Your nine-attempt math should be plain in the logs. "Number of attempts, where available" gets fuzzy when the SDK retries inside what the job runner sees as one try, so I log the count at each layer separately. Then a slow request tells me which layer spent the time.