Kafka Streams Fault Tolerance and Exactly-Once Processing

An order-reporting application reads an event, updates a store's order count, and publishes the new total. Then its process crashes. When another instance takes over, it must decide where to resume and which count to restore.

If those decisions disagree, the recovered application can count an order twice or miss it entirely. Keeping the input records available is necessary, but recovery also needs a consistent relationship between processing progress, state, and output.

This chapter explains how Kafka Streams recovers tasks and how exactly-once processing coordinates their committed results. We'll follow one count update through different failures, configure the processing mode, and define the boundary beyond which the application still needs to coordinate its work.