I'd bet you're running a "stacked" control plane and if you moved etcd off to its own machines you'd localize the churn to be that. Which, if correct, doesn't exactly refute your "kubernetes is power and memory hungry" claim but would more clearly distinguish why

I had high hopes back in my earlier days of them offering pluggable KV stores so folks could make tradeoffs if etcd doesn't align with their goals and risks but as best I can tell they've declared it a non-goal