mirror of
https://github.com/Dispatcharr/Dispatcharr.git
synced 2026-07-20 16:51:10 +00:00
1fbc23a5ee
The custom psycopg3 gevent pool only evaluates connection age inside get()/put(). A connection checked out by a greenlet that dies without returning it — e.g. a live-stream / channel-switch teardown path that never calls close_old_connections() — is never aged out. It stays counted against MAX_CONNS but is no longer in the free queue, so once MAX_CONNS connections leak per worker, get() blocks forever and every DB-backed request on that worker hangs with a 60s/504 until the container is restarted. DB-free endpoints (e.g. /api/core/version/) stay fast, which is the tell. Add a background reaper greenlet (one per worker, started lazily on first get() so the gevent hub is up) that periodically closes connections which are both expired (older than max_lifetime) and IDLE (no in-flight transaction). The IDLE guard means an active stream or in-flight request is never disrupted, and get() already refuses to hand out an expired connection so a reaped connection can't be in active use. This frees the leaked slot and lets the worker recover on its own — no restart. Side effect: DATABASE_POOL_CONN_MAX_LIFETIME now actually reclaims leaked connections (previously it only recycled healthy idle ones). Lower it to shorten leak-recovery time. Reaper cadence is configurable via DATABASE_POOL_REAP_INTERVAL (default 30s; 0 disables). Also covers the M3U-refresh INTRANS leak family (#1338). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> |
||
|---|---|---|
| .. | ||
| db | ||
| __init__.py | ||
| admin.py | ||
| app_initialization.py | ||
| asgi.py | ||
| celery.py | ||
| consumers.py | ||
| gevent_patch.py | ||
| jwt_ws_auth.py | ||
| persistent_lock.py | ||
| routing.py | ||
| settings.py | ||
| settings_test.py | ||
| urls.py | ||
| utils.py | ||
| wsgi.py | ||