Stream.get_stream() was never called in the preview path of
generate_stream_url(), so the channel_stream and stream_profile
Redis keys were never written. This caused the stream_id to be
missing from channel metadata, which silently broke two things:
- Stats page showed no stream name or logo
- stream_stats were never saved to the database
Fix: replace the manual profile-selection loop in the preview path
with a direct call to Stream.get_stream(), matching the channel path
exactly. This also fixes a pre-existing non-atomic slot reservation
(read-then-check vs INCR-first) that could over-allocate connections
on concurrent previews.
Regression introduced in 49b7b9e2.
Restructure cleanup in refresh_single_m3u_account so del statements
run before gc.collect(), enabling cyclic garbage collection. Add worker_max_memory_per_child as a safety net against memory fragmentation.
Closes#1012, Closes#1053
- Add search, channel, and status filter controls to DVR page (matching Guide/VODs pattern)
- Fix post_save signal to schedule recordings for currently-playing programs immediately instead of skipping them
- Remove duplicate evaluate_series_rules.delay() call from series rule creation endpoint to prevent race condition that created two recordings for the same program
Frontend:
- RecordingDetailsModal: fix ReferenceError crash (editing state used before declaration in useMemo — TDZ in production bundles broke all tile clicks)
- RecordingDetailsModal: fix focus-stealing on description edit by calling Movie()/Series() as plain functions instead of <Movie />/<Series />(inline component refs change every render, causing unmount/remount)
- RecordingDetailsModal: suppress sub_title in modal title after custom save
- ErrorBoundary: log caught errors and show message for easier debugging
- guideUtils: filter terminal recordings from mapRecordingsByProgramId so Guide red dots clear after recording finishes or rule is deleted
- RecordingDetailsModalUtils.test: add missing end_time to test fixtures (filterByUpcoming now uses end_time instead of start_time)
Backend:
- tasks.py: output file collision avoidance (check .mkv/.ts before writing)
- tasks.py: pre-restart TS segment concatenation (direct concat→MKV remux)
- tasks.py: recovery pipeline — recover crashed "recording" status, remux expired recordings, revoke stale tasks, deterministic task IDs, early lock release, 10-min lock TTL for large remux operations
- tasks.py: poster resolver skips external API when title matches channel name
- celery.py: trigger recover_recordings_on_startup via worker_ready signal
Tests:
- Renamed test_dvr_fixes.py → test_recording_pipeline.py (clarity)
- Renamed test_dvr_retry.py → test_db_retry.py (clarity), fixed mock paths for close_old_connections (resolved 3 test failures), replaced synthetic test
- test_recording_stop_cancel: replaced brittle time.sleep(0.5) with deterministic thread-capture pattern
- Idempotency guard: fail closed on DB errors instead of proceeding, preventing duplicate tasks from overwriting valid recordings (#641)
- Remux sanity checks: reject MKV output that is <50% of a previous MKV (duplicate-task overwrite) or <10% of the source TS (corrupt first attempt); preserve .ts for manual recovery on all failure paths (#619)
- Recurring rule timezone fix: use user's configured timezone for date window calculation instead of UTC, fixing silent recording loss for users in UTC-negative timezones after 4pm local time (#1042)
- Include .ts path in final remux failure log for easier recovery
Backend:
- Add terminal status guard to stop() endpoint -prevents overwriting completed/interrupted/failed with "stopped" (returns 409 Conflict)
- Include currently-airing programs in series rule evaluation by filtering on end_time instead of start_time
- Clean up future recordings when deleting a series rule while preserving previously recorded episodes
- Add logo proxy negative cache to prevent repeated requests to unreachable hosts from blocking Daphne workers
- Add URL validation with HEAD-then-GET fallback and per-worker cache
- Add EPG time-slot matching (80% overlap threshold) to enrich manual recordings with program metadata
- Add multi-source poster resolution pipeline (EPG, VOD, TMDB, OMDb, TVMaze, iTunes) with URL validation at each stage
- Add editable recording metadata endpoint with user_edited flag to prevent EPG auto-overwrite
- Add refresh-artwork endpoint to re-run poster resolution on demand
- Batch-fetch PeriodicTask names in schedule_upcoming to avoid N+1
