Deepview: Virtual Disk Failure Diagnosis and Pattern Detection for Azure
summary
Deepview is a production Azure system that links compute, network and storage paths, implements Lasso regression plus hypothesis testing, and runs on a streaming pipeline into Cousteau. It processes hundreds to tens of thousands of VHD-failure events per day, cuts unclassified VHD failures from tens of thousands to under 500/day, achieves time-to-detection consistently under 10 minutes (ingest ~2.5 min, sliding window ~5 min, algorithm <20 s), and surfaces concrete patterns like ToR reboots and storage gray failures with higher precision/recall than prior tomography.
AI-written summary · about
chapters (16)
- 0:00Intro
- 1:25A New Type of Failure: VHD Failures
- 2:01How much do VHD failures impact VM availability?
- 2:30Failure Triage was slow and Inaccurate
- 3:28Deepview Approach: Global View
- 5:51Deepview Model: Include the Network
- 6:46Deepview Model: Estimate Component Health
- 11:26Evaluation
- 12:27Unplanned ToR Reboot
- 13:25Storage Cluster Gray Failure
- 13:56Network Failure Storage Clusters
- 14:31Algorithm Accuracy Comparison
- 15:07Deepview Time to Detection
- 19:01Related Work
- 19:47Conclusion
- 20:26THE ADVANCED COMPUTING SYSTEMS ASSOCIATION
related talks
36mVideo analytics at scale: Deep Learning, Computer Vision, ML & Spark
17mMulti-view Feature-based SSD Failure Prediction: What, When, and Why
31mFinding the Needle in the Haystack: Predicting Storage Device Failures in...
46mInside Azure innovations with Mark Russinovich
26mImproving fault detection & real time detailed analytics with Telemetry
27mMachine Learning's Missed Opportunity in Visual Data Management
44mGray Failure: The Achilles’ Heel of Cloud-Scale Systems
41mPredictive Maintenance with Azure Synapse