Twelve Nodes One Busy

Cluster mode routes every read through the node that owns slot 0, which acts as a coordinator and saturates regardless of how keys are distributed

A key hashes to one slot and so to one primary, and adding shards cannot split its reads; relief has to come from spreading the value itself, for example copies under several key names

Rebalancing moves slots only when the per-node key counts differ, so the busy node keeps its slots until you raise its maxmemory limit

The client library opens a single connection per node, so the three new nodes stay idle until the connection pool size is raised on every client