This may be how Kong does it but it's not really 'high performance'. The right way to do rate limiting is to limit by IP using a counting Bloom Filter or Cuckoo filter along with random samples. When you hit a false positive then you have a second normal rate limiter to 'mop up' IPs that are over the first limiter.

This doesn't give you a hard exact limit but gets the job done storing far less state. You also need to bucket by IP sub-ranges in IPV6 to stop people crap flooding you with tons of unique IP's