1119 - Kubeflow.pdf

1、Update confidential designator hereVersion number here V00000A Kubernetes Native MLOps PlatformSession:1119Building End-to-End ML Pipelines with KubeflowMatthew PrahlPrincipal Software Engineer,OpenShift AI,Red HatKubeflow Pipelines Maintainer1Update confidential designator hereVersion number here V

2、00000A Unified,Cloud-Native Platform:Industry standard for building AI/ML workflows on Kubernetes.Designed for reproducibility,portability,and scalability across any cloud or on-prem environment.Maximize your investment in existing Kubernetes and OpenShift clusters.Encompasses the entire MLOps lifec

3、ycle:from data preparation and experimentation to training and deployment.Integrates with many of the best open-source tools(e.g.,TensorFlow,PyTorch,Spark,KServe).Kubeflow Pipelines:The orchestrator that ties all the pieces together for end-to-end MLOps workflows.Why Kubeflow?2Update confidential de

4、signator hereVersion number here V00000Community:Originally created at Google,but has grown into a much wider community.Backed by contributions from major players like Red Hat,Apple,Capital One,Nutanix,DHL,and many others.Kubeflow Community3Update confidential designator hereVersion number here V000

5、004Kubeflow ProjectsUpdate confidential designator hereVersion number here V00000Kubeflow NotebooksJupyter Notebooks running on a Kubernetes clusterUseful for experimenting on a small scaleFamiliar experience for most data scientistsKubeflow PipelinesWraps other Kubeflow projects when distributed wo

6、rkloads are neededUser provides any custom steps(components)to fill in the gaps or for business logicMost steps are written in native Python code but it can be any container commandLocal mode can be used to run Pipelines in subprocesses or DockerOrchestration and Execution5Update confidential design