Examples with SdkComponents org.apache.beam.runners.core.construction.SdkComponents used on opensource projects

Example 56 with SdkComponents

use of org.apache.beam.runners.core.construction.SdkComponents 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());
}

Example 57 with SdkComponents

use of org.apache.beam.runners.core.construction.SdkComponents in project beam by apache.

the class DataflowRunner method run.

@Override
public DataflowPipelineJob run(Pipeline pipeline) {
    if (useUnifiedWorker(options)) {
        // non-null if useUnifiedWorker is true
        List<String> experiments = options.getExperiments();
        if (!experiments.contains("use_runner_v2")) {
            experiments.add("use_runner_v2");
        }
        if (!experiments.contains("use_unified_worker")) {
            experiments.add("use_unified_worker");
        }
        if (!experiments.contains("beam_fn_api")) {
            experiments.add("beam_fn_api");
        }
        if (!experiments.contains("use_portable_job_submission")) {
            experiments.add("use_portable_job_submission");
        }
        options.setExperiments(ImmutableList.copyOf(experiments));
    }
    logWarningIfPCollectionViewHasNonDeterministicKeyCoder(pipeline);
    if (containsUnboundedPCollection(pipeline)) {
        options.setStreaming(true);
    }
    LOG.info("Executing pipeline on the Dataflow Service, which will have billing implications " + "related to Google Compute Engine usage and other Google Cloud Services.");
    DataflowPipelineOptions dataflowOptions = options.as(DataflowPipelineOptions.class);
    String workerHarnessContainerImageURL = DataflowRunner.getContainerImageForJob(dataflowOptions);
    // This incorrectly puns the worker harness container image (which implements v1beta3 API)
    // with the SDK harness image (which implements Fn API).
    // 
    // The same Environment is used in different and contradictory ways, depending on whether
    // it is a v1 or v2 job submission.
    RunnerApi.Environment defaultEnvironmentForDataflow = Environments.createDockerEnvironment(workerHarnessContainerImageURL);
    // The SdkComponents for portable an non-portable job submission must be kept distinct. Both
    // need the default environment.
    SdkComponents portableComponents = SdkComponents.create();
    portableComponents.registerEnvironment(defaultEnvironmentForDataflow.toBuilder().addAllDependencies(getDefaultArtifacts()).addAllCapabilities(Environments.getJavaCapabilities()).build());
    RunnerApi.Pipeline portablePipelineProto = PipelineTranslation.toProto(pipeline, portableComponents, false);
    // Note that `stageArtifacts` has to be called before `resolveArtifact` because
    // `resolveArtifact` updates local paths to staged paths in pipeline proto.
    List<DataflowPackage> packages = stageArtifacts(portablePipelineProto);
    portablePipelineProto = resolveArtifacts(portablePipelineProto);
    portablePipelineProto = applySdkEnvironmentOverrides(portablePipelineProto, options);
    if (LOG.isDebugEnabled()) {
        LOG.debug("Portable pipeline proto:\n{}", TextFormat.printer().printToString(portablePipelineProto));
    }
    // Stage the portable pipeline proto, retrieving the staged pipeline path, then update
    // the options on the new job
    // TODO: add an explicit `pipeline` parameter to the submission instead of pipeline options
    LOG.info("Staging portable pipeline proto to {}", options.getStagingLocation());
    byte[] serializedProtoPipeline = portablePipelineProto.toByteArray();
    DataflowPackage stagedPipeline = options.getStager().stageToFile(serializedProtoPipeline, PIPELINE_FILE_NAME);
    dataflowOptions.setPipelineUrl(stagedPipeline.getLocation());
    // Now rewrite things to be as needed for v1 (mutates the pipeline)
    // This way the job submitted is valid for v1 and v2, simultaneously
    replaceV1Transforms(pipeline);
    // Capture the SdkComponents for look up during step translations
    SdkComponents dataflowV1Components = SdkComponents.create();
    dataflowV1Components.registerEnvironment(defaultEnvironmentForDataflow.toBuilder().addAllDependencies(getDefaultArtifacts()).addAllCapabilities(Environments.getJavaCapabilities()).build());
    RunnerApi.Pipeline dataflowV1PipelineProto = PipelineTranslation.toProto(pipeline, dataflowV1Components, true);
    if (LOG.isDebugEnabled()) {
        LOG.debug("Dataflow v1 pipeline proto:\n{}", TextFormat.printer().printToString(dataflowV1PipelineProto));
    }
    // Set a unique client_request_id in the CreateJob request.
    // This is used to ensure idempotence of job creation across retried
    // attempts to create a job. Specifically, if the service returns a job with
