Skip to content

Forecasting

Hybernate uses a Holt-Winters double seasonal model to learn your workload's demand patterns and predict future resource needs. This page covers the mathematical model, how it builds confidence over time, and how it handles pattern shifts.

The Idea

Most workloads follow predictable patterns:

  • Daily: Traffic peaks during business hours, drops overnight
  • Weekly: Weekdays are busier than weekends

Hybernate learns these two patterns independently by feeding hourly CPU observations into a Holt-Winters model. Once it has enough data and confidence, it uses the learned patterns to:

  • Confirm that a workload is genuinely idle (not just in a temporary lull)
  • Recommend how many replicas a workload needs in the next hour

How It Works

The model tracks four components:

Component What it represents Smoothing parameter
Level The baseline demand right now \( \alpha = 0.1 \)
Trend Whether demand is growing or shrinking \( \beta = 0.01 \)
Daily factors 24 multipliers, one per hour of day \( \gamma_1 = 0.05 \)
Weekly factors 168 multipliers, one per hour of week \( \gamma_2 = 0.01 \)

Each hour, the model records the actual CPU observation \( Y(t) \), compares it to what it predicted (for confidence scoring), and updates all four components using exponential smoothing:

\[ \begin{aligned} L(t) &= \alpha \cdot \frac{Y(t)}{D(t) \cdot W(t)} + (1 - \alpha) \cdot \bigl(L(t{-}1) + T(t{-}1)\bigr) \\[6pt] T(t) &= \beta \cdot \bigl(L(t) - L(t{-}1)\bigr) + (1 - \beta) \cdot T(t{-}1) \\[6pt] D(t) &= \gamma_1 \cdot \frac{Y(t)}{L(t) \cdot W(t)} + (1 - \gamma_1) \cdot D(t - s_1) \\[6pt] W(t) &= \gamma_2 \cdot \frac{Y(t)}{L(t) \cdot D(t)} + (1 - \gamma_2) \cdot W(t - s_2) \end{aligned} \]

Where \( s_1 = 24 \) (daily season length) and \( s_2 = 168 \) (weekly season length).

The forecast for \( h \) hours ahead is:

\[ F(t+h) = \bigl(L(t) + h \cdot T(t)\bigr) \times D(t+h \bmod s_1) \times W(t+h \bmod s_2) \]

Phase Lifecycle

The engine doesn't start making decisions immediately. It progresses through phases as it collects data and builds confidence:

Observing → DailySuggesting → DailyActive → WeeklySuggesting → FullyActive

Phase Requires Behavior
Observing < 24 data points Collecting data. No predictions yet.
DailySuggesting 24+ data points Daily predictions available but not yet trusted. Dry run is always forced. Predictions are logged but never applied.
DailyActive Daily confidence >= threshold Daily predictions drive decisions. Weekly patterns still learning.
WeeklySuggesting 168+ data points Weekly predictions available but not yet trusted.
FullyActive Weekly confidence >= threshold Both daily and weekly patterns drive decisions.

The confidence threshold is configurable per workload via spec.prediction.confidence (default: 85%).

Confidence Scoring

Confidence is measured using Mean Absolute Percentage Error (MAPE) over a rolling 24-hour window:

\[ C = 1 - \frac{1}{n} \sum_{i=1}^{n} \frac{|F(i) - Y(i)|}{Y(i)} \]

A confidence of 85% means predictions are on average within 15% of actual values.

  • Confidence starts at 0% and climbs as the model learns
  • The scorer needs a full 24-hour window before it reports a meaningful score
  • Each season (daily and weekly) has its own independent confidence score

Anomaly Detection

Traffic patterns change. A marketing campaign, a new feature launch, or a seasonal shift can invalidate what the model has learned.

Hybernate detects these regime changes using z-score anomaly detection:

  1. Each hour, the prediction error is converted to a z-score using the running mean \( \mu \) and standard deviation \( \sigma \) (computed via Welford's online algorithm):

    \[ z(t) = \frac{|Y(t) - F(t)| - \mu}{\sigma} \]

  2. If \( z(t) > 3.0 \), the observation is flagged as an anomaly

  3. If 3 or more anomalies occur within a 24-hour window, the engine declares a regime change

On regime change, the engine demotes its phase:

  • FullyActive → WeeklySuggesting
  • WeeklySuggesting → DailySuggesting
  • DailySuggesting → Observing

This forces the model to re-earn confidence with the new pattern before making decisions again.

Persistence

The entire engine state (model parameters, seasonal factors, confidence scores, anomaly detector) is serialized to JSON and stored in the ManagedWorkload CR's status. This means:

  • The engine survives operator restarts
  • No external database or persistent volume is needed
  • Each workload has its own independent engine

Wall-Clock Alignment

Seasonal slots are keyed to wall-clock time (UTC), not to a running counter. This means:

  • If a workload is paused for 6 hours, the model doesn't lose alignment. The next observation goes into the correct hour-of-day slot
  • Monday 9am always maps to the same slot, regardless of gaps

Tuning

The default smoothing parameters work well for most workloads:

Parameter Default Effect of increasing
\( \alpha \) 0.1 Level reacts faster to demand changes
\( \beta \) 0.01 Trend detection is more aggressive
\( \gamma_1 \) 0.05 Daily pattern adapts faster
\( \gamma_2 \) 0.01 Weekly pattern adapts faster

Lower values produce more stable, smoother forecasts. Higher values make the model more reactive but noisier. The defaults are conservative and prioritize stability over reactivity.