Runs N programs concurrently over the same inputs and reduces their predictions. Invoked directly from a forward/2 body, like Dsxir.Predictor.Parallel; it is not a declared predictor.

Members fan out under Dsxir.TaskSupervisor via Task.Supervisor.async_stream_nolink/4, replaying the caller's Dsxir.Settings.snapshot/0 in each worker. Member failures are tolerated and surfaced via telemetry; the reduction runs over the survivors. When every member fails, run/3 raises Dsxir.Errors.Framework.PredictorError with reason: :all_failed.

Recognised opts

• :reduce_fn — ([Dsxir.Prediction.t()] -> Dsxir.Prediction.t()). When nil (default), run/3 returns the raw prediction list; otherwise it returns the reduced prediction.
• :size — sample this many programs per call. Defaults to all.
• :max_concurrency — defaults to System.schedulers_online().
• :timeout — per-member timeout in ms. Defaults to 30_000.

Return shape

Unlike the rest of the predictor namespace, run/3 returns a bare %Dsxir.Prediction{} (with a reducer) or a [%Dsxir.Prediction{}] (without one) rather than a {program, prediction} tuple — there is no single program identity across N members.

Iron Law of OTP

Workers are short-lived Tasks with no state across calls; concurrency is the purpose (independent members); fault isolation is required, hence async_stream_nolink so a member crash does not link back to the caller, and on_timeout: :kill_task reaps slow members. No process is held; the reducer is pure.