Preface: For full transparency, I used AI (Claude) to help dig into this
issue we're having and write some of the explanation below.
Our local Laravel application (with Docker) takes between 2-6s seconds
(re)building `app.css` after each Blade file update. After checking with
Claude, there seems to be some potential unnecessary traversing of big
(gitignored) directories, specifically `vendor` and `storage`.
As far as I understand it, two scans happen in tailwindcss:
- a first scan, the content scan. This one reads the project files
(auto-detect). It respects .gitignore, so it correctly skips vendor/ and
storage/.
- a second scan, the glob resolution (resolve_globs() in the oxide
crate). It builds watch patterns that the Vite plugin registers, so a
file save can trigger a CSS rebuild. It walks the project again, but
this walk does not respect .gitignore. It only skips a hardcoded list of
directory names from `ignored-content-dirs.txt` (node_modules, .git,
venv, ...). vendor/ and storage/ are not on that list, so it descends
into them and stats every file. These directories can never produce a
watch pattern, because the content scan never visited them. That work is
perhaps wasted?
Maybe why this went unnoticed: on a native filesystem a stat takes about
1µs, so 57k of them is around 0.06s. A Docker bind mount multiplies that
by roughly 100. And the big directory in a typical JS project is
`node_modules`, which the name list already catches; `vendor/` is the
PHP ecosystem's `node_modules`, and it is not on that list. (#19616 /
#19632 fixed the same problem for `scanner.files`; the CLI never reads
`scanner.globs`, so it was never affected.)
The fix: the second walk now skips any directory the scan never visited,
the same pruning the first walk (line 73) already does.
Measured on our local project: `get_globs()` after `scan()` went from
9.9s to somewhere between 0.2 and 1.0s across runs, with identical
output (35 globs). A template save now reloads in about 0.3s instead of
4.
## Test plan
- `cargo test` in `crates/oxide`: 66 tests pass, including the glob
assertions for nested and gitignored directories in `tests/scanner.rs`.
- Compared the sorted `get_globs()` output before and after on the
project above: identical.
I've also set up a repo that reproduces the issue in a fresh Laravel
app:
https://github.com/marickvantuil/tailwind-scan-reproduce
The results:
| Environment | vendor/ | scan | get_globs | syscalls under vendor/ |
| --- | --- | --- | --- | --- |
| GitHub Actions (native FS) | present | 5.0 ms | 48.8 ms | 11,410 |
| GitHub Actions (native FS) | absent | 4.5 ms | 0.9 ms | n/a |
| macOS, Docker bind mount | present | 15.8 ms | 599.4 ms | 11,410 |
| macOS, Docker bind mount | absent | 16.5 ms | 6.5 ms | n/a |
---------
Co-authored-by: Marick van Tuil <marick@dolphiq.nl>
Co-authored-by: greptile-apps[bot] <165735046+greptile-apps[bot]@users.noreply.github.com>
Co-authored-by: Robin Malfait <malfait.robin@gmail.com>
b86a6e0a0d