Kubeflow From the End User’s Perspective: The Good, The Bad, and The Ugly
summary
Survey of 50+ Kubeflow users finds concrete wins and concrete blockers: Kubeflow uses Kubernetes for large-scale training and TensorFlow Serving for inference, but resource handling breaks down — many teams run on a single server with 8 GPU not a GPU farm, while multimodel pipelines can require a 16-model pipeline needing sequential GPUs, forcing unrealistic allocations. Pain points: poor experiment/model tracking and model DB, clumsy data-labeling collaboration, unreliable resource estimation and risky autoscaling; teams add ~20% buffer instead of safe autoscale.
AI-written summary · about
chapters (5)
- 0:00Start
- 24:48Data Label for Supervised Learning
- 34:40Cloud Native Intelligence
- 35:55How Can We Estimate How Much Resource the Model Needs during Training
- 40:08Optimization to Cpu and Gpu
related talks
36mIntroducing Kubeflow (w. Special Guests Tensorflow and Apache Spark)
21mIntroducing Kubeflow (w. Special Guests Tensorflow and Apache Spark)
36mKubeflow: Tensorflow on Kubernetes
46mKubeflow, TensorFlow on Kubernetes
31mKubeflow - Data Science on Steroids
35mKubeflow: Data Scientist ready?
51mMachine Learning using Kubeflow and Kubernetes
14mKubeflow: Where ML Meets Kubernetes