fix(monitoring): stabilize post-deploy target coverage

scripts/generate_monitoring.py

@@ -16,11 +16,14 @@ Phase O-5 Wave C.1 (2026-04-02 ogt)
 
 import argparse
 import json
+import os
 import subprocess
 import sys
+import time
 from datetime import datetime
-
-import requests
+from typing import Callable
+from urllib.error import HTTPError, URLError
+from urllib.request import urlopen
 
 # ============================================================
 # 設定
@@ -28,6 +31,8 @@ import requests
 
 PROMETHEUS_URL = "http://192.168.0.110:9090"
 COVERAGE_THRESHOLD = 70  # CI 模式: 覆蓋率低於此值則 exit 1
+DEFAULT_STABILIZATION_ATTEMPTS = 3
+DEFAULT_STABILIZATION_SLEEP_SECONDS = 10.0
 
 # 已知服務清單 (job名稱 → 說明)
 EXPECTED_JOBS = {
@@ -52,13 +57,27 @@ KNOWN_DOWN_TARGETS = {
 }
 
 
+def _int_env(name: str, default: int) -> int:
+    try:
+        return max(1, int(os.environ.get(name, default)))
+    except ValueError:
+        return default
+
+
+def _float_env(name: str, default: float) -> float:
+    try:
+        return max(0.0, float(os.environ.get(name, default)))
+    except ValueError:
+        return default
+
+
 def get_prometheus_targets() -> dict:
     """查詢 Prometheus targets API"""
     try:
-        resp = requests.get(f"{PROMETHEUS_URL}/api/v1/targets", timeout=10)
-        resp.raise_for_status()
-        return resp.json()["data"]
-    except requests.RequestException as e:
+        with urlopen(f"{PROMETHEUS_URL}/api/v1/targets", timeout=10) as resp:
+            payload = json.loads(resp.read().decode("utf-8"))
+        return payload["data"]
+    except (HTTPError, URLError, TimeoutError, json.JSONDecodeError, KeyError) as e:
         print(f"❌ 無法連接 Prometheus ({PROMETHEUS_URL}): {e}", file=sys.stderr)
         sys.exit(1)
 
@@ -133,6 +152,76 @@ def build_report(jobs: dict) -> dict:
     }
 
 
+def build_report_from_targets(targets_data: dict) -> dict:
+    """從 Prometheus targets API payload 建立覆蓋率報告"""
+    return build_report(analyze_targets(targets_data))
+
+
+def report_needs_stabilization(report: dict) -> bool:
+    """是否需要重查，避免 post-deploy 瞬間 scrape 狀態造成 false red."""
+    return bool(report["real_down_jobs"] or report["missing_expected"])
+
+
+def stabilization_reason(report: dict) -> str:
+    parts: list[str] = []
+    if report["real_down_jobs"]:
+        parts.append(f"real_down={','.join(report['real_down_jobs'])}")
+    if report["missing_expected"]:
+        parts.append(f"missing_expected={','.join(report['missing_expected'])}")
+    return "; ".join(parts) if parts else "stable"
+
+
+def build_stabilized_report(
+    fetch_targets: Callable[[], dict],
+    attempts: int,
+    sleep_seconds: float,
+    emit_status: bool = True,
+) -> dict:
+    """重查 Prometheus targets，讓 CI gate 避開 rollout/scrape freshness 瞬間值."""
+    attempts = max(1, attempts)
+    sleep_seconds = max(0.0, sleep_seconds)
+
+    report: dict | None = None
+    for attempt in range(1, attempts + 1):
+        report = build_report_from_targets(fetch_targets())
+        needs_retry = report_needs_stabilization(report)
+        status = "stable"
+        if needs_retry and attempt < attempts:
+            status = "retrying"
+        elif needs_retry:
+            status = "failed"
+        elif attempt > 1:
+            status = "cleared"
+
+        report["stabilization"] = {
+            "attempt": attempt,
+            "attempts": attempts,
+            "sleep_seconds": sleep_seconds,
+            "status": status,
+            "reason": stabilization_reason(report),
+        }
+
+        if not needs_retry or attempt == attempts:
+            if emit_status and attempt > 1 and not needs_retry:
+                print(
+                    "✅ Prometheus target stabilization cleared transient coverage drift",
+                    file=sys.stderr,
+                )
+            return report
+
+        if emit_status:
+            print(
+                "⏳ Prometheus target stabilization "
+                f"{attempt}/{attempts}: {stabilization_reason(report)}",
+                file=sys.stderr,
+            )
+        time.sleep(sleep_seconds)
+
+    if report is None:
+        raise RuntimeError("monitoring report stabilization did not run")
+    return report
+
+
 def print_human_report(report: dict) -> None:
     """輸出人可讀格式報告"""
     s = report["summary"]
@@ -161,11 +250,15 @@ def print_human_report(report: dict) -> None:
             status = "❌ DOWN"
         print(f"  {status:<30} {job:<25} {desc}")
 
