What does enabling Maglev consistent hashing in Cilium's kube-proxy replacement give you?

Answer: All nodes pick the same backend for a connection, with little churn

Explanation

By default each node picks backends on its own, so if a connection's packets start arriving at a different node, for example after an upstream ECMP change, they can land on another backend. With Maglev every node computes the same lookup table from the backend list. Any node then maps a connection to the same backend without syncing state, and adding or removing a backend moves few existing connections.