Know your data the stats behind the Alters
summary
Explains concrete rules: use median for alerts because a spike barely moves it while the arithmetic mean lingers; compares arithmetic, geometric and harmonic means with examples and shows how zeros break geometric/harmonic; examines sampling and OpenTelemetry traces (one customer produces 43 TB/hour of tracing) and how trace sampling hides long tails; cites a concrete metric threshold—3.7s is the shopping‑cart abandonment point—and challenges AIs on the exponential distribution math.
AI-written summary · about
chapters (20)
- 0:00Statistics in alerting systems
- 1:25Mean median mode explained
- 3:13Numbers misleading in monitoring
- 5:18Types of means and uses
- 7:45Means in DevOps
- 10:11Means manipulated to mislead
- 12:00Geometric harmonic means avoid zeros
- 13:03Median best alerting metric
- 15:29Data variability interquartile range
- 17:57Outliers with 1.5 IQR rule
- 18:40Mode in logs user behavior
- 19:43Descriptive vs inferential stats
- 20:46Sampling risks in tracing metrics
- 22:54Recency bias in alerts
- 24:19Key distributions in monitoring
- 26:06Exponential distribution latency
- 27:33Probability vs statistics
- 28:38Bayesian thinking Monty Hall
- 30:01Correlation pitfalls confounding
- 31:25Careful data analysis
related talks
38mAdd ALL the things: abstract algebra meets analytics
35mThe Math behind the Incident Aftermath: A Practical Guide to Measuring...
25mQuantiles at Scale: Choosing the Right Estimation Algorithms for Observability
34mOptimal Observability
39mData Mesh: Data Analytics for Engineers
48mHyper Speed! When Big Data Blooms
35mBuilding a Real Time Metrics Database
39mUsing Monoids for Large Scale Business Stats