-    if report["known_down"]:
+    known_down_present = [
+        (job, reason)
+        for job, reason in report["known_down"].items()
+        if job in report["jobs"] and report["jobs"][job]["down"]
+    ]
+    if known_down_present:
         print(f"\n⚠️  已知 DOWN (不影響覆蓋率)")
-        for job, reason in report["known_down"].items():
-            if job in report["jobs"]:
-                print(f"  {job:<30} {reason}")
+        for job, reason in known_down_present:
+            print(f"  {job:<30} {reason}")
 
     if report["real_down_jobs"]:
         print(f"\n🔴 需處理的 DOWN targets")
@@ -178,6 +271,15 @@ def print_human_report(report: dict) -> None:
         for job in report["missing_expected"]:
             print(f"  {job}: {report['expected_jobs'][job]}")
 
+    stabilization = report.get("stabilization")
+    if stabilization and stabilization["attempt"] > 1:
+        print(f"\n⏱️  Prometheus target 穩定化")
+        print(
+            "  "
+            f"{stabilization['status']} after "
+            f"{stabilization['attempt']}/{stabilization['attempts']} attempts"
+        )
+
     pct = s["expected_coverage_pct"]
     threshold = COVERAGE_THRESHOLD
     if pct >= threshold and not report["real_down_jobs"]:
@@ -196,11 +298,34 @@ def main() -> None:
         action="store_true",
         help=f"CI 模式: 覆蓋率 < {COVERAGE_THRESHOLD}% 則 exit 1",
     )
+    parser.add_argument(
+        "--stabilization-attempts",
+        type=int,
+        default=_int_env(
+            "AWOOOI_MONITORING_TARGET_STABILIZATION_ATTEMPTS",
+            DEFAULT_STABILIZATION_ATTEMPTS,
+        ),
+        help="CI 模式: Prometheus target 狀態重查次數",
+    )
+    parser.add_argument(
+        "--stabilization-sleep-seconds",
+        type=float,
+        default=_float_env(
+            "AWOOOI_MONITORING_TARGET_STABILIZATION_SLEEP_SECONDS",
+            DEFAULT_STABILIZATION_SLEEP_SECONDS,
+        ),
+        help="CI 模式: Prometheus target 重查間隔秒數",
+    )
     args = parser.parse_args()
 
-    targets_data = get_prometheus_targets()
-    jobs = analyze_targets(targets_data)
-    report = build_report(jobs)
+    if args.check:
+        report = build_stabilized_report(
+            get_prometheus_targets,
+            attempts=args.stabilization_attempts,
+            sleep_seconds=args.stabilization_sleep_seconds,
+        )
+    else:
+        report = build_report_from_targets(get_prometheus_targets())
 
     if args.json:
         print(json.dumps(report, ensure_ascii=False, indent=2))