Developer guide

Run, observe, recover, and compare experiments from an existing TauGrid workspace

This guide is for developers and researchers who already have access to a TauGrid workspace. If a platform owner has not yet provisioned one, start with the Platform admin guide instead, or ask your platform team for a workspace connection.

Get started

Your platform team should give you access to a research repository containing the non-secret tau/workspace.connection.yaml descriptor. You do not need to configure or inspect it before running a target: tau run discovers it automatically, obtains the required cluster credentials, and verifies the live workspace contract. The descriptor selects a Kubernetes context and declares how Tau should access it. Most clusters use an existing kubeconfig context; AKS platforms can instead include optional Azure metadata so Tau obtains cluster-user credentials automatically. These are platform-supplied connection settings, not values researchers should copy from kubectl output or edit during onboarding.

From the repository, run tau workspace connection to verify the complete read-only connection before submitting work. Use --offline when only local repository validation is appropriate.

The examples in the TauGrid source repository, including examples/aks-cpu-quickstart, examples/kind-smoke, and examples/ray-tune-smoke, contain reusable workload configurations but do not contain a workspace connection. Kubernetes contexts, workspace names, and any provider-specific access metadata are environment-specific, so TauGrid does not commit them to the public examples. Copy an example into a Tau-enabled research repository, or add the descriptor supplied by the platform team, before submitting it.

Continue with run your first target.

Concepts


Run your first target

Validate, submit, observe, and retrieve one repository target

Share research with a teammate

View the same runs, queue state, logs, and charts from two machines

Developer concepts

Stable objects and lifecycle behind TauGrid commands

Serve a trained model

Render, deploy, inspect, and remove an online endpoint