Custom stream processing

Port states, AbstractInHandler and AbstractOutHandler

In order to interact with a port (Inlet or Outlet) of the stage we need to be able to receive events and generate new events belonging to the port. From the GraphStageLogic the following operations are available on an output port:

• push(out,elem) pushes an element to the output port. Only possible after the port has been pulled by downstream.
• complete(out) closes the output port normally.
• fail(out,exception) closes the port with a failure signal.

The events corresponding to an output port can be received in an AbstractOutHandler instance registered to the output port using setHandler(out,handler). This handler has two callbacks:

• onPull() is called when the output port is ready to emit the next element, push(out, elem) is now allowed to be called on this port.
• onDownstreamFinish() is called once the downstream has cancelled and no longer allows messages to be pushed to it. No more onPull() will arrive after this event. If not overridden this will default to stopping the stage.

Also, there are two query methods available for output ports:

• isAvailable(out) returns true if the port can be pushed.
• isClosed(out) returns true if the port is closed. At this point the port can not be pushed and will not be pulled anymore.

The relationship of the above operations, events and queries are summarized in the state machine below. Green shows the initial state while orange indicates the end state. If an operation is not listed for a state, then it is invalid to call it while the port is in that state. If an event is not listed for a state, then that event cannot happen in that state.

The following operations are available for input ports:

• pull(in) requests a new element from an input port. This is only possible after the port has been pushed by upstream.
• grab(in) acquires the element that has been received during an onPush(). It cannot be called again until the port is pushed again by the upstream.
• cancel(in) closes the input port.

The events corresponding to an input port can be received in an AbstractInHandler instance registered to the input port using setHandler(in, handler). This handler has three callbacks:

• onPush() is called when the input port has now a new element. Now it is possible to acquire this element using grab(in) and/or call pull(in) on the port to request the next element. It is not mandatory to grab the element, but if it is pulled while the element has not been grabbed it will drop the buffered element.
• onUpstreamFinish() is called once the upstream has completed and no longer can be pulled for new elements. No more onPush() will arrive after this event. If not overridden this will default to stopping the stage.
• onUpstreamFailure() is called if the upstream failed with an exception and no longer can be pulled for new elements. No more onPush() will arrive after this event. If not overridden this will default to failing the stage.

Also, there are three query methods available for input ports:

• isAvailable(in) returns true if a data element can be grabbed from the port
• hasBeenPulled(in) returns true if the port has been already pulled. Calling pull(in) in this state is illegal.
• isClosed(in) returns true if the port is closed. At this point the port can not be pulled and will not be pushed anymore.

The relationship of the above operations, events and queries are summarized in the state machine below. Green shows the initial state while orange indicates the end state. If an operation is not listed for a state, then it is invalid to call it while the port is in that state. If an event is not listed for a state, then that event cannot happen in that state.

Finally, there are two methods available for convenience to complete the stage and all of its ports:

• completeStage() is equivalent to closing all output ports and cancelling all input ports.
• failStage(exception) is equivalent to failing all output ports and cancelling all input ports.

In some cases it is inconvenient and error prone to react on the regular state machine events with the signal based API described above. For those cases there is an API which allows for a more declarative sequencing of actions which will greatly simplify some use cases at the cost of some extra allocations. The difference between the two APIs could be described as that the first one is signal driven from the outside, while this API is more active and drives its surroundings.

The operations of this part of the :class:GraphStage API are:

• emit(out, elem) and emitMultiple(out, Iterable(elem1, elem2)) replaces the OutHandler with a handler that emits one or more elements when there is demand, and then reinstalls the current handlers
• read(in)(andThen) and readN(in, n)(andThen) replaces the InHandler with a handler that reads one or more elements as they are pushed and allows the handler to react once the requested number of elements has been read.
• abortEmitting() and abortReading() which will cancel an ongoing emit or read

Note that since the above methods are implemented by temporarily replacing the handlers of the stage you should never call setHandler while they are running emit or read as that interferes with how they are implemented. The following methods are safe to call after invoking emit and read (and will lead to actually running the operation when those are done): complete(out), completeStage(), emit, emitMultiple, abortEmitting() and abortReading()

An example of how this API simplifies a stage can be found below in the second version of the :class:Duplicator.

