Refactor tests for VOD provider failover logic to improve clarity and remove unnecessary components. Introduce new tests for order candidates and ensure no database access occurs during processing.
- Added calls to `close_old_connections()` in `stream_vod()` and `build_vod_stats_data()` to prevent connection leaks during long-lived streaming responses and background stats refreshes.
- Improved connection handling in various components to ensure proper resource cleanup and prevent blocking issues.
Issue: When `process_movie_batch` / `process_series_batch`
(apps/vod/tasks.py) creates duplicate vod_movie / vod_episode records
during a refresh, existing M3U*Relation rows get repointed to the new
records. The old UUIDs that external players (Emby / Jellyfin /
ChannelsDVR) cached in `.strm` URLs are left orphaned, and the proxy
404s on every subsequent request — even though the same request
already carries a stable `stream_id` that uniquely identifies a live
relation.
This patch makes `_get_content_and_relation` use a soft UUID lookup
and, if it misses, fall back to resolving content via
`M3UMovieRelation.stream_id` / `M3UEpisodeRelation.stream_id`. The
fallback respects the strictest match first: when
`preferred_m3u_account_id` is also present, that account is queried
first; otherwise the highest-priority active relation matching the
stream_id wins (same ordering the existing relation-selection logic
uses).
This is read-side graceful degradation, complementary to the
write-side root fix proposed in closed#973 (re-use the relation's
existing movie/series during refresh rather than rematching by TMDB
ID). The two PRs together would fully resolve#961: #973 prevents new
orphaning, this PR makes any URL that ships a stream_id survive past
orphaning. URLs without a stream_id (e.g. XC-compat
`/movie/<user>/<pass>/<id>.mkv`) are NOT covered by this patch — they
need the root fix.
Series-UUID lookups (`/proxy/vod/series/<UUID>`) are left UUID-only by
design — players cache episode and movie URLs, not series URLs. Same
pattern can be added as a follow-up if needed.
Tests: 7 new mock-based regression tests covering both branches,
verified against `:dev`. UUID-first happy path is unchanged
(M3U*Relation table never queried when the UUID resolves).
Related: #961 (open), closed#973 (root fix abandoned for inactivity).
Three race conditions in multi_worker_connection_manager could leave
the Redis profile_connections:<id> counter permanently elevated with no
active streams, causing all VOD requests to 503 "All profiles at capacity".
Bug 1 — decrement_active_streams() return value was ignored
All three generator exit paths (normal, GeneratorExit, Exception) called
decrement_active_streams() and unconditionally set decremented = True
regardless of whether the lock was acquired. On lock contention the decrement
was silently skipped, active_streams remained > 0, the subsequent
has_active_streams() check returned True, and _decrement_profile_connections()
was never called. The counter was then stuck until manual DEL.
Fix: add decrement_active_streams_and_check() which performs the decrement and
the "are there remaining streams?" check atomically under a single lock,
eliminating the race window. All three exit paths and the finally block now
use this method and propagate its success/remaining return values.
Bug 2 — non-atomic GET-then-DECR in _decrement_profile_connections()
The previous implementation read the counter with GET then conditionally
called DECR. Two concurrent decrements could both pass the > 0 guard and
both fire, driving the counter to -1. A subsequent _check_and_reserve_profile_slot()
INCR would then produce 0 which passes the <= max_streams check, allowing
an extra stream to bypass the limit on the next request.
Fix: replace GET-then-DECR with a direct DECR (matching the INCR-first
pattern already used by _check_and_reserve_profile_slot) and clamp the
result to 0 if it goes negative.
Bug 3 — has_active_streams() read state without holding the lock
The separate has_active_streams() call after decrement_active_streams()
released its lock left a window where another concurrent stream could
increment active_streams back to 1, causing the profile decrement to be
skipped. This is resolved as a consequence of Bug 1's fix: the new
decrement_active_streams_and_check() method reads active_streams while
the lock is still held, eliminating the window entirely.
Adds tests covering all three scenarios in
apps/proxy/vod_proxy/tests/test_profile_connections.py.
Fixes#1121.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
