Introduction to Using and Use Cases of KubeFlow
summary
Kubeflow 0.4→0.6 demo builds a product recommender end-to-end using Jupyter notebooks, exported to TFJob and TensorFlow Serving, with storage on MinIO and deployment on OpenShift. Pipeline automation uses Argo (Airflow mentioned) to schedule continuous monthly retrains; custom components and a dual-serving switch solve safe model rollouts to avoid serving downtime. All code and deployment steps live in a GitHub repo; the example runs, requires manual Docker image builds and small custom glue for Kafka-based request/response.
AI-written summary · about
chapters (25)
- 0:00Intro
- 0:56What are containers?
- 2:03What is kubernetes
- 3:36Kubernetes desired state management
- 5:55What is kubeflow?
- 6:51Kubeflow components
- 7:22What's new in 0.6?
- 8:15Meet Kubeflow
- 8:55Cluster view
- 10:33Motivating example
- 10:48Collaborative filtering
- 11:12Rating Matrix
- 11:29Using Minio as a centralized storage
- 13:12Kubeflow options for machine learning and model serving
- 14:01Using Jupiter for creating implementation
- 16:46Converting implementation to TF Job
- 17:42Running TF Job
- 18:40Deploying TF-serving
- 19:23Using TF Serving in streaming applications
- 20:13Concept drift
- 21:41Continuous model updates implementation
- 22:21Additional custom components
- 23:14Argo workflow
- 25:21Bringing it all together
- 27:10Try this yourself
related talks
46mKubeflow, TensorFlow on Kubernetes
36mIntroducing Kubeflow (w. Special Guests Tensorflow and Apache Spark)
14mKubeflow: Where ML Meets Kubernetes
29mKubeflow: TensorFlow on Kubernetes
31mKubeflow - Data Science on Steroids
35mKubeflow: Data Scientist ready?
36mIntroducing Kubeflow (w. Special Guests Tensorflow and Apache Spark)
29mMachine Learning on Kubernetes Birds of a Feather