Examples with Coder org.apache.beam.model.pipeline.v1.RunnerApi.Coder used on opensource projects

Example 26 with Coder

use of org.apache.beam.model.pipeline.v1.RunnerApi.Coder in project beam by apache.

the class FlinkStreamingPortablePipelineTranslator method translateFlatten.

private <T> void translateFlatten(String id, RunnerApi.Pipeline pipeline, StreamingTranslationContext context) {
    RunnerApi.PTransform transform = pipeline.getComponents().getTransformsOrThrow(id);
    Map<String, String> allInputs = transform.getInputsMap();
    if (allInputs.isEmpty()) {
        // create an empty dummy source to satisfy downstream operations
        // we cannot create an empty source in Flink, therefore we have to
        // add the flatMap that simply never forwards the single element
        long shutdownAfterIdleSourcesMs = context.getPipelineOptions().getShutdownSourcesAfterIdleMs();
        DataStreamSource<WindowedValue<byte[]>> dummySource = context.getExecutionEnvironment().addSource(new ImpulseSourceFunction(shutdownAfterIdleSourcesMs));
        DataStream<WindowedValue<T>> result = dummySource.<WindowedValue<T>>flatMap((s, collector) -> {
        // never return anything
        }).returns(new CoderTypeInformation<>(WindowedValue.getFullCoder((Coder<T>) VoidCoder.of(), GlobalWindow.Coder.INSTANCE), context.getPipelineOptions()));
        context.addDataStream(Iterables.getOnlyElement(transform.getOutputsMap().values()), result);
    } else {
        DataStream<T> result = null;
        // Determine DataStreams that we use as input several times. For those, we need to uniquify
        // input streams because Flink seems to swallow watermarks when we have a union of one and
        // the same stream.
        HashMultiset<DataStream<T>> inputCounts = HashMultiset.create();
        for (String input : allInputs.values()) {
            DataStream<T> current = context.getDataStreamOrThrow(input);
            inputCounts.add(current, 1);
        }
        for (String input : allInputs.values()) {
            DataStream<T> current = context.getDataStreamOrThrow(input);
            final int timesRequired = inputCounts.count(current);
            if (timesRequired > 1) {
                current = current.flatMap(new FlatMapFunction<T, T>() {

                    private static final long serialVersionUID = 1L;

                    @Override
                    public void flatMap(T t, Collector<T> collector) {
                        collector.collect(t);
                    }
                });
            }
            result = (result == null) ? current : result.union(current);
        }
        context.addDataStream(Iterables.getOnlyElement(transform.getOutputsMap().values()), result);
    }
}

Example 27 with Coder

use of org.apache.beam.model.pipeline.v1.RunnerApi.Coder in project beam by apache.

the class FlinkStreamingPortablePipelineTranslator method translateExecutableStage.

