Search in sources :

Example 6 with WindowedValueCoder

use of org.apache.beam.sdk.util.WindowedValue.WindowedValueCoder in project beam by apache.

the class CreateExecutableStageNodeFunction method apply.

@Override
public Node apply(MutableNetwork<Node, Edge> input) {
    for (Node node : input.nodes()) {
        if (node instanceof RemoteGrpcPortNode || node instanceof ParallelInstructionNode || node instanceof InstructionOutputNode) {
            continue;
        }
        throw new IllegalArgumentException(String.format("Network contains unknown type of node: %s", input));
    }
    // Fix all non output nodes to have named edges.
    for (Node node : input.nodes()) {
        if (node instanceof InstructionOutputNode) {
            continue;
        }
        for (Node successor : input.successors(node)) {
            for (Edge edge : input.edgesConnecting(node, successor)) {
                if (edge instanceof DefaultEdge) {
                    input.removeEdge(edge);
                    input.addEdge(node, successor, MultiOutputInfoEdge.create(new MultiOutputInfo().setTag(idGenerator.getId())));
                }
            }
        }
    }
    RunnerApi.Components.Builder componentsBuilder = RunnerApi.Components.newBuilder();
    componentsBuilder.mergeFrom(this.pipeline.getComponents());
    // Default to use the Java environment if pipeline doesn't have environment specified.
    if (pipeline.getComponents().getEnvironmentsMap().isEmpty()) {
        String envId = Environments.JAVA_SDK_HARNESS_ENVIRONMENT.getUrn() + idGenerator.getId();
        componentsBuilder.putEnvironments(envId, Environments.JAVA_SDK_HARNESS_ENVIRONMENT);
    }
    // By default, use GlobalWindow for all languages.
    // For java, if there is a IntervalWindowCoder, then use FixedWindow instead.
    // TODO: should get real WindowingStategy from pipeline proto.
    String globalWindowingStrategyId = "generatedGlobalWindowingStrategy" + idGenerator.getId();
    String intervalWindowEncodingWindowingStrategyId = "generatedIntervalWindowEncodingWindowingStrategy" + idGenerator.getId();
    SdkComponents sdkComponents = SdkComponents.create(pipeline.getComponents(), null);
    try {
        registerWindowingStrategy(globalWindowingStrategyId, WindowingStrategy.globalDefault(), componentsBuilder, sdkComponents);
        registerWindowingStrategy(intervalWindowEncodingWindowingStrategyId, WindowingStrategy.of(FixedWindows.of(Duration.standardSeconds(1))), componentsBuilder, sdkComponents);
    } catch (IOException exc) {
        throw new RuntimeException("Could not convert default windowing stratey to proto", exc);
    }
    Map<Node, String> nodesToPCollections = new HashMap<>();
    ImmutableMap.Builder<String, NameContext> ptransformIdToNameContexts = ImmutableMap.builder();
    ImmutableMap.Builder<String, Iterable<SideInputInfo>> ptransformIdToSideInputInfos = ImmutableMap.builder();
    ImmutableMap.Builder<String, Iterable<PCollectionView<?>>> ptransformIdToPCollectionViews = ImmutableMap.builder();
    // A field of ExecutableStage which includes the PCollection goes to worker side.
    Set<PCollectionNode> executableStageOutputs = new HashSet<>();
    // A field of ExecutableStage which includes the PCollection goes to runner side.
    Set<PCollectionNode> executableStageInputs = new HashSet<>();
    for (InstructionOutputNode node : Iterables.filter(input.nodes(), InstructionOutputNode.class)) {
        InstructionOutput instructionOutput = node.getInstructionOutput();
        String coderId = "generatedCoder" + idGenerator.getId();
        String windowingStrategyId;
        try (ByteString.Output output = ByteString.newOutput()) {
            try {
                Coder<?> javaCoder = CloudObjects.coderFromCloudObject(CloudObject.fromSpec(instructionOutput.getCodec()));
                Coder<?> elementCoder = ((WindowedValueCoder<?>) javaCoder).getValueCoder();
                sdkComponents.registerCoder(elementCoder);
                RunnerApi.Coder coderProto = CoderTranslation.toProto(elementCoder, sdkComponents);
                componentsBuilder.putCoders(coderId, coderProto);
                // For now, Dataflow runner harness only deal with FixedWindow.
                if (javaCoder instanceof FullWindowedValueCoder) {
                    FullWindowedValueCoder<?> windowedValueCoder = (FullWindowedValueCoder<?>) javaCoder;
                    Coder<?> windowCoder = windowedValueCoder.getWindowCoder();
                    if (windowCoder instanceof IntervalWindowCoder) {
                        windowingStrategyId = intervalWindowEncodingWindowingStrategyId;
                    } else if (windowCoder instanceof GlobalWindow.Coder) {
                        windowingStrategyId = globalWindowingStrategyId;
                    } else {
                        throw new UnsupportedOperationException(String.format("Dataflow portable runner harness doesn't support windowing with %s", windowCoder));
                    }
                } else {
                    throw new UnsupportedOperationException("Dataflow portable runner harness only supports FullWindowedValueCoder");
                }
            } catch (IOException e) {
