Examples with RawUnionValue org.apache.beam.sdk.transforms.join.RawUnionValue used on opensource projects

Example 6 with RawUnionValue

use of org.apache.beam.sdk.transforms.join.RawUnionValue in project beam by apache.

the class FlinkExecutableStageFunctionTest method outputsAreTaggedCorrectly.

@Test
public void outputsAreTaggedCorrectly() throws Exception {
    WindowedValue<Integer> three = WindowedValue.valueInGlobalWindow(3);
    WindowedValue<Integer> four = WindowedValue.valueInGlobalWindow(4);
    WindowedValue<Integer> five = WindowedValue.valueInGlobalWindow(5);
    Map<String, Integer> outputTagMap = ImmutableMap.of("one", 1, "two", 2, "three", 3);
    // We use a real StageBundleFactory here in order to exercise the output receiver factory.
    StageBundleFactory stageBundleFactory = new StageBundleFactory() {

        private boolean once;

        @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 (once) {
                        return;
                    }
                    // Emit all values to the runner when the bundle is closed.
                    receiverFactory.create("one").accept(three);
                    receiverFactory.create("two").accept(four);
                    receiverFactory.create("three").accept(five);
                    once = true;
                }
            };
        }

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

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

        @Override
        public void close() throws Exception {
        }
    };
    // Wire the stage bundle factory into our context.
    when(stageContext.getStageBundleFactory(any())).thenReturn(stageBundleFactory);
    FlinkExecutableStageFunction<Integer> function = getFunction(outputTagMap);
    function.open(new Configuration());
    if (isStateful) {
        function.reduce(Collections.emptyList(), collector);
    } else {
        function.mapPartition(Collections.emptyList(), collector);
    }
    // Ensure that the tagged values sent to the collector have the correct union tags as specified
    // in the output map.
    verify(collector).collect(new RawUnionValue(1, three));
    verify(collector).collect(new RawUnionValue(2, four));
    verify(collector).collect(new RawUnionValue(3, five));
    verifyNoMoreInteractions(collector);
}

Example 7 with RawUnionValue

use of org.apache.beam.sdk.transforms.join.RawUnionValue in project beam by apache.

the class DoFnOperatorTest method pushbackDataCheckpointing.

void pushbackDataCheckpointing(TestHarnessFactory<TwoInputStreamOperatorTestHarness<WindowedValue<String>, RawUnionValue, WindowedValue<String>>> harnessFactory) throws Exception {
    TwoInputStreamOperatorTestHarness<WindowedValue<String>, RawUnionValue, WindowedValue<String>> testHarness = harnessFactory.create();
    testHarness.open();
    IntervalWindow firstWindow = new IntervalWindow(new Instant(0), new Instant(100));
    IntervalWindow secondWindow = new IntervalWindow(new Instant(0), new Instant(500));
    // push in main-input elements
    WindowedValue<String> helloElement = valueInWindow("Hello", new Instant(0), firstWindow);
    WindowedValue<String> worldElement = valueInWindow("World", new Instant(1000), firstWindow);
    testHarness.processElement1(new StreamRecord<>(helloElement));
    testHarness.processElement1(new StreamRecord<>(worldElement));
    // snapshot state, throw away the operator, then restore and verify that we still match
    // main-input elements to the side-inputs that we sent earlier
    OperatorSubtaskState snapshot = testHarness.snapshot(0, 0);
    testHarness = harnessFactory.create();
    testHarness.initializeState(snapshot);
    testHarness.open();
    // push in some side inputs for both windows
    testHarness.processElement2(new StreamRecord<>(new RawUnionValue(1, valuesInWindow(PCollectionViewTesting.materializeValuesFor(view1.getPipeline().getOptions(), View.asIterable(), "hello", "ciao"), new Instant(0), firstWindow))));
    testHarness.processElement2(new StreamRecord<>(new RawUnionValue(2, valuesInWindow(PCollectionViewTesting.materializeValuesFor(view2.getPipeline().getOptions(), View.asIterable(), "foo", "bar"), new Instant(0), secondWindow))));
    assertThat(stripStreamRecordFromWindowedValue(testHarness.getOutput()), containsInAnyOrder(helloElement, worldElement));
    testHarness.close();
}

