Search in sources :

Example 6 with ReadInstruction

use of com.google.api.services.dataflow.model.ReadInstruction in project beam by apache.

the class MapTaskToNetworkFunctionTest method testParallelEdgeFlatten.

@Test
public void testParallelEdgeFlatten() {
    // /---\
    // Read --> Read.out --> Flatten
    // \---/
    InstructionOutput readOutput = createInstructionOutput("Read.out");
    ParallelInstruction read = createParallelInstruction("Read", readOutput);
    read.setRead(new ReadInstruction());
    FlattenInstruction flattenInstruction = new FlattenInstruction();
    flattenInstruction.setInputs(ImmutableList.of(// Read.out
    createInstructionInput(0, 0), // Read.out
    createInstructionInput(0, 0), // Read.out
    createInstructionInput(0, 0)));
    InstructionOutput flattenOutput = createInstructionOutput("Flatten.out");
    ParallelInstruction flatten = createParallelInstruction("Flatten", flattenOutput);
    flatten.setFlatten(flattenInstruction);
    MapTask mapTask = new MapTask();
    mapTask.setInstructions(ImmutableList.of(read, flatten));
    mapTask.setFactory(Transport.getJsonFactory());
    Network<Node, Edge> network = new MapTaskToNetworkFunction(IdGenerators.decrementingLongs()).apply(mapTask);
    assertNetworkProperties(network);
    assertEquals(4, network.nodes().size());
    assertEquals(5, network.edges().size());
    ParallelInstructionNode readNode = get(network, read);
    InstructionOutputNode readOutputNode = getOnlySuccessor(network, readNode);
    assertEquals(readOutput, readOutputNode.getInstructionOutput());
    ParallelInstructionNode flattenNode = getOnlySuccessor(network, readOutputNode);
    // Assert that the three parallel edges are maintained
    assertEquals(3, network.edgesConnecting(readOutputNode, flattenNode).size());
    InstructionOutputNode flattenOutputNode = getOnlySuccessor(network, flattenNode);
    assertEquals(flattenOutput, flattenOutputNode.getInstructionOutput());
}
Also used : ParallelInstruction(com.google.api.services.dataflow.model.ParallelInstruction) InstructionOutputNode(org.apache.beam.runners.dataflow.worker.graph.Nodes.InstructionOutputNode) MapTask(com.google.api.services.dataflow.model.MapTask) InstructionOutputNode(org.apache.beam.runners.dataflow.worker.graph.Nodes.InstructionOutputNode) ParallelInstructionNode(org.apache.beam.runners.dataflow.worker.graph.Nodes.ParallelInstructionNode) Node(org.apache.beam.runners.dataflow.worker.graph.Nodes.Node) InstructionOutput(com.google.api.services.dataflow.model.InstructionOutput) ParallelInstructionNode(org.apache.beam.runners.dataflow.worker.graph.Nodes.ParallelInstructionNode) ReadInstruction(com.google.api.services.dataflow.model.ReadInstruction) FlattenInstruction(com.google.api.services.dataflow.model.FlattenInstruction) Edge(org.apache.beam.runners.dataflow.worker.graph.Edges.Edge) DefaultEdge(org.apache.beam.runners.dataflow.worker.graph.Edges.DefaultEdge) MultiOutputInfoEdge(org.apache.beam.runners.dataflow.worker.graph.Edges.MultiOutputInfoEdge) Test(org.junit.Test)

Example 7 with ReadInstruction

use of com.google.api.services.dataflow.model.ReadInstruction 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 8 with ReadInstruction

use of com.google.api.services.dataflow.model.ReadInstruction in project beam by apache.

the class DeduceNodeLocationsFunction method executesInSdkHarness.

/**
 * Returns true iff the {@link ParallelInstruction} represents an instruction that executes within
 * the SDK harness as part of the fused sub-graph. This excludes special instructions related to
 * {@link GroupByKey} such as window assignment and window merging as they cross the Beam Fn API
 * with a specialized protocol.
 */
private static boolean executesInSdkHarness(ParallelInstruction parallelInstruction) {
    if (parallelInstruction.getRead() != null) {
        ReadInstruction readInstruction = parallelInstruction.getRead();
        CloudObject sourceSpec = CloudObject.fromSpec(CloudSourceUtils.flattenBaseSpecs(readInstruction.getSource()).getSpec());
        return !RUNNER_SOURCES.contains(sourceSpec.getClassName());
    } else if (parallelInstruction.getParDo() != null) {
        ParDoInstruction parDoInstruction = parallelInstruction.getParDo();
        CloudObject userFnSpec = CloudObject.fromSpec(parDoInstruction.getUserFn());
        return SDK_COMPATIBLE_PAR_DO_FNS.contains(userFnSpec.getClassName());
    }
    return false;
}
Also used : ParDoInstruction(com.google.api.services.dataflow.model.ParDoInstruction) CloudObject(org.apache.beam.runners.dataflow.util.CloudObject) ReadInstruction(com.google.api.services.dataflow.model.ReadInstruction)

