In addition to normal alerting, SQL Governor has a predictive alerting feature that lets you know when something is about to go wrong. To be exact, in version 16 there are four different types of predictive alerts, which are explained in more detail below.
Linear regression-based alerts
When linear regression-based alerts are enabled for a performance counter, SQL Governor predicts the value of that counter 1-48 hours into the future by looking at the same amount of historical data and calculating a linear trendline.
For example, if you have set linear predictions for 8 hours, SQL Governor takes the last 8 hours of monitoring data, calculates a trendline based on linear regression and applies it 8 hours into the future. If at any point the alert threshold is exceeded, a predictive alert will be generated.
These alerts are visible on the Monitoring page and on the dashboard.
Anomaly detection-based alerts
Anomaly detection looks for data points that are higher (or lower, for reversed performance counters) than usual at a certain time of day and uses those data points to predict values for the next several days. Unlike pattern-oriented alerts and linear regression-based alerts, anomaly detection may send an alert that is expected to occur later than 48 hours from now.
Note that in the current version, anomaly detection-based alerts are available only for instance-level performance counters.
These alerts are visible on the dashboard.
Early warnings
Early warnings identifies workload performance counters which are statistically going into a worse direction. The change is calculated from the preceding month while the baseline is calculated from three months before. The following statistical components are inspected:
- Regression slope
- Workload volatility
- Skewness
- Kurtosis
Based on these components, the algorithm identifies performance counters which are at risk. We want to understand why the workload is headed into a worse direction before it starts to negatively affect the system performance.
These alerts are visible on the dashboard.
Query anomaly alerts
Query anomaly predictions show individual queries that are likely to cause problems in the near future. For each forecast, its validity (how likely it is to happen) and the time left until it's predicted to happen are shown.
These alerts are visible on the dashboard.