cache.scala

  // none of the remaining methods are actually required by the cache directive - but they must be implemented to
  // satisfy the spray Cache trait. The equivalent methods offered by ConcurrentLinkedHashMap appeared to be async, but
  // with Caffeine we need to use the synchronous view.

  def remove(key: Any) = {
    Option(store.synchronous().asMap().remove(key)).map(value => Future.successful(value))
  }

  def clear(): Unit = store.synchronous().invalidateAll()

  def keys: Set[Any] = store.synchronous().asMap().keySet().asScala.toSet

  def ascendingKeys(limit: Option[Int] = None) = ???

  def size = store.synchronous().asMap().size()def remove(key: Any) = {
    Option(store.synchronous().asMap().remove(key)).map(value => Future.successful(value))
  }

Option(store.synchronous().asMap().remove(key)).map(value => Future.successful(value))

This does look unfortunate. Since AsyncLoadingCache is for conveniency, it may be a bad match and you might prefer to use Cache[K, Future[V]]. Then it will behave exactly like CLHM, but be unaware of the asynchronous nature of the value. That then requires your code to be intelligent, e.g. clean-up failed futures. Various features may not behave ideally, such as expiration starting at the insertion time of the Future rather than when it has materialized. None of these are too critical and Spray's Cache was okay with that, so a direct port may prefer that rather than fake an async call by blocking.

Though I don't know why removal should care about the old value. Most likely you can scan over usages and get away with an invalidate instead.