Custom linear processing stages using GraphStage

Graph stages allows for custom linear processing stages through letting them have one input and one output and using FlowShape as their shape.

Such a stage can be illustrated as a box with two flows as it is seen in the illustration below. Demand flowing upstream leading to elements flowing downstream.

To illustrate these concepts we create a small GraphStage that implements the map transformation.

Map calls push(out) from the onPush() handler and it also calls pull() from the onPull handler resulting in the conceptual wiring above, and fully expressed in code below:

public class Map<A, B> extends GraphStage<FlowShape<A, B>> {

  private final Function<A, B> f;

  public Map(Function<A, B> f) {
    this.f = f;
  }

  public final Inlet<A> in = Inlet.create("Map.in");
  public final Outlet<B> out = Outlet.create("Map.out");

  private final FlowShape<A, B> shape = FlowShape.of(in, out);
  @Override
  public FlowShape<A,B> shape() {
    return shape;
  }

  @Override
  public GraphStageLogic createLogic(Attributes inheritedAttributes) {
    return new GraphStageLogic(shape) {

      {
        setHandler(in, new AbstractInHandler() {
          @Override
          public void onPush() throws Exception {
            push(out, f.apply(grab(in)));
          }
        });
        setHandler(out, new AbstractOutHandler() {
          @Override
          public void onPull() throws Exception {
            pull(in);
          }
        });
      }
    };
  }

}

Map is a typical example of a one-to-one transformation of a stream where demand is passed along upstream elements passed on downstream.

To demonstrate a many-to-one stage we will implement filter. The conceptual wiring of Filter looks like this:

As we see above, if the given predicate matches the current element we are propagating it downwards, otherwise we return the “ball” to our upstream so that we get the new element. This is achieved by modifying the map example by adding a conditional in the onPush handler and decide between a pull(in) or push(out) call (and of course not having a mapping f function).

public final class Filter<A> extends GraphStage<FlowShape<A, A>> {

  private final Predicate<A> p;

  public Filter(Predicate<A> p) {
    this.p = p;
  }

  public final Inlet<A> in = Inlet.create("Filter.in");
  public final Outlet<A> out = Outlet.create("Filter.out");

  private final FlowShape<A, A> shape = FlowShape.of(in, out);

  @Override
  public FlowShape<A, A> shape() {
    return shape;
  }

  public GraphStageLogic createLogic(Attributes inheritedAttributes) {
    return new GraphStageLogic(shape) {
      {

        setHandler(in, new AbstractInHandler() {
          @Override
          public void onPush() {
            A elem = grab(in);
            if (p.test(elem)) {
              push(out, elem);
            } else {
              pull(in);
            }
          }
        });

        setHandler(out, new AbstractOutHandler() {
          @Override
          public void onPull() throws Exception {
            pull(in);
          }
        });
      }
    };
  }
}

To complete the picture we define a one-to-many transformation as the next step. We chose a straightforward example stage that emits every upstream element twice downstream. The conceptual wiring of this stage looks like this:

This is a stage that has state: an option with the last element it has seen indicating if it has duplicated this last element already or not. We must also make sure to emit the extra element if the upstream completes.

public class Duplicator<A> extends GraphStage<FlowShape<A, A>> {

  public final Inlet<A> in = Inlet.create("Duplicator.in");
  public final Outlet<A> out = Outlet.create("Duplicator.out");

  private final FlowShape<A, A> shape = FlowShape.of(in, out);

  @Override
  public FlowShape<A, A> shape() {
    return shape;
  }

  public GraphStageLogic createLogic(Attributes inheritedAttributes) {
    return new GraphStageLogic(shape) {
      // Again: note that all mutable state
      // MUST be inside the GraphStageLogic
      Option<A> lastElem = Option.none();

      {
        setHandler(in, new AbstractInHandler() {
          @Override
          public void onPush() {
            A elem = grab(in);
            lastElem = Option.some(elem);
            push(out, elem);
          }

          @Override
          public void onUpstreamFinish() {
            if (lastElem.isDefined()) {
              emit(out, lastElem.get());
            }
            complete(out);
          }
        });


        setHandler(out, new AbstractOutHandler() {
          @Override
          public void onPull() throws Exception {
            if (lastElem.isDefined()) {
              push(out, lastElem.get());
              lastElem = Option.none();
            } else {
              pull(in);
            }
          }
        });
      }
    };
  }
}

In this case a pull from downstream might be consumed by the stage itself rather than passed along upstream as the stage might contain an element it wants to push. Note that we also need to handle the case where the upstream closes while the stage still has elements it wants to push downstream. This is done by overriding onUpstreamFinish in the AbstractInHandler and provide custom logic that should happen when the upstream has been finished.

