NVMe SSD Failures in the Field: the Fail-Stop and the Fail-Slow
summary
Logs from 1.8 million NVMe SSDs at Alibaba drive these claims: infant mortality disappears (failure rates in months one–three are at or below long‑term rates), extremely low write amplification is rare but deadly (3.3% of drives with WAF<1), high WAF no longer predicts failure, and spatially correlated failures shift from minute‑scale to day–to–month scale. NVMe shows pervasive fail‑slow dozens× more likely than HDDs, typical slow events average hundreds of microseconds, top 1% exceed 11 ms, and only ~0.01% of slow drives later become fail‑stop within five months.
AI-written summary · about
chapters (6)
- 0:00Start
- 3:41How Do Fader Rates Vary as the Device Gets Older and More Worn
- 11:34Internal Node Failure Distribution
- 11:46Short-Term Failures
- 17:37Transition from Fail Slow to Full Stop
- 18:39Three Major Reliability Changes in Nvme Ssd
related talks
25mPerformance Analysis of NVMe SSDs and their Implication on Real World Databases
14mOperational Characteristics of SSDs in Enterprise Storage Systems: A Large-Scale Field...
45mBig Data and SSDs
29mLarge-Scale Benchmark Results of NVMe Over Fabric Technology
52mHow ZNS SSD's Are Changing the Landscape
18mRL-Watchdog: A Fast and Predictable SSD Liveness Watchdog on Storage Systems
12mFusionRAID: Achieving Consistent Low Latency for Commodity SSD Arrays
16mZNS: Avoiding the Block Interface Tax for Flash-based SSDs