Chaos engineering tests a resilience hypothesis with controlled failure injection. Examples include killing one instance, adding latency to a dependency, interrupting broker connectivity, or simulating a zone outage. Establish a safety boundary, observe impact, and stop the experiment if the blast radius exceeds the plan.

Test recovery after the fault ends: can queues drain, replicas catch up, clients reconnect, and the system return to its SLO? Random disruption without a hypothesis creates noise. Use staging or carefully controlled production experiments with authorization and rollback.

Record findings as concrete improvements to timeouts, alerts, capacity, runbooks, and architecture. Re-run the experiment to prove the fix works.