Redis/Cassandra) AND an offline store (warehouse/lake) hold 'the same' features. The consistency requirement between them:

Agoda rewrote its feature store in Rust for serving. What serving property justifies that kind of investment?

A feature pipeline lags 40 minutes behind (the kafka-lag scenario, ML edition). For the fraud model consuming it, the failure mode is:

Embedding/vector pipelines add one data-engineering twist over classic features: