AI with Machine Learning for Observability in Netdata
summary
Netdata implements unsupervised anomaly detection at the edge, training a rolling model per metric every 3 hours on 6 hours of data, keeping 18 models and requiring agreement across those 18 to cut false positives. It stores anomaly events in the database for historical queries, runs a scoring engine that ranks millions of metrics for root‑cause candidates, and runs lightweight training (about two CPU cores per million metrics) to detect missed periodic jobs, trend shifts, gradual regressions and sudden spikes. rolling models every 3 hours 18-model agreement
AI-written summary · about
chapters (12)
- 0:00Introduction and Speaker Background
- 0:18Google's AI Experiment in Observability
- 0:54Challenges in Machine Learning for Observability
- 1:44Unsupervised Machine Learning in Observability
- 2:34Designing Machine Learning Models for Observability
- 3:32Implementing Anomaly Detection
- 4:10Optimizing Machine Learning for Real-Time Data
- 4:54Storing and Querying Anomaly Data
- 13:26Scoring Engine for Root Cause Analysis
- 18:09NetData's Unique Approach to Observability
- 18:55Simplifying Observability with NetData
- 20:07Distributed Design of NetData
related talks
06mIGNITE: Practical AI with Machine Learning for Observability
44mNetdata: Open Source, Distributed Observability Pipeline – Journey and Challenge
29mAI Driven Observability based on Open Source
19mThe Rise of AI in Anomaly Detection
42mMachine Learning In Observability Systems: From Insights To Action
33mNetdata: Open Source, Distributed Observability Pipeline – Journey and Challenges
31mLearning the learner: Using machine learning to track performance
12mModel Monitoring: Detecting and Analyzing Data Issues