The sidecar model gave us a way to ship platform concerns without touching application code. It worked. It still works, in some shapes. But the cost surface around it — CPU per pod, memory per pod, the cardinality explosion in the control plane — is harder to defend in 2026 than it was in 2018.
Our take: keep the data path centralized at the proxy, and move the platform concerns there too. Most teams don't need per-pod authority; they need predictable, observable per-route behavior.