When is caching a bad idea for API rate limiting?

Hey everyone, I've been thinking about common caching strategies and how they apply to specific problems. We often hear about caching to reduce database load or speed up content delivery, but what about its role in API rate limiting? I've seen discussions where caching stores and checks rate limit counters. My gut feeling is that for truly critical and frequently updated counters, especially in a distributed system, a simple in-memory cache might introduce consistency issues or even race conditions that lead to inaccurate rate limiting. On the other hand, if we hit a persistent store for every single API request just to check a counter, that could become a bottleneck itself. What are your thoughts? Is there a sweet spot where caching helps, or does it generally make rate limiting more complex than it needs to be?