- A flaky node in a k3s cluster caused a cascade of failures, traced to slow SSD sync writes.
- Consumer SSDs' marketed IOPS (at QD32, no sync) don't reflect real-world sync workloads like etcd's 4K writes with fsync.
- The WD Red SA500 in a Proxmox VM performed poorly on QD1 fsync (134 IOPS), below etcd's 500+ IOPS requirement, causing cluster-wide issues.
- Slow disks compound through replicated storage (Longhorn), where one slow replica drags down the whole volume's latency.