Skip to content

Adaptive Ensemble (Surge Profiles)

Ruptura v6.1 introduces adaptive ensemble weighting — the engine continuously learns which prediction model fits your infrastructure's traffic patterns best, and shifts weight toward it automatically.

The five models

Model Strengths Weaknesses
CA-ILR (Dual-scale) O(1) update, detects acceleration, edge-native Linear assumption within window
ARIMA Strong on stationary series with trends Computationally heavier, assumes linearity
Holt-Winters Excellent on seasonal / periodic patterns Requires clear seasonality
MAD (Median Absolute Deviation) Robust to outliers No temporal memory
EWMA (Exponentially Weighted Moving Average) Smooth, reacts to recent data Slow to detect sudden shifts

Weight adaptation algorithm

Every 60 seconds, Ruptura evaluates prediction accuracy over the past 1-hour sliding window:

error_model_i = MAE(predicted_i, actual) over last 1 hour

weight_i = (1 / error_i) / Σ_j (1 / error_j)

Models with lower recent prediction error receive higher weight. Weights are normalised to sum to 1.0.

The final ensemble prediction is:

prediction(t) = Σ_i weight_i(t) × prediction_i(t)

Enable adaptive weighting

# ruptura.yaml
ensemble:
  adaptive: true

When false (default), all five models are weighted equally at 0.20 each.

Monitoring weights via API

GET /api/v2/ensemble/{host}

{
  "host": "web-01",
  "updated_at": "2026-04-28T10:00:00Z",
  "weights": {
    "ca_ilr":       0.35,
    "arima":        0.22,
    "holt_winters": 0.18,
    "mad":          0.14,
    "ewma":         0.11
  }
}

Surge profiles — how Ruptura handles load spikes

Ruptura does not require you to pre-define "surge profiles" or maintenance windows. The dissipative fatigue formula (λ healing) and the adaptive ensemble work together to avoid false alarms during planned load events:

  1. At spike onset — ILR_burst slope increases, R rises.
  2. If the spike is brief — fatigue dissipates (λ recovery), R falls back.
  3. If the spike sustains — R stays elevated, ensemble weights shift toward models that track the new pattern.
  4. Weights rebalance — once the spike ends, MAE improves for baseline-tracking models and they regain weight.

This means a nightly backup job gradually "teaches" the ensemble that high CPU at 02:00 UTC is expected — no manual intervention required.

Prometheus self-metrics for ensemble

rpt_ensemble_weight{host="web-01",model="ca_ilr"}       0.35
rpt_ensemble_weight{host="web-01",model="arima"}         0.22
rpt_ensemble_weight{host="web-01",model="holt_winters"}  0.18
rpt_ensemble_weight{host="web-01",model="mad"}           0.14
rpt_ensemble_weight{host="web-01",model="ewma"}          0.11