* fix(server): compare integrity paths independently of unicode normalization
The untracked-file check compared on-disk paths against the database
byte for byte. macOS writes decomposed (NFD) filenames while an upload
stores the composed (NFC) form, so every asset with a non-ASCII name
came back from getAssetPathsByPaths under a different byte sequence,
was never removed from the candidate set, and was reported as an
untracked file. One reporter saw 14k false positives on a 55k library.
Look the batch up under both normal forms and key the comparison on
NFC, so a path that differs only in normalization is recognised as
tracked.
* test(server): cover normalized integrity lookup
low hanging fruits for the speedup of immich startup
1. don't wait 1s if we actually got the lock
2. move full import from the top of the file to be imported lazily in the branch
* Update sorting
Updates the sorting to sort
1. First by timeCreatedAt, but then
2. By filename.
* Push tsts
* Sync SQL
* Formatting fixes
---------
Co-authored-by: Daniel Dietzler <36593685+danieldietzler@users.noreply.github.com>
* fix(server): kill the child process when a spawned duplex stream is destroyed
A failed database backup left pg_dump running. The stream was destroyed but
the child was not, so the dump sat idle in transaction holding AccessShareLock
on every table until someone terminated it by hand. One report had it survive
five days and block DDL from another application on the same Postgres.
Destroying the duplex now kills the child, which also covers pipeline()
tearing every stream down when one of them fails.
The backup service additionally holds references to the streams it spawns so
the error path can destroy them. createWriteStream can throw before pipeline()
takes ownership, which is how the reported ENOENT on an unmounted backup volume
left nothing referencing the running pg_dump.
* chore(server): drop the explanatory comments from the backup teardown fix
Requested in review. The reasoning lives in the commit message and the PR
body, which is where it belongs.
`upsertExif` writes derived audio, video and keyframe metadata with an
`ON CONFLICT DO UPDATE`, but every column in those three update lists
assigns the stored row back to itself:
bitrate: ref('asset_audio.bitrate')
`SET bitrate = asset_audio.bitrate` is a self-assignment, so once a row
exists nothing can change it. Re-running metadata extraction re-reads the
file, builds a fresh snapshot, and then quietly discards it. The rest of the
file already uses `excluded` for this, as does `plugin.repository.ts`.
`asset_video.frameCount` had a second problem: it is supplied on insert but
was missing from the update list, so it would have stayed stale even after
the reference was corrected.
Extraction is meant to be repeatable. Probing improves between releases, a
file can be repaired or replaced, and a fix to how a stream is chosen is
worthless if it cannot reach the assets that were already imported. An
`upsert` that silently degrades to insert-only defeats all of that.
Metadata extraction is the only caller that passes these three objects, and
it passes a complete snapshot or nothing: each object is built from a single
probe behind a guard, and `upsertExif` skips the branch entirely when the
object is absent. So taking the incoming row cannot write partial values
over good ones. The other three callers pass `exif` only, and every other
reference to these tables is a read.
* implement updated rename tag feature
* Fix tag update handler
* Ensure toast displays updated tag name
* Ensure toast displays updated tag name
* Handle updates to tag closure table, prevent slash in API filename, test case updates
* delete child tags on parent delete
* Update frontend for better handling of nested tags
* pr checklist fixes
* fix unit test for tag controller
* Hide path for top-level tags
* regenerate SQL files
* name updates optional for tag update operations
* e2e tests for new tag api changes
* changes from code review, additional test coverage
* test fixes, web component refactor
* test assets
* test asset updates
* open-api rebuild
* update unit tests
* fix tag dto tests
* Update web/src/lib/modals/TagEditModal.svelte
Fix missing color bug
Co-authored-by: Ray <3608878+Pecacheu@users.noreply.github.com>
* fix tag e2e test
* remove closure table bug fixes
* remove nullish from tag name update schema
* regenerate openapi and sql
* fix test and e2e test
* simplify tag repository update process
* minor fix and sql regenerate
* fix tests, regen open-api, update dto
* Remove unused function
* Remove accidental test asset ref
* re-add test assets
* pin submodule to main branch commit
* Update test cases based on feedback
* Server fixes based on feedback
* regenerate sql
---------
Co-authored-by: Ray <3608878+Pecacheu@users.noreply.github.com>
* fix: download archives via html POST forms
* refactor: fix css of download panel, add translations and toasts
* style: run lint, formatter, check
* chore: generate open-api specs
* refactor(web): simplify download manager, use SvelteMap
* refactor(server): parse fields from x-www-form-urlencoded request
* refactor: move content-disposition to ImmichReadStream
* Pass asset ids as a comma-separated string
This fixes a bug where the 1000 parameter limit is being hit when
downloading 1000+ assets.
* Add download controller test
* Fix lint warnings
Aditya Raj Singh