nirbhay.sh
../reading

The Design and Implementation of a Log-Structured File System

author
Rosenblum & Ousterhout
type
paper
status
read
logged
2026-09-14
source
original →

The argument is one sentence long and everything else follows from it: memory is getting cheaper, so reads will increasingly be absorbed by the page cache, so you should stop optimising for them and optimise writes instead.

Once you accept that, the design is forced. Buffer writes in memory, flush them sequentially, never update in place. Every LSM-shaped storage engine since — LevelDB, RocksDB, Cassandra — is a descendant of this.

The part worth reading twice is the cleaner. Writing sequentially is easy; reclaiming the fragmented space afterwards without destroying your write throughput is the actual problem, and the segment-cleaning policy is where the paper spends its real effort. Compaction in modern engines has exactly the same shape and exactly the same tuning difficulties.