Feature · /forecast
容量预测
提前 30 天看到崩溃 — 不是 30 分钟。
In one line
每个基础指标的 30/90/365 天预测,附「距离突破还有 X 天」。
What you get
- 30 / 90 / 365-day forecasts on every base metric
- Confidence intervals visualized (p10 / p50 / p90)
- 'Days until' projections: 'storage hits 90% in 23 days at current trend'
- Schedule-aware (knows you have weekly batch jobs, holidays, sale spikes)
- Alert when forecast crosses a threshold, not when reality crosses one
FAQ
What's the underlying model?+
Prophet for low-cardinality, ETSformer/N-BEATS for high-cardinality, ensembled per metric. Re-trained nightly per workspace.
Related reading
AI
Anomaly detection on database metrics: why thresholds fail and what works
A walk through forecast bands, change-point detection, multi-variate anomaly, and the seasonality math that makes 'p99 over 200ms' the wrong alert by default — with the Postgres example that broke our last threshold.
Postgres
Why your Postgres p99 latency lies — and what to track instead
p99 over 1m windows is the most-displayed and most-misleading number on every DBM dashboard. Here's the histogram math, the seasonality math, and a saner default.