BaseChannelListener.listen() only called pubsub.unsubscribe() in its
finally block: the pubsub connection and the listener's own Redis client
were never released, leaking a connection for every install/uninstall
(_watcher) and every recreation of the app-status listener. The Celery
task notifiers were likewise never closed, and used the deprecated
Redis.close() instead of aclose().
- add aclose() to BaseChannelNotifier and BaseChannelListener
- listen() now closes the pubsub and its Redis client in finally
- close the notifiers in the app_manage / app_status_update tasks and
switch to aclose()
- back off in the app-status watch loop so a Redis outage no longer
spins, and drop a leftover debug print
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
app_manage_task_impl released APP_MANAGE_LOCK_KEY and broadcast a
FINISHED message from its finally block unconditionally, including on
the early-return path where acquire_lock reported the lock as already
held by a running install. A duplicate install/uninstall request would
therefore delete the running task's lock (allowing concurrent
management of the same app) and send FINISHED, which stopped that
task's AppManageListener before it was done.
Track whether this task actually acquired the lock and only
release + finish when it did.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
