ML Ops design patterns with Kubeflow Pipelines
summary
Shows concrete MLOps patterns implemented with Kubeflow Pipelines using the Python SDK on a standalone GKE install: pipelines log to a metadata server, support Keras Tuner distributed mode (launches tuner workers as Kubernetes jobs), use step-level caching and versioning, spin up a TensorBoard server and TensorFlow Serving instances, and trigger retraining when tfdv drift exceeds thresholds. Example runs tune and then train the top three hyperparameter sets in parallel, the cluster autoscaled to run jobs, and models deploy only when eval meets the accuracy gate.
AI-written summary · about
chapters (5)
related talks
22mBuilding a MLOps Platform with Kubeflow on GKE
54mScaling Machine Learning Pipelines in Cloud
06mMLOps, Kubeflow, and Tekton
51mMachine Learning using Kubeflow and Kubernetes
14mMLOps: DevOps for AI with KubeFlow Pipelines
39mOperating Deep Learning Pipelines Anywhere Using Kubeflow
29mML Engineering Kubeflow
30mEnd-to-End MLOps with MLflow and Kubeflow