Levee is configured with a success-rate target and timeout. It then continuously monitors the workload performance characteristics to detect downstream capacity exhaustion or failures. It can also spot a surge from growing concurrency before failures arrive. It runs in-process, uses a small, fixed amount of memory, has zero dependencies, and processes millions of requests per second.
In a deterministic 10-node mesh simulation, Levee outscored carefully tuned static rate limiters/breakers deployed throughout the mesh, while recording fewer failures and no node crashes.