Predictions
Predictions forecast a monitor’s next response times, derive a predicted status from its thresholds, flag unusual values, and can alert you before the monitor itself changes status.
Reading a monitor’s forecast
Open the monitor, tab Prediction. For each probe, the Prediction status box shows:
| Element | How to read it |
|---|---|
| Grey line | The latest checks (up to 60). A red dot on the last one: an anomaly, a check unusual for this monitor. |
| Blue dashed line | The forecast for the next checks: the median of the latest 60, so it stays flat. |
| Blue bands | Where 8 (darker) and 19 (lighter) checks out of 20 fell at that step ahead on this monitor’s history. Usually wider above than below. |
| Dotted lines | The monitor’s minor / major / critical thresholds. |
| Next 5 checks: about … | The forecast and its 80 % range, in words. |
| Measured on this monitor’s history | How large its usual error is, compared with simply repeating the last check, and how often the ranges held. |
| Since then, N forecasts checked | The same, measured on the forecasts made since training against what actually happened. |
| No new check since … | The checks stopped: the forecast is not current and raises no alert. |
A forecast is an estimate: trust it as far as these measures say, and treat a forecast breach as a risk, not a certainty.
Everything about it now lives in the Predictive analytics section:
| You want to | Read |
|---|---|
| Set it up, step by step | Tutorial: your first forecast |
| Understand forecasts, training and alerts | How it works |
| Look up a setting or the API | Configuration |
| Fix a problem | Troubleshooting |
Upgrading from an earlier release
The prediction service now reads InfluxDB with its own settings and trains its models itself: no more
train.py to run by hand, and the InfluxDB fields of the Webserver are only needed for an old prediction service. Requests are signed by the Webserver. Follow the tutorial from step 1 to move an existing installation over.