Availability domains
Monitor and recover media/TURN, room authorization, agent workers, STT/TTS, frame sampling, reasoning, retrieval, policy, database, storage and notifications independently.
Backpressure order
- Preserve room signalling and two-way audio.
- Reduce video bitrate, resolution, frame rate and AI sampling.
- Pause nonessential visual analysis and deep research.
- Switch to audio plus explicit stills.
- Queue nonurgent cases and preserve accepted evidence.
- Protect critical/officer-priority sessions.
- Never bypass exact policy verification to increase throughput.
Failure modes
| Failure | Response |
|---|---|
| SFU/TURN issue | Reconnect, alternate TURN/SFU, preserve case, offer audio/queued fallback |
| Video sampler unavailable | Declare visual analysis unavailable; request explicit still or human join |
| STT unavailable | Typed/recorded fallback or hold with explanation |
| Reasoning provider unavailable | Fallback model, queue, or officer route |
| Policy service unavailable | Block exact treatment output |
| Database unavailable | Do not acknowledge durable actions; use controlled reconnect |
Disaster recovery
Back up PostgreSQL and selected-evidence storage, export media configuration and room policies, maintain provider exit/runbooks, and test restoration. Active calls may need reconnection; durable case state and accepted evidence must survive.
SLOs and observability
Track join success, time to audio/video, packet loss, jitter, TURN usage, reconnects, degradation mode, accepted-frame latency, room abandonment, officer join time and case completion.