Playground
Use the Data Modelling Agent to run evals against branch changes before you merge.
Playground is a Data Modelling Agent workflow for testing changes before you merge them. It runs evals against your branch and compares the results with main, helping you catch regressions in dbt models, semantic definitions, playbooks, Org Context, and evals themselves.
Admin only
Running evals is available only to workspace admins.
Progressive rollout
Playground is being rolled out gradually. If it isn't available in your workspace yet, ask your Sundial contact.
Run evals on a branch
Start a Data Modelling Agent conversation in the Sundial UI or through the Sundial MCP server, then ask:
- "Run the evals on my branch changes."
- "Run the evals against my branch before I merge."
Be explicit that you want to test the branch. Asking only to "run the evals" runs them against main.
The agent reviews the change and suggests the evals most likely to be affected. You can confirm or edit that list before anything is built or graded.
Prerequisites
- GitHub content sync must be active. Evals must live in the connected repository under
sundial/evals/; branch runs include only these Git-backed evals. - dbt SQL changes require a writable dev materialization dataset. If you use a customer-owned warehouse, create this dataset and grant Sundial write access. Semantic, playbook, Org Context, and eval-only changes do not require it.
Review the result
The verdict compares the branch with main:
- Improved: the branch performed better.
- Regressed: the branch performed worse. Fix the change and rerun the relevant evals before merging.
- Inconclusive: the results do not show a clear difference. Review the individual eval results before deciding whether to merge.
The Evals page records which branch each historical run used. Treat a regression as a merge blocker; merge only after the branch results are acceptable.