Frontend:
- Fix series modal showing only one episode — preserve _group_count from categorized prop when merging with store version so the episode list renders correctly
- Fix recording card button positioning — use flex column layout with marginTop auto so buttons align at the bottom regardless of content
- Add delete fallback in RecurringRuleModal for orphaned recordings from deleted rules
- Use end_time filter in upcoming episodes list to include currently-airing episodes
- Add inline metadata editing (title/description) in details modal
- Add refresh artwork button in details modal
- Replace hardcoded /logo.png fallbacks with imported default logo
- Add React.memo with custom comparator to RecordingCard
Tests:
- Add 21 EPG time-slot matching tests (overlap threshold, edge cases)
- Add 16 recording metadata tests (validation, artwork, logo cache)
- Add 17 URL validation tests (HEAD/GET fallback, caching, eviction)
Add exponential backoff retry (_db_retry helper) for transient DB errors during final metadata save, startup recovery, and initial TS proxy
connection. Add FFmpeg remux retry for transient I/O errors. Retry same proxy base before falling back to next candidate on initial connection failure.
Fix DVR card logos not displaying: channel summary API returns logo_id (integer) but frontend expected a nested logo object with cache_url.
Add getChannelLogoUrl() helper that handles both data shapes. Backend artwork prefetch and run_recording now fall back to the channel's own logo when no show-specific artwork is found, ensuring poster_logo_id is always populated.
Prevent incorrect artwork by skipping external API searches (TVMaze, iTunes) when no program title is available - channel names like
"USA A&E SD*" returned unrelated results from fuzzy matching.
Add 13 tests for retry logic (test_dvr_retry.py)
- Add reconnection loop to DVR streaming: on transient connectivity loss (ReadTimeout, ConnectionError, ChunkedEncodingError), the recording task retries the same TS proxy base up to 5 times with a 2-second delay, appending to the existing file. Counter resets on successful data receipt. TS container format tolerates brief discontinuities; MKV remux normalises timestamps.
- Extract _check_recording_cancelled() helper to consolidate repeated stop/delete DB checks across the streaming and reconnection paths.
- Fix comskip treating exit code 1 (no commercials detected) as a fatal error.
Use check=False and handle return codes explicitly: 0 = commercials found,
1 = no commercials (skipped), negative = signal, other = real error.
- Add human-readable reason messages for comskip 'skipped' notifications in
WebSocket handler.
- Remove redundant conditional Channel import in stream_generator._cleanup() that shadowed the module-level import, causing UnboundLocalError when the conditional branch did not execute.
File changes:
Backend
apps/channels/api_views.py — Added extend() action that uses queryset .update() to bypass signals, letting the running task pick up the new end_time via polling. Moved WS recording_stopped notification to synchronous (before HTTP response). Refactored destroy() to delete DB row first, send WS event immediately, defer cleanup to background thread.
apps/channels/signals.py — Added pre_save guard to skip task revocation when status is "recording". Added post_save reentrancy guard to skip artwork prefetch for internal field-only saves (custom_properties, task_id, end_time).
apps/channels/tasks.py — Added 2-second polling loop in streaming that checks for stop, deletion, and extended end_time. Added refresh_from_db() + merge before metadata save to prevent overwriting concurrent artwork prefetch. Replaced inline PeriodicTask cleanup with revoke_task()/_dvr_task_name(). Consolidated redundant DB queries in error path. Fixed TOCTOU os.path.exists + os.remove. Preserved "stopped" status in final-status logic instead of overwriting with "completed".
core/utils.py — Made send_websocket_update gevent-safe by detecting monkey-patching and offloading async_to_sync to gevent's native threadpool, fixing the event loop deadlock that caused cross-recording interference.
apps/proxy/ts_proxy/client_manager.py — Moved _trigger_stats_update to a background thread so remove_client() returns immediately.
apps/proxy/ts_proxy/stream_generator.py — Added Redis-based health check for non-owner workers in _should_send_keepalive(), so keepalives fire correctly even when stream_manager is None.
