fix: skip unhealthy direct ollama probes
All checks were successful
CD Pipeline / deploy (push) Successful in 1m10s
All checks were successful
CD Pipeline / deploy (push) Successful in 1m10s
This commit is contained in:
@@ -324,3 +324,9 @@
|
||||
- V10.626 新增 `OLLAMA_HOST_PRIMARY_PROXY` / `OLLAMA_HOST_SECONDARY_PROXY`,預設為 `http://192.168.0.110:11435` / `http://192.168.0.110:11436`。
|
||||
- `resolve_ollama_host()` 順序調整為 GCP-A direct → GCP-A via 110 proxy → GCP-B direct → GCP-B via 110 proxy → 111;proxy rescue 是同順位入口救援,不代表 direct GCP host 已恢復。
|
||||
- 近 24 小時 `ai_calls` 只有 `ollama_secondary=51`、`gcp_ollama=3`、`nim=1`,沒有 Gemini provider;Gemini hard disabled / fallback disabled 的紅線仍有效。
|
||||
|
||||
## 30. 2026-06-18 V10.627 Resolver 讀 host_health 跳過 direct timeout
|
||||
|
||||
- V10.626 已能在 GCP-A direct timeout 後走 110 proxy,但 cache refresh 仍會先等一次 direct `/api/version` timeout。
|
||||
- V10.627 新增 direct-only host health skip:`resolve_ollama_host()` 會讀最近 `host_health_probes`,若 GCP-A/GCP-B direct 在視窗內已 unhealthy,先跳過 direct endpoint,改試同順位 110 proxy;proxy rescue 不吃這個 skip,避免因 direct unhealthy 誤跳過可用 proxy。
|
||||
- 新增 `OLLAMA_RESOLVE_HOST_HEALTH_SKIP_ENABLED=true` 與 `OLLAMA_RESOLVE_HOST_HEALTH_SKIP_WINDOW_MINUTES=20`;DB 讀取失敗 fail-open,回到原本網路探測。
|
||||
|
||||
Reference in New Issue
Block a user