Data Latency (Stack View)

Stack-wise, latency decomposes per hop: event capture → stream/batch transport → transformation → serving layer, each adding delay against each consumer’s tolerance (fraud: minutes; media buyers: near-real-time clicks; finance: settled dailies).

Architecture practice

Design tiered paths rather than one pipeline serving all: streaming lanes for operational triggers (webhooks, fraud signals), batch lanes for conformed analytics — labeled honestly in every UI they feed. Publish freshness SLOs, alert on breach, and resist the “make everything real-time” demand: streaming’s cost and complexity buy value only where decisions actually wait on it. Latency honesty, as with its analytics twin entry, is mostly a communication discipline.