Examples with RemoteBundle org.apache.beam.runners.fnexecution.control.RemoteBundle used on opensource projects

Example 6 with RemoteBundle

use of org.apache.beam.runners.fnexecution.control.RemoteBundle in project beam by apache.

the class ExecutableStageDoFnOperatorTest method testEnsureDeferredStateCleanupTimerFiring.

private void testEnsureDeferredStateCleanupTimerFiring(boolean withCheckpointing) throws Exception {
    TupleTag<Integer> mainOutput = new TupleTag<>("main-output");
    DoFnOperator.MultiOutputOutputManagerFactory<Integer> outputManagerFactory = new DoFnOperator.MultiOutputOutputManagerFactory(mainOutput, VoidCoder.of(), new SerializablePipelineOptions(FlinkPipelineOptions.defaults()));
    StringUtf8Coder keyCoder = StringUtf8Coder.of();
    WindowingStrategy windowingStrategy = WindowingStrategy.of(FixedWindows.of(Duration.millis(1000)));
    KvCoder<String, Integer> kvCoder = KvCoder.of(keyCoder, VarIntCoder.of());
    ExecutableStageDoFnOperator<Integer, Integer> operator = getOperator(mainOutput, Collections.emptyList(), outputManagerFactory, windowingStrategy, keyCoder, WindowedValue.getFullCoder(kvCoder, windowingStrategy.getWindowFn().windowCoder()));
    @SuppressWarnings("unchecked") RemoteBundle bundle = Mockito.mock(RemoteBundle.class);
    when(stageBundleFactory.getBundle(any(), any(), any(), any(), any(), any())).thenReturn(bundle);
    KV<String, String> timerInputKey = KV.of("transformId", "timerId");
    AtomicBoolean timerInputReceived = new AtomicBoolean();
    IntervalWindow window = new IntervalWindow(new Instant(0), new Instant(1000));
    IntervalWindow.IntervalWindowCoder windowCoder = IntervalWindow.IntervalWindowCoder.of();
    WindowedValue<KV<String, Integer>> windowedValue = WindowedValue.of(KV.of("one", 1), window.maxTimestamp(), ImmutableList.of(window), PaneInfo.NO_FIRING);
    FnDataReceiver receiver = Mockito.mock(FnDataReceiver.class);
    FnDataReceiver<Timer> timerReceiver = Mockito.mock(FnDataReceiver.class);
    doAnswer((invocation) -> {
        timerInputReceived.set(true);
        return null;
    }).when(timerReceiver).accept(any());
    when(bundle.getInputReceivers()).thenReturn(ImmutableMap.of("input", receiver));
    when(bundle.getTimerReceivers()).thenReturn(ImmutableMap.of(timerInputKey, timerReceiver));
    KeyedOneInputStreamOperatorTestHarness<ByteBuffer, WindowedValue<KV<String, Integer>>, WindowedValue<Integer>> testHarness = new KeyedOneInputStreamOperatorTestHarness(operator, operator.keySelector, new CoderTypeInformation<>(FlinkKeyUtils.ByteBufferCoder.of(), FlinkPipelineOptions.defaults()));
    testHarness.open();
    Lock stateBackendLock = Whitebox.getInternalState(operator, "stateBackendLock");
    stateBackendLock.lock();
    KeyedStateBackend<ByteBuffer> keyedStateBackend = operator.getKeyedStateBackend();
    ByteBuffer key = FlinkKeyUtils.encodeKey(windowedValue.getValue().getKey(), keyCoder);
    keyedStateBackend.setCurrentKey(key);
    DoFnOperator.FlinkTimerInternals timerInternals = Whitebox.getInternalState(operator, "timerInternals");
    Object doFnRunner = Whitebox.getInternalState(operator, "doFnRunner");
    Object delegate = Whitebox.getInternalState(doFnRunner, "delegate");
    Object stateCleaner = Whitebox.getInternalState(delegate, "stateCleaner");
    Collection<?> cleanupQueue = Whitebox.getInternalState(stateCleaner, "cleanupQueue");
    // create some state which can be cleaned up
    assertThat(testHarness.numKeyedStateEntries(), is(0));
    StateNamespace stateNamespace = StateNamespaces.window(windowCoder, window);
    // State from the SDK Harness is stored as ByteStrings
