Audience observation: race conditions happen under load, which pushes receivers to scale out, which makes the atomic/central locking needed for idempotency harder to implement — so overloaded receivers are least likely to support idempotency. Is that the speaker's experience? Answer: yes, this is exactly CAP-theorem territory (consistency/availability/partitioning). The solution is sharding: use a deterministic function (e.g. a hash) so the same entity always lands on the same disjoint partition, which can then independently enforce atomicity. Not the only problem that will surface, but the canonical answer for this one.