Problem
A post-incident review just found that an event a downstream service swore it never received was, in fact, acknowledged by the cluster, and then lost when the broker that accepted it died a few seconds later with nothing else holding a copy yet. Build a distributed event-streaming platform: topics, partitions, producers, and consumer groups, with durable retention and replay, at a real-world scale of roughly 5M events/sec in production. This is the final problem in the entire curriculum; design it from a blank canvas so an acknowledged write can never quietly disappear, the cluster tolerates losing a broker, one partition getting far more traffic than any other doesn't become a permanent bottleneck, and data old enough to be rarely read can move to cheaper storage without losing the ability to replay it.