    BagState<ByteString> state = operator.keyedStateInternals.state(stateNamespace, StateTags.bag(stateId, ByteStringCoder.of()));
    state.add(ByteString.copyFrom("userstate".getBytes(Charsets.UTF_8)));
    assertThat(testHarness.numKeyedStateEntries(), is(1));
    // user timer that fires after the end of the window and after state cleanup
    TimerInternals.TimerData userTimer = TimerInternals.TimerData.of("", TimerReceiverFactory.encodeToTimerDataTimerId(timerInputKey.getKey(), timerInputKey.getValue()), stateNamespace, window.maxTimestamp(), window.maxTimestamp(), TimeDomain.EVENT_TIME);
    timerInternals.setTimer(userTimer);
    // start of bundle
    testHarness.processElement(new StreamRecord<>(windowedValue));
    verify(receiver).accept(windowedValue);
    // move watermark past user timer while bundle is in progress
    testHarness.processWatermark(new Watermark(window.maxTimestamp().plus(Duration.millis(1)).getMillis()));
    // Output watermark is held back and timers do not yet fire (they can still be changed!)
    assertThat(timerInputReceived.get(), is(false));
    assertThat(operator.getCurrentOutputWatermark(), is(BoundedWindow.TIMESTAMP_MIN_VALUE.getMillis()));
    // The timer fires on bundle finish
    operator.invokeFinishBundle();
    assertThat(timerInputReceived.getAndSet(false), is(true));
    // Move watermark past the cleanup timer
    testHarness.processWatermark(new Watermark(window.maxTimestamp().plus(Duration.millis(2)).getMillis()));
    operator.invokeFinishBundle();
    // Cleanup timer has fired and cleanup queue is prepared for bundle finish
    assertThat(testHarness.numEventTimeTimers(), is(0));
    assertThat(testHarness.numKeyedStateEntries(), is(1));
    assertThat(cleanupQueue, hasSize(1));
    // Cleanup timer are rescheduled if a new timer is created during the bundle
    TimerInternals.TimerData userTimer2 = TimerInternals.TimerData.of("", TimerReceiverFactory.encodeToTimerDataTimerId(timerInputKey.getKey(), timerInputKey.getValue()), stateNamespace, window.maxTimestamp(), window.maxTimestamp(), TimeDomain.EVENT_TIME);
    operator.setTimer(Timer.of(windowedValue.getValue().getKey(), "", windowedValue.getWindows(), window.maxTimestamp(), window.maxTimestamp(), PaneInfo.NO_FIRING), userTimer2);
    assertThat(testHarness.numEventTimeTimers(), is(1));
    if (withCheckpointing) {
        // Upon checkpointing, the bundle will be finished.
        testHarness.snapshot(0, 0);
    } else {
        operator.invokeFinishBundle();
    }
    // Cleanup queue has been processed and cleanup timer has been re-added due to pending timers
    // for the window.
    assertThat(cleanupQueue, hasSize(0));
    verifyNoMoreInteractions(receiver);
    assertThat(testHarness.numKeyedStateEntries(), is(2));
    assertThat(testHarness.numEventTimeTimers(), is(2));
    // No timer has been fired but bundle should be ended
    assertThat(timerInputReceived.get(), is(false));
    assertThat(Whitebox.getInternalState(operator, "bundleStarted"), is(false));
    // Allow user timer and cleanup timer to fire by triggering watermark advancement
    testHarness.setProcessingTime(testHarness.getProcessingTime() + 1);
    assertThat(timerInputReceived.getAndSet(false), is(true));
    assertThat(cleanupQueue, hasSize(1));
    // Cleanup will be executed after the bundle is complete because there are no more pending
    // timers for the window
    operator.invokeFinishBundle();
    assertThat(cleanupQueue, hasSize(0));
    assertThat(testHarness.numKeyedStateEntries(), is(0));
    testHarness.close();
    verifyNoMoreInteractions(receiver);
}