Frontend
frontend/src/WebSocket.jsx — Added 400ms debounced scheduleRecordingFetch() replacing all inline fetchRecordings() calls. Added recording_extended and recording_updated event handlers. recording_cancelled now does surgical removeRecording() by ID when available.
frontend/src/api.js — Added extendRecording(id, extraMinutes).
frontend/src/components/cards/RecordingCard.jsx — Added Extend menu (+15m/+30m/+1h) for in-progress recordings. Widened Comskip button to stopped/interrupted statuses.
frontend/src/components/forms/ProgramRecordingModal.jsx — Removed manual fetchRecordings() calls, relying on WS-driven debounced refresh.
frontend/src/components/forms/RecordingDetailsModal.jsx — Added stopped to playable statuses. Widened Comskip button. Removed manual fetchRecordings().
frontend/src/components/forms/RecurringRuleModal.jsx — Removed all manual fetchRecordings() calls, relying on WS events.
frontend/src/pages/Guide.jsx — Removed isLoading subscription that caused loading flash on debounced refresh. Removed manual fetchRecordings() after saving a series rule.
frontend/src/store/channels.jsx — Stripped isLoading/error state from fetchRecordings to prevent full-page loading indicators. Added no-op guards in removeRecording.
frontend/src/utils/cards/RecordingCardUtils.js — Added extendRecordingById wrapper.
Tests
apps/channels/tests/test_recording_extend.py (new) — 14 tests: extend endpoint validation, stacked extensions, signal bypass, pre_save guard behavior.
apps/channels/tests/test_recording_stop_cancel.py (new) — 18 tests: stop endpoint, DVR client teardown, cancel was_in_progress flag, run_recording race guards, signal reentrancy guards.
apps/channels/tests/test_ts_proxy_keepalive.py (new) — 21 tests: keepalive owner/non-owner logic, stats update error handling, client removal non-blocking, timing invariants.
apps/channels/tests/test_recording_scheduling.py — Updated mock assertion to match new _stop_dvr_clients call signature.
Summary:
Adds an Extend action for in-progress recordings that bypasses Django signals and lets the running task dynamically pick up the new end_time via its 2-second DB polling loop. Fixes cross-recording interference caused by send_websocket_update deadlocking the gevent event loop. Fixes an artwork race condition where run_recording overwrote concurrent prefetch data. Adds Redis-based keepalive health checks for non-owner workers to prevent DVR read timeouts during source transitions. Replaces manual fetchRecordings() calls throughout the frontend with a debounced WS-driven refresh. 53 new tests.
Fixes#940: Duplicate DVR task executions were caused by ClockedSchedule objects being shared across PeriodicTasks, triggering run_recording multiple
times for the same recording. Added an idempotency guard in run_recording that exits early if status is already 'recording', 'completed', or 'stopped'.
revoke_task() now deletes the PeriodicTask and orphaned ClockedSchedule on execution rather than relying on Celery revocation alone.
Fixes#454: Added POST /api/channels/recordings/{id}/stop/ action that marks the recording as stopped and terminates only the specific DVR proxy client (identified via User-Agent: Dispatcharr-DVR/recording-{id}) without disrupting
simultaneous recordings on the same or other channels. The API returns
immediately; DVR client teardown and task revocation happen in a background thread to avoid 504 timeouts on the recordings list endpoint.
