"Stamp each write with the current time and keep the latest" feels obvious — and it silently deletes data, because two machines can never agree on what time it is.
You add 24 more workers to fix Kafka lag — and they all sit idle. Meanwhile one huge message is stuck at the front of a partition, and nothing behind it moves.
“We sold 3 tickets but only had 1.” The code is inside a transaction, looks right, works a hundred times locally, and only breaks in production under load.