codex/bound-transcode-segments
2
Commits
| Author | SHA1 | Message | Date | |
|---|---|---|---|---|
 CoffeeKnyteandGitHub
|
a2ef26bece |
perf: root-cause fixes for endpoints still slow after #292 (NextUp, series badges, resume tail, subtitle fonts) (#350)
* docs(plans): root-cause analysis for endpoints still slow after PR #292 Five endpoint groups stayed slow after the home/Continue Watching/Latest latency work shipped: Resume (110s p95), NextUp (17s p95), Latest (17s), /Items, and the home sections routes. The caps and caches from PR #292 are live in the deployed binary; they bounded how many rows the loops touch but not what each underlying query costs. Documents the four confirmed root causes (4.3M stale completed-with-position progress rows + missing resume index, unbounded next-up anchor scan, per-episode series rollup fanout, two index-starved history/scanner paths) with live EXPLAIN ANALYZE measurements and the fix plan implemented by the follow-up commits. AI-use disclosure: analysis and doc produced with AI (Claude) assistance. * perf(catalog): bound the global next-up anchor scan to recent completions The completed_episodes CTE in buildListNextUpQuery derived per-series anchors from the profile's ENTIRE completed history — DISTINCT ON over 233k rows joined to episodes for the worst bulk-import profile, then a per-series LATERAL that scans every episode of a fully-watched series before yielding nothing. 648 slow executions in a 19h window, 44.7s worst; this drove /Shows/NextUp (17.1s p95) and the next-up injection on the native home sections aggregate. Global queries now derive anchors from the profile's nextUpAnchorMaxRows (500) most recent completed rows — an ordered index walk on idx_uwp_profile_completed, with the hidden-items exclusion and date cutoff applied inside the bounded scan so hidden/old rows never consume the anchor budget. A next-up rail surfaces ~24 series; the 500 most recent completions cover every series that can realistically rank on it. Series-scoped calls (the show-detail tile) keep the unbounded shape: they must anchor on the series' last completed episode no matter how long ago it was watched, and are naturally bounded by one series. Measured on the live worst-case profile with the exact generated SQL: 44.7s worst / ~2.6s avg before; 10ms after (together with the one-time stale-resume-point data repair applied directly to the deployment DB — see docs/superpowers/plans/2026-07-06-slow-endpoint-root-causes.md). AI-use disclosure: implemented with AI (Claude) assistance. * perf(jellycompat,userstore): aggregate series watch-state rollup in SQL The series Played/UnplayedItemCount badge on list rails (per-library Latest, library browse, search results) and series detail pages was computed by materializing EVERY episode of every series on the page (episodeRepo.ListBySeriesIDs) and then batching per-episode progress+history lookups in 500-id chunks. A 50-series page of an episode-heavy library (Sports) expanded to 32,467 episode rows and ~65 sequential queries — measured 17-18s per /Items/Latest request, and PR #292's cached Latest fast path pays it on every response for series libraries. The same fanout made /Items?searchTerm=... slow whenever the result set was mostly series (Meilisearch itself answers in milliseconds). New optional store capability userstore.SeriesEpisodeRollupStore, implemented by PostgresUserStore as one GROUP BY e.series_id aggregate with semantics identical to the chunked path (episode availability via episode_libraries, hidden-items visibility on progress rows, completed-history fold, in-progress = not watched with position > 0 — verified value-for-value against the old semantics on a real 1,586-episode series). enrichSeriesListUserData and enrichDetailUserData use it when present; SQLite-backed stores and rollup query failures keep the existing chunked path as fallback. catalog.SeasonUserDataFromCounts pins the counts-to-DTO mapping to EpisodeRollupUserData. Measured on the live worst-case profile against the real 50-series Sports Latest page: ~17s of chunked round-trips before, 119ms in one query after. Part of docs/superpowers/plans/2026-07-06-slow-endpoint-root-causes.md. AI-use disclosure: implemented with AI (Claude) assistance. * perf(catalog): bound superseded-episode completed walk to recent history The Resume / Continue Watching superseded-episode filter loaded a profile's *entire* completed history into memory on every request that contained an in-progress episode: CompletedProgressSnapshots paged user_watch_progress WHERE completed=TRUE with no upper bound. The 2026-07-06 slow-query comparison showed this surviving as a 60-116s