KubeFlow’s Serverless Component: 10x Faster, a 1/10 of the Effort
Talkat KubeCon + CloudNativeCon North America 2019 · on Tuesday, 2019-11-19· at 16:25· Room 6F - San Diego Convention Center, Upper Level·
summary
Claims a Nuclio + mlrun stack makes pipeline steps up to 10x faster and 1/10 the effort by converting Jupyter cells to functions, auto-building containers, and wiring artifacts/metrics automatically. Examples show ~30% throughput gain for simple ingress, GPU model serving up to 4x improvement, and parallel training with Horovod via the MPI operator (a half-hour run on 4 worker pods). Functions support multiple triggers, zero-to-idle autoscaling and explicit worker counts (demo used 16 workers and a 60-worker parallel run) and deploys through KFServing.
AI-written summary · about
chapters (17)
- 0:00Introduction
- 0:24Data Science and ML Ops
- 1:14Plumbing
- 2:23ML Pipeline Example
- 4:51Introducing New Clio
- 6:47Resource Utilization
- 8:00Data Preparation and Training
- 9:31Automation
- 14:04Running ML Functions
- 17:45Orit Jupiter
- 18:53KubeFlow Pipeline
- 23:22KubeFlow Distributed Tensorflow
- 24:21KubeFlow Tasks
- 25:27KubeFlow Machine Learning
- 26:32ML Run UI
- 28:33Artifact Path
- 29:43Questions
related talks
54mEconomics and Best Practices of Running AI/ML Workloads on Kubernetes
31mIntroduction to Using and Use Cases of KubeFlow
29mServerless on Kubernetes: What We Learned From Customers
14mKubeflow: Where ML Meets Kubernetes
31mKubeflow Intro – Michał Jastrzębski & Ala Raddaoui, Intel
01h 00mServerless Madness on Kubernetes
31mKubeflow - Data Science on Steroids
31mManaging Edge Computing with Serverless