    // a different client_request_id, it means the returned one is a different
    // job previously created with the same job name, and that the job creation
    // has been effectively rejected. The SDK should return
    // Error::Already_Exists to user in that case.
    int randomNum = new Random().nextInt(9000) + 1000;
    String requestId = DateTimeFormat.forPattern("YYYYMMddHHmmssmmm").withZone(DateTimeZone.UTC).print(DateTimeUtils.currentTimeMillis()) + "_" + randomNum;
    // Try to create a debuggee ID. This must happen before the job is translated since it may
    // update the options.
    maybeRegisterDebuggee(dataflowOptions, requestId);
    JobSpecification jobSpecification = translator.translate(pipeline, dataflowV1PipelineProto, dataflowV1Components, this, packages);
    if (!isNullOrEmpty(dataflowOptions.getDataflowWorkerJar()) && !useUnifiedWorker(options)) {
        List<String> experiments = firstNonNull(dataflowOptions.getExperiments(), Collections.emptyList());
        if (!experiments.contains("use_staged_dataflow_worker_jar")) {
            dataflowOptions.setExperiments(ImmutableList.<String>builder().addAll(experiments).add("use_staged_dataflow_worker_jar").build());
        }
    }
    Job newJob = jobSpecification.getJob();
    try {
        newJob.getEnvironment().setSdkPipelineOptions(MAPPER.readValue(MAPPER_WITH_MODULES.writeValueAsBytes(options), Map.class));
    } catch (IOException e) {
        throw new IllegalArgumentException("PipelineOptions specified failed to serialize to JSON.", e);
    }
    newJob.setClientRequestId(requestId);
    DataflowRunnerInfo dataflowRunnerInfo = DataflowRunnerInfo.getDataflowRunnerInfo();
    String version = dataflowRunnerInfo.getVersion();
    checkState(!"${pom.version}".equals(version), "Unable to submit a job to the Dataflow service with unset version ${pom.version}");
    LOG.info("Dataflow SDK version: {}", version);
    newJob.getEnvironment().setUserAgent((Map) dataflowRunnerInfo.getProperties());
    // must be verified.
    if (!isNullOrEmpty(options.getGcpTempLocation())) {
        newJob.getEnvironment().setTempStoragePrefix(dataflowOptions.getPathValidator().verifyPath(options.getGcpTempLocation()));
    }
    newJob.getEnvironment().setDataset(options.getTempDatasetId());
    if (options.getWorkerRegion() != null) {
        newJob.getEnvironment().setWorkerRegion(options.getWorkerRegion());
    }
    if (options.getWorkerZone() != null) {
        newJob.getEnvironment().setWorkerZone(options.getWorkerZone());
    }
    if (options.getFlexRSGoal() == DataflowPipelineOptions.FlexResourceSchedulingGoal.COST_OPTIMIZED) {
        newJob.getEnvironment().setFlexResourceSchedulingGoal("FLEXRS_COST_OPTIMIZED");
    } else if (options.getFlexRSGoal() == DataflowPipelineOptions.FlexResourceSchedulingGoal.SPEED_OPTIMIZED) {
        newJob.getEnvironment().setFlexResourceSchedulingGoal("FLEXRS_SPEED_OPTIMIZED");
    }
    // Represent the minCpuPlatform pipeline option as an experiment, if not already present.
    if (!isNullOrEmpty(dataflowOptions.getMinCpuPlatform())) {
        List<String> experiments = firstNonNull(dataflowOptions.getExperiments(), Collections.emptyList());
        List<String> minCpuFlags = experiments.stream().filter(p -> p.startsWith("min_cpu_platform")).collect(Collectors.toList());
        if (minCpuFlags.isEmpty()) {
            dataflowOptions.setExperiments(ImmutableList.<String>builder().addAll(experiments).add("min_cpu_platform=" + dataflowOptions.getMinCpuPlatform()).build());
        } else {
            LOG.warn("Flag min_cpu_platform is defined in both top level PipelineOption, " + "as well as under experiments. Proceed using {}.", minCpuFlags.get(0));
        }
    }
    newJob.getEnvironment().setExperiments(ImmutableList.copyOf(firstNonNull(dataflowOptions.getExperiments(), Collections.emptyList())));
    // Set the Docker container image that executes Dataflow worker harness, residing in Google
    // Container Registry. Translator is guaranteed to create a worker pool prior to this point.
    // For runner_v1, only worker_harness_container is set.
    // For runner_v2, both worker_harness_container and sdk_harness_container are set to the same
    // value.
    String containerImage = getContainerImageForJob(options);
    for (WorkerPool workerPool : newJob.getEnvironment().getWorkerPools()) {
        workerPool.setWorkerHarnessContainerImage(containerImage);
    }
    configureSdkHarnessContainerImages(options, portablePipelineProto, newJob);
    newJob.getEnvironment().setVersion(getEnvironmentVersion(options));
    if (hooks != null) {
        hooks.modifyEnvironmentBeforeSubmission(newJob.getEnvironment());
    }
    // will be downloaded from GCS by the service.