Example 9 with ReadInstruction

use of com.google.api.services.dataflow.model.ReadInstruction in project beam by apache.

the class RegisterNodeFunction 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())));
                }
            }
        }
    }
    // We start off by replacing all edges within the graph with edges that have the named
    // outputs from the predecessor step. For ParallelInstruction Source nodes and RemoteGrpcPort
    // nodes this is a generated port id. All ParDoInstructions will have already
    ProcessBundleDescriptor.Builder processBundleDescriptor = ProcessBundleDescriptor.newBuilder().setId(idGenerator.getId()).setStateApiServiceDescriptor(stateApiServiceDescriptor);
    // For intermediate PCollections we fabricate, we make a bogus WindowingStrategy
    // TODO: create a correct windowing strategy, including coders and environment
    SdkComponents sdkComponents = SdkComponents.create(pipeline.getComponents(), null);
    // Default to use the Java environment if pipeline doesn't have environment specified.
    if (pipeline.getComponents().getEnvironmentsMap().isEmpty()) {
        sdkComponents.registerEnvironment(Environments.JAVA_SDK_HARNESS_ENVIRONMENT);
    }
    String fakeWindowingStrategyId = "fakeWindowingStrategy" + idGenerator.getId();
    try {
        RunnerApi.MessageWithComponents fakeWindowingStrategyProto = WindowingStrategyTranslation.toMessageProto(WindowingStrategy.globalDefault(), sdkComponents);
        processBundleDescriptor.putWindowingStrategies(fakeWindowingStrategyId, fakeWindowingStrategyProto.getWindowingStrategy()).putAllCoders(fakeWindowingStrategyProto.getComponents().getCodersMap()).putAllEnvironments(fakeWindowingStrategyProto.getComponents().getEnvironmentsMap());
    } 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();
    ImmutableMap.Builder<String, NameContext> pcollectionIdToNameContexts = ImmutableMap.builder();
    ImmutableMap.Builder<InstructionOutputNode, String> instructionOutputNodeToCoderIdBuilder = ImmutableMap.builder();
    // 2. Generate new PCollectionId and register it with ProcessBundleDescriptor.
    for (InstructionOutputNode node : Iterables.filter(input.nodes(), InstructionOutputNode.class)) {
        InstructionOutput instructionOutput = node.getInstructionOutput();
        String coderId = "generatedCoder" + idGenerator.getId();
        instructionOutputNodeToCoderIdBuilder.put(node, coderId);
        try (ByteString.Output output = ByteString.newOutput()) {
            try {
                Coder<?> javaCoder = CloudObjects.coderFromCloudObject(CloudObject.fromSpec(instructionOutput.getCodec()));
                sdkComponents.registerCoder(javaCoder);
                RunnerApi.Coder coderProto = CoderTranslation.toProto(javaCoder, sdkComponents);
                processBundleDescriptor.putCoders(coderId, coderProto);
            } 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());
                processBundleDescriptor.putCoders(coderId, RunnerApi.Coder.newBuilder().setSpec(RunnerApi.FunctionSpec.newBuilder().setPayload(output.toByteString())).build());
            }
        } catch (IOException e) {
            throw new IllegalArgumentException(String.format("Unable to encode coder %s for output %s", instructionOutput.getCodec(), instructionOutput), e);
        }
        // Generate new PCollection ID and map it to relevant node.
        // Will later be used to fill PTransform inputs/outputs information.
        String pcollectionId = "generatedPcollection" + idGenerator.getId();
        processBundleDescriptor.putPcollections(pcollectionId, RunnerApi.PCollection.newBuilder().setCoderId(coderId).setWindowingStrategyId(fakeWindowingStrategyId).build());
        nodesToPCollections.put(node, pcollectionId);
        pcollectionIdToNameContexts.put(pcollectionId, NameContext.create(null, instructionOutput.getOriginalName(), instructionOutput.getSystemName(), instructionOutput.getName()));
    }
    processBundleDescriptor.putAllCoders(sdkComponents.toComponents().getCodersMap());
    Map<InstructionOutputNode, String> instructionOutputNodeToCoderIdMap = instructionOutputNodeToCoderIdBuilder.build();
    for (ParallelInstructionNode node : Iterables.filter(input.nodes(), ParallelInstructionNode.class)) {
        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();
        if (parallelInstruction.getParDo() != null) {
            ParDoInstruction parDoInstruction = parallelInstruction.getParDo();
            CloudObject userFnSpec = CloudObject.fromSpec(parDoInstruction.getUserFn());
            String userFnClassName = userFnSpec.getClassName();
            if ("CombineValuesFn".equals(userFnClassName) || "KeyedCombineFn".equals(userFnClassName)) {
                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);
