Start from the workload
A module describes a Kubernetes workload and its forecasting, resource, and execution policies. Metrics supply inputs, models predict change, and recommendations combine forecasts with constraints. Actual outcomes feed the next evaluation.
Modules, metrics, and models
| Concept | What it manages |
|---|---|
| Workload module | Target, templates, intervals, thresholds, bounds, mode, and fallback. |
| Metric template | Queries, variables, meaning, units, and scope, rendered for each module. |
| Dataset | Sampling, aggregation, conversion, precision, and gap handling. |
| Model | Central activation, parameters, weights, budgets, and adoption thresholds. |
| Recommendation snapshot | Resource and replica advice with sources, constraints, version, and expiry. |
Four operating modes
- Observe: collect, forecast, and compare without applying resource changes.
- Recommend: review resource and replica advice and its reasoning.
- Admission: apply authorized resource advice to new Pods.
- Auto: let the Operator adjust replicas within bounds, steps, and cooldowns.
Usage boundaries
The platform manages a single cluster, not cross-cluster migration. Read historical observations, capacity estimates, future forecasts, and actual Ready capacity separately. Estimated per-Pod capacity is not a measured maximum.
Keep one replica controller per workload. Admission affects new Pods; running resource configurations require creation or rolling updates.
Configure models and actions separately
Prediction mode determines how advice is calculated: no prediction, a fixed algorithm, or dynamic algorithms. Operating mode determines permitted actions: Observe, Recommend, Admission, or Auto. Configure them independently to compare models while keeping cluster resources unchanged.
New modules start with prediction disabled. Once enabled, fixed mode runs the selected method; dynamic mode uses enabled methods and combines their votes for dynamic replica decisions. See models and prediction modes for eligibility rules.