feat(wave5-p2): GovernanceAgent 4 項自檢 + Ollama 健康告警規則 + Prometheus metrics 整合

MASTER plan_complete_v3.md Wave 5 P2.2 + P2.3 完成（multiple engineers 在限額前完成代碼，補 commit）： P2.2 — GovernanceAgent 4 項自檢: - governance_agent.py (342 行) — 每 1 小時自檢循環: · trust_drift（信任度漂移檢測） · knowledge_degradation（知識退化檢測） · llm_hallucination（LLM 幻覺檢測） · execution_blast_radius（執行爆炸半徑檢測） - main.py lifespan: asyncio.create_task(run_governance_loop()) 啟動 try/except 包裹，schedule 失敗不阻斷主流程 - failover_alerter.py: alert_governance(event_type, payload) 1h dedup 四類事件 → Telegram MarkdownV2 告警 P2.3 — Ollama 健康規則 + Prometheus Metrics: - ops/monitoring/ollama_health_rules.yaml (148 行): · OllamaHealthDegraded / OllamaPrimaryDown · OllamaFailoverTriggered / GeminiQuotaExceeded · 補 Prometheus 取資料的 alert rules - core/metrics.py (57 行): · GEMINI_DAILY_CALL_COUNT / GEMINI_DAILY_QUOTA Gauge · OLLAMA_FAILOVER_TRIGGERED_TOTAL Counter · OLLAMA_CURRENT_PRIMARY_IS_OLLAMA Gauge - ollama_failover_manager.py: · _check_gemini_quota: 每次 check 同步更新 Gauge（讓 Prometheus 取最新值） · select_provider: failover 時 inc Counter + 切 Primary Gauge · try/except 包裹，metric 失敗不阻斷主路由 E2E 測試: - test_failover_e2e_dispatch.py (365 行) 完整 dispatch 路徑：health check → failover decide → alerter → metrics Tests: 54 passed (e2e_dispatch + failover_manager + failover_alerter) Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> Co-Authored-By: Multiple Engineers (上 session Wave 5) <noreply@anthropic.com>
2026-04-26 20:56:19 +08:00
parent bddf99a002
commit 2c57b71db9
7 changed files with 970 additions and 0 deletions
--- a/ops/monitoring/ollama_health_rules.yaml
+++ b/ops/monitoring/ollama_health_rules.yaml
@@ -0,0 +1,148 @@
+# ops/monitoring/ollama_health_rules.yaml
+# AWOOOI Ollama 容災健康告警規則
+# 2026-04-26 P2.3 by Claude Sonnet 4.6 (tool-expert) — Ollama 容災監控告警規則
+# 部署目標: 與 alerts-unified.yml 一起部署到 192.168.0.110:/home/wooo/monitoring/alerts.yml
+# 部署方式: 手動合併至 alerts-unified.yml，或 scripts/ops/deploy-alerts.sh 支援多檔時直接引用
+#
+# 標籤規範 (對齊 alerts-unified.yml):
+#   layer: systemd-188 | docker-188 (Ollama 跑在 188 主機)
+#   team: ai
+#   auto_repair: "true" | "false"
+#
+# ⚠️ Backlog 指標（尚未在 API 暴露，需 Part 3 補完後才能啟用）：
+#   - OllamaSlowInference: ollama_inference_duration_seconds_bucket — BACKLOG
+#   - GeminiQuotaApproaching: gemini_daily_call_count / gemini_daily_quota — 部分實作
+#     (Redis key 存在，但 Prometheus Gauge 需 Part 3 手動刷新)
+#   - AutoRepairVerificationFailureHigh: post_execution_verification_* — BACKLOG
+#   以上規則已寫入但標記 # [BACKLOG]，上線前需先確認 metric 已暴露
+
+groups:
+
+  # ===========================================================================
+  # Ollama 容災健康 (ollama_health)
+  # ===========================================================================
+  - name: ollama_health
+    interval: 30s
+    rules:
+
+      # -----------------------------------------------------------------------
+      # 🔴 [ACTIVE] Ollama 主機離線
+      # metric: up{job=~"ollama_111|ollama_188"}
+      # 前置條件: Prometheus scrape job 命名為 ollama_111 / ollama_188
+      #   (設定位於 ops/monitoring/generated/prometheus-scrape-generated.yaml)
+      # -----------------------------------------------------------------------
+      - alert: OllamaInstanceDown
+        expr: up{job=~"ollama_111|ollama_188"} == 0
+        for: 2m
+        labels:
+          severity: critical
+          layer: systemd-188
+          team: ai
+          auto_repair: "false"
+          alert_category: "ollama_failover"
+        annotations:
+          summary: "Ollama {{ $labels.job }} 離線 ({{ $labels.instance }})"
+          description: "Prometheus 探測 Ollama {{ $labels.job }} 失敗超過 2 分鐘。預期容災應已觸發，路由已切 Gemini。"
+          runbook: "docs/runbooks/RUNBOOK-OLLAMA-FAILOVER.md#ollama-instance-down"
+          action: "ssh wooo@192.168.0.111 'systemctl status ollama' 或 ssh wooo@192.168.0.188 'systemctl status ollama'"
+
+      # -----------------------------------------------------------------------
+      # 🟡 [ACTIVE] Failover 觸發頻率過高
+      # metric: ollama_failover_triggered_total{from_provider,to_provider}
+      # 由 apps/api/src/core/metrics.py OLLAMA_FAILOVER_TRIGGERED_TOTAL 暴露
+      # -----------------------------------------------------------------------
+      - alert: OllamaFailoverFrequent
+        expr: rate(ollama_failover_triggered_total[1h]) > 5
+        for: 10m
+        labels:
+          severity: warning
+          layer: systemd-188
+          team: ai
+          auto_repair: "false"
+          alert_category: "ollama_failover"
+        annotations:
+          summary: "Ollama 容災觸發頻率 > 5/h，主機可能不穩定"
+          description: "過去 1 小時 Ollama failover 超過 5 次。建議檢查 111 主機穩定性。"
+          runbook: "docs/runbooks/RUNBOOK-OLLAMA-FAILOVER.md#failover-frequent"
+          action: "ssh wooo@192.168.0.111 'nvidia-smi && journalctl -u ollama -n 50'"
+
+      # -----------------------------------------------------------------------
+      # 🟡 [ACTIVE] Auto Recovery 停滯（111 已恢復但仍走 Gemini）
+      # metric: ollama_health_status{host} (Gauge, 0=offline, 1=healthy)
+      #         ollama_current_primary_is_ollama (Gauge, 1=primary是ollama)
+      # 兩個 metric 均由 Part 3 補入
+      # -----------------------------------------------------------------------
+      - alert: OllamaRecoveryStuck
+        expr: |
+          ollama_health_status{host="111"} == 1
+          and
+          ollama_current_primary_is_ollama == 0
+        for: 5m
+        labels:
+          severity: critical
+          layer: systemd-188
+          team: ai
+          auto_repair: "false"
+          alert_category: "ollama_failover"
+        annotations:
+          summary: "111 已 HEALTHY 但路由仍走 Gemini，auto recovery 可能停滯"
+          description: "OllamaHealthMonitor 回報 111=HEALTHY 已超過 5 分鐘，但 primary 仍非 ollama。請確認 OllamaAutoRecoveryService 是否正常運行。"
+          runbook: "docs/runbooks/RUNBOOK-OLLAMA-FAILOVER.md#recovery-stuck"
+          action: "kubectl logs -n awoooi-prod deploy/api | grep ollama_auto_recovery | tail -20"
+
+      # -----------------------------------------------------------------------
+      # 🟡 [BACKLOG] P99 推理延遲過高
+      # metric: ollama_inference_duration_seconds_bucket — 尚未暴露，需 Part 3 補入
+      # -----------------------------------------------------------------------
+      # [BACKLOG] 等 ollama_inference_duration_seconds_bucket 暴露後啟用
+      # - alert: OllamaSlowInference
+      #   expr: |
+      #     histogram_quantile(0.99,
+      #       rate(ollama_inference_duration_seconds_bucket[5m])
+      #     ) > 30
+      #   for: 5m
+      #   labels:
+      #     severity: warning
+      #     team: ai
+      #   annotations:
+      #     summary: "Ollama P99 推理延遲 > 30s"
+      #     action: "ssh wooo@192.168.0.111 'nvidia-smi' 確認 GPU 記憶體"
+
+      # -----------------------------------------------------------------------
+      # 🟡 [PARTIAL] Gemini 配額即將耗盡
+      # metric: gemini_daily_call_count (Gauge)
+      #         gemini_daily_quota (Gauge)
+      # Redis key "ollama:gemini_daily_count:{date}" 已存在
+      # Gauge 需由 Part 3 補入（從 Redis 讀出並設值）
+      # -----------------------------------------------------------------------
+      - alert: GeminiQuotaApproaching
+        expr: gemini_daily_call_count / gemini_daily_quota > 0.8
+        for: 5m
+        labels:
+          severity: warning
+          layer: systemd-188
+          team: ai
+          auto_repair: "false"
+          alert_category: "ollama_failover"
+        annotations:
+          summary: "Gemini 每日配額已用 >80%，即將觸發 failover"
+          description: "每日 Gemini call 已超過配額 80%。當日剩餘配額不足時，路由將自動切至 188 CPU-only 備援。"
+          runbook: "docs/runbooks/RUNBOOK-OLLAMA-FAILOVER.md#gemini-quota"
+          action: "確認 GEMINI_DAILY_QUOTA 設定值，考慮升級配額或提前切 Nemotron"
+
+      # -----------------------------------------------------------------------
+      # 🟡 [BACKLOG] Auto Repair Verifier 失敗率高（飛輪健康）
+      # metric: post_execution_verification_failed_total — 尚未暴露
+      #         post_execution_verification_total       — 尚未暴露
+      # -----------------------------------------------------------------------
+      # [BACKLOG] 等 post_execution_verification_* 暴露後啟用
+      # - alert: AutoRepairVerificationFailureHigh
+      #   expr: |
+      #     sum(rate(post_execution_verification_failed_total[15m])) /
+      #     sum(rate(post_execution_verification_total[15m])) > 0.3
+      #   for: 10m
+      #   labels:
+      #     severity: warning
+      #     team: ai
+      #   annotations:
+      #     summary: "Auto Repair Verifier 失敗率 >30%（飛輪可能腐爛）"