Kubeflow Katib is a Kubernetes-based, portable service for automated hyperparameter optimization. It provides a practical alternative to rebuilding a generic HPO system when experiments require remote execution and reusable service infrastructure. section 5.0 section 5.4.4
A general HPO service organizes the optimization loop around several responsibilities:
- Runs or coordinates training trials with selected hyperparameter values. section 5.3.2
- Monitors training and records trial metrics and hyperparameter values in a trial history database. section 5.3.2
- Passes completed trial history to a hyperparameter suggestion component, which produces the next candidates. section 5.3.2
Choosing a service over a library
When library-based HPO encounters practical limitations, switch to a service-oriented workflow. Katib is the recommended general-purpose option, while Hyperopt, Optuna, and Ray Tune support local library or framework workflows. section 5.0 section 5.4.4
Use the general HPO service design to trace how training, metric history, trial completion, and candidate suggestion fit into one optimization workflow. section 5.3.2
