Analytics

The Analytics page (/analytics) is where you look ahead and look back: forecasts of where load is going, anomalies that have already happened, and scaling decisions that prepare your workloads for what's coming. It brings together data from the predictive scaler and the anomaly detector.

Summary cards

• Active Forecasts — workloads with a current forecast.
• Critical Anomalies — unacknowledged anomalies at critical severity.
• Pending Decisions — scaling decisions waiting for approval.
• Total Anomalies — anomalies in the selected time range.

Use the filters at the top of the page to narrow everything by cluster, namespace and time range.

Forecasts

Each workload with enough history gets a forecast card: a line showing expected load and a shaded band showing the range it will most likely fall within (95% confidence).

How forecasting works

The predictive scaler uses Holt's linear trend method (double exponential smoothing). It tracks two things over the metric's history — the current level and the current trend — and projects them forward:

Level:    L(t) = α·y(t) + (1-α)·(L(t-1) + T(t-1))
Trend:    T(t) = β·(L(t) - L(t-1)) + (1-β)·T(t-1)
Forecast: ŷ(t+h) = L(t) + h·T(t)          (h steps ahead)

• α (alpha) controls how quickly the level reacts to new data; β (beta) does the same for the trend. The defaults (0.3 and 0.1) suit most workloads and can be tuned per deployment — see predictive-scaler configuration.
• The confidence band is ±1.96 × RMSE of the model's recent errors.
• A workload needs at least 24 data points before its first forecast. Until then, its card explains that it's still collecting data.

Anomalies

A table of anomalies detected across your clusters, newest first.

Column	Description
Metric	What was measured, for example cpu_usage_pct or crash_loop_count.
Resource	The affected resource or namespace.
Value / Z-score	The observed value, and how many standard deviations it was from normal.
Baseline	The rolling average the value was compared against.
Severity	Low, medium, high or critical.
Healing	Whether a self-healing action ran, and its result.
Detected	How long ago it was detected.

Filter by severity, and select an anomaly to see its detail, the remediation that ran (if any), and links to related incidents. Select Acknowledge to mark that you've seen it; acknowledged anomalies drop out of the critical count.

Scaling decisions

When a forecast shows a workload will need more (or less) capacity, the predictive scaler proposes a scaling decision. Each card shows:

• the workload (namespace and name);
• current → recommended replicas, with an up or down indicator;
• the reason and the model's confidence;
• the status: pending, approved, applied or rejected.

Approve applies the recommendation: if the workload has a HorizontalPodAutoscaler, KubeOpera raises its minimum replicas so the HPA and the forecast work together; otherwise it scales the Deployment directly. Reject records that you declined it and leaves the cluster unchanged. Both actions are recorded with who took them.

Decisions can also be handled by AI: the Cost Optimizer agent can review pending decisions and approve or reject them on your behalf when your role and policy allow it — every such action appears in the agent run's history.

Next steps

• Predictive scaler — configuration and API.
• Anomaly detector — alert rules and self-healing.
• Reactive AI Pipeline — how anomalies feed automatic remediation.