Root cause: stale NATS JetStream consumers accumulated across API
restarts, causing 13+ consumers to fight over messages in a single
Python async event loop (100% CPU).
Fixes:
- Add performance indexes on devices(tenant_id, hostname),
devices(tenant_id, status), key_access_log(tenant_id, created_at)
— drops devices seq_scans from 402k to 6 per interval
- Remove redundant ORDER BY t.name from fleet summary SQL
(tenant name sort is client-side, was forcing a cross-table sort)
- Bump NATS memory limit from 128MB to 256MB (was at 118/128)
- Increase dev poll interval from 60s to 120s for 400+ device fleet
The stream purge + restart brought API CPU from 100% to 0.3%.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
1042319a08