GenServer that periodically samples local CPU via :cpu_sup and fetches remote node CPU via :erpc.multicall/5, storing all results in NodeCpuCache.

Remote fetches run in parallel — :erpc.multicall/5 applies the timeout per-call, so poll wall-time stays bounded by @remote_timeout regardless of cluster size.

:cpu_sup (from :os_mon) is used instead of scheduler wall time because enabling :erlang.system_flag(:scheduler_wall_time, true) is a VM-wide side effect with measurable overhead. :os_mon samples the OS directly and introduces no global BEAM flag.

Startup jitter

The first poll is scheduled with a random delay in [0, poll_startup_jitter] ms (default 60_000). When many nodes boot simultaneously — e.g. a cluster-wide deploy — this prevents every poller from issuing its initial multicall in the same instant and overwhelming remote NodeCpuCache.get_local/1 workers. Set :poll_startup_jitter to 0 to disable (useful in tests).

Telemetry

Events are emitted under the [:rpc_load_balancer, :least_cpu, :poll] prefix and can be attached via :telemetry or wired into Telemetry.Metrics by downstream applications.

• [:rpc_load_balancer, :least_cpu, :poll, :start | :stop | :exception] — span around a full poll cycle (local sample + remote fetches). Metadata: %{load_balancer_name: name}.

• [:rpc_load_balancer, :least_cpu, :poll, :remote_error] — emitted for each remote that failed within a multicall cycle. Measurements: %{count: 1}. Metadata: %{load_balancer_name: name, node: remote_node, kind: :exit | :error}. :exit covers connectivity-class failures (timeout, noconnection); :error covers remote raises and unexpected shapes.