Example 8 with RawUnionValue

use of org.apache.beam.sdk.transforms.join.RawUnionValue in project beam by apache.

the class DoFnOperatorTest method testSideInputs.

public void testSideInputs(boolean keyed) throws Exception {
    WindowedValue.ValueOnlyWindowedValueCoder<String> windowedValueCoder = WindowedValue.getValueOnlyCoder(StringUtf8Coder.of());
    TupleTag<String> outputTag = new TupleTag<>("main-output");
    ImmutableMap<Integer, PCollectionView<?>> sideInputMapping = ImmutableMap.<Integer, PCollectionView<?>>builder().put(1, view1).put(2, view2).build();
    Coder<String> keyCoder = null;
    if (keyed) {
        keyCoder = StringUtf8Coder.of();
    }
    DoFnOperator<String, String, String> doFnOperator = new DoFnOperator<>(new IdentityDoFn<String>(), "stepName", windowedValueCoder, outputTag, Collections.<TupleTag<?>>emptyList(), new DoFnOperator.DefaultOutputManagerFactory<String>(), WindowingStrategy.globalDefault(), sideInputMapping, /* side-input mapping */
    ImmutableList.<PCollectionView<?>>of(view1, view2), /* side inputs */
    PipelineOptionsFactory.as(FlinkPipelineOptions.class), keyCoder);
    TwoInputStreamOperatorTestHarness<WindowedValue<String>, RawUnionValue, String> testHarness = new TwoInputStreamOperatorTestHarness<>(doFnOperator);
    if (keyed) {
        // we use a dummy key for the second input since it is considered to be broadcast
        testHarness = new KeyedTwoInputStreamOperatorTestHarness<>(doFnOperator, new StringKeySelector(), new DummyKeySelector(), BasicTypeInfo.STRING_TYPE_INFO);
    }
    testHarness.open();
    IntervalWindow firstWindow = new IntervalWindow(new Instant(0), new Instant(100));
    IntervalWindow secondWindow = new IntervalWindow(new Instant(0), new Instant(500));
    // test the keep of sideInputs events
    testHarness.processElement2(new StreamRecord<>(new RawUnionValue(1, valuesInWindow(ImmutableList.of("hello", "ciao"), new Instant(0), firstWindow))));
    testHarness.processElement2(new StreamRecord<>(new RawUnionValue(2, valuesInWindow(ImmutableList.of("foo", "bar"), new Instant(0), secondWindow))));
    // push in a regular elements
    WindowedValue<String> helloElement = valueInWindow("Hello", new Instant(0), firstWindow);
    WindowedValue<String> worldElement = valueInWindow("World", new Instant(1000), firstWindow);
    testHarness.processElement1(new StreamRecord<>(helloElement));
    testHarness.processElement1(new StreamRecord<>(worldElement));
    // test the keep of pushed-back events
    testHarness.processElement2(new StreamRecord<>(new RawUnionValue(1, valuesInWindow(ImmutableList.of("hello", "ciao"), new Instant(1000), firstWindow))));
    testHarness.processElement2(new StreamRecord<>(new RawUnionValue(2, valuesInWindow(ImmutableList.of("foo", "bar"), new Instant(1000), secondWindow))));
    assertThat(this.<String>stripStreamRecordFromWindowedValue(testHarness.getOutput()), contains(helloElement, worldElement));
    testHarness.close();
}

Example 9 with RawUnionValue

use of org.apache.beam.sdk.transforms.join.RawUnionValue in project beam by apache.

the class DoFnOperatorTest method testSideInputs.

