The component that limits the throughput of the whole system, identifying it is the first step of any capacity-planning exercise.
A bottleneck is the slowest or most-saturated component in a request path, the one that determines the overall throughput of the system. In a 5-service chain where 4 services handle 10K rps each and one handles 2K, the chain handles 2K. The hard part is identifying which is the bottleneck under live load: it's often the dependency that's quietly running at 95% of its capacity while everything else is at 30%, hidden by averages but obvious in percentile latency.
Capacity planning that doesn't identify the bottleneck wastes money, scaling the wrong tier doesn't change throughput. Worse, removing one bottleneck shifts the load and exposes the next one, sometimes catastrophically. A disciplined practice is to load-test the system end-to-end before scaling, find the next bottleneck, and decide whether to address it or accept it.
See the part of the platform that handles bottleneck in production.