Ownership is claimed with a Lua check-and-set keyed on the instance holding
the websocket, routing prefers a live controller over a local copy, the
session list deduplicates by owner, removal is ownership-checked, undelivered
routed messages surface, single-session lookups stop scanning the keyspace and
directory writes leave the request path bounded by a timeout.
Publish every session to valkey with the instance holding it, and route a
command for a non-local session to its owner over a per-instance pub/sub
channel. Entries expire, so a dead instance leaves the directory.
Next Up walked every series in an unpaginated loop and read user data one
episode at a time, so the Home screen row cost two queries per series once the
cache was gone.
- read the played state of every candidate episode in one query
- read the versions the resume check and the last played date need in one query each
- default PrefetchedUserData on IUserBaseItemComparer so plugin comparers still compile
- cover the bounded query count and the plugin comparer with tests
- share one database across the replica tests
The user data cache and the item's in-memory rows were both filled once per
replica and never invalidated, so a pod serving a playback tick read its own
stale resume position and wrote it back over the position another pod had just
saved, silently losing resume points, played state, favourites and ratings.
- drop the user item data cache
- read single and batched user data from the database on every query
- prefetch user data for in-memory sorts and filters so each stays one query
- cover the lost update against real PostgreSQL with two manager instances
Point Jellyfin.Database.Tests.PostgreSQL at the server postgres-migration-chain
already runs, give every test its own database, and fix the two tests that only
ever failed against real PostgreSQL.
Add a Redis pub/sub invalidation bus, publish from the configuration and library-option write paths, and drop the matching local cache entry on the instances that did not write. No-op without a Redis connection string, and fails open when it is unreachable.
Add the eight provider, live TV and plugin-update task keys to
ScanLeaderOptions.GatedTaskKeys and widen the test's task-key discovery to
every assembly that declares an IScheduledTask.
Jellyfin:ScanLeader:Enabled defaults to false and nothing sets it, so
leader election never runs and every replica executes the timer-driven
library tasks concurrently - the exact behaviour the lease prevents.
- Enable gating by default when a Redis connection string is configured
- Honour an explicit Enabled setting either way
- Carry the effective decision onto the bound options the task worker reads
- Log at startup whether gating is active
- Warn when gating is enabled but no Redis connection string is configured
PostgreSQL has no min(uuid) aggregate, so every query that picked a group
representative with MIN over the item id failed with 42883: library browse,
search, Recently Added, the by-name endpoints and Upcoming.
- Pick the representative with an anti-join on (primary version, id)
- Cover the collapse with a repository test against a real PostgreSQL
The startup configuration only reads JELLYFIN_ prefixed environment
variables, so the bare Jellyfin__TranscodeStore__* form used by the chart,
the manifests and the README is dropped and the Redis store is never
registered. Nothing logs the selected store, so the fallback is invisible.
- Read bare Jellyfin__* variables into the Jellyfin:* configuration root
- Keep an explicit JELLYFIN_ variable winning over the bare form
- Log the selected transcode session store at startup
- Ping Redis once at startup and log an unreachable store at Error
- Log endpoints only, never the connection string
The device cache was filled once at construction, so a token minted by one
replica was unknown to every other replica already running and a token revoked
on one replica stayed valid on the others until they restarted.
- drop the eager device and device options dictionaries
- read devices and device options from the database on every query
- push the device query filters and ordering into SQL
- open the request's database context only for the api key fallback
- cover both directions against real PostgreSQL with two manager instances
A server handed in through JELLYFIN_TEST_POSTGRES outlives the run, so a second
run finds the databases the first one created. Also name the failures in
Jellyfin.Database.Tests.PostgreSQL the CI step steps around.
Every PostgreSQL test carries Category=RequiresDocker and the pipeline filters
that category out, so the provider production runs on was never exercised and
two upgrade-path bugs reached it in a row. Nothing covered the whole startup
sequence either - code and schema migrations interleaved the way the migration
service orders them.
- run both database stages through JellyfinMigrationService against a seeded
library and against a fresh install
- take the server from JELLYFIN_TEST_POSTGRES when it is set, else start a
container
- add a pipeline step that runs the PostgreSQL tests on every push and pull
request, with the server inside the step: the kubernetes backend has no
docker daemon, and a service container deadlocks on the workspace volume
MigrateRatingLevels issued an ExecuteUpdate while the SELECT DISTINCT reader
was still open on the same connection. SQLite tolerates that, Npgsql does not,
so the AppInitialisation stage aborted and every PostgreSQL instance
crash-looped on startup.
- read the distinct ratings into a list before the update loop
- cover the migration against a real PostgreSQL, with NULL and empty ratings
An element the enum cannot parse throws, and the configuration manager
catches that and returns defaults, so one bad preset discarded every other
encoding setting.
- serialize EncoderPreset through a string surrogate in XML
- fall back to auto for an empty, unknown or out-of-range value
- accept only the two id shapes the uuid cast parses, so misplaced hyphens
null out instead of aborting the migration
- clear an OwnerId that already is the detached placeholder, matching the
SQLite chain, so CleanupOrphanedExtras cannot delete the item
- recreate the placeholder item before the repoint if it is missing
- compare the migrated schema against the model, constraints included
The v12.0 rebase replaced the PostgreSQL provider's initial migration instead of
adding to it, so an existing database kept the pre-12.0 schema while the code
migrations ran against it and startup aborted on a missing NormalizedUsername.
- restore 20260305010333_InitialPostgreSql as the baseline
- add 20260306000000_UpgradeToServer12Schema carrying it to the 12.0 model
- convert OwnerId and PrimaryVersionId to uuid with explicit casts, clear
unparseable ids and repoint dangling owners at the placeholder item
- drop orphaned permissions and preferences before UserId becomes non-nullable
- add 20260524120336_AddUniqueNormalizedUsernameIndex after the code migration
that fills the column in
- read an empty or unknown encoding.xml EncoderPreset as the default
- cover both paths against a real PostgreSQL and guard the migration ordering
The chart sets the lease duration but not the retention window that keeps an
orphaned session record available for takeover.
- add ha.transcodeStore.sessionRetentionSeconds and pass it to the statefulset
The branch is a fresh rebase onto the v12.0 tag, so it shares no history with
main and cannot be merged normally.
- record main as a parent without changing a single file
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
The workspace volume is 9.8G and the clone plus build output already fill 5.7G,
so a shared package cache there leaves too little free space to test against.
- restore, build and test in one step with the cache on ephemeral storage
- exclude BackupServiceTests, which requires 5GiB free on the workspace
- print workspace free space before the test run
The backup service test needs 5GiB free on the workspace and the apt mirror
occasionally serves a half-synced index, so both fail the pipeline at random.
- request and limit ephemeral storage on every step
- retry apt-get update and verify fontconfig is loadable before testing
- report workspace free space before the test run
- run the full test filter again
BackupService refuses to write a backup with less than 5GiB free, which the
pipeline workspace volume does not have, so the suite fails on disk size alone.
- exclude BackupServiceTests from the test filter
- keep restore, build and test as separate steps
- install fontconfig for the Skia tests before running them
A solution-wide test run reports one exit code, so a failing suite cannot be
identified without the pipeline logs.
- run the suites carrying HA tests as separate named steps
- run the remaining upstream suites in one step after installing fontconfig
unkin-agent
benvin