There are three ways this can be achieved:
1. Post Materialization with SourceQueue
You can use Source.queue
that materializes the Flow into a SourceQueue
:
case class Weather(zipCode : String, temperature : Double, raining : Boolean)
val bufferSize = 100
//if the buffer fills up then this strategy drops the oldest elements
//upon the arrival of a new element.
val overflowStrategy = akka.stream.OverflowStrategy.dropHead
val queue = Source.queue(bufferSize, overflowStrategy)
.filter(!_.raining)
.to(Sink foreach println)
.run() // in order to "keep" the queue Materialized value instead of the Sink's
queue offer Weather("02139", 32.0, true)
2. Post Materialization with Actor
There is a similar question and answer here, the gist being that you materialize the stream as an ActorRef and send messages to that ref:
val ref = Source.actorRef[Weather](Int.MaxValue, fail)
.filter(!_.raining)
.to(Sink foreach println )
.run() // in order to "keep" the ref Materialized value instead of the Sink's
ref ! Weather("02139", 32.0, true)
3. Pre Materialization with Actor
Similarly, you could explicitly create an Actor that contains a message buffer, use that Actor to create a Source, and then send that Actor messages as described in the answer here:
object WeatherForwarder {
def props : Props = Props[WeatherForwarder]
}
//see provided link for example definition
class WeatherForwarder extends Actor {...}
val actorRef = actorSystem actorOf WeatherForwarder.props
//note the stream has not been instatiated yet
actorRef ! Weather("02139", 32.0, true)
//stream already has 1 Weather value to process which is sitting in the
//ActorRef's internal buffer
val stream = Source(ActorPublisher[Weather](actorRef)).runWith{...}
与恶龙缠斗过久,自身亦成为恶龙;凝视深渊过久,深渊将回以凝视…