Why Is My App SLOw? Defining Reliability in Platform Engineering
summary
Google Cloud Serverless SRE defines reliability by testing for stationarity of per-version latency distributions, building log‑normal baselines for cohorts and computing a Z-score per request to normalize across apps. The pipeline flags requests >2σ, aggregates the fraction (>10% signals a platform regression), stores baselines for ~50% of workloads that fit, and exposes a region metric called overloaded minutes. Backtests across months recover historical incidents, the method is cheap at high QPS, and it surfaces customer‑impacting regressions reliably.
AI-written summary · about
chapters (21)
- 0:00Intro
- 2:08Serverless platform is amazing
- 5:56My app is slow
- 8:14The platform is slow
- 9:29Total (end-to-end) latency distribution
- 10:54Request delivery latency
- 12:37Goal
- 13:39Reliability in practice
- 17:09Applying to the model
- 18:39Stationarity
- 19:312-Sigma Technique
- 24:30Mechanics
- 26:41Overload score
- 28:08Impact analysis
- 29:03FAQ
- 31:56Backtesting
- 33:21Limitations
- 35:13Other applications
- 37:06Approximate cohort A/B testing
- 37:33Conclusions
- 40:02Outro
related talks
47mWhy Is My App SLOw? Defining Reliability in Platform Engineering
35mSRE 101
25mProduct Reliability for Google Maps
15mShit, our Site is Down! Let's build a culture of reliability
50mFrom Load Testing to Reliability Engineering: Making Performance Testing Predict Production Behavior
01h 00mBuilding Reliable Services at NRK TV
27mContinuous Reliability. How?
13mBeyond observability: Using reliability scores to drive results