private <InputT, OutputT> void translateExecutableStage(String id, RunnerApi.Pipeline pipeline, StreamingTranslationContext context) {
    // TODO: Fail on splittable DoFns.
    // TODO: Special-case single outputs to avoid multiplexing PCollections.
    RunnerApi.Components components = pipeline.getComponents();
    RunnerApi.PTransform transform = components.getTransformsOrThrow(id);
    Map<String, String> outputs = transform.getOutputsMap();
    final RunnerApi.ExecutableStagePayload stagePayload;
    try {
        stagePayload = RunnerApi.ExecutableStagePayload.parseFrom(transform.getSpec().getPayload());
    } catch (IOException e) {
        throw new RuntimeException(e);
    }
    String inputPCollectionId = stagePayload.getInput();
    final TransformedSideInputs transformedSideInputs;
    if (stagePayload.getSideInputsCount() > 0) {
        transformedSideInputs = transformSideInputs(stagePayload, components, context);
    } else {
        transformedSideInputs = new TransformedSideInputs(Collections.emptyMap(), null);
    }
    Map<TupleTag<?>, OutputTag<WindowedValue<?>>> tagsToOutputTags = Maps.newLinkedHashMap();
    Map<TupleTag<?>, Coder<WindowedValue<?>>> tagsToCoders = Maps.newLinkedHashMap();
    // TODO: does it matter which output we designate as "main"
    final TupleTag<OutputT> mainOutputTag = outputs.isEmpty() ? null : new TupleTag(outputs.keySet().iterator().next());
    // associate output tags with ids, output manager uses these Integer ids to serialize state
    BiMap<String, Integer> outputIndexMap = createOutputMap(outputs.keySet());
    Map<String, Coder<WindowedValue<?>>> outputCoders = Maps.newHashMap();
    Map<TupleTag<?>, Integer> tagsToIds = Maps.newHashMap();
    Map<String, TupleTag<?>> collectionIdToTupleTag = Maps.newHashMap();
    // order output names for deterministic mapping
    for (String localOutputName : new TreeMap<>(outputIndexMap).keySet()) {
        String collectionId = outputs.get(localOutputName);
        Coder<WindowedValue<?>> windowCoder = (Coder) instantiateCoder(collectionId, components);
        outputCoders.put(localOutputName, windowCoder);
        TupleTag<?> tupleTag = new TupleTag<>(localOutputName);
        CoderTypeInformation<WindowedValue<?>> typeInformation = new CoderTypeInformation(windowCoder, context.getPipelineOptions());
        tagsToOutputTags.put(tupleTag, new OutputTag<>(localOutputName, typeInformation));
        tagsToCoders.put(tupleTag, windowCoder);
        tagsToIds.put(tupleTag, outputIndexMap.get(localOutputName));
        collectionIdToTupleTag.put(collectionId, tupleTag);
    }
    final SingleOutputStreamOperator<WindowedValue<OutputT>> outputStream;
    DataStream<WindowedValue<InputT>> inputDataStream = context.getDataStreamOrThrow(inputPCollectionId);
    CoderTypeInformation<WindowedValue<OutputT>> outputTypeInformation = !outputs.isEmpty() ? new CoderTypeInformation(outputCoders.get(mainOutputTag.getId()), context.getPipelineOptions()) : null;
    ArrayList<TupleTag<?>> additionalOutputTags = Lists.newArrayList();
    for (TupleTag<?> tupleTag : tagsToCoders.keySet()) {
        if (!mainOutputTag.getId().equals(tupleTag.getId())) {
            additionalOutputTags.add(tupleTag);
        }
    }
    final Coder<WindowedValue<InputT>> windowedInputCoder = instantiateCoder(inputPCollectionId, components);
    final boolean stateful = stagePayload.getUserStatesCount() > 0 || stagePayload.getTimersCount() > 0;
    final boolean hasSdfProcessFn = stagePayload.getComponents().getTransformsMap().values().stream().anyMatch(pTransform -> pTransform.getSpec().getUrn().equals(PTransformTranslation.SPLITTABLE_PROCESS_SIZED_ELEMENTS_AND_RESTRICTIONS_URN));
    Coder keyCoder = null;
    KeySelector<WindowedValue<InputT>, ?> keySelector = null;
    if (stateful || hasSdfProcessFn) {
        // Stateful/SDF stages are only allowed of KV input.
        Coder valueCoder = ((WindowedValue.FullWindowedValueCoder) windowedInputCoder).getValueCoder();
        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()));
        }
        if (stateful) {
            keyCoder = ((KvCoder) valueCoder).getKeyCoder();
            keySelector = new KvToByteBufferKeySelector(keyCoder, new SerializablePipelineOptions(context.getPipelineOptions()));
        } else {
            // as the key.
            if (!(((KvCoder) valueCoder).getKeyCoder() instanceof KvCoder)) {
                throw new IllegalStateException(String.format(Locale.ENGLISH, "The element coder for splittable DoFn '%s' must be KVCoder(KvCoder, DoubleCoder) but is: %s", inputPCollectionId, valueCoder.getClass().getSimpleName()));
            }
            keyCoder = ((KvCoder) ((KvCoder) valueCoder).getKeyCoder()).getKeyCoder();
            keySelector = new SdfByteBufferKeySelector(keyCoder, new SerializablePipelineOptions(context.getPipelineOptions()));
        }
        inputDataStream = inputDataStream.keyBy(keySelector);
    }
    DoFnOperator.MultiOutputOutputManagerFactory<OutputT> outputManagerFactory = new DoFnOperator.MultiOutputOutputManagerFactory<>(mainOutputTag, tagsToOutputTags, tagsToCoders, tagsToIds, new SerializablePipelineOptions(context.getPipelineOptions()));
    DoFnOperator<InputT, OutputT> doFnOperator = new ExecutableStageDoFnOperator<>(transform.getUniqueName(), windowedInputCoder, Collections.emptyMap(), mainOutputTag, additionalOutputTags, outputManagerFactory, transformedSideInputs.unionTagToView, new ArrayList<>(transformedSideInputs.unionTagToView.values()), getSideInputIdToPCollectionViewMap(stagePayload, components), context.getPipelineOptions(), stagePayload, context.getJobInfo(), FlinkExecutableStageContextFactory.getInstance(), collectionIdToTupleTag, getWindowingStrategy(inputPCollectionId, components), keyCoder, keySelector);
    final String operatorName = generateNameFromStagePayload(stagePayload);
    if (transformedSideInputs.unionTagToView.isEmpty()) {
        outputStream = inputDataStream.transform(operatorName, outputTypeInformation, doFnOperator);
    } else {
        DataStream<RawUnionValue> sideInputStream = transformedSideInputs.unionedSideInputs.broadcast();
        if (stateful || hasSdfProcessFn) {
            // We have to manually construct the two-input transform because we're not
            // allowed to have only one input keyed, normally. Since Flink 1.5.0 it's
            // possible to use the Broadcast State Pattern which provides a more elegant
            // way to process keyed main input with broadcast state, but it's not feasible
            // here because it breaks the DoFnOperator abstraction.
            TwoInputTransformation<WindowedValue<KV<?, InputT>>, RawUnionValue, WindowedValue<OutputT>> rawFlinkTransform = new TwoInputTransformation(inputDataStream.getTransformation(), sideInputStream.getTransformation(), transform.getUniqueName(), doFnOperator, outputTypeInformation, inputDataStream.getParallelism());
            rawFlinkTransform.setStateKeyType(((KeyedStream) inputDataStream).getKeyType());
            rawFlinkTransform.setStateKeySelectors(((KeyedStream) inputDataStream).getKeySelector(), null);
            outputStream = new SingleOutputStreamOperator(inputDataStream.getExecutionEnvironment(), // we have to cheat around the ctor being protected
            rawFlinkTransform) {
            };
        } else {
            outputStream = inputDataStream.connect(sideInputStream).transform(operatorName, outputTypeInformation, doFnOperator);
        }
    }
    // Assign a unique but consistent id to re-map operator state
    outputStream.uid(transform.getUniqueName());
    if (mainOutputTag != null) {
        context.addDataStream(outputs.get(mainOutputTag.getId()), outputStream);
    }
    for (TupleTag<?> tupleTag : additionalOutputTags) {
        context.addDataStream(outputs.get(tupleTag.getId()), outputStream.getSideOutput(tagsToOutputTags.get(tupleTag)));
    }
}