                throw new IllegalArgumentException(String.format("Unable to encode coder %s for output %s", instructionOutput.getCodec(), instructionOutput), e);
            } catch (Exception e) {
                // Coder probably wasn't a java coder
                OBJECT_MAPPER.writeValue(output, instructionOutput.getCodec());
                componentsBuilder.putCoders(coderId, RunnerApi.Coder.newBuilder().setSpec(RunnerApi.FunctionSpec.newBuilder().setPayload(output.toByteString())).build());
                // For non-java coder, hope it's GlobalWindows by default.
                // TODO(BEAM-6231): Actually discover the right windowing strategy.
                windowingStrategyId = globalWindowingStrategyId;
            }
        } catch (IOException e) {
            throw new IllegalArgumentException(String.format("Unable to encode coder %s for output %s", instructionOutput.getCodec(), instructionOutput), e);
        }
        // TODO(BEAM-6275): Set correct IsBounded on generated PCollections
        String pcollectionId = node.getPcollectionId();
        RunnerApi.PCollection pCollection = RunnerApi.PCollection.newBuilder().setCoderId(coderId).setWindowingStrategyId(windowingStrategyId).setIsBounded(RunnerApi.IsBounded.Enum.BOUNDED).build();
        nodesToPCollections.put(node, pcollectionId);
        componentsBuilder.putPcollections(pcollectionId, pCollection);
        // is set
        if (isExecutableStageOutputPCollection(input, node)) {
            executableStageOutputs.add(PipelineNode.pCollection(pcollectionId, pCollection));
        }
        if (isExecutableStageInputPCollection(input, node)) {
            executableStageInputs.add(PipelineNode.pCollection(pcollectionId, pCollection));
        }
    }
    componentsBuilder.putAllCoders(sdkComponents.toComponents().getCodersMap());
    Set<PTransformNode> executableStageTransforms = new HashSet<>();
    Set<TimerReference> executableStageTimers = new HashSet<>();
    List<UserStateId> userStateIds = new ArrayList<>();
    Set<SideInputReference> executableStageSideInputs = new HashSet<>();
    for (ParallelInstructionNode node : Iterables.filter(input.nodes(), ParallelInstructionNode.class)) {
        ImmutableMap.Builder<String, PCollectionNode> sideInputIds = ImmutableMap.builder();
        ParallelInstruction parallelInstruction = node.getParallelInstruction();
        String ptransformId = "generatedPtransform" + idGenerator.getId();
        ptransformIdToNameContexts.put(ptransformId, NameContext.create(null, parallelInstruction.getOriginalName(), parallelInstruction.getSystemName(), parallelInstruction.getName()));
        RunnerApi.PTransform.Builder pTransform = RunnerApi.PTransform.newBuilder();
        RunnerApi.FunctionSpec.Builder transformSpec = RunnerApi.FunctionSpec.newBuilder();
        List<String> timerIds = new ArrayList<>();
        if (parallelInstruction.getParDo() != null) {
            ParDoInstruction parDoInstruction = parallelInstruction.getParDo();
            CloudObject userFnSpec = CloudObject.fromSpec(parDoInstruction.getUserFn());
            String userFnClassName = userFnSpec.getClassName();
            if (userFnClassName.equals("CombineValuesFn") || userFnClassName.equals("KeyedCombineFn")) {
                transformSpec = transformCombineValuesFnToFunctionSpec(userFnSpec);
                ptransformIdToPCollectionViews.put(ptransformId, Collections.emptyList());
            } else {
                String parDoPTransformId = getString(userFnSpec, PropertyNames.SERIALIZED_FN);
                RunnerApi.PTransform parDoPTransform = pipeline.getComponents().getTransformsOrDefault(parDoPTransformId, null);
                // TODO: only the non-null branch should exist; for migration ease only
                if (parDoPTransform != null) {
                    checkArgument(parDoPTransform.getSpec().getUrn().equals(PTransformTranslation.PAR_DO_TRANSFORM_URN), "Found transform \"%s\" for ParallelDo instruction, " + " but that transform had unexpected URN \"%s\" (expected \"%s\")", parDoPTransformId, parDoPTransform.getSpec().getUrn(), PTransformTranslation.PAR_DO_TRANSFORM_URN);
                    RunnerApi.ParDoPayload parDoPayload;
                    try {
                        parDoPayload = RunnerApi.ParDoPayload.parseFrom(parDoPTransform.getSpec().getPayload());
                    } catch (InvalidProtocolBufferException exc) {
                        throw new RuntimeException("ParDo did not have a ParDoPayload", exc);
                    }
                    // user timers and user state.
                    for (Map.Entry<String, RunnerApi.TimerFamilySpec> entry : parDoPayload.getTimerFamilySpecsMap().entrySet()) {
                        timerIds.add(entry.getKey());
                    }
                    for (Map.Entry<String, RunnerApi.StateSpec> entry : parDoPayload.getStateSpecsMap().entrySet()) {
                        UserStateId.Builder builder = UserStateId.newBuilder();
                        builder.setTransformId(parDoPTransformId);
                        builder.setLocalName(entry.getKey());
                        userStateIds.add(builder.build());
                    }
