Evolving Deep Learning Platform with Knative
summary
Describes how Baidu used Knative components—Eventing, Serving, and Build—to rebuild its internal deep learning platform: wiring Eventing to automate pipelines from training to inference, using Serving for smart routing and autoscaling of inference services, and using Build to automate creation of training job container images, resulting in a reported 20% reduction in resource consumption.
AI-written summary · about
chapters (17)
- 0:00Intro
- 2:05Why DLP on Kubernetes
- 3:50What we did
- 5:33Why serverless
- 7:53What' s changed in 2018/2019
- 9:49Our current stack
- 12:28Knative components
- 15:25Our old architecture
- 16:17Knative build - DLP example
- 17:09Tektoncd-pipeline with buildkit
- 18:19What is Knative serving
- 20:57Knative serving - Ingress
- 23:45Knative serving - custom autoscale class
- 25:10Knative serving - cold start
- 26:20Knative with container instance - old solution
- 30:17Compute and Network on Edge
- 32:32Knative eventing - CloudEvents
related talks
39mEvolving Deep Learning Platform with Knative
22mYi Wang, Tech Lead of AI Platform, Baidu, NYC 2017
50mBuilding Next-Generation Apps with Knative
01h 00mEventing with Knative and Cloud Run: From Basics to Advanced
52mServerless with Knative
18mKeynote: Optimizing AI Workload Scheduling: Bilibili's Journey To an Efficient Cloud Native AI Platform
17mKeynote: Scaling Deep Learning Models in Production Using Kubernetes
45mServerless with Knative