                    }
                    ImmutableList.Builder<PCollectionView<?>> pcollectionViews = ImmutableList.builder();
                    for (Map.Entry<String, SideInput> sideInputEntry : parDoPayload.getSideInputsMap().entrySet()) {
                        pcollectionViews.add(transformSideInputForRunner(pipeline, parDoPTransform, sideInputEntry.getKey(), sideInputEntry.getValue()));
                        transformSideInputForSdk(pipeline, parDoPTransform, sideInputEntry.getKey(), processBundleDescriptor, pTransform);
                    }
                    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));
                }
                // Add side input information for batch pipelines
                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));
        }
        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);
        processBundleDescriptor.putTransforms(ptransformId, pTransform.build());
    }
    // Add the PTransforms representing the remote gRPC nodes
    for (RemoteGrpcPortNode node : Iterables.filter(input.nodes(), RemoteGrpcPortNode.class)) {
        RunnerApi.PTransform.Builder pTransform = RunnerApi.PTransform.newBuilder();
        Set<Node> predecessors = input.predecessors(node);
        Set<Node> successors = input.successors(node);
        if (predecessors.isEmpty() && !successors.isEmpty()) {
            Node instructionOutputNode = Iterables.getOnlyElement(successors);
            pTransform.putOutputs("generatedOutput" + idGenerator.getId(), nodesToPCollections.get(instructionOutputNode));
            pTransform.setSpec(RunnerApi.FunctionSpec.newBuilder().setUrn(DATA_INPUT_URN).setPayload(node.getRemoteGrpcPort().toBuilder().setCoderId(instructionOutputNodeToCoderIdMap.get(instructionOutputNode)).build().toByteString()).build());
        } else if (!predecessors.isEmpty() && successors.isEmpty()) {
            Node instructionOutputNode = Iterables.getOnlyElement(predecessors);
            pTransform.putInputs("generatedInput" + idGenerator.getId(), nodesToPCollections.get(instructionOutputNode));
            pTransform.setSpec(RunnerApi.FunctionSpec.newBuilder().setUrn(DATA_OUTPUT_URN).setPayload(node.getRemoteGrpcPort().toBuilder().setCoderId(instructionOutputNodeToCoderIdMap.get(instructionOutputNode)).build().toByteString()).build());
        } else {
            throw new IllegalStateException("Expected either one input OR one output " + "InstructionOutputNode for this RemoteGrpcPortNode");
        }
        processBundleDescriptor.putTransforms(node.getPrimitiveTransformId(), pTransform.build());
    }
    return RegisterRequestNode.create(RegisterRequest.newBuilder().addProcessBundleDescriptor(processBundleDescriptor).build(), ptransformIdToNameContexts.build(), ptransformIdToSideInputInfos.build(), ptransformIdToPCollectionViews.build(), pcollectionIdToNameContexts.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) RegisterRequestNode(org.apache.beam.runners.dataflow.worker.graph.Nodes.RegisterRequestNode) InstructionOutputNode(org.apache.beam.runners.dataflow.worker.graph.Nodes.InstructionOutputNode) ParallelInstructionNode(org.apache.beam.runners.dataflow.worker.graph.Nodes.ParallelInstructionNode) Node(org.apache.beam.runners.dataflow.worker.graph.Nodes.Node) 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) Structs.getString(org.apache.beam.runners.dataflow.util.Structs.getString) ByteString(org.apache.beam.vendor.grpc.v1p43p2.com.google.protobuf.ByteString) RunnerApi(org.apache.beam.model.pipeline.v1.RunnerApi) SideInput(org.apache.beam.model.pipeline.v1.RunnerApi.SideInput) 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) DataflowPortabilityPCollectionView(org.apache.beam.runners.dataflow.worker.DataflowPortabilityPCollectionView) PCollectionView(org.apache.beam.sdk.values.PCollectionView) ImmutableMap(org.apache.beam.vendor.guava.v26_0_jre.com.google.common.collect.ImmutableMap) Map(java.util.Map) HashMap(java.util.HashMap) RemoteGrpcPortNode(org.apache.beam.runners.dataflow.worker.graph.Nodes.RemoteGrpcPortNode) ProcessBundleDescriptor(org.apache.beam.model.fnexecution.v1.BeamFnApi.ProcessBundleDescriptor) 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) NameContext(org.apache.beam.runners.dataflow.worker.counters.NameContext) InvalidProtocolBufferException(org.apache.beam.vendor.grpc.v1p43p2.com.google.protobuf.InvalidProtocolBufferException) IOException(java.io.IOException) InvalidProtocolBufferException(org.apache.beam.vendor.grpc.v1p43p2.com.google.protobuf.InvalidProtocolBufferException) IOException(java.io.IOException) ParallelInstruction(com.google.api.services.dataflow.model.ParallelInstruction) CloudObject(org.apache.beam.runners.dataflow.util.CloudObject) 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 10 with ReadInstruction