Example 28 with Coder

use of org.apache.beam.model.pipeline.v1.RunnerApi.Coder in project beam by apache.

the class SamzaDoFnRunners method createPortable.

/**
 * Create DoFnRunner for portable runner.
 */
@SuppressWarnings("unchecked")
public static <InT, FnOutT> DoFnRunner<InT, FnOutT> createPortable(String transformId, String bundleStateId, Coder<WindowedValue<InT>> windowedValueCoder, ExecutableStage executableStage, Map<?, PCollectionView<?>> sideInputMapping, SideInputHandler sideInputHandler, SamzaStoreStateInternals.Factory<?> nonKeyedStateInternalsFactory, SamzaTimerInternalsFactory<?> timerInternalsFactory, SamzaPipelineOptions pipelineOptions, DoFnRunners.OutputManager outputManager, StageBundleFactory stageBundleFactory, TupleTag<FnOutT> mainOutputTag, Map<String, TupleTag<?>> idToTupleTagMap, Context context, String transformFullName) {
    // storing events within a bundle in states
    final BagState<WindowedValue<InT>> bundledEventsBag = nonKeyedStateInternalsFactory.stateInternalsForKey(null).state(StateNamespaces.global(), StateTags.bag(bundleStateId, windowedValueCoder));
    final StateRequestHandler stateRequestHandler = SamzaStateRequestHandlers.of(transformId, context.getTaskContext(), pipelineOptions, executableStage, stageBundleFactory, (Map<RunnerApi.ExecutableStagePayload.SideInputId, PCollectionView<?>>) sideInputMapping, sideInputHandler);
    final SamzaExecutionContext executionContext = (SamzaExecutionContext) context.getApplicationContainerContext();
    final DoFnRunner<InT, FnOutT> underlyingRunner = new SdkHarnessDoFnRunner<>(timerInternalsFactory, WindowUtils.getWindowStrategy(executableStage.getInputPCollection().getId(), executableStage.getComponents()), outputManager, stageBundleFactory, idToTupleTagMap, bundledEventsBag, stateRequestHandler);
    return pipelineOptions.getEnableMetrics() ? DoFnRunnerWithMetrics.wrap(underlyingRunner, executionContext.getMetricsContainer(), transformFullName) : underlyingRunner;
}

Example 29 with Coder

use of org.apache.beam.model.pipeline.v1.RunnerApi.Coder in project beam by apache.

the class DataflowPipelineTranslatorTest method testStreamingSplittableParDoTranslation.

