“This is why write-amplification of sequential IO loads is close to 1 but random IO is typically much higher.”
NVMe drives can’t overwrite a page or erase less than a whole block, so random writes leave every block half stale. Garbage collection then rewrites valid pages, and random write performance can fall off a cliff around two-thirds full. AWS local instance store NVMe comes with zero over-provisioning. Kafka and BookKeeper got the design right for spinning disks, and it still holds.