The prediction hot path: route, admit, cache, hook, dispatch, record.

Everything in this module runs in the calling process — a Phoenix controller, an Oban job, a Task. Nothing here is a GenServer, and that is the point: MLServe adds no serialisation point between your request and the model.

The order of operations is deliberate:

route lookup (1 ETS read)
  → admission control (atomic counter)
    → telemetry span opens
      → cache lookup
        → preprocess hook
          → dispatch
          → postprocess hook
        → cache write
    → telemetry span closes
  → counters settled

Cache, hooks and validation all happen before dispatch, so a worker is occupied only for the duration of actual inference. A preprocess hook that queries Postgres for stored features runs on the caller's own scheduler time, never on a GPU worker's.

Dispatch strategies