/**
 * Smoke test to fail fast if translation of a splittable ParDo in streaming breaks.
 */
@Test
public void testStreamingSplittableParDoTranslation() throws Exception {
    DataflowPipelineOptions options = buildPipelineOptions();
    DataflowRunner runner = DataflowRunner.fromOptions(options);
    options.setStreaming(true);
    DataflowPipelineTranslator translator = DataflowPipelineTranslator.fromOptions(options);
    Pipeline pipeline = Pipeline.create(options);
    PCollection<String> windowedInput = pipeline.apply(Create.of("a")).apply(Window.into(FixedWindows.of(Duration.standardMinutes(1))));
    windowedInput.apply(ParDo.of(new TestSplittableFn()));
    runner.replaceV1Transforms(pipeline);
    SdkComponents sdkComponents = createSdkComponents(options);
    RunnerApi.Pipeline pipelineProto = PipelineTranslation.toProto(pipeline, sdkComponents, true);
    Job job = translator.translate(pipeline, pipelineProto, sdkComponents, runner, Collections.emptyList()).getJob();
    // The job should contain a SplittableParDo.ProcessKeyedElements step, translated as
    // "SplittableProcessKeyed".
    List<Step> steps = job.getSteps();
    Step processKeyedStep = null;
    for (Step step : steps) {
        if ("SplittableProcessKeyed".equals(step.getKind())) {
            assertNull(processKeyedStep);
            processKeyedStep = step;
        }
    }
    assertNotNull(processKeyedStep);
    @SuppressWarnings({ "unchecked", "rawtypes" }) DoFnInfo<String, Integer> fnInfo = (DoFnInfo<String, Integer>) SerializableUtils.deserializeFromByteArray(jsonStringToByteArray(getString(processKeyedStep.getProperties(), PropertyNames.SERIALIZED_FN)), "DoFnInfo");
    assertThat(fnInfo.getDoFn(), instanceOf(TestSplittableFn.class));
    assertThat(fnInfo.getWindowingStrategy().getWindowFn(), Matchers.<WindowFn>equalTo(FixedWindows.of(Duration.standardMinutes(1))));
    assertThat(fnInfo.getInputCoder(), instanceOf(StringUtf8Coder.class));
    Coder<?> restrictionCoder = CloudObjects.coderFromCloudObject((CloudObject) Structs.getObject(processKeyedStep.getProperties(), PropertyNames.RESTRICTION_CODER));
    assertEquals(KvCoder.of(SerializableCoder.of(OffsetRange.class), VoidCoder.of()), restrictionCoder);
}

Example 30 with Coder

use of org.apache.beam.model.pipeline.v1.RunnerApi.Coder in project beam by apache.

the class ProcessBundleHandlerTest method setupProcessBundleHandlerForSimpleRecordingDoFn.