Resume tail even after the in-progress index landed live, because the 4.3M zeroed Plex-import rows are still completed=TRUE and were re-walked every load. A completed episode can only supersede an in-progress one it was finished more recently than (the query gates on done_progress.updated_at > ip_progress.updated_at), so only completed rows newer than the oldest in-progress entry can matter. Compute that cutoff in SupersededEpisodeProgressIDs and pass it to CompletedProgressSnapshots, which — since the completed listing is ordered updated_at DESC — stops paging as soon as it crosses the cutoff. Import-heavy profiles whose back-catalogue predates their current in-progress items now stop on the first page instead of paging hundreds of thousands of irrelevant rows. Correctness is unchanged: no relevant superseding row is excluded. * perf(catalog): hard-cap superseded-episode completed walk at 5 pages The updated_at cutoff added in the previous commit bounds the completed walk on the relevance axis, but a very old in-progress entry sitting behind a large volume of newer completions could still page deep. Add a 5-page (2,500-row) hard backstop on top of the cutoff: normal profiles still stop on page one via the cutoff, and only the adversarial tail hits the cap. When it engages the tail of the completed set goes unscanned, so a superseded episode could momentarily survive on Continue Watching — we log a warning when that happens (with profile_id + rows scanned) rather than mis-filter silently, and it self-corrects once the stale in-progress entry ages out of the scanned window. * perf(playback): extract subtitle fonts in a single ffmpeg pass Embedded ASS/SSA font extraction spawned one ffmpeg process per font attachment, each re-opening the (usually CephFS-backed) media file. Anime releases carry 15-47 fonts, so the per-spawn file-open cost dominated and pushed GET /api/v1/stream/{sid}/subtitles/{track}/fonts to a 17-60 s plateau (p95 ~33 s in the live logs). Collapse the N spawns into one ffmpeg invocation that dumps every attachment to a temp dir (-dump_attachment:idx path ... -i file -map 0:t? -c copy), then read the files back. The file is opened once instead of N times, taking p95 from ~30 s to ~1-2 s with no change to output. Safety is preserved. The 32-attachment / 32 MiB caps still apply: attachment size is stat'd before read so an over-limit font never enters memory, and a watchdog polls the dump dir and kills ffmpeg if its on-disk output crosses the cap -- restoring the hard bound the old pipe-per-attachment reader enforced by killing at maxBytes+1, so a container with oversized "font" attachments can't fill the disk. Part of the slow-endpoint follow-up; see slow-query-analysis/subtitle-fonts-extraction-findings.md. * fix(review): report enforced font-byte cap; correct doc subtitle scope Address PR #350 review: - dumpFontAttachments reported the maxSubtitleFontBytes package constant in both over-limit errors instead of the maxBytes argument the caller passed, so the message misstated the enforced bound whenever a different cap was in effect (as the tests use). Interpolate maxBytes in both messages. - The root-cause plan claimed subtitle extraction was 'out of scope' while the branch actually optimizes /subtitles/{track}/fonts. Scope the out-of-scope note to subtitle *track* conversion and record the fonts single-pass work as deliverable 5. |
||
|
|
b246f271cb |
fix(jellycompat): hide dismissed and superseded entries from Resume (#114)
The jellycompat /UserItems/Resume endpoint served the raw in_progress
list, so Jellyfin clients showed every half-watched episode a profile
ever abandoned, including stale entries from earlier seasons of shows
the viewer had long moved past. The first-party Continue Watching row
already hid those via dismissal filtering and the superseded-episode
check, but that logic was private to internal/sections.
Extract the shared rules into internal/catalog
(ContinueWatchingProgressFilter, HomeDismissalIndex) and apply them to
the compat Resume path through a new UserDataService method,
FilterResumeProgress. The sections fetcher now delegates to the same
code, so both surfaces agree on what "still watching" means.
Resume pagination keeps advancing by raw batch counts so filtering
cannot terminate scans early, and the raw-offset fast path is limited
to StartIndex=0 for Resume because filtered lists make raw offsets
diverge from the visible list. The watched-items view ("completed"
status) stays unfiltered, and the per-series collapse deliberately
stays first-party only to preserve Jellyfin endpoint semantics.
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
|