Kernel bypass: the operational bill
DPDK and io_uring both remove the socket layer from the fast path, and both move complexity into the process. You gain deterministic latency and lose the ability to inspect traffic with ordinary tools.
In practice the deciding factor is not throughput but debugging. A team that cannot tcpdump a production problem will spend more time on incidents than it saves in microseconds.
Our rule: bypass the kernel only for the component whose latency budget is under 50 microseconds. Everything else stays on the normal stack, where the tooling is.