changelog Published on 2026-11-15

xfmpress v0.4.13: write phase 5.3 s to 0.9 s on repeat builds

6 min read # xfmpress# changelog

v0.4.13 started with unglamorous profiling: counting filesystem calls on a 1,000-page site. The count revealed two pure wastes in the write phase — a flood of system calls doomed to fail, and files being rewritten despite byte-identical content. After the fix, the write phase of a repeat build on that site drops from 5.3 s to 0.9 s, while a cold build (with dist removed) is unchanged.

What this release solves

Repeat builds are the daily reality: change one word, rewrite a thousand pages. Every build flushed all pages to disk (~1.5 s) and mkdir'd their directories (~1.1 s) even when the bytes were identical. On top of that, pruneStale tried rmdir on every directory in dist on every build — 1,000 page directories meant 1,000 failing rmdir calls, ~3.3 s sequentially on NTFS. Together, that's the couple of seconds you wait after every save.

What changed

Fixed

  • The pruneStale empty-directory sweep now runs on demand. A directory can only become empty because a file inside it was deleted, so the sweep runs only when something was actually deleted, and only over the ancestors of those files — deepest first, in parallel. No deletions means no rmdir calls at all.
  • Unchanged pages are no longer rewritten. Output is compared per file (size first, then content) and skipped when it already matches; a skipped file needs no directory either. Assets use the same idea with a make rule (size equal and destination not older than source → leave it alone), so large images are not even read.

Added

  • xfm build reports 未变化 N 个文件(内容一致,未重写) when files were skipped, so you can see exactly how much the build actually wrote.
  • bench/bench.js now reports cold (dist removed) and repeat builds separately instead of mixing them into one median — their cost structures are entirely different.

Do you need to do anything

No configuration changes; upgrade and the speed-up applies. The only visible differences are faster repeat builds and one extra line in the build output counting skipped files. If your CI always starts from a clean dist, this particular win won't apply (cold builds are unchanged) — but for the local writing loop, the save-to-preview wait shrinks noticeably. Everyday build usage in the quick start.

Comments