void testSideInputs(boolean keyed) throws Exception {
    Coder<WindowedValue<String>> coder = WindowedValue.getValueOnlyCoder(StringUtf8Coder.of());
    TupleTag<String> outputTag = new TupleTag<>("main-output");
    ImmutableMap<Integer, PCollectionView<?>> sideInputMapping = ImmutableMap.<Integer, PCollectionView<?>>builder().put(1, view1).put(2, view2).build();
    Coder<String> keyCoder = StringUtf8Coder.of();
    KeySelector<WindowedValue<String>, ByteBuffer> keySelector = null;
    if (keyed) {
        keySelector = value -> FlinkKeyUtils.encodeKey(value.getValue(), keyCoder);
    }
    DoFnOperator<String, String> doFnOperator = new DoFnOperator<>(new IdentityDoFn<>(), "stepName", coder, Collections.emptyMap(), outputTag, Collections.emptyList(), new DoFnOperator.MultiOutputOutputManagerFactory<>(outputTag, coder, new SerializablePipelineOptions(FlinkPipelineOptions.defaults())), WindowingStrategy.of(FixedWindows.of(Duration.millis(100))), sideInputMapping, /* side-input mapping */
    ImmutableList.of(view1, view2), /* side inputs */
    FlinkPipelineOptions.defaults(), keyed ? keyCoder : null, keyed ? keySelector : null, DoFnSchemaInformation.create(), Collections.emptyMap());
    TwoInputStreamOperatorTestHarness<WindowedValue<String>, RawUnionValue, WindowedValue<String>> testHarness = new TwoInputStreamOperatorTestHarness<>(doFnOperator);
    if (keyed) {
        // we use a dummy key for the second input since it is considered to be broadcast
        testHarness = new KeyedTwoInputStreamOperatorTestHarness<>(doFnOperator, keySelector, null, new CoderTypeInformation<>(FlinkKeyUtils.ByteBufferCoder.of(), FlinkPipelineOptions.defaults()));
    }
    testHarness.open();
    IntervalWindow firstWindow = new IntervalWindow(new Instant(0), new Instant(100));
    IntervalWindow secondWindow = new IntervalWindow(new Instant(0), new Instant(500));
    // test the keep of sideInputs events
    testHarness.processElement2(new StreamRecord<>(new RawUnionValue(1, valuesInWindow(PCollectionViewTesting.materializeValuesFor(view1.getPipeline().getOptions(), View.asIterable(), "hello", "ciao"), new Instant(0), firstWindow))));
    testHarness.processElement2(new StreamRecord<>(new RawUnionValue(2, valuesInWindow(PCollectionViewTesting.materializeValuesFor(view2.getPipeline().getOptions(), View.asIterable(), "foo", "bar"), new Instant(0), secondWindow))));
    // push in a regular elements
    WindowedValue<String> helloElement = valueInWindow("Hello", new Instant(0), firstWindow);
    WindowedValue<String> worldElement = valueInWindow("World", new Instant(1000), firstWindow);
    testHarness.processElement1(new StreamRecord<>(helloElement));
    testHarness.processElement1(new StreamRecord<>(worldElement));
    // test the keep of pushed-back events
    testHarness.processElement2(new StreamRecord<>(new RawUnionValue(1, valuesInWindow(PCollectionViewTesting.materializeValuesFor(view1.getPipeline().getOptions(), View.asIterable(), "hello", "ciao"), new Instant(1000), firstWindow))));
    testHarness.processElement2(new StreamRecord<>(new RawUnionValue(2, valuesInWindow(PCollectionViewTesting.materializeValuesFor(view2.getPipeline().getOptions(), View.asIterable(), "foo", "bar"), new Instant(1000), secondWindow))));
    assertThat(stripStreamRecordFromWindowedValue(testHarness.getOutput()), contains(helloElement, worldElement));
    testHarness.close();
}

Example 10 with RawUnionValue

use of org.apache.beam.sdk.transforms.join.RawUnionValue in project beam by apache.

the class FlinkBatchPortablePipelineTranslator method translateExecutableStage.

