Statistics for Engineers
summary
Walks through concrete tactics: use histograms and heatmaps to preserve distributional detail, avoid reporting average latency (mean is non-robust), prefer medians and percentiles for body and tail, and implement latency SLOs from full histograms (example SLO: P99 under 100 ms over 20 days). Recommend mergeable structures like HDR Histogram or t-digest (Prometheus’ 10-bucket histograms can yield ~300% error). Shows sampling math: Bernoulli sampling error scales with 1/(n·p), 10% sampling on 100 RPS gives ~30% error; provides a GitHub-backed sampling simulator used at Zalando.
AI-written summary · about
chapters (10)
- 0:00Introduction
- 1:58Visualizations
- 8:35Statistics
- 13:46Average
- 14:02Median
- 17:11Percentile
- 20:01Merge
- 22:21Latency SLOs
- 24:10Sampling
- 30:25Summary
related talks
38mKnow your data: The stats behind your alerts
32mLearning Statistics Will Save Your Life
41mKnow your data: The stats behind your alerts
53mEmbedded Linux Quality Assurance: How to Not Lie with Statistics
54mPerformance tests and other lies
35mHonest Performance Testing
27mThe Subtle Art of Lying with Statistics
40mHow to Lie With Stats and Charts