AI for Observability

What It Covers

Matt Ryer and Yassir Akinci from Grafana Labs discuss practical AI applications in observability, distinguishing between useful machine learning implementations and AI-for-AI's-sake marketing approaches.

Key Questions Answered

• •Forecasting over static alerts: Dynamic alerting using seasonal pattern recognition outperforms static thresholds by learning from historical data to predict expected behavior with confidence bounds for anomaly detection.
• •AI as translation layer: GenAI excels at explaining complex technical data like flame graphs in human language, making observability tools accessible to new developers without deep domain expertise.
• •System-level anomaly detection: Multi-metric analysis across entire service dependencies identifies root causes faster than single-metric monitoring, focusing operators on actual problem sources rather than downstream symptoms.
• •Structured data requirements: Dashboard generation from natural language fails without proper metric relationships and label structures - successful AI implementation requires well-organized underlying data architecture and context.
• •Practical ML over GenAI: Traditional machine learning techniques like outlier detection and change point analysis solve most observability problems more reliably than large language models for numerical data.