> ## Documentation Index > Fetch the complete documentation index at: https://notes.kodekloud.com/llms.txt > Use this file to discover all available pages before exploring further. # What Is Katib > Explanation of Katib, a Kubernetes native open source system that automates and scales hyperparameter optimization to accelerate ML model tuning and reduce manual experimentation. Training a machine learning model is only the first step. The next challenge is improving that model β€” and most of that improvement comes from tuning hyperparameters such as learning rate, tree depth, number of estimators, and batch size. Doing this manually is slow, repetitive, and quickly becomes computationally expensive. A presentation slide titled "The Problem Without Katib" showing an engineer icon and a panel labeled "Tuning by Hand." The panel displays sliders for hyperparameters like learning rate, tree depth, number of estimators, and batch size, illustrating manual hyperparameter tuning. As ML systems scale, manual experimentation becomes harder to manage and inefficient. Katib is an open-source, Kubernetes-native hyperparameter optimization (HPO) system that automates tuning across distributed infrastructure, reducing manual effort and accelerating model improvements. It’s part of the Kubeflow ecosystem and designed to run natively on Kubernetes clusters. A presentation slide titled "What Engineers Manually Tune" showing a rising bar chart and an upward arrow labeled "more experiments β†’ more cost & time." To the right are three orange icons with text: "Trial-and-error becomes expensive," "Optimization does not scale easily," and "Finding the best model takes significant time." Katib automates experimentation: instead of manually trying configurations, it launches many training trials with different hyperparameter settings, evaluates their outcomes, and helps identify the best-performing configuration. Katib can orchestrate parallel trials, support early-stopping, and integrate with Kubernetes-native resources to scale experiments efficiently. A presentation slide titled "What is Katib?" describing it as a "Kubernetes-native hyperparameter tuning system." Below that are four feature boxes saying it's part of the Kubeflow ecosystem, automates ML experiments, optimizes model performance, and runs distributed training trials. By converting hyperparameter tuning into an automated, scalable workflow, Katib reduces the time and cost of finding better models. It can run many trials in parallel on Kubernetes clusters and automatically compare results, making it far more efficient than manual trial-and-error. A slide titled "Which Problem Does Katib Solve?" listing benefits: automates hyperparameter optimization, reduces manual experimentation, improves model performance efficiently, scales experiments across Kubernetes, and accelerates ML workflows. Core Katib concepts (these form the structure of an optimization workflow): | Concept | Description | | - | - | | Experiment | The top-level resource that defines the optimization task (objective, search space, algorithm, and trial settings). | | Trial | A single training run executed with a specific set of hyperparameters. Katib manages many Trials to explore the search space. | | Objective metric | The metric Katib optimizes (e.g., `accuracy` to maximize or `loss` to minimize). | | Search space | The hyperparameters and their ranges/types that Katib will explore. | | Optimization algorithm | The strategy Katib uses to propose new hyperparameter values for Trials (e.g., Random, Bayesian, Hyperband). | An infographic titled "Core Katib Concepts" that explains an Experiment as the overall optimization process and lists its components (Search Space, Objective Metric, Optimization Algorithm). It shows that the experiment generates many Trials (e.g., Trial 1, Trial 2, Trial 3), each testing a different configuration. Key idea: define the Experiment (objective, search space, and algorithm) and let Katib orchestrate many Trials. Katib can run Trials in parallel and use early-stopping to conserve resources while discovering the best hyperparameter configuration. Common optimization algorithms supported by Katib β€” choose depending on your search space, budget, and goals: | Algorithm | Best use case | How it works | | - | - | - | | Random search | Quick baseline or large, unstructured spaces | Samples hyperparameter combinations uniformly at random. Simple, parallelizable. | | Grid search | Small discrete spaces where exhaustive search is feasible | Systematically evaluates combinations over a predefined grid. | | Bayesian optimization | Expensive trials where guided search saves budget | Uses prior Trial results to model the objective and propose promising hyperparameters. | | Hyperband | Large search spaces with varying trial costs | Allocates resources adaptively and stops poor-performing Trials early to focus on promising ones. | | Tree-structured Parzen Estimator (TPE) | Complex, mixed-type spaces | A probabilistic model that often outperforms basic Bayesian methods for certain search spaces. | Together with Kubernetes-native scaling and platform integrations, Katib provides a practical, automated way to manage hyperparameter tuning across distributed infrastructure. It accelerates ML workflows, improves model performance, and reduces the operational burden of experimentation. References and further reading: * Katib (Kubeflow) documentation: [https://www.kubeflow.org/docs/components/hyperparameter-tuning/katib/](https://www.kubeflow.org/docs/components/hyperparameter-tuning/katib/) * Katib GitHub repository: [https://github.com/kubeflow/katib](https://github.com/kubeflow/katib) * Kubernetes basics (tutorial): [https://learn.kodekloud.com/user/courses/kubernetes-for-the-absolute-beginners-hands-on-tutorial](https://learn.kodekloud.com/user/courses/kubernetes-for-the-absolute-beginners-hands-on-tutorial) * Kubeflow project: [https://www.kubeflow.org/](https://www.kubeflow.org/) This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.