                    // To facilitate the creation of Set executableStageSideInputs.
                    for (String sideInputTag : parDoPayload.getSideInputsMap().keySet()) {
                        String sideInputPCollectionId = parDoPTransform.getInputsOrThrow(sideInputTag);
                        RunnerApi.PCollection sideInputPCollection = pipeline.getComponents().getPcollectionsOrThrow(sideInputPCollectionId);
                        pTransform.putInputs(sideInputTag, sideInputPCollectionId);
                        PCollectionNode pCollectionNode = PipelineNode.pCollection(sideInputPCollectionId, sideInputPCollection);
                        sideInputIds.put(sideInputTag, pCollectionNode);
                    }
                    // To facilitate the creation of Map(ptransformId -> pCollectionView), which is
                    // required by constructing an ExecutableStageNode.
                    ImmutableList.Builder<PCollectionView<?>> pcollectionViews = ImmutableList.builder();
                    for (Map.Entry<String, RunnerApi.SideInput> sideInputEntry : parDoPayload.getSideInputsMap().entrySet()) {
                        pcollectionViews.add(RegisterNodeFunction.transformSideInputForRunner(pipeline, parDoPTransform, sideInputEntry.getKey(), sideInputEntry.getValue()));
                    }
                    ptransformIdToPCollectionViews.put(ptransformId, pcollectionViews.build());
                    transformSpec.setUrn(PTransformTranslation.PAR_DO_TRANSFORM_URN).setPayload(parDoPayload.toByteString());
                } else {
                    // legacy path - bytes are the FunctionSpec's payload field, basically, and
                    // SDKs expect it in the PTransform's payload field
                    byte[] userFnBytes = getBytes(userFnSpec, PropertyNames.SERIALIZED_FN);
                    transformSpec.setUrn(ParDoTranslation.CUSTOM_JAVA_DO_FN_URN).setPayload(ByteString.copyFrom(userFnBytes));
                }
                if (parDoInstruction.getSideInputs() != null) {
                    ptransformIdToSideInputInfos.put(ptransformId, forSideInputInfos(parDoInstruction.getSideInputs(), true));
                }
            }
        } else if (parallelInstruction.getRead() != null) {
            ReadInstruction readInstruction = parallelInstruction.getRead();
            CloudObject sourceSpec = CloudObject.fromSpec(CloudSourceUtils.flattenBaseSpecs(readInstruction.getSource()).getSpec());
            // TODO: Need to plumb through the SDK specific function spec.
            transformSpec.setUrn(JAVA_SOURCE_URN);
            try {
                byte[] serializedSource = Base64.getDecoder().decode(getString(sourceSpec, SERIALIZED_SOURCE));
                ByteString sourceByteString = ByteString.copyFrom(serializedSource);
                transformSpec.setPayload(sourceByteString);
            } catch (Exception e) {
                throw new IllegalArgumentException(String.format("Unable to process Read %s", parallelInstruction), e);
            }
        } else if (parallelInstruction.getFlatten() != null) {
            transformSpec.setUrn(PTransformTranslation.FLATTEN_TRANSFORM_URN);
        } else {
            throw new IllegalArgumentException(String.format("Unknown type of ParallelInstruction %s", parallelInstruction));
        }
        // predecessor in a ParDo. This PCollection is called the "main input".
        for (Node predecessorOutput : input.predecessors(node)) {
            pTransform.putInputs("generatedInput" + idGenerator.getId(), nodesToPCollections.get(predecessorOutput));
        }
        for (Edge edge : input.outEdges(node)) {
            Node nodeOutput = input.incidentNodes(edge).target();
            MultiOutputInfoEdge edge2 = (MultiOutputInfoEdge) edge;
            pTransform.putOutputs(edge2.getMultiOutputInfo().getTag(), nodesToPCollections.get(nodeOutput));
        }
        pTransform.setSpec(transformSpec);
        PTransformNode pTransformNode = PipelineNode.pTransform(ptransformId, pTransform.build());
        executableStageTransforms.add(pTransformNode);
        for (String timerId : timerIds) {
            executableStageTimers.add(TimerReference.of(pTransformNode, timerId));
        }
        ImmutableMap<String, PCollectionNode> sideInputIdToPCollectionNodes = sideInputIds.build();
        for (String sideInputTag : sideInputIdToPCollectionNodes.keySet()) {
            SideInputReference sideInputReference = SideInputReference.of(pTransformNode, sideInputTag, sideInputIdToPCollectionNodes.get(sideInputTag));
            executableStageSideInputs.add(sideInputReference);
        }
        executableStageTransforms.add(pTransformNode);
    }
    if (executableStageInputs.size() != 1) {
        throw new UnsupportedOperationException("ExecutableStage only support one input PCollection");
    }
    PCollectionNode executableInput = executableStageInputs.iterator().next();
    RunnerApi.Components executableStageComponents = componentsBuilder.build();
    // Get Environment from ptransform, otherwise, use JAVA_SDK_HARNESS_ENVIRONMENT as default.
    Environment executableStageEnv = getEnvironmentFromPTransform(executableStageComponents, executableStageTransforms);
    if (executableStageEnv == null) {
