Cleanup never matched a live session because the controller registered empty
manifest and segment paths, renewal was a read-modify-write that could revert a
takeover, and the takeover script compared an ISO date to a number, so it errored.
- populate the session record's manifest and segment paths from the playlist path
- renew the lease via a Lua compare-and-set on the owning pod
- store the lease expiry as unix milliseconds so the scripts can compare it
- retain the session record past its lease so an orphan can still be taken over
- test the Redis store against a real Redis, including the renew-vs-takeover race
- drop the live stream record nothing ever read back
Timer-driven library tasks fire on every replica, so a library refresh or a
database optimise runs once per pod against the same shared library.
- add IScanLeaderLease with a Redis TTL implementation and a no-op default
- skip timer-driven runs of the gated tasks on instances without the lease
- treat an unreachable Redis as holding the lease so tasks never stop running
- leave manual and API-triggered runs ungated
Live stream ownership is only tracked in process memory, so no replica can tell
which pod holds a stream open once that pod is gone.
- persist a LiveStreamSession record when the live stream mappings change
- delete the record when the stream is closed
- keep closing the stream when the store is unreachable
A transcode that is not in the store is invisible to the other replicas, so the
HLS entry points have to publish and hold the lease themselves.
- register the play session in the store when ffmpeg starts
- renew the lease on a background loop and delete the session when it ends
- shorten segments and bound the playlist window when resuming a stored session
- add RecoverySegmentLengthSeconds and RecoverySegmentBufferCount encoding options
Every replica runs the transcode cleanup task against the same shared transcode
directory, so one pod deletes segments another pod is still streaming.
- read the active sessions from ITranscodeSessionStore before deleting
- keep files matching an active session manifest path or segment prefix
- abort the sweep when the store cannot be reached
Transcode state lives only in the process that started ffmpeg, so a pod restart
drops every in-flight HLS stream with no way for a peer to pick it up.
- add ITranscodeSessionStore plus the TranscodeSession and LiveStreamSession records
- add RedisTranscodeSessionStore with TTL leases and an atomic Lua takeover script
- add NullTranscodeSessionStore for single-instance deployments
- pick the store from Jellyfin:TranscodeStore:RedisConnectionString at startup
Multi-replica deployments cannot share a SQLite file, so the database has to
move to a server engine before the rest of the HA work is usable.
- add Jellyfin.Database.Providers.PostgreSQL with an EF Core Npgsql provider
- register the provider and a pooled NpgsqlDataSource when DatabaseType is Jellyfin-PostgreSQL
- accept postgresql:// URIs and POSTGRES_CONNECTION_STRING alongside CustomProviderOptions
- add container-backed provider, CRUD, concurrency and migration tests
Dolby vision files having unexpected transfers now also marked as invalid, and it will now get its base video range from its base layer color transfer, as not all invalid dolby vision files are HDR now.
Since the SkiaSharp 3 update the MatrixConvolution image filter used in
SkiaEncoder.ResizeImage no longer has a fast CPU path: on the software
rasterizer it takes about 4.5 seconds per megapixel-sized image, which
turns every cold image request into a multi-second operation and makes
first-time loads of a library view take minutes.
Draw the resize without the paint filter and apply the identical 3x3
kernel (same weights, clamped edges, alpha included) directly on the
resized pixels instead. This drops a cold 1000x1500 -> 663x995 poster
render from ~4.6s to well under a second; the convolution pass itself
takes ~86ms. Output is visually unchanged.
unkin-agent
Cody Robibero