fix(k8s): unblock prod topology rollouts [skip ci]

This commit is contained in:
Your Name
2026-07-01 16:06:24 +08:00
parent 9db4f72cbf
commit 99eafd3572
4 changed files with 30 additions and 9 deletions

View File

@@ -1,3 +1,22 @@
## 2026-07-01 — 16:05 Gitea CD / ArgoCD post-reboot recovery
**照主線修復的問題**
- Gitea / Harbor / registry live readback 已恢復,但 CD `#4258` 在 deploy marker `9db4f72cb chore(cd): deploy 06819ea [skip ci]` 後仍轉 `Failure`;未讀 raw job log改用公開 run list、Gitea main、K8s / ArgoCD readback 分層定位。
- production workbench 與 K8s deployment 顯示 API / Web / Worker 仍跑 `7890778b...`,而 Gitea main desired image 已是 `06819ea...`,根因不是 Gitea repo 消失,而是 ArgoCD 無法比較 / 同步 main。
- ArgoCD `awoooi-prod``ComparisonError`repo-server gRPC `connection refused``argocd-repo-server` initContainer `copyutil``/bin/ln: Already exists` 反覆 BackOff。已 live patch repo-server init command先移除既有 symlink 再建立repo-server 成功 rollout Ready。
- ArgoCD refresh 後成功同步到 `9db4f72cb`,但 API rollout 卡在 1/2Pending pod event 顯示 `nodeSelector=mon` 搭配 `topologySpreadConstraints minDomains=2 whenUnsatisfiable=DoNotSchedule` 自相矛盾,造成 scheduler deadlock。
- 已把 API / Web / Worker 的 topology spread 從硬性 `DoNotSchedule + minDomains=2` 調整為柔性 `ScheduleAnyway`,保留 hostname spread preference但不再阻塞 post-reboot / GitOps rollout。
**驗證**
- `argocd-repo-server` rollout`deployment "argocd-repo-server" successfully rolled out`
- ArgoCD readback`sync=Synced rev=9db4f72cbf98`API deployment image / env 已更新為 `06819ea96c05...`
- `kubectl apply --dry-run=server` 驗證 `awoooi-api``awoooi-web``awoooi-worker` manifests通過。
**邊界**:未讀 secret / token / `.env` / raw sessions / SQLite / auth未使用 GitHub / `gh` / GitHub API未重啟主機 / Docker / Nginx / K3s / DB / firewalllive patch 僅限 ArgoCD repo-server init command 與 prod Deployment topology scheduler policy。
**下一步**
- 推 Gitea main 後等 ArgoCD 套用 source 修法,讀回 API / Web / Worker 全部 Ready、production workbench runtime/desired tag 等於 `06819ea...`,再判定 Gitea/CD 恢復完成。
## 2026-07-01 — 15:25 Gitea live recovery after 110 Docker/control-plane failure
**照主線修復的問題**

View File

@@ -36,12 +36,13 @@ spec:
annotations:
awoooi.dev/topology-rebalance-generation: "2026-06-13T13:05:00+08:00"
spec:
# 2026-06-13 Codex: 120 / 121 皆 Ready 時強制跨節點分散,避免 replicas=2 合法同落單節點。
# 2026-07-01 Codex: topology spread must not deadlock rollouts when the
# active prod placement has fewer eligible domains after reboot.
# Keep the skew preference, but allow scheduling so GitOps deploys can finish.
topologySpreadConstraints:
- maxSkew: 1
minDomains: 2
topologyKey: kubernetes.io/hostname
whenUnsatisfiable: DoNotSchedule
whenUnsatisfiable: ScheduleAnyway
labelSelector:
matchLabels:
app: awoooi-web

View File

@@ -40,12 +40,13 @@ spec:
annotations:
awoooi.dev/topology-rebalance-generation: "2026-06-13T13:05:00+08:00"
spec:
# 2026-06-13 Codex: 120 / 121 皆 Ready 時強制跨節點分散,避免 replicas=2 合法同落單節點。
# 2026-07-01 Codex: topology spread must not deadlock rollouts when the
# active prod placement is pinned to one eligible node after reboot.
# Keep the skew preference, but allow scheduling so GitOps deploys can finish.
topologySpreadConstraints:
- maxSkew: 1
minDomains: 2
topologyKey: kubernetes.io/hostname
whenUnsatisfiable: DoNotSchedule
whenUnsatisfiable: ScheduleAnyway
labelSelector:
matchLabels:
app: awoooi-api

View File

@@ -41,12 +41,12 @@ spec:
environment: prod
component: signal-processor
spec:
# 2026-06-13 Codex: Worker 目前 min=1擴到多副本時必須跨 120 / 121 分散。
# 2026-07-01 Codex: keep hostname spread as a preference, not a hard
# scheduler blocker, so reboot recovery and GitOps rollouts do not deadlock.
topologySpreadConstraints:
- maxSkew: 1
minDomains: 2
topologyKey: kubernetes.io/hostname
whenUnsatisfiable: DoNotSchedule
whenUnsatisfiable: ScheduleAnyway
labelSelector:
matchLabels:
app: awoooi-worker