SolarWinds AI: Smarter Database Monitoring

Anomaly detection in a database, usually powered by machine learning, is a method of identifying unusual events in a database. Though databases can have outliers (and anomalies are outliers in most cases), not all outliers are anomalies. An anomaly detection tool can help DBAs more easily find “unusual” or “unexpected” instances based on database performance baselines, defining unusual and unexpected as “statistically improbable.”

Anomaly detection in database monitoring is ideal for the following:

• Finding unusual metric values to identify undetected issues
• Finding changes in important metrics for database administrators to investigate
• Zeroing in on where to search when trying to diagnose a detected problem 
• Reducing the need to recalibrate thresholds

DBAs can use anomaly detection in database monitoring to help drill down into what matters more quickly. A database anomaly detection tool can also alert DBAs to unusual changes potentially indicating a database performance problem before it grows into a bottleneck.

Machine learning helps improve anomaly detection in a couple key areas:

• Increases Accuracy: Machine learning makes anomaly-based monitoring for databases more accurate. To effectively highlight unusual events in your database, the anomaly detection tool must first establish a baseline of your database performance. However, these baselines change over time as you reoptimize and table tune. Machine learning can automatically account for these changes, recalibrate, and create new baselines, so DBAs have the most accurate data to work with. Also, machine learning can make the anomaly detection tool smarter over time, so its alerts become more accurate.
• Reduces Complexity: Machine learning uses automation and fast analysis to help database administrators break down large data sets so they can use them to create actionable tasks. It also helps reduce complexity in troubleshooting, as machine learning can point DBAs to where performance issues may be. This makes it so they don’t have to search the entire database for them.

Anomaly-based database monitoring in SolarWinds^® Database Performance Analyzer (DPA) is built to help inform performance optimization efforts in two major ways:

1. The machine learning algorithm learns what “normal” is for your database and predicts wait times. The algorithm requires a minimum of three days of data to start “learning” and can use up to 90 days of historical learning. DPA collects the data from the algorithm. 
2. Based on the learning data, DPA’s algorithm calculates the amount of wait time the database is likely to experience during each hour of the next 90 days and the standard deviation for the entire data set. (The standard deviation is used to calculate thresholds.) When enough data becomes available, DPA makes predictions about daily and weekly seasonality and patterns of predictable fluctuations during the day.

DPA compares the actual wait time for an hour-long period and compares it to the predicted wait time, looking for discrepancies. If the actual wait time is above a critical threshold, DPA can do the following:

• Trigger a Database Instance Wait Time Anomaly alert if one has been configured
• Change the color of the wait time meter on the DPA homepage
• Display yellow or red segments on the bars in the Anomaly Detection charts