Why Is My App SLOw? Defining Reliability in Platform Engineering
summary
Platform engineers get a concrete method: use structured logs, F1 and Monarch to build per-cohort baselines (mean and stddev) from 30 days of data, convert latencies to Z‑scores, bucket 15‑minute windows and compute the 2-Sigma technique J‑score (fraction of requests with Z>2). Cohorts under 20 QPS are ignored; normal/log‑normal/Box‑Cox cover ~45%/60%/70% of workloads. Alerts fire when J-score above 10% (50th percentile cluster rising past ~5% and a 30th percentile spike), and backtesting shows it finds known incidents and new regressions.
AI-written summary · about
chapters (20)
- 0:00Intro
- 3:27Serverless platform is amazing
- 7:52"My app is slow"
- 9:56The platform is slow
- 12:46Total (end-to-end) latency distribution
- 14:26Request delivery latency
- 16:12Goal
- 18:47Reliability in practice
- 24:26Applying to the model
- 26:09Stationarity
- 27:202-Sigma Technique
- 33:32Mechanics
- 34:50Overload score
- 37:25Impact analysis
- 38:36FAQ
- 40:45Backtesting
- 42:07Limitations
- 44:04Other applications
- 44:51Conclusions
- 46:47Outro
related talks
40mWhy Is My App SLOw? Defining Reliability in Platform Engineering
30mReliability Engineering for Humans
01h 10mPrincipled Performance Analytics
15mShit, our Site is Down! Let's build a culture of reliability
45mPerformance-by-Design: Embedding Intelligent Scaling and Guardrails Into Platform Engineering
28mEngineering Data Reliably Using SLO Theory
13mBeyond observability: Using reliability scores to drive results
37mMonitoring the Health of Your App