The session map is memory-only, so session directories surviving a restart
can never be served and never be evicted -- and the work root is a tmpfs,
so that is leaked RAM until reboot. The restart that ships the TTFB fix
would have stranded 1.56 GB.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Playback failed in Jellyfin the day after deployment, for uncached videos
only. The handler blocked on the full download-and-mux before sending
anything at all -- not even response headers -- so a 46-minute upload sat
silent for 79 seconds and the client gave up. The proxy counted it a
success, which is why /healthz and doctor both looked fine.
Yesterday's "DirectPlay verified" only ever ran against videos already
pulled during testing, so the first-play path was never exercised.
The output is already a fragmented MP4, so it is readable while being
written; a finished file only buys a correct duration and working seeks.
Cap the wait at --first-byte-grace (12s, explicit in the unit) and stream
whatever has not muxed by then. Measured: 156s -> 12.0s TTFB on a
66-minute upload, with ranges on a complete file unchanged.
The cost is a first play with no seek bar when the grace is missed. Every
later play of that video is perfect.
Also fixes a hang found while testing this: the streaming loop waited on
"finished and good" rather than "finished", so a producer that died after
writing some bytes held the connection for the full 45s stall timeout and
then dropped it.
The stale yt-dlp warning pointed at youtube-automate's venv, the tree we
are decommissioning; it now names ytstream's. /healthz reports the serving
mode and grace. Eight new proxy tests, 353 passing.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Phases 1-4 of plan.md §13. Forked from youtube-automate as planned rather than
written from scratch: naming, NFO, auth, the admin UI, settings and the DB layer
came across largely unchanged, download.py is gone, and the pieces that only make
sense for a streaming library are new.
api.py YouTube Data API v3 client. The whole metadata path.
strm.py Materialising: a .strm, an .nfo and a thumbnail. Replaces the
330-line download.py, because the job is writing a URL to a file.
subsync.py The subscription mirror, most of which is refusals.
reap.py Retention, rewritten around the 30-day window and min_keep_videos.
discovery.py RSS polling plus an API-backed, resumable, bounded backfill.
proxy/ The verified PoC, moved in with a systemd unit.
330 tests, all passing, no network and no yt-dlp in any of them. The suite leans
towards the failure paths, because that is where this design can actually hurt
someone: a 403 that looks like an unsubscribe, a video that ages out and comes
back, a title that never arrives. tests/test_proxy.py replaces the two standalone
scripts under proxy/ and now drives the real make_handler(mgr, ...) rather than the
PoC's make_handler(path, done), so routing and video-id validation are covered too.
Ran it end to end against the live API and it found three real bugs.
The first was mine and the tests caught it: strm.remove pruned empty directories
up to the media root, so a channel directory whose tvshow.nfo happened to be
missing would be deleted along with the season. It only looked safe because
tvshow.nfo normally stops the walk. The prune boundary is now the channel
directory explicitly.
The other two only showed up against real data, and they compounded. The backfill
inserted rows with no title and left the RSS poll to fill them in — but RSS returns
15 entries, which for Pitch Side spans 23 days against a 30-day window, so five of
twenty episodes were named after their video ids. Worse, strm.materialise wrote
that fallback back to the database as the title, which made the row look titled and
permanently disabled the repair path. Both fixed: playlistItems.list now requests
snippet as well as contentDetails, which costs the same single quota unit and
carries the title alongside the exact publish date, and the fallback is used for the
filename without being persisted. A title that does arrive late now also removes the
badly-named files and re-queues, so the episode is rewritten rather than keeping its
video-id name forever. Verified against the live API: all twenty Pitch Side episodes
now carry real titles.
Measured on the real account: 119 subscriptions queued for approval and none added
on the first sync, then a two-channel run backfilled and materialised 26 episodes in
under seven seconds.
Two deliberate departures from plan.md, both recorded there:
min_keep_videos defaults to 5 rather than being left as an open question. Without
it 52 of 117 measured channels are empty Jellyfin series that flicker in and out as
their single video crosses the retention line, and the plan already recommended it.
The Jellyfin refresh is a bare /Library/Refresh with a comment explaining why it
must stay that way. A normal scan makes zero media probes; FullRefresh does probe,
and at 400 episodes that is 400 cold starts.
Not yet done: no systemd units are installed (needs root — deploy/deploy.sh), the
admin UI has no routes for sources or the approval queue yet, and nothing has been
pointed at the real media root.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude