mule/runtime/watchman

The queue’s drain sentinel — a port of Oban.Queue.Watchman.

Erlang supervisors terminate children in reverse start order, so within the one_for_all queue supervisor the watchman goes down FIRST, while the producer and the running jobs are still alive. Its trapped parent exit is the drain trigger: pause the producer (InitiateShutdown), wait up to shutdown_grace_period for the running jobs to finish, emit [:mule, :queue, :shutdown], then exit shutdown. The queue supervisor’s child spec gives the watchman a shutdown timeout of grace + margin (watchman.ex’s shutdown: shutdown + interval trick) so the supervisor waits the drain out.

gleam_otp actors have no terminate callback; the trapped-exit message is the substitute, exactly as in the producer and peer. The drain arrives on a dedicated Subject(String) (Elixir’s raw {:drained, queue} send), so the producer never imports this module’s message type.

Divergence (parity §9): elapsed is the real monotonic milliseconds between the shutdown call and the drain/timeout — watchman.ex reports the constant 0 on drain and the full grace period on timeout, even when the drain took most of it.

pub type Message {
  ParentExit
}

pub fn start(
  config config: config.Config,
  queue queue: String,
) -> Result(
  actor.Started(process.Subject(Message)),
  actor.StartError,
)