ClusterLoad

ClusterLoad: per-node CPU/memory headroom for capacity-aware RPC routing.

Each node samples its own resources with local/1 (backed by the System.* builtins). Samples are shared across the cluster (e.g. gossiped or pulled on demand) and collected into a candidate list; pick_by/2 then routes to the node your scoring function rates highest. Raw metrics are exposed — you choose what "headroom" means.

Wire transfer: to_ints/1 / from_ints/2 marshal a NodeLoad to/from a flat Int list (node_id stays caller-side, since the caller knows which node it queried), suitable for MessagePack or any int-array codec.

  let me   = ClusterLoad.local("node-a")
  -- ... collect peers' NodeLoad into `candidates` ...
  match ClusterLoad.pick_by(candidates, ClusterLoad.spare_cpu_milli) do
    Some(node) -> -- route the expensive call to node.node_id
    None       -> -- cluster view empty; fall back to local
  end