private static <InputT> void translateExecutableStage(PTransformNode transform, RunnerApi.Pipeline pipeline, BatchTranslationContext context) {
    // TODO: Fail on splittable DoFns.
    // TODO: Special-case single outputs to avoid multiplexing PCollections.
    RunnerApi.Components components = pipeline.getComponents();
    Map<String, String> outputs = transform.getTransform().getOutputsMap();
    // Mapping from PCollection id to coder tag id.
    BiMap<String, Integer> outputMap = createOutputMap(outputs.values());
    // Collect all output Coders and create a UnionCoder for our tagged outputs.
    List<Coder<?>> unionCoders = Lists.newArrayList();
    // Enforce tuple tag sorting by union tag index.
    Map<String, Coder<WindowedValue<?>>> outputCoders = Maps.newHashMap();
    for (String collectionId : new TreeMap<>(outputMap.inverse()).values()) {
        PCollectionNode collectionNode = PipelineNode.pCollection(collectionId, components.getPcollectionsOrThrow(collectionId));
        Coder<WindowedValue<?>> coder;
        try {
            coder = (Coder) WireCoders.instantiateRunnerWireCoder(collectionNode, components);
        } catch (IOException e) {
            throw new RuntimeException(e);
        }
        outputCoders.put(collectionId, coder);
        unionCoders.add(coder);
    }
    UnionCoder unionCoder = UnionCoder.of(unionCoders);
    TypeInformation<RawUnionValue> typeInformation = new CoderTypeInformation<>(unionCoder, context.getPipelineOptions());
    RunnerApi.ExecutableStagePayload stagePayload;
    try {
        stagePayload = RunnerApi.ExecutableStagePayload.parseFrom(transform.getTransform().getSpec().getPayload());
    } catch (IOException e) {
        throw new RuntimeException(e);
    }
    String inputPCollectionId = stagePayload.getInput();
    Coder<WindowedValue<InputT>> windowedInputCoder = instantiateCoder(inputPCollectionId, components);
    DataSet<WindowedValue<InputT>> inputDataSet = context.getDataSetOrThrow(inputPCollectionId);
    final FlinkExecutableStageFunction<InputT> function = new FlinkExecutableStageFunction<>(transform.getTransform().getUniqueName(), context.getPipelineOptions(), stagePayload, context.getJobInfo(), outputMap, FlinkExecutableStageContextFactory.getInstance(), getWindowingStrategy(inputPCollectionId, components).getWindowFn().windowCoder(), windowedInputCoder);
    final String operatorName = generateNameFromStagePayload(stagePayload);
    final SingleInputUdfOperator taggedDataset;
    if (stagePayload.getUserStatesCount() > 0 || stagePayload.getTimersCount() > 0) {
        Coder valueCoder = ((WindowedValue.FullWindowedValueCoder) windowedInputCoder).getValueCoder();
        // Stateful stages are only allowed of KV input to be able to group on the key
        if (!(valueCoder instanceof KvCoder)) {
            throw new IllegalStateException(String.format(Locale.ENGLISH, "The element coder for stateful DoFn '%s' must be KvCoder but is: %s", inputPCollectionId, valueCoder.getClass().getSimpleName()));
        }
        Coder keyCoder = ((KvCoder) valueCoder).getKeyCoder();
        Grouping<WindowedValue<InputT>> groupedInput = inputDataSet.groupBy(new KvKeySelector<>(keyCoder));
        boolean requiresTimeSortedInput = requiresTimeSortedInput(stagePayload, false);
        if (requiresTimeSortedInput) {
            groupedInput = ((UnsortedGrouping<WindowedValue<InputT>>) groupedInput).sortGroup(WindowedValue::getTimestamp, Order.ASCENDING);
        }
        taggedDataset = new GroupReduceOperator<>(groupedInput, typeInformation, function, operatorName);
    } else {
        taggedDataset = new MapPartitionOperator<>(inputDataSet, typeInformation, function, operatorName);
    }
    for (SideInputId sideInputId : stagePayload.getSideInputsList()) {
        String collectionId = stagePayload.getComponents().getTransformsOrThrow(sideInputId.getTransformId()).getInputsOrThrow(sideInputId.getLocalName());
        // Register under the global PCollection name. Only ExecutableStageFunction needs to know the
        // mapping from local name to global name and how to translate the broadcast data to a state
        // API view.
        taggedDataset.withBroadcastSet(context.getDataSetOrThrow(collectionId), collectionId);
    }
    for (String collectionId : outputs.values()) {
        pruneOutput(taggedDataset, context, outputMap.get(collectionId), outputCoders.get(collectionId), collectionId);
    }
    if (outputs.isEmpty()) {
        // NOTE: After pipeline translation, we traverse the set of unconsumed PCollections and add a
        // no-op sink to each to make sure they are materialized by Flink. However, some SDK-executed
        // stages have no runner-visible output after fusion. We handle this case by adding a sink
        // here.
        taggedDataset.output(new DiscardingOutputFormat<>()).name("DiscardingOutput");
    }
}