Example 7 with RemoteBundle

use of org.apache.beam.runners.fnexecution.control.RemoteBundle in project beam by apache.

the class ExecutableStageDoFnOperatorTest method outputsAreTaggedCorrectly.

@Test
public void outputsAreTaggedCorrectly() throws Exception {
    WindowedValue.ValueOnlyWindowedValueCoder<Integer> coder = WindowedValue.getValueOnlyCoder(VarIntCoder.of());
    TupleTag<Integer> mainOutput = new TupleTag<>("main-output");
    TupleTag<Integer> additionalOutput1 = new TupleTag<>("output-1");
    TupleTag<Integer> additionalOutput2 = new TupleTag<>("output-2");
    ImmutableMap<TupleTag<?>, OutputTag<?>> tagsToOutputTags = ImmutableMap.<TupleTag<?>, OutputTag<?>>builder().put(additionalOutput1, new OutputTag<WindowedValue<String>>(additionalOutput1.getId()) {
    }).put(additionalOutput2, new OutputTag<WindowedValue<String>>(additionalOutput2.getId()) {
    }).build();
    ImmutableMap<TupleTag<?>, Coder<WindowedValue<?>>> tagsToCoders = ImmutableMap.<TupleTag<?>, Coder<WindowedValue<?>>>builder().put(mainOutput, (Coder) coder).put(additionalOutput1, coder).put(additionalOutput2, coder).build();
    ImmutableMap<TupleTag<?>, Integer> tagsToIds = ImmutableMap.<TupleTag<?>, Integer>builder().put(mainOutput, 0).put(additionalOutput1, 1).put(additionalOutput2, 2).build();
    DoFnOperator.MultiOutputOutputManagerFactory<Integer> outputManagerFactory = new DoFnOperator.MultiOutputOutputManagerFactory(mainOutput, tagsToOutputTags, tagsToCoders, tagsToIds, new SerializablePipelineOptions(FlinkPipelineOptions.defaults()));
    WindowedValue<Integer> zero = WindowedValue.valueInGlobalWindow(0);
    WindowedValue<Integer> three = WindowedValue.valueInGlobalWindow(3);
    WindowedValue<Integer> four = WindowedValue.valueInGlobalWindow(4);
    WindowedValue<Integer> five = WindowedValue.valueInGlobalWindow(5);
    // We use a real StageBundleFactory here in order to exercise the output receiver factory.
    StageBundleFactory stageBundleFactory = new StageBundleFactory() {

        private boolean onceEmitted;

        @Override
        public RemoteBundle getBundle(OutputReceiverFactory receiverFactory, TimerReceiverFactory timerReceiverFactory, StateRequestHandler stateRequestHandler, BundleProgressHandler progressHandler, BundleFinalizationHandler finalizationHandler, BundleCheckpointHandler checkpointHandler) {
            return new RemoteBundle() {

                @Override
                public String getId() {
                    return "bundle-id";
                }

                @Override
                public Map<String, FnDataReceiver> getInputReceivers() {
                    return ImmutableMap.of("input", input -> {
                    /* Ignore input*/
                    });
                }

                @Override
                public Map<KV<String, String>, FnDataReceiver<Timer>> getTimerReceivers() {
                    return Collections.emptyMap();
                }

                @Override
                public void requestProgress() {
                    throw new UnsupportedOperationException();
                }

                @Override
                public void split(double fractionOfRemainder) {
                    throw new UnsupportedOperationException();
                }