        executableStageEnv = Environments.JAVA_SDK_HARNESS_ENVIRONMENT;
    }
    Set<UserStateReference> executableStageUserStateReference = new HashSet<>();
    for (UserStateId userStateId : userStateIds) {
        executableStageUserStateReference.add(UserStateReference.fromUserStateId(userStateId, executableStageComponents));
    }
    ExecutableStage executableStage = ImmutableExecutableStage.ofFullComponents(executableStageComponents, executableStageEnv, executableInput, executableStageSideInputs, executableStageUserStateReference, executableStageTimers, executableStageTransforms, executableStageOutputs, DEFAULT_WIRE_CODER_SETTINGS);
    return ExecutableStageNode.create(executableStage, ptransformIdToNameContexts.build(), ptransformIdToSideInputInfos.build(), ptransformIdToPCollectionViews.build());
}
Also used : HashMap(java.util.HashMap) MultiOutputInfo(com.google.api.services.dataflow.model.MultiOutputInfo) ByteString(org.apache.beam.vendor.grpc.v1p43p2.com.google.protobuf.ByteString) PCollectionNode(org.apache.beam.runners.core.construction.graph.PipelineNode.PCollectionNode) InstructionOutputNode(org.apache.beam.runners.dataflow.worker.graph.Nodes.InstructionOutputNode) PTransformNode(org.apache.beam.runners.core.construction.graph.PipelineNode.PTransformNode) ParallelInstructionNode(org.apache.beam.runners.dataflow.worker.graph.Nodes.ParallelInstructionNode) Node(org.apache.beam.runners.dataflow.worker.graph.Nodes.Node) ExecutableStageNode(org.apache.beam.runners.dataflow.worker.graph.Nodes.ExecutableStageNode) PipelineNode(org.apache.beam.runners.core.construction.graph.PipelineNode) RemoteGrpcPortNode(org.apache.beam.runners.dataflow.worker.graph.Nodes.RemoteGrpcPortNode) InstructionOutput(com.google.api.services.dataflow.model.InstructionOutput) PTransformNode(org.apache.beam.runners.core.construction.graph.PipelineNode.PTransformNode) ArrayList(java.util.ArrayList) Structs.getString(org.apache.beam.runners.dataflow.util.Structs.getString) ByteString(org.apache.beam.vendor.grpc.v1p43p2.com.google.protobuf.ByteString) RehydratedComponents(org.apache.beam.runners.core.construction.RehydratedComponents) SdkComponents(org.apache.beam.runners.core.construction.SdkComponents) RunnerApi(org.apache.beam.model.pipeline.v1.RunnerApi) SideInputReference(org.apache.beam.runners.core.construction.graph.SideInputReference) ImmutableExecutableStage(org.apache.beam.runners.core.construction.graph.ImmutableExecutableStage) ExecutableStage(org.apache.beam.runners.core.construction.graph.ExecutableStage) HashSet(java.util.HashSet) DefaultEdge(org.apache.beam.runners.dataflow.worker.graph.Edges.DefaultEdge) MultiOutputInfoEdge(org.apache.beam.runners.dataflow.worker.graph.Edges.MultiOutputInfoEdge) ImmutableMap(org.apache.beam.vendor.guava.v26_0_jre.com.google.common.collect.ImmutableMap) ParDoInstruction(com.google.api.services.dataflow.model.ParDoInstruction) PCollectionView(org.apache.beam.sdk.values.PCollectionView) Environment(org.apache.beam.model.pipeline.v1.RunnerApi.Environment) GlobalWindow(org.apache.beam.sdk.transforms.windowing.GlobalWindow) ImmutableMap(org.apache.beam.vendor.guava.v26_0_jre.com.google.common.collect.ImmutableMap) Map(java.util.Map) HashMap(java.util.HashMap) IntervalWindowCoder(org.apache.beam.sdk.transforms.windowing.IntervalWindow.IntervalWindowCoder) RemoteGrpcPortNode(org.apache.beam.runners.dataflow.worker.graph.Nodes.RemoteGrpcPortNode) TimerReference(org.apache.beam.runners.core.construction.graph.TimerReference) ImmutableList(org.apache.beam.vendor.guava.v26_0_jre.com.google.common.collect.ImmutableList) ParallelInstructionNode(org.apache.beam.runners.dataflow.worker.graph.Nodes.ParallelInstructionNode) SdkComponents(org.apache.beam.runners.core.construction.SdkComponents) ReadInstruction(com.google.api.services.dataflow.model.ReadInstruction) InstructionOutputNode(org.apache.beam.runners.dataflow.worker.graph.Nodes.InstructionOutputNode) FullWindowedValueCoder(org.apache.beam.sdk.util.WindowedValue.FullWindowedValueCoder) NameContext(org.apache.beam.runners.dataflow.worker.counters.NameContext) InvalidProtocolBufferException(org.apache.beam.vendor.grpc.v1p43p2.com.google.protobuf.InvalidProtocolBufferException) IOException(java.io.IOException) PCollectionNode(org.apache.beam.runners.core.construction.graph.PipelineNode.PCollectionNode) InvalidProtocolBufferException(org.apache.beam.vendor.grpc.v1p43p2.com.google.protobuf.InvalidProtocolBufferException) IOException(java.io.IOException) ParallelInstruction(com.google.api.services.dataflow.model.ParallelInstruction) WindowedValueCoder(org.apache.beam.sdk.util.WindowedValue.WindowedValueCoder) FullWindowedValueCoder(org.apache.beam.sdk.util.WindowedValue.FullWindowedValueCoder) CloudObject(org.apache.beam.runners.dataflow.util.CloudObject) UserStateId(org.apache.beam.model.pipeline.v1.RunnerApi.ExecutableStagePayload.UserStateId) UserStateReference(org.apache.beam.runners.core.construction.graph.UserStateReference) Edge(org.apache.beam.runners.dataflow.worker.graph.Edges.Edge) MultiOutputInfoEdge(org.apache.beam.runners.dataflow.worker.graph.Edges.MultiOutputInfoEdge) DefaultEdge(org.apache.beam.runners.dataflow.worker.graph.Edges.DefaultEdge)