This example can be simplified by replacing the usage of a mutable state with calls to emitMultiple which will replace the handlers, emit each of multiple elements and then reinstate the original handlers:

public class Duplicator2<A> extends GraphStage<FlowShape<A, A>> {

  public final Inlet<A> in = Inlet.create("Duplicator.in");
  public final Outlet<A> out = Outlet.create("Duplicator.out");

  private final FlowShape<A, A> shape = FlowShape.of(in, out);

  @Override
  public FlowShape<A, A> shape() {
    return shape;
  }

  @Override
  public GraphStageLogic createLogic(Attributes inheritedAttributes) {
    return new GraphStageLogic(shape) {

      {
        setHandler(in, new AbstractInHandler() {
          @Override
          public void onPush() {
            A elem = grab(in);
            // this will temporarily suspend this handler until the two elems
            // are emitted and then reinstates it
            emitMultiple(out, Arrays.asList(elem, elem).iterator());
          }
        });

        setHandler(out, new AbstractOutHandler() {
          @Override
          public void onPull() throws Exception {
            pull(in);
          }
        });
      }
    };
  }
}

Finally, to demonstrate all of the stages above, we put them together into a processing chain, which conceptually would correspond to the following structure:

In code this is only a few lines, using the via use our custom stages in a stream:

CompletionStage<String> resultFuture = Source.from(Arrays.asList(1,2,3,4,5))
        .via(new Filter<Integer>((n) -> n % 2 == 0))
        .via(new Duplicator<Integer>())
        .via(new Map<Integer, Integer>((n) -> n / 2))
        .runWith(sink, mat);

If we attempt to draw the sequence of events, it shows that there is one "event token" in circulation in a potential chain of stages, just like our conceptual "railroad tracks" representation predicts.

Completion

Completion handling usually (but not exclusively) comes into the picture when processing stages need to emit a few more elements after their upstream source has been completed. We have seen an example of this in our first Duplicator implementation where the last element needs to be doubled even after the upstream neighbor stage has been completed. This can be done by overriding the onUpstreamFinish method in AbstractInHandler.

Stages by default automatically stop once all of their ports (input and output) have been closed externally or internally. It is possible to opt out from this behavior by invoking setKeepGoing(true) (which is not supported from the stage’s constructor and usually done in preStart). In this case the stage must be explicitly closed by calling completeStage() or failStage(exception). This feature carries the risk of leaking streams and actors, therefore it should be used with care.

Using timers

It is possible to use timers in GraphStages by using TimerGraphStageLogic as the base class for the returned logic. Timers can be scheduled by calling one of scheduleOnce(key,delay), schedulePeriodically(key,period) or schedulePeriodicallyWithInitialDelay(key,delay,period) and passing an object as a key for that timer (can be any object, for example a String). The onTimer(key) method needs to be overridden and it will be called once the timer of key fires. It is possible to cancel a timer using cancelTimer(key) and check the status of a timer with isTimerActive(key). Timers will be automatically cleaned up when the stage completes.

Timers can not be scheduled from the constructor of the logic, but it is possible to schedule them from the preStart() lifecycle hook.

In this sample the stage toggles between open and closed, where open means no elements are passed through. The stage starts out as closed but as soon as an element is pushed downstream the gate becomes open for a duration of time during which it will consume and drop upstream messages:

// each time an event is pushed through it will trigger a period of silence
public class TimedGate<A> extends GraphStage<FlowShape<A, A>> {

  private final FiniteDuration silencePeriod;

  public TimedGate(FiniteDuration silencePeriod) {
    this.silencePeriod = silencePeriod;
  }

  public final Inlet<A> in = Inlet.create("TimedGate.in");
  public final Outlet<A> out = Outlet.create("TimedGate.out");

  private final FlowShape<A, A> shape = FlowShape.of(in, out);
  @Override
  public FlowShape<A, A> shape() {
    return shape;
  }

