Datenbank-Kapazitätsprognosen, die 30 Tage im Voraus warnen

Linear extrapolation pages you 12 hours before the disk fills. Useful for the on-call, useless for provisioning — you can’t resize an EBS volume on Saturday in 12 hours and have it land cleanly. The actually-useful forecast is one that pages 30 days out, with the math right enough to trust. Here’s what works.

Why linear extrapolation fails

• Seasonality. Disk grows faster on weekdays. A linear fit on 7 days extrapolates the wrong slope.
• Step changes. Last week’s deploy doubled write rate. Linear fit smooths it; the real curve has a kink.
• Non-stationarity. Growth rate itself changes over time (acquisition spike, seasonal product launch).

The model stack that works

Three models, ensemble:

• Prophet for the seasonality + holiday + changepoint backbone. Fast to fit per-series, robust on noisy data.
• ETSformer / N-BEATS for high-cardinality scenarios where you fit thousands of series. Transformers handle long histories better than Prophet.
• Linear baseline as a safety floor. If the ensemble disagrees with linear by > 3×, alert the operator that something needs review.

# Prophet recipe — minimum viable forecast
from prophet import Prophet

m = Prophet(
    daily_seasonality=True,
    weekly_seasonality=True,
    yearly_seasonality='auto',
    changepoint_prior_scale=0.05,   # tune up for spiky workloads
    seasonality_prior_scale=10,
)
m.add_country_holidays(country_name='US')
m.add_regressor('deploy_count')   # exogenous
m.fit(history_df)

future = m.make_future_dataframe(periods=30, freq='D')
future['deploy_count'] = predicted_deploys(future)
fcst = m.predict(future)

# Use yhat_lower / yhat_upper bounds, not yhat alone — the band is what alerts.

Exogenous variables that move the needle

• Day-of-week / business-day flag — the single biggest accuracy gain.
• Holidays — country-specific. Black Friday, Lunar New Year, regional holidays for B2C.
• Deploy events — regimes change after deploys. Inject as event markers; Prophet handles them as “holidays”.
• Marketing campaign flags — if the team can post events to the metrics pipeline, you get free correlation.

Evaluating forecasts honestly

• Backtest with sliding window. Fit on weeks 1-4, predict week 5, score. Slide.
• Score with MAPE for level metrics (disk, connections), SMAPE for noisier metrics (QPS).
• Track calibration of bands: if your 90% interval contains the actual value 70% of the time, your bands are too narrow.

Alerting on forecast breaches

The alert isn’t “disk is full.” It’s “disk will be full in N days.”

# Pseudo-rule
if forecast.crosses_threshold(metric='disk_used',
                              threshold=0.85 * disk_total,
                              within_days=30):
    page(severity='warning',
         message=f"Disk on {host} forecast to breach 85% in {days} days")

Useful: tier severity by lead time. 30-day = warning (planning), 7-day = high (provision now), 24h = critical (page on-call).

FAQ