Additional changes:
- destroy() now only calls _stop_dvr_clients() for in-progress recordings, preventing accidental stream termination when deleting completed recordings
- WebSocket events differentiated: recording_stopped, recording_cancelled (in-progress cancel), and recording_cancelled (delete) with was_in_progress flag so the frontend shows distinct "Recording deleted" vs "Recording cancelled" notifications
- Frontend: Stop and Cancel split into separate buttons with confirmation
modals; stopped recordings move immediately to Previously Recorded via
optimistic status bucketing; removed dead 'Stopped' badge state
- 27 new unit tests covering scheduling, idempotency, revocation, and DVR client isolation
- Fix Postgres version check failing with restricted DB users (use $POSTGRES_DB instead of hardcoded 'postgres')
- Fix DVR recording broken in modular mode (respect DISPATCHARR_PORT instead of hardcoding 9191)
- Remove flushdb() from wait_for_redis.py to prevent Redis data loss on container restart
- Add DISPATCHARR_PORT to celery environment in docker-compose.yml
- Add depends_on health conditions for proper service startup ordering
- Add extra_hosts for host.docker.internal resolution on Linux
- Harden celery entrypoint with timeouts for JWT wait (120s) and migration wait (300s)
- Replace fragile showmigrations grep with migrate --check
- Add unit tests for DVR port resolution and flushdb removal regression
- Register Lua scripts for optimized chunk retrieval in StreamBuffer.
- Implement atomic Lua binary search to find the oldest available chunk in Redis.
- Update StreamGenerator to handle expired chunks more efficiently, reducing TOCTOU race conditions.
The redirect path in stream_ts() released the connection slot via release_stream() but did not reset the connection_allocated flag. Could have could caused a redundant no-op call in the exception handler.
If ProxyServer() raises during singleton construction, _instance was left
as _INITIALIZING permanently, causing all subsequent get_instance() callers
to spin in an infinite gevent.sleep() loop. Wrap construction in try/except
and reset _instance to None on failure so callers can retry.
Also replace the non-atomic setnx() + expire() pair in try_acquire_ownership()
with a single atomic SET NX EX call, consistent with the approach already used
in extend_ownership() and eliminating the race window where a crash between the
two calls could leave a key with no TTL (permanent ownership lock).
- Use atomic SET NX EX instead of separate SETNX + EXPIRE to prevent
zombie locks if the process crashes between the two calls
- Replace time.sleep() with gevent.sleep() in get_instance() spin-wait
to avoid blocking the greenlet hub
- Defer cleanup when ownership is lost but clients are still connected,
giving the new owner time to establish its stream before we tear down
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Three interrelated bugs cause TS proxy streams to terminate prematurely
in multi-worker uWSGI deployments:
1. Double ProxyServer instantiation: ProxyConfig.ready() in apps/proxy/apps.py
calls TSProxyServer() directly, bypassing get_instance(). The subsequent
TSProxyConfig.ready() call to get_instance() creates a second instance.
Each instance starts its own cleanup thread, but only one holds channel
data — the orphaned cleanup thread cannot extend ownership.
2. Redis flushdb() on every client init: RedisClient.get_client() in
core/utils.py calls flushdb() whenever a new connection is created.
Celery autoscale workers spawning mid-stream nuke all Redis keys
including ownership, client records, and channel metadata.
3. No recovery from expired ownership: get_channel_owner() has a TOCTOU
bug (two separate GET calls in a lambda). extend_ownership() silently
fails when keys expire. Non-owner cleanup unconditionally kills streams
even when the worker holds the stream_manager.
Fixes:
- Use TSProxyServer.get_instance() in ProxyConfig.ready()
- Remove flushdb() from Redis client initialization
- Use sentinel pattern for gevent-safe singleton (threading.Lock does not
work with gevent greenlets)
- Single GET in get_channel_owner() to avoid TOCTOU race
- Re-acquire expired ownership keys in extend_ownership()
- Attempt re-acquisition before cleanup in non-owner path
Relates to #992, #980
Replace N+1 update_or_create() loop with bulk_create(update_conflicts=True)
wrapped in transaction.atomic(). On large M3U accounts with many groups,
the previous approach issued individual DB queries per group/category,
which could take minutes on slower hardware like Synology NAS devices.
