Keynote: Adapting Kubernetes for Machine Learning Workflows
summary
Bloomberg adapts Kubernetes with custom resources and controllers to run TensorFlow, Spark, JVM and Python jobs, adding first‑class data citizenship (automatic secrets and HDFS mounts), multi‑tenant controls and hardware aware scheduling. Jobs use a t-shirt size API (small/large/GPU) that maps to GPUs or nodes, a TensorFlow CRD and a hyperparameter tuning operator, and the system runs production ML across NLP, CV and time‑series workloads without forcing practitioners to manage containers.
AI-written summary · about
chapters (13)
- 0:00Start
- 0:40Outline
- 1:45Natural Language Processing & Events
- 3:07Applications Contingent on ML
- 5:07Embarking on a New Problem: What?
- 5:48Embarking on a New Problem: How?
- 6:48The Infinite Model Development Loop
- 9:24Kubernetes Life Cycle
- 10:37Deployment Resource (YAML)
- 11:01Deployment Resource: Spec (Replicas)
- 11:12Deployment Resource: Spec (Pod Template)
- 12:43Custom Resources
- 15:56Achieving Nirvana (Reprise)
related talks
12mKeynote: Building Enterprise AI/ML Platform on Kubernetes
42mThe Challenges Managing a Kubernetes-Based Machine Learning Infrastructure
15mKeynote: Machine Learning on Kubernetes Made Easy With Kubeflow
10mKeynote: Machine Learning using Kubeflow and Kubernetes
27mHarnessing Kubernetes for Scalable AI/ML Workloads
22mKeynote: Kubeflow ML on Kubernetes
42mKubernetes for Machine Learning: Managing and Scaling AI Workloads
26mMachine Learning Meets K8s