                @Override
                public void close() throws Exception {
                    if (onceEmitted) {
                        return;
                    }
                    // Emit all values to the runner when the bundle is closed.
                    receiverFactory.create(mainOutput.getId()).accept(three);
                    receiverFactory.create(additionalOutput1.getId()).accept(four);
                    receiverFactory.create(additionalOutput2.getId()).accept(five);
                    onceEmitted = true;
                }
            };
        }

        @Override
        public ProcessBundleDescriptors.ExecutableProcessBundleDescriptor getProcessBundleDescriptor() {
            return processBundleDescriptor;
        }

        @Override
        public InstructionRequestHandler getInstructionRequestHandler() {
            return null;
        }

        @Override
        public void close() {
        }
    };
    // Wire the stage bundle factory into our context.
    when(stageContext.getStageBundleFactory(any())).thenReturn(stageBundleFactory);
    ExecutableStageDoFnOperator<Integer, Integer> operator = getOperator(mainOutput, ImmutableList.of(additionalOutput1, additionalOutput2), outputManagerFactory);
    OneInputStreamOperatorTestHarness<WindowedValue<Integer>, WindowedValue<Integer>> testHarness = new OneInputStreamOperatorTestHarness<>(operator);
    long watermark = testHarness.getCurrentWatermark() + 1;
    testHarness.open();
    testHarness.processElement(new StreamRecord<>(zero));
    testHarness.processWatermark(watermark);
    watermark++;
    testHarness.processWatermark(watermark);
    assertEquals(watermark, testHarness.getCurrentWatermark());
    // watermark hold until bundle complete
    assertEquals(0, testHarness.getOutput().size());
    // triggers finish bundle
    testHarness.close();
    assertThat(stripStreamRecordFromWindowedValue(testHarness.getOutput()), contains(three));
    assertThat(testHarness.getSideOutput(tagsToOutputTags.get(additionalOutput1)), contains(new StreamRecord<>(four)));
    assertThat(testHarness.getSideOutput(tagsToOutputTags.get(additionalOutput2)), contains(new StreamRecord<>(five)));
}

Example 8 with RemoteBundle

use of org.apache.beam.runners.fnexecution.control.RemoteBundle in project beam by apache.

the class FlinkExecutableStageFunction method mapPartition.

/**
 * For non-stateful processing via a simple MapPartitionFunction.
 */
@Override
public void mapPartition(Iterable<WindowedValue<InputT>> iterable, Collector<RawUnionValue> collector) throws Exception {
    ReceiverFactory receiverFactory = new ReceiverFactory(collector, outputMap);
    if (sdfStateInternals != null) {
        sdfTimerInternals.advanceProcessingTime(Instant.now());
        sdfTimerInternals.advanceSynchronizedProcessingTime(Instant.now());
    }
    try (RemoteBundle bundle = stageBundleFactory.getBundle(receiverFactory, stateRequestHandler, progressHandler, finalizationHandler, bundleCheckpointHandler)) {
        processElements(iterable, bundle);
    }
    if (sdfTimerInternals != null) {
        // Finally, advance the processing time to infinity to fire any timers.
        sdfTimerInternals.advanceProcessingTime(BoundedWindow.TIMESTAMP_MAX_VALUE);
        sdfTimerInternals.advanceSynchronizedProcessingTime(BoundedWindow.TIMESTAMP_MAX_VALUE);
        // Now we fire the SDF timers and process elements generated by timers.
        while (sdfTimerInternals.hasPendingTimers()) {
            try (RemoteBundle bundle = stageBundleFactory.getBundle(receiverFactory, stateRequestHandler, progressHandler, finalizationHandler, bundleCheckpointHandler)) {
                List<WindowedValue<InputT>> residuals = new ArrayList<>();
                TimerInternals.TimerData timer;
                while ((timer = sdfTimerInternals.removeNextProcessingTimer()) != null) {
                    WindowedValue stateValue = sdfStateInternals.state(timer.getNamespace(), StateTags.value(timer.getTimerId(), inputCoder)).read();
                    residuals.add(stateValue);
                }
                processElements(residuals, bundle);
            }
        }
    }
}

Example 9 with RemoteBundle

use of org.apache.beam.runners.fnexecution.control.RemoteBundle in project beam by apache.

the class FlinkExecutableStageFunction method reduce.

