An In-Depth Study of Correlated Failures in Production SSD-Based Data Centers

on YouTube ↗12m

" summary

Analysis of nearly one million SSDs from Alibaba across 11 drive models shows intra-node and intra-rack correlated failures are common and group sizes sometimes exceed redundancy limits. SMART attributes are weak predictors (max SRCC 0.23), while drive model density, age, many same-model drives per node/rack, and write-dominant workloads increase correlation. Trace-driven simulation finds erasure coding outperforms replication but all schemes degrade under correlated failures and lazy recovery worsens data-loss risk compared to eager recovery.

AI-written summary · about

" chapters (15)

1. 0:00Intro
2. 0:34Motivation
3. 1:08Our Contribution
4. 1:53Dataset
5. 2:36Analysis Methodology
6. 3:20Correlations Among Failures
7. 4:22Impact of Drive Models
8. 5:26Impact of Lithography
9. 5:59Impact of Age
10. 6:25Impact of Capacity
11. 6:53Impact of SMART Attributes
12. 7:34Impact of Applications
13. 8:52Case Study: Redundancy Protection
14. 9:34Simulation Results
15. 10:59Conclusion