  @Override
  public GraphStageLogic createLogic(Attributes inheritedAttributes) {
    return new TimerGraphStageLogic(shape) {

      private boolean open = false;

      {
        setHandler(in, new AbstractInHandler() {
          @Override
          public void onPush() throws Exception {
            A elem = grab(in);
            if (open) pull(in);
            else {
              push(out, elem);
              open = true;
              scheduleOnce("key", silencePeriod);
            }
          }
        });
        setHandler(out, new AbstractOutHandler() {
          @Override
          public void onPull() throws Exception {
            pull(in);
          }
        });
      }

      @Override
      public void onTimer(Object key) {
        if (key.equals("key")) {
          open = false;
        }
      }
    };
  }
}

Using asynchronous side-channels

In order to receive asynchronous events that are not arriving as stream elements (for example a completion of a future or a callback from a 3rd party API) one must acquire a AsyncCallback by calling getAsyncCallback() from the stage logic. The method getAsyncCallback takes as a parameter a callback that will be called once the asynchronous event fires. It is important to not call the callback directly, instead, the external API must call the invoke(event) method on the returned AsyncCallback. The execution engine will take care of calling the provided callback in a thread-safe way. The callback can safely access the state of the GraphStageLogic implementation.

Sharing the AsyncCallback from the constructor risks race conditions, therefore it is recommended to use the preStart() lifecycle hook instead.

This example shows an asynchronous side channel graph stage that starts dropping elements when a future completes:

// will close upstream in all materializations of the stage instance
// when the completion stage completes
public class KillSwitch<A> extends GraphStage<FlowShape<A, A>> {

  private final CompletionStage<Done> switchF;

  public KillSwitch(CompletionStage<Done> switchF) {
    this.switchF = switchF;
  }

  public final Inlet<A> in = Inlet.create("KillSwitch.in");
  public final Outlet<A> out = Outlet.create("KillSwitch.out");

  private final FlowShape<A, A> shape = FlowShape.of(in, out);
  @Override
  public FlowShape<A, A> shape() {
    return shape;
  }

  @Override
  public GraphStageLogic createLogic(Attributes inheritedAttributes) {
    return new GraphStageLogic(shape) {

      {
        setHandler(in, new AbstractInHandler() {
          @Override
          public void onPush() {
            push(out, grab(in));
          }
        });
        setHandler(out, new AbstractOutHandler() {
          @Override
          public void onPull() throws Exception {
            pull(in);
          }
        });
      }

      @Override
      public void preStart() {
        AsyncCallback<Done> callback = createAsyncCallback(new Procedure<Done>() {
          @Override
          public void apply(Done param) throws Exception {
            completeStage();
          }
        });

        ExecutionContext ec = system.dispatcher();
        switchF.thenAccept(callback::invoke);
      }
    };
  }
}

Integration with actors

This section is a stub and will be extended in the next release This is an experimental feature*

It is possible to acquire an ActorRef that can be addressed from the outside of the stage, similarly how AsyncCallback allows injecting asynchronous events into a stage logic. This reference can be obtained by calling getStageActorRef(receive) passing in a function that takes a Pair of the sender ActorRef and the received message. This reference can be used to watch other actors by calling its watch(ref) or unwatch(ref) methods. The reference can be also watched by external actors. The current limitations of this ActorRef are:

• they are not location transparent, they cannot be accessed via remoting.
• they cannot be returned as materialized values.
• they cannot be accessed from the constructor of the GraphStageLogic, but they can be accessed from the preStart() method.

Custom materialized values

Custom stages can return materialized values instead of Unit by inheriting from GraphStageWithMaterializedValue instead of the simpler GraphStage. The difference is that in this case the method createLogicAndMaterializedValue(inheritedAttributes) needs to be overridden, and in addition to the stage logic the materialized value must be provided

In this sample the materialized value is a future containing the first element to go through the stream:

public class FirstValue<A> extends GraphStageWithMaterializedValue<FlowShape<A, A>, CompletionStage<A>> {

  public final Inlet<A> in = Inlet.create("FirstValue.in");
  public final Outlet<A> out = Outlet.create("FirstValue.out");

  private final FlowShape<A, A> shape = FlowShape.of(in, out);
  @Override
  public FlowShape<A, A> shape() {
    return shape;
  }

