Perseus: A Fail-Slow Detection Framework for Cloud Storage Systems
summary
Perseus detects fail-slow drives non‑intrusively by modeling latency versus throughput with a light regression model, using outlier removal, an adaptive upper bound, sliding‑window event aggregation and a risk score. Deployed across 248K drives during 10 months, it flags 304 fail-slow cases, isolates them to cut node‑level 99.99th tail latency by 48%, and ships a public benchmark drawn from 41K normal and 315 verified fail‑slow drives for root‑cause analysis.
AI-written summary · about
chapters (7)
- 0:00Start
- 0:1421st USENIX Conference on File and Storage Technologi
- 4:39FAST? Failed Attempt: Threshold Filtering
- 11:56Outlier Detection
- 12:14Building Regression Model
- 12:39Identifying Fail-Slow Event
- 13:16Evaluating Risk
related talks
09mTest of Time Award: Analysis of Data Corruption in the Storage Stack
01h 01mThe Performance Investigator's Field Guide
12mSRE Principles for High-Speed Infrastructure Monitoring
28mA Study of SSD Reliability in Large Scale Enterprise Storage Deployments
27mMaking Disk Failure Predictions SMARTer!
35mAnomaly Detection for Cloud Native Storage
46mLearning security monitoring through failure
22mFailing Over without Falling Over
