book

Kubeflow for Machine Learning

by Trevor Grant, Holden Karau, Boris Lublinsky, Richard Liu, Ilan Filonenko

October 2020

Intermediate to advanced

261 pages

6h 19m

English

O'Reilly Media, Inc.

Book available

Read now

Unlock full access

Our Assumption About YouYour Responsibility as a PractitionerConventions Used in This BookCode ExamplesUsing Code ExamplesO’Reilly Online LearningHow to Contact the AuthorsHow to Contact UsAcknowledgmentsGrievances
Model Development Life CycleWhere Does Kubeflow Fit In?Why Containerize?Why Kubernetes?Kubeflow’s Design and Core ComponentsData Exploration with NotebooksData/Feature PreparationTrainingHyperparameter TuningModel ValidationInference/PredictionPipelinesComponent OverviewAlternatives to KubeflowClipper (RiseLabs)MLflow (Databricks)OthersIntroducing Our Case StudiesModified National Institute of Standards and TechnologyMailing List DataProduct RecommenderCT ScansConclusion
Getting Set Up with KubeflowInstalling Kubeflow and Its DependenciesSetting Up Local KubernetesSetting Up Your Kubeflow Development EnvironmentCreating Our First Kubeflow ProjectTraining and Deploying a ModelTraining and Monitoring ProgressTest QueryGoing Beyond a Local DeploymentConclusion
Getting Around the Central DashboardNotebooks (JupyterHub)Training OperatorsKubeflow PipelinesHyperparameter TuningModel InferenceMetadataComponent SummarySupport ComponentsMinIOIstioKnativeApache SparkKubeflow Multiuser IsolationConclusion
Getting Started with PipelinesExploring the Prepackaged Sample PipelinesBuilding a Simple Pipeline in PythonStoring Data Between StepsIntroduction to Kubeflow Pipelines ComponentsArgo: the Foundation of PipelinesWhat Kubeflow Pipelines Adds to Argo WorkflowBuilding a Pipeline Using Existing ImagesKubeflow Pipeline ComponentsAdvanced Topics in PipelinesConditional Execution of Pipeline StagesRunning Pipelines on ScheduleConclusion
Deciding on the Correct ToolingLocal Data and Feature PreparationFetching the DataData Cleaning: Filtering Out the JunkFormatting the DataFeature PreparationCustom ContainersDistributed ToolingTensorFlow ExtendedDistributed Data Using Apache SparkDistributed Feature Preparation Using Apache SparkPutting It Together in a PipelineUsing an Entire Notebook as a Data Preparation Pipeline StageConclusion
Kubeflow ML MetadataProgrammatic QueryKubeflow Metadata UIUsing MLflow’s Metadata Tools with KubeflowCreating and Deploying an MLflow Tracking ServerLogging Data on RunsUsing the MLflow UIConclusion
Building a Recommender with TensorFlowGetting StartedStarting a New Notebook SessionTensorFlow TrainingDeploying a TensorFlow Training JobDistributed TrainingUsing GPUsUsing Other Frameworks for Distributed TrainingTraining a Model Using Scikit-LearnStarting a New Notebook SessionData PreparationScikit-Learn TrainingExplaining the ModelExporting ModelIntegration into PipelinesConclusion
Model ServingModel Serving RequirementsModel MonitoringModel Accuracy, Drift, and ExplainabilityModel Monitoring RequirementsModel UpdatingModel Updating RequirementsSummary of Inference RequirementsModel Inference in KubeflowTensorFlow ServingReviewSeldon CoreDesigning a Seldon Inference GraphTesting Your ModelServing RequestsMonitoring Your ModelsReviewKFServingServerless and the Service PlaneData PlaneExample WalkthroughPeeling Back the Underlying InfrastructureReviewConclusion

The Denoising CT Scans ExampleData Prep with PythonDS-SVD with Apache SparkVisualizationThe CT Scan Denoising PipelineSharing the PipelineConclusion
AutoML: An OverviewHyperparameter Tuning with Kubeflow KatibKatib ConceptsInstalling KatibRunning Your First Katib ExperimentPrepping Your Training CodeConfiguring an ExperimentRunning the ExperimentKatib User InterfaceTuning Distributed Training JobsNeural Architecture SearchAdvantages of Katib over Other FrameworksConclusion
Google CloudTPU-Accelerated InstancesDataflow for TFX
Building Streaming Applications Leveraging Model ServingStream Processing Engines and LibrariesIntroducing CloudflowBuilding Batch Applications Leveraging Model Serving

Content preview from Kubeflow for Machine Learning

Appendix C. Using Model Serving in Applications

In Chapter 8 you learned different approaches for exposing model servers provided by Kubeflow. As described there, Kubeflow provides several ways of deploying trained models and providing both REST and gRPC interfaces for running model inference. However, it falls short in providing support for using these models in custom applications. Here we will present some of the approaches to building applications by leveraging model servers exposed by Kubeflow.

When it comes to applications leveraging model inference, they can be broadly classified into two categories: real time and batch applications. In the real time/stream applications model, inference is done on data directly as it is produced or received. In this case, typically only one request is available at a time and it can be used for inferencing as it arrives. In the batch scenarios all of the data is available up front and can be used for inference either sequentially or in parallel. We will start from the streaming use case and then take a look at possible batch implementations.

Building Streaming Applications Leveraging Model Serving

The majority of today’s streaming applications leverage Apache Kafka as the data backbone of a system. The two possible options for implementing streaming applications themselves are: usage of stream processing engines and usage of stream processing libraries.