What it is
Messages travel through a producer, a broker, a consumer and an external system, and any hop can crash, drop or repeat. Each delivery strategy decides when to commit the offset, whether to retry and where to deduplicate. This sandbox injects faults at named points (after poll, after process, after commit, a lost ack, a rebalance, a poison message) and shows which strategies lose, duplicate or reorder messages.
When to use it
Use it to see the outcome matrix of strategies against faults, to replay a single failure step by step, and to search automatically for the fewest faults that break a property. It backs the rule of thumb: delivery is at-least-once, and exactly-once is an effect you build with idempotence or with a transaction that covers the sink.
Gotchas & limits
One partition, one consumer group and a few messages with deterministic timing; no replication, leader changes, disk or real network. The counter-example search is bounded and not a proof. Defaults follow the Apache Kafka 4.2 documentation checked on 2026-10-22; behaviour changes between versions, and the page ranks no products.