fix(deploy): migrate 跨 pod 串行化 + rescale 迁移幂等守卫(测试环境×10⁶事故复盘)

事故:连推 3 commit 触发 3 轮 rolling 部署,多个新 pod 并发跑 migrate 且崩溃重跑,
积分 ×10 rescale 被交错重放——ai.0025 执行 6 次(单价/任务计价 ×10⁶),accounts.0008
执行 2 次(限额 ×100),billing.0003/0004 从未完成(django_migrations 漏记录)。
测试库数据已按精确倍率手工修复并补记迁移记录(备份于本机)。

两层防复发:
1. docker-entrypoint 用 MySQL GET_LOCK('airshelf_migrate') 串行化 migrate,
   后到 pod 等锁,拿到时迁移已被记录 → 自然 no-op;
2. 三个 rescale 迁移加 airshelf_rescale_marker 幂等标记(与数据变更同事务提交):
   记录丢失/崩溃重跑时,标记在 → 跳过,不会重复 ×10。

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

core/backend/docker-entrypoint.sh

@@ -3,10 +3,39 @@ set -e
 
 # Only the web (gunicorn) container should run migrations / collectstatic.
 # The celery worker shares this image but skips DB schema mutation to avoid races.
+#
+# ⚠️ migrate 必须跨 pod 串行化(MySQL GET_LOCK):2026-07-07 事故——连推 3 个 commit 触发
+# 3 轮 rolling 部署,多个新 pod 并发跑 migrate,数据迁移(积分 ×10 rescale)被交错重放 6 次,
+# 测试库单价被乘成 ×10⁶(¥2 → 2,000,000)。锁把并发压成串行;后到者拿到锁时迁移已被
+# 先到者记录,migrate 自然 no-op。锁超时 600s 拿不到 → 快速失败重启,绝不裸跑。
 case "$1" in
   gunicorn)
-    echo "[entrypoint] running migrations..."
-    python manage.py migrate --noinput
+    echo "[entrypoint] running migrations (serialized via DB advisory lock)..."
+    python - <<'PYEOF'
+import os
+
+import django
+
+os.environ.setdefault("DJANGO_SETTINGS_MODULE", "airshelf.settings.production")
+django.setup()
+
+from django.core.management import call_command
+from django.db import connection
+
+if connection.vendor == "mysql":
+    with connection.cursor() as cursor:
+        cursor.execute("SELECT GET_LOCK('airshelf_migrate', 600)")
+        acquired = cursor.fetchone()[0]
+    if not acquired:
+        raise SystemExit("[entrypoint] FATAL: migrate advisory lock timeout (another pod stuck?)")
+    try:
+        call_command("migrate", interactive=False)
+    finally:
+        with connection.cursor() as cursor:
+            cursor.execute("SELECT RELEASE_LOCK('airshelf_migrate')")
+else:
+    call_command("migrate", interactive=False)
+PYEOF
     echo "[entrypoint] collecting static..."
     python manage.py collectstatic --noinput
     ;;