Example 7 with WindowedValueCoder

use of org.apache.beam.sdk.util.WindowedValue.WindowedValueCoder in project beam by apache.

the class FlinkStreamingPortablePipelineTranslator method getSideInputIdToPCollectionViewMap.

private static LinkedHashMap<RunnerApi.ExecutableStagePayload.SideInputId, PCollectionView<?>> getSideInputIdToPCollectionViewMap(RunnerApi.ExecutableStagePayload stagePayload, RunnerApi.Components components) {
    RehydratedComponents rehydratedComponents = RehydratedComponents.forComponents(components);
    LinkedHashMap<RunnerApi.ExecutableStagePayload.SideInputId, PCollectionView<?>> sideInputs = new LinkedHashMap<>();
    // for PCollectionView compatibility, not used to transform materialization
    ViewFn<Iterable<WindowedValue<?>>, ?> viewFn = (ViewFn) new PCollectionViews.MultimapViewFn<>((PCollectionViews.TypeDescriptorSupplier<Iterable<WindowedValue<Void>>>) () -> TypeDescriptors.iterables(new TypeDescriptor<WindowedValue<Void>>() {
    }), (PCollectionViews.TypeDescriptorSupplier<Void>) TypeDescriptors::voids);
    for (RunnerApi.ExecutableStagePayload.SideInputId sideInputId : stagePayload.getSideInputsList()) {
        // TODO: local name is unique as long as only one transform with side input can be within a
        // stage
        String sideInputTag = sideInputId.getLocalName();
        String collectionId = components.getTransformsOrThrow(sideInputId.getTransformId()).getInputsOrThrow(sideInputId.getLocalName());
        RunnerApi.WindowingStrategy windowingStrategyProto = components.getWindowingStrategiesOrThrow(components.getPcollectionsOrThrow(collectionId).getWindowingStrategyId());
        final WindowingStrategy<?, ?> windowingStrategy;
        try {
            windowingStrategy = WindowingStrategyTranslation.fromProto(windowingStrategyProto, rehydratedComponents);
        } catch (InvalidProtocolBufferException e) {
            throw new IllegalStateException(String.format("Unable to hydrate side input windowing strategy %s.", windowingStrategyProto), e);
        }
        Coder<WindowedValue<Object>> coder = instantiateCoder(collectionId, components);
        // side input materialization via GBK (T -> Iterable<T>)
        WindowedValueCoder wvCoder = (WindowedValueCoder) coder;
        coder = wvCoder.withValueCoder(IterableCoder.of(wvCoder.getValueCoder()));
        sideInputs.put(sideInputId, new RunnerPCollectionView<>(null, new TupleTag<>(sideInputTag), viewFn, // TODO: support custom mapping fn
        windowingStrategy.getWindowFn().getDefaultWindowMappingFn(), windowingStrategy, coder));
    }
    return sideInputs;
}
Also used : TupleTag(org.apache.beam.sdk.values.TupleTag) LinkedHashMap(java.util.LinkedHashMap) RunnerApi(org.apache.beam.model.pipeline.v1.RunnerApi) ViewFn(org.apache.beam.sdk.transforms.ViewFn) WindowedValue(org.apache.beam.sdk.util.WindowedValue) PCollectionViews(org.apache.beam.sdk.values.PCollectionViews) InvalidProtocolBufferException(org.apache.beam.vendor.grpc.v1p43p2.com.google.protobuf.InvalidProtocolBufferException) TypeDescriptors(org.apache.beam.sdk.values.TypeDescriptors) RunnerPCollectionView(org.apache.beam.runners.core.construction.RunnerPCollectionView) PCollectionView(org.apache.beam.sdk.values.PCollectionView) WindowedValueCoder(org.apache.beam.sdk.util.WindowedValue.WindowedValueCoder) TypeDescriptor(org.apache.beam.sdk.values.TypeDescriptor) RehydratedComponents(org.apache.beam.runners.core.construction.RehydratedComponents)