    if (hasExperiment(options, "upload_graph")) {
        DataflowPackage stagedGraph = options.getStager().stageToFile(DataflowPipelineTranslator.jobToString(newJob).getBytes(UTF_8), DATAFLOW_GRAPH_FILE_NAME);
        newJob.getSteps().clear();
        newJob.setStepsLocation(stagedGraph.getLocation());
    }
    if (!isNullOrEmpty(options.getDataflowJobFile()) || !isNullOrEmpty(options.getTemplateLocation())) {
        boolean isTemplate = !isNullOrEmpty(options.getTemplateLocation());
        if (isTemplate) {
            checkArgument(isNullOrEmpty(options.getDataflowJobFile()), "--dataflowJobFile and --templateLocation are mutually exclusive.");
        }
        String fileLocation = firstNonNull(options.getTemplateLocation(), options.getDataflowJobFile());
        checkArgument(fileLocation.startsWith("/") || fileLocation.startsWith("gs://"), "Location must be local or on Cloud Storage, got %s.", fileLocation);
        ResourceId fileResource = FileSystems.matchNewResource(fileLocation, false);
        String workSpecJson = DataflowPipelineTranslator.jobToString(newJob);
        try (PrintWriter printWriter = new PrintWriter(new BufferedWriter(new OutputStreamWriter(Channels.newOutputStream(FileSystems.create(fileResource, MimeTypes.TEXT)), UTF_8)))) {
            printWriter.print(workSpecJson);
            LOG.info("Printed job specification to {}", fileLocation);
        } catch (IOException ex) {
            String error = String.format("Cannot create output file at %s", fileLocation);
            if (isTemplate) {
                throw new RuntimeException(error, ex);
            } else {
                LOG.warn(error, ex);
            }
        }
        if (isTemplate) {
            LOG.info("Template successfully created.");
            return new DataflowTemplateJob();
        }
    }
    String jobIdToUpdate = null;
    if (options.isUpdate()) {
        jobIdToUpdate = getJobIdFromName(options.getJobName());
        newJob.setTransformNameMapping(options.getTransformNameMapping());
        newJob.setReplaceJobId(jobIdToUpdate);
    }
    if (options.getCreateFromSnapshot() != null && !options.getCreateFromSnapshot().isEmpty()) {
        newJob.setCreatedFromSnapshotId(options.getCreateFromSnapshot());
    }
    Job jobResult;
    try {
        jobResult = dataflowClient.createJob(newJob);
    } catch (GoogleJsonResponseException e) {
        String errorMessages = "Unexpected errors";
        if (e.getDetails() != null) {
            if (Utf8.encodedLength(newJob.toString()) >= CREATE_JOB_REQUEST_LIMIT_BYTES) {
                errorMessages = "The size of the serialized JSON representation of the pipeline " + "exceeds the allowable limit. " + "For more information, please see the documentation on job submission:\n" + "https://cloud.google.com/dataflow/docs/guides/deploying-a-pipeline#jobs";
            } else {
                errorMessages = e.getDetails().getMessage();
            }
        }
        throw new RuntimeException("Failed to create a workflow job: " + errorMessages, e);
    } catch (IOException e) {
        throw new RuntimeException("Failed to create a workflow job", e);
    }
    // Use a raw client for post-launch monitoring, as status calls may fail
    // regularly and need not be retried automatically.
    DataflowPipelineJob dataflowPipelineJob = new DataflowPipelineJob(DataflowClient.create(options), jobResult.getId(), options, jobSpecification != null ? jobSpecification.getStepNames() : Collections.emptyMap(), portablePipelineProto);
    // depending on whether this is a reload or not.
    if (jobResult.getClientRequestId() != null && !jobResult.getClientRequestId().isEmpty() && !jobResult.getClientRequestId().equals(requestId)) {
        // If updating a job.
        if (options.isUpdate()) {
            throw new DataflowJobAlreadyUpdatedException(dataflowPipelineJob, String.format("The job named %s with id: %s has already been updated into job id: %s " + "and cannot be updated again.", newJob.getName(), jobIdToUpdate, jobResult.getId()));
        } else {
            throw new DataflowJobAlreadyExistsException(dataflowPipelineJob, String.format("There is already an active job named %s with id: %s. If you want to submit a" + " second job, try again by setting a different name using --jobName.", newJob.getName(), jobResult.getId()));
        }
    }
    LOG.info("To access the Dataflow monitoring console, please navigate to {}", MonitoringUtil.getJobMonitoringPageURL(options.getProject(), options.getRegion(), jobResult.getId()));
    LOG.info("Submitted job: {}", jobResult.getId());
    LOG.info("To cancel the job using the 'gcloud' tool, run:\n> {}", MonitoringUtil.getGcloudCancelCommand(options, jobResult.getId()));
    return dataflowPipelineJob;
}

