Exactly-Once, Except When It Isn'tKafka · hard · ~14 minThe team shipped a Kafka Streams-style consume-transform-produce service with transactions ('EOS mode') for the ledger pipeline. Two complaints since: the analytics warehouse fed from the OUTPUT topic still shows occasional duplicates, and downstream consumer latency jumped noticeably.Both are explainable — and one of them is by design. Untangle what transactions bought, what they didn't, and what the latency is.Evidence

Evidence

Config

Checkpoints

Why does the warehouse still see duplicates despite EOS?

Where does the added ~120ms of consumer latency come from?

What did the fencing incident demonstrate?

Give the senior summary: what did EOS buy this pipeline, what still needed engineering, and how do you tune the latency tradeoff?