Image::handleImageTransform derives the output filename from the current
media_path and applies the encoder's output extension. When that differs from
what is already stored (heic/avif -> jpg, or a thumbnail regenerated to a new
extension), the new file landed at a different path and the previous file was
left orphaned in the media directory — the source of the leftover _thumb files
under public/m/_v2.
Capture the path each transform supersedes and delete it after a successful
write (only when the new output path differs, so we never delete what we just
wrote). Remove the stale MediaDeleteLeafCleanupTest whose source change is not
in the tree.
Move all Eloquent models from the app/ root directory to app/Models/
for consistency with modern Laravel conventions. The project already had
54 models in App\Models; this migrates the remaining 52 legacy models.
Changes:
- Move 52 model files from app/ to app/Models/
- Update namespace declarations in each model
- Update all ~1000 import references across the codebase
- Add Relation::morphMap() in AppServiceProvider for backward
compatibility with existing polymorphic database records
- Add missing HasSnowflakePrimary imports for models that relied
on same-namespace resolution
Convert all 273 short facade alias imports (e.g. 'use Cache;') to their
fully-qualified class names (e.g. 'use Illuminate\Support\Facades\Cache;')
across 193 files.
This resolves 643 PHPStan 'class.notFound' errors caused by the static
analyzer being unable to resolve global aliases, and aligns with modern
Laravel conventions. It also unblocks removing the aliases array from
config/app.php in a future change.
All 107 tests pass.
Blurhash::generate() allocates one PHP array per pixel of the source. At
roughly 255 bytes per pixel (measured: 224 MB peak for a 720x1280 frame) a
1920x1080 frame approaches half a gigabyte.
Image thumbnails survive this because they are capped at 640x640 in
Image::__construct() *and* run under that constructor's
ini_set('memory_limit', '1024M'). Video thumbnails get neither: FFmpeg saves
them at the source video's resolution, and VideoThumbnail never raises the
limit. So a video whose frame is 1080p or larger exhausts memory_limit.
That is a PHP fatal, not an \Exception, which has three consequences:
- the catch block in VideoThumbnail::handle() does not catch it
- the job never lands in failed_jobs, so nothing reports a problem
- MediaStoragePipeline::dispatch() on the last line of handle() never runs
The video therefore stays on local disk permanently while images beside it
replicate normally. Reported in #2652 (2021-02-13) and diagnosed correctly in
that thread on 2021-11-04.
Two changes:
1. Blurhash::generate() downscales to 128px on the long edge before sampling.
The result is a 4x4-component DCT, so full-resolution sampling adds
essentially nothing: measured against the full-resolution hash, mean
per-channel deviation of the decoded 24x24 preview is ~7.5/255 at a 32px
sample, ~4.5/255 at 64px, ~2.5/255 at 128px, and no better at 256px. Peak
memory for the frame above drops from 224 MB to 6 MB.
This removes the ceiling for every caller rather than moving it, which is
all that raising memory_limit would have done. Existing stored hashes are
not recomputed, so nothing already published changes appearance.
2. VideoThumbnail wraps the blurhash in its own try/catch, so a decorative
step can no longer skip the replication dispatch. Change 1 covers the
fatal; this covers any ordinary exception.
Verified on a live instance with S3 cloud storage: a 1920x1080 video that
previously stranded now generates a blurhash, uploads original and thumbnail
to the bucket, sets cdn_url/thumbnail_url/replicated_at, and removes the local
copies. Existing images re-hash to visually identical previews.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Smartphone photos stored in landscape orientation with an EXIF rotation
tag were being saved to S3 in the wrong orientation. Image.php read the
raw pixel dimensions without first applying the EXIF tag, so portrait
photos (e.g. 4032×3024 with Orientation=6) were classified and resized
as landscape (1920×1080).
Calling orient() immediately after read() physically rotates the image
to match its EXIF orientation tag before any dimension checks or
scaling. This ensures portrait photos remain portrait after processing.
Intervention Image v3 reference:
https://image.intervention.io/v3/modifying/orientation