Observability’s Sixth Sense: Detecting Anomalies in Metrics
summary
Claims that a single unified anomaly score per timeseries replaces hundreds of brittle thresholds, produced by models that run as online (incremental) or offline (full retrain) jobs and stored back into Victoria Metrics. Uses univariate stats, Prophet-style full-seasonality retrains and multivariate Isolation Forests; models auto-retrain on schedules, emit a 0–>1+ severity score, and power alerts that reduce MTTR, cut alert fatigue, and surface contextual, multivariate, and regime-shift anomalies. Demo shows a UI to configure readers, writers, schedulers, and filters.
AI-written summary · about
chapters (14)
- 0:00Observability and anomaly detection
- 1:06Static threshold alerting problems
- 4:41Anomaly types in production
- 6:48Normal behavior models
- 9:57Anomaly score explained
- 12:30Time series anomaly basics
- 13:15VM anomaly detection system
- 16:24VM anomaly components
- 17:47ML-powered VM anomaly rules
- 19:09Univariate vs multivariate models
- 21:18Offline vs online training
- 23:51Domain knowledge in models
- 27:04Demo: VM anomaly
- 33:27Q&A
related talks
40mObservability’s Sixth Sense: Detecting Anomalies in Metrics
35mObservability’s Sixth Sense: Detecting Anomalies in Metrics
28mThe Sixth Sense: Finding “Dead” Metrics and Silent Alerts in Your Monitoring
34mAnomaly Detection in Time Series from Scratch Using...
38mWatch Everything, Watch Anything: Anomaly Detection
12mModel Monitoring: Detecting and Analyzing Data Issues
34mML-Bound Anomaly Detection: Experiences Beyond Commodity Observability
31mFirst Day Foresight: Anomaly Detection for Observability