What is online feature computation vs precomputed feature architecture in classification systems?

Online feature computation calculates features at request time, while precomputed features are computed in advance and stored for fast retrieval.

Classification systems must balance latency, freshness, and cost when designing feature pipelines. Online feature computation ensures up-to-date data but increases inference latency. Precomputed features reduce latency by storing results in a feature store or cache but may become stale. Large-scale systems often use a hybrid approach: frequently changing features computed online, and stable aggregations precomputed in batch pipelines. This separation is critical for maintaining consistency and meeting SLA requirements in production ML systems.