Read disturbance in NAND flash SSDs has been characterized at the device level for years: repeated reads to a page disturb neighboring cells, triggering internal error-correction and refresh operations. What this study (SIGMETRICS 2026, POMACS) finds is that those device-level mitigations have measurable, non-trivial performance consequences at the system level. The I/O tail latency spikes and throughput degradation are visible to applications -- not just to the SSD firmware. Prior work stopped at the device boundary; this paper crosses it.
The population for whom this matters shifted when AI inference went 24/7. Traditional enterprise storage workloads had mixed read-write ratios and burst patterns that kept read disturbance events rare. Large-scale inference clusters running continuous batch decode are read-dominant, high-intensity, and never idle. The read patterns that NAND flash vendors did not design the device-level mitigations to handle are now the steady-state in AI data centers.
The validation gap is direct: storage validation suites used for AI inference infrastructure were not designed to characterize read disturbance under continuous AI read patterns. Teams qualifying storage for inference clusters are clearing tests written for a different workload shape. The 24-month consequence is that some fraction of SSD performance degradation events in inference clusters will be attributed to the wrong root cause -- firmware bugs, network jitter, thermal throttling -- until read disturbance becomes a standard item in AI storage validation suites. The Onur Mutlu group has the track record on flash reliability to make this finding load-bearing; take it seriously.