Example 58 with SdkComponents

use of org.apache.beam.runners.core.construction.SdkComponents in project beam by apache.

the class StreamingDataflowWorkerTest method runMergeSessionsActions.

// Helper for running tests for merging sessions based upon Actions consisting of GetWorkResponse
// and expected timers and holds in the corresponding commit. All GetData requests are responded
// to with empty state, relying on user worker caching to keep data written.
private void runMergeSessionsActions(List<Action> actions) throws Exception {
    Coder<KV<String, String>> kvCoder = KvCoder.of(StringUtf8Coder.of(), StringUtf8Coder.of());
    Coder<WindowedValue<KV<String, String>>> windowedKvCoder = FullWindowedValueCoder.of(kvCoder, IntervalWindow.getCoder());
    KvCoder<String, List<String>> groupedCoder = KvCoder.of(StringUtf8Coder.of(), ListCoder.of(StringUtf8Coder.of()));
    Coder<WindowedValue<KV<String, List<String>>>> windowedGroupedCoder = FullWindowedValueCoder.of(groupedCoder, IntervalWindow.getCoder());
    CloudObject spec = CloudObject.forClassName("MergeWindowsDoFn");
    SdkComponents sdkComponents = SdkComponents.create();
    sdkComponents.registerEnvironment(Environments.JAVA_SDK_HARNESS_ENVIRONMENT);
    addString(spec, PropertyNames.SERIALIZED_FN, StringUtils.byteArrayToJsonString(WindowingStrategyTranslation.toMessageProto(WindowingStrategy.of(Sessions.withGapDuration(Duration.millis(10))).withMode(AccumulationMode.DISCARDING_FIRED_PANES).withTrigger(Repeatedly.forever(AfterWatermark.pastEndOfWindow().withLateFirings(AfterPane.elementCountAtLeast(1)))).withAllowedLateness(Duration.standardMinutes(60)), sdkComponents).toByteArray()));
    addObject(spec, WorkerPropertyNames.INPUT_CODER, CloudObjects.asCloudObject(windowedKvCoder, /*sdkComponents=*/
    null));
    ParallelInstruction mergeWindowsInstruction = new ParallelInstruction().setSystemName("MergeWindows-System").setName("MergeWindowsStep").setOriginalName("MergeWindowsOriginal").setParDo(new ParDoInstruction().setInput(new InstructionInput().setProducerInstructionIndex(0).setOutputNum(0)).setNumOutputs(1).setUserFn(spec)).setOutputs(Arrays.asList(new InstructionOutput().setOriginalName(DEFAULT_OUTPUT_ORIGINAL_NAME).setSystemName(DEFAULT_OUTPUT_SYSTEM_NAME).setName("output").setCodec(CloudObjects.asCloudObject(windowedGroupedCoder, /*sdkComponents=*/
    null))));
    List<ParallelInstruction> instructions = Arrays.asList(makeWindowingSourceInstruction(kvCoder), mergeWindowsInstruction, makeSinkInstruction(groupedCoder, 1));
    FakeWindmillServer server = new FakeWindmillServer(errorCollector);
    StreamingDataflowWorker worker = makeWorker(instructions, createTestingPipelineOptions(server), false);
    Map<String, String> nameMap = new HashMap<>();
    nameMap.put("MergeWindowsStep", "MergeWindows");
    worker.addStateNameMappings(nameMap);
    worker.start();
    // Respond to any GetData requests with empty state.
