Batch vs Stream & Latency Budgets

Latency budgets, not buzzwords

'Real-time' is meaningless without a number. A streaming system is characterised by:

• End-to-end latency — produce → consume committed (target: ms to seconds).
• Processing semantics — at-most-once, at-least-once, exactly-once.
• Backpressure handling — what happens when consumers slow down?
• Replayability — can a new consumer replay 30 days of history without resharding?

CAP & PACELC remind us: under partition you choose Availability vs Consistency; even in steady state you trade Latency vs Consistency. Kafka leans CP (with acks=all) but lets you opt back into AP by tuning replication and acks.