Setup: HAOS VM on Proxmox (Celeron J4005, 2 cores), HA 2026.8.3, 2000 entities, MariaDB recorder, Alexa via Nabu Casa Cloud.
One evening, with 350 entities going through reconnects, the core process pinned at 74–106% CPU for hours (33 threads on 2 vCPUs). Symptoms: automations ran late and out of order, UI slow and sometimes frozen. The log was flooded with 415 “Timeout sending report to Alexa” errors in 40 minutes — state_report calls failing and retrying while the main loop starved.
Disk was fine (69%), MariaDB was idle, network/TLS to Alexa didn’t had any latency (7–16 ms), MQTT was OK. strace showed the CPU going to futex contention (thread queuing).
Fix: reloading the Cloud integration purged the queued Alexa reports. Within a minute: CPU went from 100% to 15%, API latency back from 300 ms to 3 ms, automations worked fine again.
Questions:
- Do Alexa state_report retries retry unbounded? With many entities unavailable, this seems to grow into a self-sustaining queue. Can retries be rate-limited, or unavailable entities excluded from reports?
- Is ~2000 entities simply too much for a 2-core Celeron? What would you trim first?