    for (int i = 0; i < 1000; ++i) {
        server.addDataFnToOffer(EMPTY_DATA_RESPONDER);
    }
    for (int i = 0; i < actions.size(); ++i) {
        Action action = actions.get(i);
        server.addWorkToOffer(action.response);
        Map<Long, Windmill.WorkItemCommitRequest> result = server.waitForAndGetCommits(1);
        WorkItemCommitRequest actualOutput = result.get(i + 1L);
        assertThat(actualOutput, Matchers.not(Matchers.nullValue()));
        verifyTimers(actualOutput, action.expectedTimers);
        verifyHolds(actualOutput, action.expectedHolds);
    }
}

Example 59 with SdkComponents

use of org.apache.beam.runners.core.construction.SdkComponents in project beam by apache.

the class StreamingDataflowWorkerTest method testAssignWindows.

@Test
public void testAssignWindows() throws Exception {
    Duration gapDuration = Duration.standardSeconds(1);
    CloudObject spec = CloudObject.forClassName("AssignWindowsDoFn");
    SdkComponents sdkComponents = SdkComponents.create();
    sdkComponents.registerEnvironment(Environments.JAVA_SDK_HARNESS_ENVIRONMENT);
    addString(spec, PropertyNames.SERIALIZED_FN, StringUtils.byteArrayToJsonString(WindowingStrategyTranslation.toMessageProto(WindowingStrategy.of(FixedWindows.of(gapDuration)), sdkComponents).toByteArray()));
    ParallelInstruction addWindowsInstruction = new ParallelInstruction().setSystemName("AssignWindows").setName("AssignWindows").setOriginalName("AssignWindowsOriginal").setParDo(new ParDoInstruction().setInput(new InstructionInput().setProducerInstructionIndex(0).setOutputNum(0)).setNumOutputs(1).setUserFn(spec)).setOutputs(Arrays.asList(new InstructionOutput().setOriginalName(DEFAULT_OUTPUT_ORIGINAL_NAME).setSystemName(DEFAULT_OUTPUT_SYSTEM_NAME).setName("output").setCodec(CloudObjects.asCloudObject(WindowedValue.getFullCoder(StringUtf8Coder.of(), IntervalWindow.getCoder()), /*sdkComponents=*/
    null))));
    List<ParallelInstruction> instructions = Arrays.asList(makeSourceInstruction(StringUtf8Coder.of()), addWindowsInstruction, makeSinkInstruction(StringUtf8Coder.of(), 1));
    FakeWindmillServer server = new FakeWindmillServer(errorCollector);
    int timestamp1 = 0;
    int timestamp2 = 1000000;
    server.addWorkToOffer(makeInput(timestamp1, timestamp1));
    server.addWorkToOffer(makeInput(timestamp2, timestamp2));
    StreamingDataflowWorker worker = makeWorker(instructions, createTestingPipelineOptions(server), false);
    worker.start();
    Map<Long, Windmill.WorkItemCommitRequest> result = server.waitForAndGetCommits(2);
    assertThat(result.get((long) timestamp1), equalTo(setMessagesMetadata(PaneInfo.NO_FIRING, intervalWindowBytes(WINDOW_AT_ZERO), makeExpectedOutput(timestamp1, timestamp1)).build()));
    assertThat(result.get((long) timestamp2), equalTo(setMessagesMetadata(PaneInfo.NO_FIRING, intervalWindowBytes(WINDOW_AT_ONE_SECOND), makeExpectedOutput(timestamp2, timestamp2)).build()));
}

