The Valkey transport layer used 100ms busy-polling loops (Task.Delay(100))
across ~90 concurrent loops in 45+ services, generating ~900 idle
commands/sec and burning ~58% CPU while the system was completely idle.
Replace polling with Redis Pub/Sub notifications:
- Publishers fire PUBLISH after each XADD (fire-and-forget)
- Consumers SUBSCRIBE and wait on SemaphoreSlim with 30s fallback timeout
- Applies to both ValkeyMessageQueue (INotifiableQueue) and ValkeyEventStream
- Non-Valkey transports fall back to 1s polling via QueueWaitExtensions
Increase heartbeat interval from 10s to 45s across all transport options,
with corresponding health threshold adjustments (stale: 135s, degraded: 90s).
Expected idle CPU reduction: ~58% → ~3-5%.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
841add4f27