ratelimit.scala

import scala.concurrent.ExecutionContext.Implicits.global

class ApiService[T](limit: Long, timeFrame: FiniteDuration) {
  // previous code ...
  
  private def makeRequest(): Unit = requestQueue.pollFirst() match {
    case null => // do nothing

    case req  =>
      requestCount += 1
      val res = req.f()
      res.onComplete {
        // re-queue at the head of the queue on failure and
        // attempt to make another request
        case Failure(_: Throwable) => this.synchronized {
          requestQueue.offerFirst(req)
          tryRequest()
        }

        // complete the promise for the future we gave back to the
        // user and attempt to make anothe request on success
        case Success(result)       =>
          req.p.complete(Success(result))
      }
      tryRequest()
  }

}