Example 60 with SdkComponents

use of org.apache.beam.runners.core.construction.SdkComponents in project beam by apache.

the class StreamingDataflowWorkerTest method testMergeWindowsCaching.

@Test
public // the first processing having is_new_key set.
void testMergeWindowsCaching() throws Exception {
    Coder<KV<String, String>> kvCoder = KvCoder.of(StringUtf8Coder.of(), StringUtf8Coder.of());
    Coder<WindowedValue<KV<String, String>>> windowedKvCoder = FullWindowedValueCoder.of(kvCoder, IntervalWindow.getCoder());
    KvCoder<String, List<String>> groupedCoder = KvCoder.of(StringUtf8Coder.of(), ListCoder.of(StringUtf8Coder.of()));
    Coder<WindowedValue<KV<String, List<String>>>> windowedGroupedCoder = FullWindowedValueCoder.of(groupedCoder, IntervalWindow.getCoder());
    CloudObject spec = CloudObject.forClassName("MergeWindowsDoFn");
    SdkComponents sdkComponents = SdkComponents.create();
    sdkComponents.registerEnvironment(Environments.JAVA_SDK_HARNESS_ENVIRONMENT);
    addString(spec, PropertyNames.SERIALIZED_FN, StringUtils.byteArrayToJsonString(WindowingStrategyTranslation.toMessageProto(WindowingStrategy.of(FixedWindows.of(Duration.standardSeconds(1))).withTimestampCombiner(TimestampCombiner.EARLIEST), sdkComponents).toByteArray()));
    addObject(spec, WorkerPropertyNames.INPUT_CODER, CloudObjects.asCloudObject(windowedKvCoder, /*sdkComponents=*/
    null));
    ParallelInstruction mergeWindowsInstruction = new ParallelInstruction().setSystemName("MergeWindows-System").setName("MergeWindowsStep").setOriginalName("MergeWindowsOriginal").setParDo(new ParDoInstruction().setInput(new InstructionInput().setProducerInstructionIndex(0).setOutputNum(0)).setNumOutputs(1).setUserFn(spec)).setOutputs(Arrays.asList(new InstructionOutput().setOriginalName(DEFAULT_OUTPUT_ORIGINAL_NAME).setSystemName(DEFAULT_OUTPUT_SYSTEM_NAME).setName("output").setCodec(CloudObjects.asCloudObject(windowedGroupedCoder, /*sdkComponents=*/
    null))));
    List<ParallelInstruction> instructions = Arrays.asList(makeWindowingSourceInstruction(kvCoder), mergeWindowsInstruction, // Use multiple stages in the maptask to test caching with multiple stages.
    makeDoFnInstruction(new PassthroughDoFn(), 1, groupedCoder), makeSinkInstruction(groupedCoder, 2));
    FakeWindmillServer server = new FakeWindmillServer(errorCollector);
    StreamingDataflowWorker worker = makeWorker(instructions, createTestingPipelineOptions(server), false);
    Map<String, String> nameMap = new HashMap<>();
    nameMap.put("MergeWindowsStep", "MergeWindows");
    worker.addStateNameMappings(nameMap);
    worker.start();
    server.addWorkToOffer(buildInput("work {" + "  computation_id: \"" + DEFAULT_COMPUTATION_ID + "\"" + "  input_data_watermark: 0" + "  work {" + "    key: \"" + DEFAULT_KEY_STRING + "\"" + "    sharding_key: " + DEFAULT_SHARDING_KEY + "    cache_token: 1" + "    work_token: 1" + "    is_new_key: 1" + "    message_bundles {" + "      source_computation_id: \"" + DEFAULT_SOURCE_COMPUTATION_ID + "\"" + "      messages {" + "        timestamp: 0" + "        data: \"" + dataStringForIndex(0) + "\"" + "      }" + "    }" + "  }" + "}", intervalWindowBytes(WINDOW_AT_ZERO)));
