rlmesh.TelemetryRow

classfrom rlmesh import TelemetryRow[source]

One aggregated metric series from a Model.run() eval.

TelemetryRow(
    op: str,
    component: str,
    metric: str,
    unit: str,
    count: int,
    avg: float,
    p50: float,
    p95: float,
    p99: float,
)

The runtime aggregates every measurement per (op, component, metric) over the whole run. rpc.total is the client-observed round trip of an op; endpoint.total the peer’s own handling wall, split by endpoint.decode / endpoint.user / endpoint.encode; endpoint.queue the wait before the op’s handler ran (the model’s permit, gate, and handler lock; the env’s lock) and predict.in_flight the model endpoint’s slot occupancy at dispatch; predict.adapter the RLMesh adapter work (observation assembly + action apply) inside the handler’s own time, so the model’s own forward is endpoint.user minus it; held.episodes / held.bytes the per-episode frame-stack state the adapter engine was holding when the predict was stamped (what grows with concurrent episodes and shrinks on episode-end GC; a zero is a real sample); lane.skew a vector env’s straggler dispersion, only from an env that times its own lanes; runner.round one full predict -> step -> transform loop iteration (subtract the per-op rows for the driver’s own residual). A metric a peer never stamps produces no row.