Repository navigation
v1.14.0
What's new since v1.13.4
Upgrade notes
Back up your database before upgrading. Four migrations run, and three of them change data rather than schema:
| Revision | What it does |
|---|---|
tm2_drop_same_lang |
Deletes rows. Removes every translation_memory entry whose source and target language are equal — those are damage, not cache. |
i198_reset_counters |
Zeroes the four provider rate counters. Pre-reset values are preserved in provider_stats_pre_i198; download history and contribution share stay comparable across the upgrade. |
i201_failure_kind |
Schema only — adds the classified failure column. |
tm3_strip_soft |
Rewrites rows. Removes invented soft line breaks from translation-memory entries, the same rule 1.13.4 applied to hard breaks. Irreversible: the original output is not kept. |
Two things worth expecting rather than discovering: a provider needs 50 searches after the reset before the delivers-nothing verdict may fire, and a provider that was broken before the upgrade and fixed by it can read healthy: false on purely pre-fix history until its next successful search. Both are cosmetic and clear themselves.
Added
- Sublarr can run under a path prefix. A Base URL in the settings lets Sublarr live at
https://host/sublarr/behind a reverse proxy — API, WebSocket and static assets all follow it. - One gate for disk-heavy media work. Remuxing and extraction used to run as parallel as the queue allowed; after a reboot that meant tens of gigabytes read at once. A single process-wide limit now governs all media I/O.
- Provider health reflects whether searches actually found anything. A provider that answers politely and never returns a candidate is no longer reported as healthy, and the panel now says why a provider is unhealthy instead of rendering every failure the same way.
- Which providers earn their place. Every provider shows its share of all subtitles downloaded, the overview sorts by it, and one that has searched plenty and never won is marked as contributing nothing — deliberately separate from health.
- Wanted items that have given up are visible as such, and exhausted items get a bounded second chance instead of sitting in the queue forever.
Fixed
- A typeset-heavy ASS no longer takes the whole app down. The sanitizer looked for drawing-mode blocks with a pattern allowed to span the entire file. When a subtitle opens drawing mode and never closes it — ordinary for karaoke and sign typesetting — the search restarted at end-of-file for every opener. A 25 MB translated ASS with 24 073 openers and no closer pinned a GIL-holding thread for 29 minutes: the container stayed up and answered nothing for all of it. The cost grows with the square of the file size, so a larger subtitle is a different order of magnitude, not a longer wait. Blocks are now found by walking each line's override tags once, which is also the correct reading, since ASS resets override tags at every event.
- Translated dialogue is no longer eaten between two drawing tags. The same pattern deleted everything between an unclosed opener and any later closer, including untouched dialogue lines in between.
- The extractor only extracts the languages your profile keeps. Every text track used to be extracted and the foreign ones trashed afterwards — on a remux with 19 subtitle tracks, thirteen full reads of the file per pass. One install logged 4 152 extract passes and 3 136 trashed sidecars in a single day.
- A sidecar is only "already extracted" if its letters agree with its name. A file named
.de.srtholding Arabic was reported as the German track on every pass. - The translator can no longer run away with the wanted search, and a subtitle-automation tick stops claiming work before its deadline.
- A source that answers empty can no longer wipe its wanted queue. A Sonarr that responds with an empty list during startup used to look like a library that had lost everything.
- Same-language "translation" can no longer happen — anywhere, including through a region subtag that used to smuggle a same-language request past the check.
- The translation memory drops entries that were never translations, and a translation may no longer invent a line break in either spelling.
- Every HTTP request now leaves a trace. Sublarr recorded no request at all, which made every field report a guess.
- Configuration is stricter and importable again. A misspelled setting is refused instead of silently stored, and
POST /config/importworks. - API fixes.
GET /api/v1/wanted?page=0answers instead of 500ing, and two hostile paths on the subtitle endpoint answer 4xx. - Duplicate scan works again, sidecar cleanup handles bracketed release names, and a database predating Alembic gets its automation queue back.
- Six bugs from a field report, each reproduced before being fixed.
Security
- socket.io-parser raised to 4.2.7 (GHSA-2m8v-j782-fhvr). A packet announcing binary attachments but carrying none leaves the parser holding its reconstruction buffer indefinitely; repeated, it exhausts memory in the browser tab. Sublarr keeps a Socket.IO connection open on every page, so the parser is always in the path. Only the transitive dependency moves —
socket.io-clientstays on 4.8.3.
Tested via
RC cycle v1.14.0-rc.4 through v1.14.0-rc.13 on a staging instance running a mirror of production data. The final 1.14.0 image itself was verified on that staging instance before promotion.
Image: ghcr.io/abrechen2/sublarr:1.14.0
Docs: https://sublarr.de/docs/