Beyond Goldilocks Reliability
summary
Claims a calibration-free stationarity model that scores the relative likelihood of observed events, runs on historical telemetry and flags gradual, days-long degradations long before alerts trigger. Implementation uses hierarchical likelihood scoring to narrow faults to subsystems (example: isolated slow I/O cluster), detects regressions at p50 not just tails, and backtests against months of incidents. Prototypes produce reliable impact assessments that found missed effects in a 12-hour postmortem and enable proactive remediation.
AI-written summary · about
chapters (16)
- 0:00Intro
- 0:36Acknowledgements
- 1:22Our Reliability Approach
- 3:43Goldilocks Reliability
- 6:25Load Bearing Assumptions
- 9:45Practical Porridge Problems
- 12:36The Trouble with Thresholds
- 14:14Mo' Porridge Mo' Problems
- 19:59Make More Models!
- 22:12Model Elephants
- 23:55Reliability, modeled as Stationarity
- 26:24Stationarity Works!
- 29:11Hierarchical Diagnostics
- 30:43Stationarity Exposes Reliability Phenomena
- 32:19Tantalizing Capabilities
- 35:26Conclusions
related talks
13mBeyond observability: Using reliability scores to drive results
47mWhy Is My App SLOw? Defining Reliability in Platform Engineering
12mHistory-based Latency Prober Tuning
35mObservability’s Sixth Sense: Detecting Anomalies in Metrics
01h 10mPrincipled Performance Analytics
47mDid My Deploy Degrade Production?
26mML Monitoring & Observability
40mWhy Is My App SLOw? Defining Reliability in Platform Engineering