- Ya
- 4 Septemba 2026, 20:51 UTC
- Mwandishi
- Kamo
- Ahadi ya
- 2ab9848
McpServerManager runs each MCP server as a child process of THIS pod and tracks them in in-memory maps; McpHealthMonitor writes what it finds there into a SHARED config row every 30 seconds. Two pods would each spawn their own copy of every configured server and then take turns overwriting that row with their own local view, so the recorded health would describe whichever pod wrote last rather than the system. A distributed lock does not help here — the state being reported is genuinely pod-local. Scaling needs the server processes moved out of the pod, or the health row made per-pod. Written down so the next person does not have to rediscover it from a confusing health display.