How would you implement the Saga pattern using Kafka?

The Saga pattern is an application-level design pattern, not a Kafka feature; Kafka supplies the event log it can be built on. In a choreography-based saga each service: 1) consumes a domain event topic (for example OrderCreated), 2) performs its local transaction, 3) publishes a result event (for example PaymentCompleted or PaymentFailed). Because there is no distributed transaction, failures are handled with compensating transactions: when a step publishes a failure event, the services that already acted consume it and undo their changes with compensating actions. Practices that fit Kafka: use a transactional producer with read_committed consumers where a consume-transform-produce step must be atomic within Kafka (this does not cover a service's own database writes); make handlers idempotent using event IDs, because delivery is at-least-once by default; route messages that cannot be processed or compensated to a dead-letter topic; and Kafka Streams can keep saga state in a KTable backed by a state store and a changelog topic.

Expected key concepts: choreography, Saga, compensating transactions, domain event, local transaction, result event, idempotent, event IDs, dead-letter topic, transactional producer, KTable