use of com.google.api.services.dataflow.model.ReadInstruction in project beam by apache.

the class IntrinsicMapTaskExecutorFactory method createReadOperation.

OperationNode createReadOperation(Network<Node, Edge> network, ParallelInstructionNode node, PipelineOptions options, ReaderFactory readerFactory, DataflowExecutionContext<?> executionContext, DataflowOperationContext operationContext) throws Exception {
    ParallelInstruction instruction = node.getParallelInstruction();
    ReadInstruction read = instruction.getRead();
    Source cloudSource = CloudSourceUtils.flattenBaseSpecs(read.getSource());
    CloudObject sourceSpec = CloudObject.fromSpec(cloudSource.getSpec());
    Coder<?> coder = CloudObjects.coderFromCloudObject(CloudObject.fromSpec(cloudSource.getCodec()));
    NativeReader<?> reader = readerFactory.create(sourceSpec, coder, options, executionContext, operationContext);
    OutputReceiver[] receivers = getOutputReceivers(network, node);
    return OperationNode.create(ReadOperation.create(reader, receivers, operationContext));
}
Also used : ParallelInstruction(com.google.api.services.dataflow.model.ParallelInstruction) CloudObject(org.apache.beam.runners.dataflow.util.CloudObject) OutputReceiver(org.apache.beam.runners.dataflow.worker.util.common.worker.OutputReceiver) ReadInstruction(com.google.api.services.dataflow.model.ReadInstruction) Source(com.google.api.services.dataflow.model.Source)

Aggregations

ReadInstruction (com.google.api.services.dataflow.model.ReadInstruction)18 ParallelInstruction (com.google.api.services.dataflow.model.ParallelInstruction)17 InstructionOutput (com.google.api.services.dataflow.model.InstructionOutput)13 CloudObject (org.apache.beam.runners.dataflow.util.CloudObject)9 DefaultEdge (org.apache.beam.runners.dataflow.worker.graph.Edges.DefaultEdge)9 Edge (org.apache.beam.runners.dataflow.worker.graph.Edges.Edge)9 MultiOutputInfoEdge (org.apache.beam.runners.dataflow.worker.graph.Edges.MultiOutputInfoEdge)9 InstructionOutputNode (org.apache.beam.runners.dataflow.worker.graph.Nodes.InstructionOutputNode)9 Node (org.apache.beam.runners.dataflow.worker.graph.Nodes.Node)9 ParallelInstructionNode (org.apache.beam.runners.dataflow.worker.graph.Nodes.ParallelInstructionNode)9 Test (org.junit.Test)9 MapTask (com.google.api.services.dataflow.model.MapTask)7 Source (com.google.api.services.dataflow.model.Source)6 ParDoInstruction (com.google.api.services.dataflow.model.ParDoInstruction)5 MultiOutputInfo (com.google.api.services.dataflow.model.MultiOutputInfo)4 WriteInstruction (com.google.api.services.dataflow.model.WriteInstruction)3 TestCountingSource (org.apache.beam.runners.dataflow.worker.testing.TestCountingSource)3 ByteString (org.apache.beam.vendor.grpc.v1p43p2.com.google.protobuf.ByteString)3 FlattenInstruction (com.google.api.services.dataflow.model.FlattenInstruction)2 IOException (java.io.IOException)2