Example 8 with WindowedValueCoder

use of org.apache.beam.sdk.util.WindowedValue.WindowedValueCoder in project beam by apache.

the class FlinkStreamingPortablePipelineTranslator method transformSideInputs.

private TransformedSideInputs transformSideInputs(RunnerApi.ExecutableStagePayload stagePayload, RunnerApi.Components components, StreamingTranslationContext context) {
    LinkedHashMap<RunnerApi.ExecutableStagePayload.SideInputId, PCollectionView<?>> sideInputs = getSideInputIdToPCollectionViewMap(stagePayload, components);
    Map<TupleTag<?>, Integer> tagToIntMapping = new HashMap<>();
    Map<Integer, PCollectionView<?>> intToViewMapping = new HashMap<>();
    List<WindowedValueCoder<KV<Void, Object>>> kvCoders = new ArrayList<>();
    List<Coder<?>> viewCoders = new ArrayList<>();
    int count = 0;
    for (Map.Entry<RunnerApi.ExecutableStagePayload.SideInputId, PCollectionView<?>> sideInput : sideInputs.entrySet()) {
        TupleTag<?> tag = sideInput.getValue().getTagInternal();
        intToViewMapping.put(count, sideInput.getValue());
        tagToIntMapping.put(tag, count);
        count++;
        String collectionId = components.getTransformsOrThrow(sideInput.getKey().getTransformId()).getInputsOrThrow(sideInput.getKey().getLocalName());
        DataStream<Object> sideInputStream = context.getDataStreamOrThrow(collectionId);
        TypeInformation<Object> tpe = sideInputStream.getType();
        if (!(tpe instanceof CoderTypeInformation)) {
            throw new IllegalStateException("Input Stream TypeInformation is no CoderTypeInformation.");
        }
        WindowedValueCoder<Object> coder = (WindowedValueCoder) ((CoderTypeInformation) tpe).getCoder();
        Coder<KV<Void, Object>> kvCoder = KvCoder.of(VoidCoder.of(), coder.getValueCoder());
        kvCoders.add(coder.withValueCoder(kvCoder));
        // coder for materialized view matching GBK below
        WindowedValueCoder<KV<Void, Iterable<Object>>> viewCoder = coder.withValueCoder(KvCoder.of(VoidCoder.of(), IterableCoder.of(coder.getValueCoder())));
        viewCoders.add(viewCoder);
    }
    // second pass, now that we gathered the input coders
    UnionCoder unionCoder = UnionCoder.of(viewCoders);
    CoderTypeInformation<RawUnionValue> unionTypeInformation = new CoderTypeInformation<>(unionCoder, context.getPipelineOptions());
    // transform each side input to RawUnionValue and union them
    DataStream<RawUnionValue> sideInputUnion = null;
    for (Map.Entry<RunnerApi.ExecutableStagePayload.SideInputId, PCollectionView<?>> sideInput : sideInputs.entrySet()) {
        TupleTag<?> tag = sideInput.getValue().getTagInternal();
        final int intTag = tagToIntMapping.get(tag);
        RunnerApi.PTransform pTransform = components.getTransformsOrThrow(sideInput.getKey().getTransformId());
        String collectionId = pTransform.getInputsOrThrow(sideInput.getKey().getLocalName());
        DataStream<WindowedValue<?>> sideInputStream = context.getDataStreamOrThrow(collectionId);
        // insert GBK to materialize side input view
        String viewName = sideInput.getKey().getTransformId() + "-" + sideInput.getKey().getLocalName();
        WindowedValueCoder<KV<Void, Object>> kvCoder = kvCoders.get(intTag);
        DataStream<WindowedValue<KV<Void, Object>>> keyedSideInputStream = sideInputStream.map(new ToVoidKeyValue(context.getPipelineOptions()));
        SingleOutputStreamOperator<WindowedValue<KV<Void, Iterable<Object>>>> viewStream = addGBK(keyedSideInputStream, sideInput.getValue().getWindowingStrategyInternal(), kvCoder, viewName, context);
        // Assign a unique but consistent id to re-map operator state
        viewStream.uid(pTransform.getUniqueName() + "-" + sideInput.getKey().getLocalName());
        DataStream<RawUnionValue> unionValueStream = viewStream.map(new FlinkStreamingTransformTranslators.ToRawUnion<>(intTag, context.getPipelineOptions())).returns(unionTypeInformation);
        if (sideInputUnion == null) {
            sideInputUnion = unionValueStream;
        } else {
            sideInputUnion = sideInputUnion.union(unionValueStream);
        }
    }
    return new TransformedSideInputs(intToViewMapping, sideInputUnion);
}
Also used : LinkedHashMap(java.util.LinkedHashMap) HashMap(java.util.HashMap) ArrayList(java.util.ArrayList) TupleTag(org.apache.beam.sdk.values.TupleTag) RunnerApi(org.apache.beam.model.pipeline.v1.RunnerApi) WindowedValue(org.apache.beam.sdk.util.WindowedValue) CoderTypeInformation(org.apache.beam.runners.flink.translation.types.CoderTypeInformation) SingletonKeyedWorkItemCoder(org.apache.beam.runners.flink.translation.wrappers.streaming.SingletonKeyedWorkItemCoder) WindowedValueCoder(org.apache.beam.sdk.util.WindowedValue.WindowedValueCoder) KvCoder(org.apache.beam.sdk.coders.KvCoder) PipelineTranslatorUtils.instantiateCoder(org.apache.beam.runners.fnexecution.translation.PipelineTranslatorUtils.instantiateCoder) IterableCoder(org.apache.beam.sdk.coders.IterableCoder) VoidCoder(org.apache.beam.sdk.coders.VoidCoder) UnionCoder(org.apache.beam.sdk.transforms.join.UnionCoder) Coder(org.apache.beam.sdk.coders.Coder) ByteArrayCoder(org.apache.beam.sdk.coders.ByteArrayCoder) UnionCoder(org.apache.beam.sdk.transforms.join.UnionCoder) RawUnionValue(org.apache.beam.sdk.transforms.join.RawUnionValue) KV(org.apache.beam.sdk.values.KV) RunnerPCollectionView(org.apache.beam.runners.core.construction.RunnerPCollectionView) PCollectionView(org.apache.beam.sdk.values.PCollectionView) WindowedValueCoder(org.apache.beam.sdk.util.WindowedValue.WindowedValueCoder) ImmutableMap(org.apache.beam.vendor.guava.v26_0_jre.com.google.common.collect.ImmutableMap) Map(java.util.Map) LinkedHashMap(java.util.LinkedHashMap) BiMap(org.apache.beam.vendor.guava.v26_0_jre.com.google.common.collect.BiMap) TreeMap(java.util.TreeMap) PipelineTranslatorUtils.createOutputMap(org.apache.beam.runners.fnexecution.translation.PipelineTranslatorUtils.createOutputMap) HashMap(java.util.HashMap)