    Map<Long, Windmill.WorkItemCommitRequest> result = server.waitForAndGetCommits(1);
    Iterable<CounterUpdate> counters = worker.buildCounters();
    // These tags and data are opaque strings and this is a change detector test.
    // The "/u" indicates the user's namespace, versus "/s" for system namespace
    String window = "/gAAAAAAAA-joBw/";
    String timerTagPrefix = "/s" + window + "+0";
    ByteString bufferTag = ByteString.copyFromUtf8(window + "+ubuf");
    ByteString paneInfoTag = ByteString.copyFromUtf8(window + "+upane");
    String watermarkDataHoldTag = window + "+uhold";
    String watermarkExtraHoldTag = window + "+uextra";
    String stateFamily = "MergeWindows";
    ByteString bufferData = ByteString.copyFromUtf8("data0");
    // Encoded form for Iterable<String>: -1, true, 'data0', false
    ByteString outputData = ByteString.copyFrom(new byte[] { (byte) 0xff, (byte) 0xff, (byte) 0xff, (byte) 0xff, 0x01, 0x05, 0x64, 0x61, 0x74, 0x61, 0x30, 0x00 });
    // These values are not essential to the change detector test
    long timerTimestamp = 999000L;
    WorkItemCommitRequest actualOutput = result.get(1L);
    // Set timer
    verifyTimers(actualOutput, buildWatermarkTimer(timerTagPrefix, 999));
    assertThat(actualOutput.getBagUpdatesList(), Matchers.contains(Matchers.equalTo(Windmill.TagBag.newBuilder().setTag(bufferTag).setStateFamily(stateFamily).addValues(bufferData).build())));
    verifyHolds(actualOutput, buildHold(watermarkDataHoldTag, 0, false));
    // No state reads
    assertEquals(0L, splitIntToLong(getCounter(counters, "WindmillStateBytesRead").getInteger()));
    // Timer + buffer + watermark hold
    assertEquals(Windmill.WorkItemCommitRequest.newBuilder(actualOutput).clearCounterUpdates().clearOutputMessages().build().getSerializedSize(), splitIntToLong(getCounter(counters, "WindmillStateBytesWritten").getInteger()));
    // Input messages
    assertEquals(VarInt.getLength(0L) + dataStringForIndex(0).length() + addPaneTag(PaneInfo.NO_FIRING, intervalWindowBytes(WINDOW_AT_ZERO)).size() + // proto overhead
    5L, splitIntToLong(getCounter(counters, "WindmillShuffleBytesRead").getInteger()));
    Windmill.GetWorkResponse.Builder getWorkResponse = Windmill.GetWorkResponse.newBuilder();
    getWorkResponse.addWorkBuilder().setComputationId(DEFAULT_COMPUTATION_ID).setInputDataWatermark(timerTimestamp + 1000).addWorkBuilder().setKey(ByteString.copyFromUtf8(DEFAULT_KEY_STRING)).setShardingKey(DEFAULT_SHARDING_KEY).setWorkToken(2).setCacheToken(1).getTimersBuilder().addTimers(buildWatermarkTimer(timerTagPrefix, timerTimestamp));
    server.addWorkToOffer(getWorkResponse.build());
    long expectedBytesRead = 0L;
    Windmill.GetDataResponse.Builder dataResponse = Windmill.GetDataResponse.newBuilder();
    Windmill.KeyedGetDataResponse.Builder dataBuilder = dataResponse.addDataBuilder().setComputationId(DEFAULT_COMPUTATION_ID).addDataBuilder().setKey(ByteString.copyFromUtf8(DEFAULT_KEY_STRING)).setShardingKey(DEFAULT_SHARDING_KEY);
    // These reads are skipped due to being cached from accesses in the first work item processing.
