use of org.apache.beam.runners.dataflow.util.CloudObject in project beam by apache.
the class InsertFetchAndFilterStreamingSideInputNodes method forNetwork.
public MutableNetwork<Node, Edge> forNetwork(MutableNetwork<Node, Edge> network) {
if (pipeline == null) {
return network;
}
RehydratedComponents rehydratedComponents = RehydratedComponents.forComponents(pipeline.getComponents());
for (ParallelInstructionNode node : ImmutableList.copyOf(Iterables.filter(network.nodes(), ParallelInstructionNode.class))) {
// to worry about it.
if (node.getParallelInstruction().getParDo() == null || !ExecutionLocation.SDK_HARNESS.equals(node.getExecutionLocation())) {
continue;
}
ParDoInstruction parDoInstruction = node.getParallelInstruction().getParDo();
CloudObject userFnSpec = CloudObject.fromSpec(parDoInstruction.getUserFn());
String parDoPTransformId = getString(userFnSpec, PropertyNames.SERIALIZED_FN);
// Skip ParDoInstruction nodes that contain payloads without side inputs.
String userFnClassName = userFnSpec.getClassName();
if ("CombineValuesFn".equals(userFnClassName) || "KeyedCombineFn".equals(userFnClassName)) {
// These nodes have CombinePayloads which have no side inputs.
continue;
}
RunnerApi.PTransform parDoPTransform = pipeline.getComponents().getTransformsOrDefault(parDoPTransformId, null);
// TODO: only the non-null branch should exist; for migration ease only
if (parDoPTransform == null) {
continue;
}
RunnerApi.ParDoPayload parDoPayload;
try {
parDoPayload = RunnerApi.ParDoPayload.parseFrom(parDoPTransform.getSpec().getPayload());
} catch (InvalidProtocolBufferException exc) {
throw new RuntimeException("ParDo did not have a ParDoPayload", exc);
}
// Skip any ParDo that doesn't have a side input.
if (parDoPayload.getSideInputsMap().isEmpty()) {
continue;
}
String mainInputPCollectionLocalName = Iterables.getOnlyElement(Sets.difference(parDoPTransform.getInputsMap().keySet(), parDoPayload.getSideInputsMap().keySet()));
RunnerApi.WindowingStrategy windowingStrategyProto = pipeline.getComponents().getWindowingStrategiesOrThrow(pipeline.getComponents().getPcollectionsOrThrow(parDoPTransform.getInputsOrThrow(mainInputPCollectionLocalName)).getWindowingStrategyId());
WindowingStrategy windowingStrategy;
try {
windowingStrategy = WindowingStrategyTranslation.fromProto(windowingStrategyProto, rehydratedComponents);
} catch (InvalidProtocolBufferException e) {
throw new IllegalStateException(String.format("Unable to decode windowing strategy %s.", windowingStrategyProto), e);
}
// Gather all the side input window mapping fns which we need to request the SDK to map
ImmutableMap.Builder<PCollectionView<?>, RunnerApi.FunctionSpec> pCollectionViewsToWindowMapingsFns = ImmutableMap.builder();
parDoPayload.getSideInputsMap().forEach((sideInputTag, sideInput) -> pCollectionViewsToWindowMapingsFns.put(RegisterNodeFunction.transformSideInputForRunner(pipeline, parDoPTransform, sideInputTag, sideInput), sideInput.getWindowMappingFn()));
Node streamingSideInputWindowHandlerNode = FetchAndFilterStreamingSideInputsNode.create(windowingStrategy, pCollectionViewsToWindowMapingsFns.build(), NameContext.create(null, node.getParallelInstruction().getOriginalName(), node.getParallelInstruction().getSystemName(), node.getParallelInstruction().getName()));
// Rewire the graph such that streaming side inputs ParDos are preceded by a
// node which filters any side inputs that aren't ready and fetches any ready side inputs.