Example 9 with WindowedValueCoder

use of org.apache.beam.sdk.util.WindowedValue.WindowedValueCoder in project beam by apache.

the class FlinkStreamingPortablePipelineTranslator method translateGroupByKey.

private <K, V> void translateGroupByKey(String id, RunnerApi.Pipeline pipeline, StreamingTranslationContext context) {
    RunnerApi.PTransform pTransform = pipeline.getComponents().getTransformsOrThrow(id);
    String inputPCollectionId = Iterables.getOnlyElement(pTransform.getInputsMap().values());
    RehydratedComponents rehydratedComponents = RehydratedComponents.forComponents(pipeline.getComponents());
    RunnerApi.WindowingStrategy windowingStrategyProto = pipeline.getComponents().getWindowingStrategiesOrThrow(pipeline.getComponents().getPcollectionsOrThrow(inputPCollectionId).getWindowingStrategyId());
    WindowingStrategy<?, ?> windowingStrategy;
    try {
        windowingStrategy = WindowingStrategyTranslation.fromProto(windowingStrategyProto, rehydratedComponents);
    } catch (InvalidProtocolBufferException e) {
        throw new IllegalStateException(String.format("Unable to hydrate GroupByKey windowing strategy %s.", windowingStrategyProto), e);
    }
    WindowedValueCoder<KV<K, V>> windowedInputCoder = (WindowedValueCoder) instantiateCoder(inputPCollectionId, pipeline.getComponents());
    DataStream<WindowedValue<KV<K, V>>> inputDataStream = context.getDataStreamOrThrow(inputPCollectionId);
    SingleOutputStreamOperator<WindowedValue<KV<K, Iterable<V>>>> outputDataStream = addGBK(inputDataStream, windowingStrategy, windowedInputCoder, pTransform.getUniqueName(), context);
    // Assign a unique but consistent id to re-map operator state
    outputDataStream.uid(pTransform.getUniqueName());
    context.addDataStream(Iterables.getOnlyElement(pTransform.getOutputsMap().values()), outputDataStream);
}
Also used : InvalidProtocolBufferException(org.apache.beam.vendor.grpc.v1p43p2.com.google.protobuf.InvalidProtocolBufferException) KV(org.apache.beam.sdk.values.KV) RunnerApi(org.apache.beam.model.pipeline.v1.RunnerApi) WindowedValueCoder(org.apache.beam.sdk.util.WindowedValue.WindowedValueCoder) WindowedValue(org.apache.beam.sdk.util.WindowedValue) KV(org.apache.beam.sdk.values.KV) RehydratedComponents(org.apache.beam.runners.core.construction.RehydratedComponents)

