Principled Performance Analytics
summary
An analytics method called Two Sigma partitions workloads into cohort-based baselines, fits per-cohort distributions (log-normal or normal approximations), and scores incoming events as z-score likelihoods, then monitors the fraction >2σ to detect regressions. It runs at production scale (examples up to 1.4T events/day), surfaces slow‑burn degradations hours earlier (one case ≈18 hours) with very low false positives in backtests, and supports hierarchical tracing of where latency concentrates for faster diagnosis.
AI-written summary · about
chapters (18)
- 0:00Introduction
- 1:19Is it working
- 3:19Reliability
- 5:40The Problem with SLOs
- 7:31A Simple Example
- 10:37How Does This Play Out
- 11:35The Prize
- 12:36Errors
- 18:25Qualitative vs Quantitative
- 19:59Performance Analytics Example
- 25:34High Level Model
- 31:36Two Sigma
- 39:24Building Detection
- 40:56FAQ
- 46:39Real Data
- 51:55Diagnosis
- 55:39Excursions
- 58:22Decoration
related talks
47mWhy Is My App SLOw? Defining Reliability in Platform Engineering
29mA New Kind of Analytics: Actionable Performance Analysis
39mBeyond Goldilocks Reliability
26mML Monitoring & Observability
47mDid My Deploy Degrade Production?
28mFive-sigma Network Events (and How to Find Them)
17mStopping Performance Regression via Changepoint Detection
35mObservability’s Sixth Sense: Detecting Anomalies in Metrics