How to Display 100,500 Metrics of a Distributed System and Not Go Insane
summary
A talk about visualizing very large numbers of metrics from distributed systems (example: 100 machines × 100 metrics). It discusses challenges of limited screen space and human attention and compares aggregation techniques for presenting many time series, arguing that integrals (area under curve) can be a better summary than arithmetic mean or median for some monitoring needs.
AI-written summary · about
chapters (12)
- 0:00Introduction
- 4:29Talk overview
- 7:17Hazelcast context
- 11:30Target audience
- 15:04Time series analysis
- 25:47Detecting metric anomalies
- 33:54Data filtering with integrals
- 40:10Integral rules and benchmarks
- 44:44Top-3 and bottom-3 filters
- 46:50Outlier filter algorithm
- 49:21Implementation results
- 51:30Q&A
related talks
47mMonitoring at Scale: Intuitive Dashboard Design
46mZabbix and Millions of Metrics: Best Practices for Large-Scale Monitoring
36mMetrics - What's your code actually doing?
26mImproving How We Observe Our Observability Data: Techniques for SREs
01h 01mFinding Metrics that matter and using them safely
06mPerformance is about more than time series charts
42mMetrics at scale
38mGiving Away Your Secrets: Opening Metrics Up to Users