    // dataBuilder
    // .addBagsBuilder()
    // .setTag(bufferTag)
    // .setStateFamily(stateFamily)
    // .addValues(bufferData);
    // dataBuilder
    // .addWatermarkHoldsBuilder()
    // .setTag(ByteString.copyFromUtf8(watermarkDataHoldTag))
    // .setStateFamily(stateFamily)
    // .addTimestamps(0);
    dataBuilder.addWatermarkHoldsBuilder().setTag(ByteString.copyFromUtf8(watermarkExtraHoldTag)).setStateFamily(stateFamily).addTimestamps(0);
    dataBuilder.addValuesBuilder().setTag(paneInfoTag).setStateFamily(stateFamily).getValueBuilder().setTimestamp(0).setData(ByteString.EMPTY);
    server.addDataToOffer(dataResponse.build());
    expectedBytesRead += dataBuilder.build().getSerializedSize();
    result = server.waitForAndGetCommits(1);
    counters = worker.buildCounters();
    actualOutput = result.get(2L);
    assertEquals(1, actualOutput.getOutputMessagesCount());
    assertEquals(DEFAULT_DESTINATION_STREAM_ID, actualOutput.getOutputMessages(0).getDestinationStreamId());
    assertEquals(DEFAULT_KEY_STRING, actualOutput.getOutputMessages(0).getBundles(0).getKey().toStringUtf8());
    assertEquals(0, actualOutput.getOutputMessages(0).getBundles(0).getMessages(0).getTimestamp());
    assertEquals(outputData, actualOutput.getOutputMessages(0).getBundles(0).getMessages(0).getData());
    ByteString metadata = actualOutput.getOutputMessages(0).getBundles(0).getMessages(0).getMetadata();
    InputStream inStream = metadata.newInput();
    assertEquals(PaneInfo.createPane(true, true, Timing.ON_TIME), PaneInfoCoder.INSTANCE.decode(inStream));
    assertEquals(Arrays.asList(WINDOW_AT_ZERO), DEFAULT_WINDOW_COLLECTION_CODER.decode(inStream, Coder.Context.OUTER));
    // Data was deleted
    assertThat("" + actualOutput.getValueUpdatesList(), actualOutput.getValueUpdatesList(), Matchers.contains(Matchers.equalTo(Windmill.TagValue.newBuilder().setTag(paneInfoTag).setStateFamily(stateFamily).setValue(Windmill.Value.newBuilder().setTimestamp(Long.MAX_VALUE).setData(ByteString.EMPTY)).build())));
    assertThat("" + actualOutput.getBagUpdatesList(), actualOutput.getBagUpdatesList(), Matchers.contains(Matchers.equalTo(Windmill.TagBag.newBuilder().setTag(bufferTag).setStateFamily(stateFamily).setDeleteAll(true).build())));
    verifyHolds(actualOutput, buildHold(watermarkDataHoldTag, -1, true), buildHold(watermarkExtraHoldTag, -1, true));
    // State reads for windowing
    assertEquals(expectedBytesRead, splitIntToLong(getCounter(counters, "WindmillStateBytesRead").getInteger()));
    // State updates to clear state
    assertEquals(Windmill.WorkItemCommitRequest.newBuilder(actualOutput).clearCounterUpdates().clearOutputMessages().build().getSerializedSize(), splitIntToLong(getCounter(counters, "WindmillStateBytesWritten").getInteger()));
    // No input messages
    assertEquals(0L, splitIntToLong(getCounter(counters, "WindmillShuffleBytesRead").getInteger()));
    CacheStats stats = worker.stateCache.getCacheStats();
    LOG.info("cache stats {}", stats);
    assertEquals(1, stats.hitCount());
    assertEquals(4, stats.missCount());
}