private ProcessBundleHandler setupProcessBundleHandlerForSimpleRecordingDoFn(List<String> dataOutput, List<Timers> timerOutput, boolean enableOutputEmbedding) throws Exception {
    DoFnWithExecutionInformation doFnWithExecutionInformation = DoFnWithExecutionInformation.of(new SimpleDoFn(), SimpleDoFn.MAIN_OUTPUT_TAG, Collections.emptyMap(), DoFnSchemaInformation.create());
    RunnerApi.FunctionSpec functionSpec = RunnerApi.FunctionSpec.newBuilder().setUrn(ParDoTranslation.CUSTOM_JAVA_DO_FN_URN).setPayload(ByteString.copyFrom(SerializableUtils.serializeToByteArray(doFnWithExecutionInformation))).build();
    RunnerApi.ParDoPayload parDoPayload = ParDoPayload.newBuilder().setDoFn(functionSpec).putTimerFamilySpecs("tfs-" + SimpleDoFn.TIMER_FAMILY_ID, TimerFamilySpec.newBuilder().setTimeDomain(RunnerApi.TimeDomain.Enum.EVENT_TIME).setTimerFamilyCoderId("timer-coder").build()).build();
    BeamFnApi.ProcessBundleDescriptor processBundleDescriptor = ProcessBundleDescriptor.newBuilder().putTransforms("2L", PTransform.newBuilder().setSpec(FunctionSpec.newBuilder().setUrn(DATA_INPUT_URN).build()).putOutputs("2L-output", "2L-output-pc").build()).putTransforms("3L", PTransform.newBuilder().setSpec(FunctionSpec.newBuilder().setUrn(PTransformTranslation.PAR_DO_TRANSFORM_URN).setPayload(parDoPayload.toByteString())).putInputs("3L-input", "2L-output-pc").build()).putPcollections("2L-output-pc", PCollection.newBuilder().setWindowingStrategyId("window-strategy").setCoderId("2L-output-coder").setIsBounded(IsBounded.Enum.BOUNDED).build()).putWindowingStrategies("window-strategy", WindowingStrategy.newBuilder().setWindowCoderId("window-strategy-coder").setWindowFn(FunctionSpec.newBuilder().setUrn("beam:window_fn:global_windows:v1")).setOutputTime(OutputTime.Enum.END_OF_WINDOW).setAccumulationMode(AccumulationMode.Enum.ACCUMULATING).setTrigger(Trigger.newBuilder().setAlways(Always.getDefaultInstance())).setClosingBehavior(ClosingBehavior.Enum.EMIT_ALWAYS).setOnTimeBehavior(OnTimeBehavior.Enum.FIRE_ALWAYS).build()).setTimerApiServiceDescriptor(ApiServiceDescriptor.newBuilder().setUrl("url").build()).putCoders("string_coder", CoderTranslation.toProto(StringUtf8Coder.of()).getCoder()).putCoders("2L-output-coder", Coder.newBuilder().setSpec(FunctionSpec.newBuilder().setUrn(ModelCoders.KV_CODER_URN).build()).addComponentCoderIds("string_coder").addComponentCoderIds("string_coder").build()).putCoders("window-strategy-coder", Coder.newBuilder().setSpec(FunctionSpec.newBuilder().setUrn(ModelCoders.GLOBAL_WINDOW_CODER_URN).build()).build()).putCoders("timer-coder", Coder.newBuilder().setSpec(FunctionSpec.newBuilder().setUrn(ModelCoders.TIMER_CODER_URN)).addComponentCoderIds("string_coder").addComponentCoderIds("window-strategy-coder").build()).build();
    Map<String, BeamFnApi.ProcessBundleDescriptor> fnApiRegistry = ImmutableMap.of("1L", processBundleDescriptor);
    Map<String, PTransformRunnerFactory> urnToPTransformRunnerFactoryMap = Maps.newHashMap(REGISTERED_RUNNER_FACTORIES);
    urnToPTransformRunnerFactoryMap.put(DATA_INPUT_URN, (PTransformRunnerFactory<Object>) (context) -> {
        context.addIncomingDataEndpoint(ApiServiceDescriptor.getDefaultInstance(), KvCoder.of(StringUtf8Coder.of(), StringUtf8Coder.of()), (input) -> {
            dataOutput.add(input.getValue());
        });
        return null;
    });
    Mockito.doAnswer((invocation) -> new BeamFnDataOutboundAggregator(PipelineOptionsFactory.create(), invocation.getArgument(1), new StreamObserver<Elements>() {

        @Override
        public void onNext(Elements elements) {
            for (Timers timer : elements.getTimersList()) {
                timerOutput.addAll(elements.getTimersList());
            }
        }

        @Override
        public void onError(Throwable throwable) {
        }

        @Override
        public void onCompleted() {
        }
    }, invocation.getArgument(2))).when(beamFnDataClient).createOutboundAggregator(any(), any(), anyBoolean());
    return new ProcessBundleHandler(PipelineOptionsFactory.create(), enableOutputEmbedding ? Collections.singleton(BeamUrns.getUrn(StandardRunnerProtocols.Enum.CONTROL_RESPONSE_ELEMENTS_EMBEDDING)) : Collections.emptySet(), fnApiRegistry::get, beamFnDataClient, null, /* beamFnStateClient */
    null, /* finalizeBundleHandler */
    new ShortIdMap(), urnToPTransformRunnerFactoryMap, Caches.noop(), new BundleProcessorCache());
}