#01 1. The Resource and Latency Tax of the Service Mesh Sidecar Model
In a conventional Istio / Linkerd setup, traffic follows this path:
Client App -> iptables -> Envoy Sidecar In -> Network -> Envoy Sidecar Out -> iptables -> Server App.
Every request crosses the user/kernel boundary four times, which not only burns a great deal of CPU but also adds at least 3~6 milliseconds of jitter to P99 latency.
With eBPF (Extended Berkeley Packet Filter) we can attach sandboxed programs inside the Linux kernel, intercept at the originating socket (sockops), map the pair into a sockmap, and have packets copied straight to the destination socket in kernel memory, skipping the entire TCP/IP stack!