Monitoring Design Principles
summary
Reframes monitoring as measuring real system behavior not vague metrics, argues for SLOs that are assessable (calculate percentiles over fixed one to five minute windows), insists 95th percentile is not a histogram and you must store histograms not just percentiles or raw samples (example: 36 million latency samples revealed five distinct modes), recommends eBPF and streaming analytics (Spark, scalable time series DBs) to capture syscall and I/O latencies at high volume, and requires alerts to include remediation, stakeholders and live telemetry outside the blast radius.
AI-written summary · about
chapters (16)
- 0:00Introduction
- 3:04What is Monitoring
- 5:29Web Monitoring
- 9:12Database Monitoring
- 12:03Real Systems Monitoring
- 14:06Monitoring has become sophisticated
- 17:12Service Level Objectives
- 23:20Heat Maps
- 30:52Time Uncertainty
- 34:08Monitoring 101
- 36:44Monitor outside the tech stack
- 38:40Do not siloed data
- 39:33Value observation of real work
- 41:06Synthesis of traffic
- 41:55percentiles
- 45:24alerts
