If you use a sharded “eventual consistency” database cluster to store the number of reset attempts per interval, then sufficient IPs can be purchased to overwhelm your “eventual” threshold unless each reset attempt compels consistency before evaluating whether the limit is reached. That would qualify as a “race condition” in language terms, even if not being the classical sort of one considered such by programmers, and present the easy described fix (take a consistent read lock rather than an inconsistent one or none at all) when checking the rate limit.