Waiters that lost the refresh lock polled the shared cache for only 750ms
before falling through to the upstream, but the refresh lock lives for 15s.
On a cold cache with an upstream slower than 750ms, every waiting pod gave up
and stampeded the origin (each also writing L2). Extend the poll window to
match the lock TTL so waiters return the entry the holder publishes instead.
Adds a regression test that fails without the fix.
The User and Status columns used hardcoded pixel widths with table-layout:
fixed, so they stayed narrow regardless of how wide the widget actually
rendered - text truncated even with plenty of unused space in the
"currently playing" column. Switch both to percentages so they scale with
the widget.
Emby's integration also hardcoded playback/metadata to null for every
session, so the widget's existing progress bar, resolution, and bitrate
display never had anything to show. Port the same TranscodingInfo/
MediaStreams parsing Jellyfin already does (same underlying /Sessions API
shape) so Emby sessions get real position/duration, resolution, bitrate,
and transcode-status data. Also fix ticksToMs treating a position of
exactly 0 as missing, and share the one implementation between the two
integrations instead of duplicating it.
The WidgetTimeProvider plumbed a server timestamp and the container's
timezone down to useWidgetNow so clock widgets would paint a real time in
the SSR HTML instead of "--:--". That traded a cosmetic first-paint win for
a wrong clock: the seeded timezone is the server's, so anyone whose timezone
differs from the container's saw an incorrect time for the whole hydration
window. Fixing that properly needs a client-set timezone cookie read during
RSC render, which is far more plumbing than the problem warrants.
Revert the clock to release/v2 behaviour ("--:--" until hydration) and drop
the provider, the @homarr/widgets/time export and their tests. The e2e clock
assertions now wait on hydration explicitly instead of the seeded paint.
Also restore the widget.beszel.getSystemStats persistence exclusion. #6289
added it deliberately to keep those large time-series payloads out of the
persisted cache, and this branch had removed it with no stated rationale.
Query-cache persistence and the background-refresh indicator are unchanged.
Review follow-ups on the session query cache:
- Cache the per-payload size verdict and reuse a single TextEncoder. The
persist subscription re-dehydrates the whole cache on every cache event,
so every widget payload was being superjson-serialized several times per
refresh cycle just to measure it.
- Skip the size measurement on the restore path. It runs before any query is
allowed to fetch and the payload was already budgeted when it was written.
- Reject stored queries whose dataUpdatedAt is not a positive finite number;
`undefined <= 0` is false, so those were hydrating with a NaN timestamp.
- Add `stop()` and tear the persister down before `queryClient.clear()`, so
the clear cannot overwrite the saved snapshot with an empty cache.
- Remove every other scope's cache when a persister is created: signing out
navigates away before the scope guard can react, which left the previous
account's dashboard data in the tab.
- Remove the entry instead of storing an empty cache.
Refresh indicator:
- Make it decorative instead of an aria-live region; every widget refreshes
on a timer, so it made assistive technology repeat "loading" forever.
Also broadens the e2e hydration-error matcher to production React's minified
error codes, hoists the single WidgetTimeProvider in ClientBoard, memoizes the
resolved local time zone, and drops the exports this branch made dead.
ajnart
Thomas
Bergasha
Thomas Camlong