pub fn queue_telemetry() -> QueueTelemetryExpand description
Snapshot the current queue telemetry counters.
This is a lock-free, best-effort snapshot: the counters are loaded
independently, so under concurrent enqueue/process the derived in_flight
may transiently disagree with any single instantaneous state. The skew is
bounded and one-directional — enqueued is loaded first (the oldest value)
and processed last (the newest), so in_flight can only underestimate
the true backlog (floored at 0 by the saturating subtraction), never
overestimate it. That is the safe bias for the consumers (load-governor
pressure classification, backpressure), which read it per-frame and recover
on the next interval; do not treat it as an exact, linearizable count.
dropped is deliberately NOT subtracted from in_flight: drop paths
reject a task before it is ever counted as enqueued, so subtracting it
would double-penalize and permanently understate the backlog after any
overload episode.