Scaling Machine Learning Pipelines in Cloud
summary
Kubeflow on Kubernetes lets teams write reusable ML workflows in Python that compile to Argo YAML and run as containerized pods, with node selectors for GPUs and cluster autoscaler adding nodes when pods queue. It demonstrates Kubeflow pipelines compiled from a Python DSL, uses Katib hyperparameter tuning and metadata stores, and exposes models via KFServing autoscaling backed by KNative/Istio. Demo runs locally with MiniKF demo, and large real-world scale (OpenAI ~2,500-node training clusters) shows the approach can scale, though operator and version management remains a maintenance burden.
AI-written summary · about
chapters (40)
- 0:00Intro
- 1:15About Salman
- 2:35Giveaway
- 3:39Agenda
- 4:37Who is this talk for
- 8:38Training Models
- 11:45Docker
- 16:58Where can we run Kubernetes
- 17:22Benefits of Kubernetes
- 19:11Scaling Pipelines with Kubernetes
- 19:35Introducing Kubernetes
- 20:25Data Science Machine Learning Engineers
- 23:03Kubernetes tools
- 25:07Operators
- 26:46Fashion mnist
- 27:04Neural Network
- 28:08Hyper Parameters
- 29:16Auto ML
- 30:39Serving
- 31:10kubeflow
- 32:33Why Kubernetes
- 33:41Data Science and Machine Learning
- 34:31QFlow Pipeline
- 37:21MiniCube
- 44:18Python Pipeline SDK
- 44:57GPU Pipeline
- 45:28GPU Worker
- 46:07Experimenting
- 46:19Running the pipeline
- 46:30Kale
- 46:43Kubernetes Extension
- 47:01Adding Labels
- 47:16Creating a Pipeline
- 47:43Kubernetes Autoscaling
- 49:07Considerations
- 50:38OpenAI
- 52:05Summary
- 52:46References
- 53:15Thanks
- 53:41Outro
related talks
51mMachine Learning using Kubeflow and Kubernetes
46mDeploying scalable Machine Learning models with Kubernetes
36mMachine learning pipeline Kubeflow, let's start
39mOperating Deep Learning Pipelines Anywhere Using Kubeflow
39mDeploy your ML Workloads on Kubernetes with Kubeflow
35mOrganizing data engineering around Kubeflow
05mCI/CD Pipelines and Machine Learning
05mLightning Talk: Build a Reproducible ML Workflow with Kubeflow Pipelines