Skip to main content

​
Runtime tuning

  • Reuse tts_config hashes: VoiceManager caches rendered audio per text + config hash, so reusing configs dramatically reduces provider round-trips.
  • Disable audio when needed: Set audio=false in the config message to use LiveKit data relaying without spinning up STT/TTS providers.
  • Balance CPU vs. quality: noise-filter improves transcripts but costs CPU; toggle the feature flag when ultra-low latency is more important than denoising.
  • Persist caches: Mount CACHE_PATH to a durable volume if you want previously synthesized clips and turn-detect assets to survive restarts.

​
Observability & testing

  • Use the integration tests in tests/ (e.g., tests/ws_tests.rs) as references when extending message formats or LiveKit behavior.
  • Emit application metrics around provider errors and queue depth to quickly detect throttling.
  • Enable structured logging inside WebSocket handlers so you can correlate STT/TTS lifecycle events with LiveKit callbacks.

​
Release hygiene

  • Regenerate the OpenAPI file with cargo run --features openapi -- openapi -o docs/openapi.yaml whenever you tweak request or response schemas.
  • Keep the documentation and server version in lockstep; expose the git SHA or semver string via / so probes can verify deployments.
  • Rotate provider credentials and authentication keys as part of your deployment pipelines; the Authentication guide documents the required env vars.

​
Operational playbook

Many outages boil down to LiveKit connectivity or provider rate limits. Monitor those dependencies explicitly and surface their status in your dashboards.
  1. Circuit breakers: Treat provider 5xx responses as signals to backoff and fall back to cached content when possible.
  2. Traffic shaping: Consider per-tenant rate limits enforced inside your JWT auth service so Sayna sees predictable load.
  3. WebSocket health: Alert on connection churn, repeated error frames, or clients that never send config.
  4. LiveKit drift: Keep Sayna and LiveKit clocks in sync; recording requests rely on accurate timestamps.