Edge mainInput = Iterables.getOnlyElement(network.inEdges(node));
InstructionOutputNode predecessor = (InstructionOutputNode) network.incidentNodes(mainInput).source();
InstructionOutputNode predecessorCopy = InstructionOutputNode.create(predecessor.getInstructionOutput(), predecessor.getPcollectionId());
network.removeEdge(mainInput);
network.addNode(streamingSideInputWindowHandlerNode);
network.addNode(predecessorCopy);
network.addEdge(predecessor, streamingSideInputWindowHandlerNode, mainInput.clone());
network.addEdge(streamingSideInputWindowHandlerNode, predecessorCopy, mainInput.clone());
network.addEdge(predecessorCopy, node, mainInput.clone());
}
return network;
}
use of org.apache.beam.runners.dataflow.util.CloudObject 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());
}
use of org.apache.beam.runners.dataflow.util.CloudObject 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));
}
use of org.apache.beam.runners.dataflow.util.CloudObject in project beam by apache.
the class CombineValuesFnFactory method create.
@Override
public ParDoFn create(PipelineOptions options, CloudObject cloudUserFn, @Nullable List<SideInputInfo> sideInputInfos, TupleTag<?> mainOutputTag, Map<TupleTag<?>, Integer> outputTupleTagsToReceiverIndices, DataflowExecutionContext<?> executionContext, DataflowOperationContext operationContext) throws Exception {
Preconditions.checkArgument(outputTupleTagsToReceiverIndices.size() == 1, "expected exactly one output for CombineValuesFn");
Object deserializedFn = SerializableUtils.deserializeFromByteArray(getBytes(cloudUserFn, PropertyNames.SERIALIZED_FN), "serialized user fn");
Preconditions.checkArgument(deserializedFn instanceof AppliedCombineFn);
AppliedCombineFn<?, ?, ?, ?> combineFn = (AppliedCombineFn<?, ?, ?, ?>) deserializedFn;
Iterable<PCollectionView<?>> sideInputViews = combineFn.getSideInputViews();
SideInputReader sideInputReader = executionContext.getSideInputReader(sideInputInfos, sideInputViews, operationContext);
// Get the combine phase, default to ALL. (The implementation
// doesn't have to split the combiner).
String phase = getString(cloudUserFn, WorkerPropertyNames.PHASE, CombinePhase.ALL);
DoFnInfo<?, ?> doFnInfo = getDoFnInfo(combineFn, sideInputReader, phase);
return new SimpleParDoFn(options, DoFnInstanceManagers.singleInstance(doFnInfo), sideInputReader, mainOutputTag, outputTupleTagsToReceiverIndices, executionContext.getStepContext(operationContext), operationContext, doFnInfo.getDoFnSchemaInformation(), doFnInfo.getSideInputMapping(), SimpleDoFnRunnerFactory.INSTANCE);
}
use of org.apache.beam.runners.dataflow.util.CloudObject in project beam by apache.
the class ConcatReaderFactory method createSourceFromDictionary.
public static Source createSourceFromDictionary(Map<String, Object> dictionary) throws Exception {
Source source = new Source();
// Set spec
CloudObject subSourceSpec = CloudObject.fromSpec(getObject(dictionary, PropertyNames.SOURCE_SPEC));
source.setSpec(subSourceSpec);
// Set encoding
CloudObject subSourceEncoding = CloudObject.fromSpec(getObject(dictionary, PropertyNames.ENCODING, null));
if (subSourceEncoding != null) {
source.setCodec(subSourceEncoding);
}
// Set base specs
List<Map<String, Object>> subSourceBaseSpecs = getListOfMaps(dictionary, WorkerPropertyNames.CONCAT_SOURCE_BASE_SPECS, null);
if (subSourceBaseSpecs != null) {
source.setBaseSpecs(subSourceBaseSpecs);
}
// Set metadata
SourceMetadata metadata = new SourceMetadata();
Boolean infinite = getBoolean(dictionary, PropertyNames.SOURCE_IS_INFINITE, null);
if (infinite != null) {
metadata.setInfinite(infinite);
}
Long estimatedSizeBytes = getLong(dictionary, PropertyNames.SOURCE_ESTIMATED_SIZE_BYTES, null);
if (estimatedSizeBytes != null) {
metadata.setEstimatedSizeBytes(estimatedSizeBytes);
}
if (estimatedSizeBytes != null || infinite != null) {
source.setMetadata(metadata);
}
// Set doesNotNeedSplitting
Boolean doesNotNeedSplitting = getBoolean(dictionary, PropertyNames.SOURCE_DOES_NOT_NEED_SPLITTING, null);
if (doesNotNeedSplitting != null) {
source.setDoesNotNeedSplitting(doesNotNeedSplitting);
}
return source;
}
Aggregations