Start with one workload
Connect a noncritical workload to your own metrics. Begin with observation, review the advice, and gradually authorize actions.
- Add a source. Configure a Metrics data source and check the connection.
- Define metric templates. Choose CPU, memory, request volume, or other inputs. Confirm queries, units, and aggregation scope.
- Create a module. Select the namespace, workload, and container. Connect templates and configure intervals, bounds, schedules, and fallback.
- Run and compare. Inspect history quality, model predictions, backtests, actual performance, and recommendations.
What to review first
| Area | What to check |
|---|---|
| History and datasets | Scope, units, sample interval, gaps, and freshness. |
| Model runs | Successful runs and reasons for insufficient history, missing parameters, or timeout. |
| Backtests and evaluation | CPU and memory forecast errors, particularly peak underestimation. |
| Recommendation details | Resource and replica advice, constraints, consensus, expiry, and fallback reasons. |
Enable automation gradually
Start in Observe and review advice in Recommend. Authorize Admission and Auto separately when ready to apply changes. Both write paths are disabled by global settings by default.
Confirm metric meaning, sufficient history, explainable advice, appropriate bounds, and fallback values. Review admission and replica scaling separately.
Move from collection to model forecasts
Confirm history collection first, then enable prediction. Choose Dynamic algorithms to compare methods, or Fixed algorithm and select a specific method. Non-baseline models need backtest promotion before their results can be adopted.
Fallback defaults are expected while prediction is disabled or No prediction is selected. Stay in Observe or Recommend while reviewing forecasts, scores, and look-ahead details. Authorize admission or scaling separately. See prediction modes.