/**
 * For stateful and timer processing via a GroupReduceFunction.
 */
@Override
public void reduce(Iterable<WindowedValue<InputT>> iterable, Collector<RawUnionValue> collector) throws Exception {
    // Need to discard the old key's state
    if (bagUserStateHandlerFactory != null) {
        bagUserStateHandlerFactory.resetForNewKey();
    }
    // Used with Batch, we know that all the data is available for this key. We can't use the
    // timer manager from the context because it doesn't exist. So we create one and advance
    // time to the end after processing all elements.
    final InMemoryTimerInternals timerInternals = new InMemoryTimerInternals();
    timerInternals.advanceProcessingTime(Instant.now());
    timerInternals.advanceSynchronizedProcessingTime(Instant.now());
    ReceiverFactory receiverFactory = new ReceiverFactory(collector, outputMap);
    TimerReceiverFactory timerReceiverFactory = new TimerReceiverFactory(stageBundleFactory, (Timer<?> timer, TimerInternals.TimerData timerData) -> {
        currentTimerKey = timer.getUserKey();
        if (timer.getClearBit()) {
            timerInternals.deleteTimer(timerData);
        } else {
            timerInternals.setTimer(timerData);
        }
    }, windowCoder);
    // First process all elements and make sure no more elements can arrive
    try (RemoteBundle bundle = stageBundleFactory.getBundle(receiverFactory, timerReceiverFactory, stateRequestHandler, progressHandler)) {
        processElements(iterable, bundle);
    }
    // Finish any pending windows by advancing the input watermark to infinity.
    timerInternals.advanceInputWatermark(BoundedWindow.TIMESTAMP_MAX_VALUE);
    // Finally, advance the processing time to infinity to fire any timers.
    timerInternals.advanceProcessingTime(BoundedWindow.TIMESTAMP_MAX_VALUE);
    timerInternals.advanceSynchronizedProcessingTime(BoundedWindow.TIMESTAMP_MAX_VALUE);
    // Now we fire the timers and process elements generated by timers (which may be timers itself)
    while (timerInternals.hasPendingTimers()) {
        try (RemoteBundle bundle = stageBundleFactory.getBundle(receiverFactory, timerReceiverFactory, stateRequestHandler, progressHandler)) {
            PipelineTranslatorUtils.fireEligibleTimers(timerInternals, bundle.getTimerReceivers(), currentTimerKey);
        }
    }
}

Example 10 with RemoteBundle

use of org.apache.beam.runners.fnexecution.control.RemoteBundle in project beam by apache.

the class FlinkExecutableStageFunctionTest method sdkErrorsSurfaceOnClose.

@Test
public void sdkErrorsSurfaceOnClose() throws Exception {
    FlinkExecutableStageFunction<Integer> function = getFunction(Collections.emptyMap());
    function.open(new Configuration());
    @SuppressWarnings("unchecked") RemoteBundle bundle = Mockito.mock(RemoteBundle.class);
    when(stageBundleFactory.getBundle(any(), any(StateRequestHandler.class), any(BundleProgressHandler.class), any(BundleFinalizationHandler.class), any(BundleCheckpointHandler.class))).thenReturn(bundle);
    @SuppressWarnings("unchecked") FnDataReceiver<WindowedValue<?>> receiver = Mockito.mock(FnDataReceiver.class);
    when(bundle.getInputReceivers()).thenReturn(ImmutableMap.of("input", receiver));
    Exception expected = new Exception();
    doThrow(expected).when(bundle).close();
    thrown.expect(is(expected));
    function.mapPartition(Collections.emptyList(), collector);
}