Example 10 with WindowedValueCoder

use of org.apache.beam.sdk.util.WindowedValue.WindowedValueCoder in project beam by apache.

the class BeamFnMapTaskExecutorFactory method createPartialGroupByKeyOperation.

<K> OperationNode createPartialGroupByKeyOperation(Network<Node, Edge> network, ParallelInstructionNode node, PipelineOptions options, DataflowExecutionContext<?> executionContext, DataflowOperationContext operationContext) throws Exception {
    ParallelInstruction instruction = node.getParallelInstruction();
    PartialGroupByKeyInstruction pgbk = instruction.getPartialGroupByKey();
    OutputReceiver[] receivers = getOutputReceivers(network, node);
    Coder<?> windowedCoder = CloudObjects.coderFromCloudObject(CloudObject.fromSpec(pgbk.getInputElementCodec()));
    if (!(windowedCoder instanceof WindowedValueCoder)) {
        throw new IllegalArgumentException(String.format("unexpected kind of input coder for PartialGroupByKeyOperation: %s", windowedCoder));
    }
    Coder<?> elemCoder = ((WindowedValueCoder<?>) windowedCoder).getValueCoder();
    if (!(elemCoder instanceof KvCoder)) {
        throw new IllegalArgumentException(String.format("unexpected kind of input element coder for PartialGroupByKeyOperation: %s", elemCoder));
    }
    @SuppressWarnings("unchecked") KvCoder<K, ?> keyedElementCoder = (KvCoder<K, ?>) elemCoder;
    CloudObject cloudUserFn = pgbk.getValueCombiningFn() != null ? CloudObject.fromSpec(pgbk.getValueCombiningFn()) : null;
    ParDoFn fn = PartialGroupByKeyParDoFns.create(options, keyedElementCoder, cloudUserFn, pgbk.getSideInputs(), Arrays.<Receiver>asList(receivers), executionContext, operationContext);
    return OperationNode.create(new ParDoOperation(fn, receivers, operationContext));
}
Also used : OutputReceiver(org.apache.beam.runners.dataflow.worker.util.common.worker.OutputReceiver) KvCoder(org.apache.beam.sdk.coders.KvCoder) ParDoFn(org.apache.beam.runners.dataflow.worker.util.common.worker.ParDoFn) ParDoOperation(org.apache.beam.runners.dataflow.worker.util.common.worker.ParDoOperation) ParallelInstruction(com.google.api.services.dataflow.model.ParallelInstruction) WindowedValueCoder(org.apache.beam.sdk.util.WindowedValue.WindowedValueCoder) CloudObject(org.apache.beam.runners.dataflow.util.CloudObject) PartialGroupByKeyInstruction(com.google.api.services.dataflow.model.PartialGroupByKeyInstruction)

Aggregations

WindowedValueCoder (org.apache.beam.sdk.util.WindowedValue.WindowedValueCoder)16 RunnerApi (org.apache.beam.model.pipeline.v1.RunnerApi)9 WindowedValue (org.apache.beam.sdk.util.WindowedValue)8 Coder (org.apache.beam.sdk.coders.Coder)7 KvCoder (org.apache.beam.sdk.coders.KvCoder)7 CloudObject (org.apache.beam.runners.dataflow.util.CloudObject)6 List (java.util.List)5 Map (java.util.Map)5 IOException (java.io.IOException)4 ArrayList (java.util.ArrayList)4 HashMap (java.util.HashMap)4 RehydratedComponents (org.apache.beam.runners.core.construction.RehydratedComponents)4 InvalidProtocolBufferException (org.apache.beam.vendor.grpc.v1p43p2.com.google.protobuf.InvalidProtocolBufferException)4 ParallelInstruction (com.google.api.services.dataflow.model.ParallelInstruction)3 Structs.getString (org.apache.beam.runners.dataflow.util.Structs.getString)3 ParDoFn (org.apache.beam.runners.dataflow.worker.util.common.worker.ParDoFn)3 KV (org.apache.beam.sdk.values.KV)3 PCollectionView (org.apache.beam.sdk.values.PCollectionView)3 TupleTag (org.apache.beam.sdk.values.TupleTag)3 ImmutableMap (org.apache.beam.vendor.guava.v26_0_jre.com.google.common.collect.ImmutableMap)3