Jian GuoKION GROUP ITKION GROUP
KION GROUP IT's production Open WebUI on Kubernetes with Helm on GKE, Redis Sentinel and websockets, for about 40,000 users.
GitHub (open-webui/open-webui), published Read the issue (opens GitHub (open-webui/open-webui) in a new tab)
Observation
KION GROUP's 40,000-user Open WebUI deployment lost pods when Redis Sentinel lookups slowed: health checks timed out and Kubernetes killed the containers with no chat load.
- Takeaway
- At company scale, load-test the Redis Sentinel path and watch connection churn, because a slow lookup can take down healthy pods through the health check.
- What happened
- Open WebUI v0.10.2 opened a new Redis connection for each operation through Sentinel, about 6.9 a second or some 25,000 an hour. When Sentinel resolution slowed, the timeouts blocked the event loop, the /health endpoint timed out, and Kubernetes killed the pod with exit code 137, an outage while Redis itself stayed healthy with no restarts or rejected connections. The report proposes caching the master client and reusing it across calls.
- Public engineering write-up (a GitHub issue filed from KION GROUP IT)
- Connection rates measured
Deployment details
KION GROUP IT's production Open WebUI on Kubernetes with Helm on GKE, Redis Sentinel and websockets, for about 40,000 users.
- Public engineering write-up (a GitHub issue filed from KION GROUP IT)
- Connection rates measured
- Team
- KION GROUP IT
- Workload
- Company-wide chat for about 40,000 users, websockets enabled
- Product version
- Open WebUI v0.10.2 with redis-py 8.0.0
- Measurement
- About 6.9 new Redis connections a second, some 25,000 an hour (Jul 20, 2026)


