Kappa architecture uses a single streaming pipeline (plus a durable log) as the main processing path. Need a recompute? Replay the log through the same stream job (or a batch job that reuses the same logic), instead of maintaining a separate batch system forever.
Lambda: stream path + batch path (two logics)
Kappa: log (Kafka) --> one processing job --> serving tables
\-- replay for backfill / recompute --/Requirements
- Durable, replayable event log
- Idempotent or versioned sinks so replay is safe
- Enough retention / cold storage of events
Interview tip: "Kappa = unify on the stream/log; Lambda = dual batch+speed." Mention trade-off: replay cost vs dual-code cost.