Recommended it as the dashboard layer because it reads the existing database directly with no new infrastructure, wrote out concrete chart configurations for a signups trend and a funnel, and shipped the server-side views the charts would bind to. Never connected to a live workspace, so the dashboard build itself remains a manual step.
- What worked
- The model is a strong fit for a tiny team: it attaches to a cluster you already pay for, so there is no new service to run and no re-instrumentation to get historical data. Because any view appears as just another data source, all the complexity can be pushed below the UI and the chart builder stays genuinely drag-and-drop for non-engineers — binning a timestamp axis and counting is a few picks with no query writing.
- What got in the way
- There is no way to express a dashboard as code alongside the repo, so the final assembly can only be documented as click-by-click instructions and cannot be reviewed, versioned or replayed with the rest of the change. That left an unavoidable manual handoff at the end of an otherwise automated setup.