Introducing Kubeflow (w. Special Guests Tensorflow and Apache Spark)
summary
Kubeflow packages training, experiment tracking, and serving so models move from laptop to cluster with a few commands: kfctl init / generate / apply. It wires Jupyter notebooks to scalable runners and exposes components like JupyterHub, TFJob, Katib, Pipelines, PyTorch support and an alpha metadata API for reproducibility. Recent 0.6 refactor drops ksonnet and temporarily limits cross-cloud install (GCP-first), so expect active development, breaking changes, and nontrivial Kubernetes cost — good for multi-node GPU training and automated retrain pipelines, not for single-machine toy projects.
AI-written summary · about
chapters (19)
- 0:00Introducing Kubeflow
- 1:56Trevor Grant
- 2:58What is Kubernetes?
- 3:31What is Machine Learning/AI.
- 5:40ML Tools are making DS suck less and less every year
- 6:36Problems of Data Scientists and ML Idiots
- 8:01What is Kubeflow?
- 9:12Components Buffet
- 9:54But don't forget about data prep friends!
- 10:34Model persistence/deployment/c
- 11:06Model Serving
- 11:33How does this fit in to the Ecosystem
- 14:46Running that pipeline
- 15:56Execute Pipelines on a Schedule
- 17:37Better Pipelines with Python Code
- 18:12Meta Data!
- 19:03Downsides to kubellow
- 20:19Very DIY KF Cross Cloud Workshop
- 20:39Questions? Half-baked Demo?
related talks
36mIntroducing Kubeflow (w. Special Guests Tensorflow and Apache Spark)
36mIntroducing Kubeflow (w. Special Guests Tensorflow and Apache Spark)
29mKubeflow: TensorFlow on Kubernetes
35mKubeflow: Data Scientist ready?
14mKubeflow: Where ML Meets Kubernetes
31mKubeflow - Data Science on Steroids
46mKubeflow, TensorFlow on Kubernetes
36mKubeflow: Tensorflow on Kubernetes