  @Override
  public Tuple2<GraphStageLogic, CompletionStage<A>> createLogicAndMaterializedValue(Attributes inheritedAttributes) {
    Promise<A> promise = Futures.promise();

    GraphStageLogic logic = new GraphStageLogic(shape) {
      {
        setHandler(in, new AbstractInHandler() {
          @Override
          public void onPush() {
            A elem = grab(in);
            promise.success(elem);
            push(out, elem);

            // replace handler with one just forwarding
            setHandler(in, new AbstractInHandler() {
              @Override
              public void onPush() {
                push(out, grab(in));
              }
            });
          }
        });

        setHandler(out, new AbstractOutHandler() {
          @Override
          public void onPull() throws Exception {
            pull(in);
          }
        });
      }
    };

    return new Tuple2(logic, promise.future());
  }
}

Using attributes to affect the behavior of a stage

This section is a stub and will be extended in the next release

Stages can access the Attributes object created by the materializer. This contains all the applied (inherited) attributes applying to the stage, ordered from least specific (outermost) towards the most specific (innermost) attribute. It is the responsibility of the stage to decide how to reconcile this inheritance chain to a final effective decision.

See Modularity, Composition and Hierarchy for an explanation on how attributes work.

Rate decoupled graph stages

Sometimes it is desirable to decouple the rate of the upstream and downstream of a stage, synchronizing only when needed.

This is achieved in the model by representing a GraphStage as a boundary between two regions where the demand sent upstream is decoupled from the demand that arrives from downstream. One immediate consequence of this difference is that an onPush call does not always lead to calling push and an onPull call does not always lead to calling pull.

One of the important use-case for this is to build buffer-like entities, that allow independent progress of upstream and downstream stages when the buffer is not full or empty, and slowing down the appropriate side if the buffer becomes empty or full.

The next diagram illustrates the event sequence for a buffer with capacity of two elements in a setting where the downstream demand is slow to start and the buffer will fill up with upstream elements before any demand is seen from downstream.

Another scenario would be where the demand from downstream starts coming in before any element is pushed into the buffer stage.

The first difference we can notice is that our Buffer stage is automatically pulling its upstream on initialization. The buffer has demand for up to two elements without any downstream demand.

The following code example demonstrates a buffer class corresponding to the message sequence chart above.

public class TwoBuffer<A> extends GraphStage<FlowShape<A, A>> {

  public final Inlet<A> in = Inlet.create("TwoBuffer.in");
  public final Outlet<A> out = Outlet.create("TwoBuffer.out");

  private final FlowShape<A, A> shape = FlowShape.of(in, out);

  @Override
  public FlowShape<A, A> shape() {
    return shape;
  }

  @Override
  public GraphStageLogic createLogic(Attributes inheritedAttributes) {
    return new GraphStageLogic(shape) {

      private final int SIZE = 2;
      private Queue<A> buffer = new ArrayDeque<>(SIZE);
      private boolean downstreamWaiting = false;

      private boolean isBufferFull() {
        return buffer.size() == SIZE;
      }

      @Override
      public void preStart() {
        // a detached stage needs to start upstream demand
        // itself as it is not triggered by downstream demand
        pull(in);
      }

      {
        setHandler(in, new AbstractInHandler() {
          @Override
          public void onPush() {
            A elem = grab(in);
            buffer.add(elem);
            if (downstreamWaiting) {
              downstreamWaiting = false;
              A bufferedElem = buffer.poll();
              push(out, bufferedElem);
            }
            if (!isBufferFull()) {
              pull(in);
            }
          }

          @Override
          public void onUpstreamFinish() {
            if (!buffer.isEmpty()) {
              // emit the rest if possible
              emitMultiple(out, buffer.iterator());
            }
            completeStage();
          }
        });


        setHandler(out, new AbstractOutHandler() {
          @Override
          public void onPull() throws Exception {
            if (buffer.isEmpty()) {
              downstreamWaiting = true;
            } else {
              A elem = buffer.poll();
              push(out, elem);
            }
            if (!isBufferFull() && !hasBeenPulled(in)) {
              pull(in);
            }
          }
        });
      }
    };

  }
}