Are DHT algorithms good for handling large amount of counting states?

Question

I wonder if a DHT algorithm like Kademlia is good at handling for my specific use case.

I want to have a service that can maintain a large amounts of kv (string, int) pairs, within these kv pairs I also want the ability to do computation like how many increments in the past minute, hour, day, etc.

I am thinking about how to handle and scale upwards to millions of request to this "system" in both reads and writes. I have a high requirement for low latency and data accuracy, fault tolerance is also important so that's why I'm thinking of DHT algos.

If I were to rank my top priority, it would be

1. Latency
2. Accuracy
3. Fault tolerance

(Ignoring scalability here due to the nature of distributed systems)

Are there better approaches to this problem? Or would DHTs be the best way to go for a distributed system like this.

Edit: In relation to the CAP theorem, I would say it would be a CA system.