The main loop snapshots managed_tags, then two steps remove a slot from
managed_tags before update_ingestor_manager compares the snapshot with the
current state and calls unsubscribe_slot for every slot that disappeared:
- manage_leases drops an extra lease ("Extra slots available"): it already
unsubscribed the slot, so the second call is redundant;
- update_slot_config finds the slot gone from Redis (deleted by the
orchestrator): nothing unsubscribed it yet.
unsubscribe_slot looked the slot up in managed_tags[slot] to find its servers,
so both cases raised KeyError, which stopped the main loop and restarted the
container (seen on dev on 2026-10-06 18:30 during a rollout, ~40 s without
reads). In the second case the subscription also stayed open on the OPC server.
The same code is on main, so production has it too.
unsubscribe_slot now asks the readers themselves: every reader (ServerReader or
the legacy OpcManager, both keep `subscriptions` by slot) holding a
subscription for the slot unsubscribes it. A slot no reader holds is a no-op.
This also covers a resubscription whose new config dropped a server: the old
server's subscription for that slot is closed too.
Tests: unit tests for the new lookup and the double call, plus two regressions
at the Ingestor level with the real unsubscribe_slot (dropped extra lease;
slot removed from Redis). Without the fix they fail with the same KeyError at
ingestor_manager.py:558. 280 unit tests.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
057c1fd1ce