Notes from the cold path

Most of what I do professionally lives on the hot path, where everything is measured and nothing is interesting. These are notes from the other side: the code that runs once per hour, once per crash, or once ever — and is therefore almost never right.

I keep them here mostly so I stop re-deriving the same conclusions every eighteen months. Corrections are welcome; the address is on the about page.

• 2026-06-14 Nobody agrees on what fsync promises

  Four filesystems, four different answers to the same question about renamed files and dirty parent directories.

• 2026-04-02 The compaction tail nobody budgets for

  Why p99 write latency on an LSM tree is a scheduling problem long before it is a disk problem.

• 2026-01-27 Reading your own page cache back

  A benchmark that measured memory bandwidth for three weeks while everyone involved believed it measured a disk.