feat(trading): 浏览器掉线兜底——任务边界自愈重建,中途掉线转 needs_human

SiteInteractor 的 Chromium 是进程级单例,此前启动后即假定永远活着:全仓唯一的
is_connected() 探活在 scraping 侧,交易侧既不探活也不重启。容器里 Chromium 崩溃
是有真实前提的(/dev/shm 不足、OOM kill、seccomp 挡 sandbox,docker-compose.yml
里已有相关注释),一旦发生,进程还活着但之后每一单都会失败,且 /health 恒返回
ok,restart: unless-stopped 永远不会被触发。

更隐蔽的一条:clear_cart / enter_checkout / pay 等处的 new_page()、context.request
写在 try 之外,掉线抛的 TargetClosedError 不是 AppError,会穿过 runner 的
except AppError 落到主循环那个只记日志的兜底里——任务一次都不上报,网关侧要干等
整个 lease_ttl(默认 300s)才被 sweep 置 stale。clear_cart 是 execute() 的 step 0,
浏览器死时最先撞上的正是它。

分三层处理:

- 任务边界自愈。_launch() 从 start() 抽出,_context_options() 统一 context 参数
  (重建必须与启动完全一致,指纹漂移就是一次风控事件);_refresh_context_if_stale
  改名 _ensure_context_ready(),先探活重建再做原有的 storage_state mtime 检查
  (顺序不可换,mtime 重建要用 self._browser)。重建前丢弃 _checkout_pages 里的
  残留确认页并记 warning——那些 Page 已随浏览器一起没了。

- 中途掉线不重建,抛 BrowserDeadError(新增,5006)。新增 _new_page() 与
  _request() 两个壳收口裸异常;_request() 只在确认浏览器真死了时才改写异常,
  站点 5xx 这类正常业务失败原样抛出。submit_order / pay 入口用
  _require_live_browser() 直接拒绝:这两步复用 enter_checkout 留存的 Page,
  重建救不回服务端订单草稿,而 pay 跑的时候订单已经真的提交了。顺带修掉一个
  误诊——浏览器死时 submit_order 原先报「未找到确认按钮」,把「浏览器崩了」
  说成「站点改版了」,两者的处置方式完全不同。

- runner 把 BrowserDeadError 转 needs_human 而非 failed,except 分支排在
  except AppError 之前(子类,顺序反了就报 failed)。掉线发生在动作中途,
  站点侧生效与否无从判断,不能给上游「明确失败」的结论。

/health 暴露 browser 状态,掉线时 degraded + HTTP 503 + code 5006,让
Dockerfile.trading 的 HEALTHCHECK 探到并重启容器。重启不会导致重复下单:网关侧
任务绝不自动重投,租约过期只置 stale 等人工 reclaim(docs/order-gateway.md §5),
重启只是恢复领新任务的能力。启动窗口期 started=False 不算掉线。

README 错误码表补 5005(此前遗漏)与 5006。

新增 15 个用例覆盖探活三态、is_connected() 自身抛错、边界重建/不重建/丢弃残留页/
重建失败、两个包装壳的分支、submit/pay 拒绝、runner 转 needs_human、/health 503。
真实 Chromium 崩溃无法在离线测试里制造,用例模拟的是 is_connected() 返回 False
这个唯一可观测信号,覆盖的是代码对该信号的反应而非崩溃本身;容器 HEALTHCHECK
真的触发重启这条链路尚未实跑验证。

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
This commit is contained in:
2026-08-20 16:01:47 +08:00
co-authored by Claude Opus 5
parent 6b6b243132
commit 3284e086ad
10 changed files with 633 additions and 56 deletions
+230 -2
View File
@@ -20,6 +20,7 @@ import pytest
from datetime import datetime, timezone
from app.shared.errors import (
BrowserDeadError,
CartOperationError,
InvalidRequestError,
NotLoggedInError,
@@ -576,10 +577,24 @@ _ORDER_DETAIL_LANDED_URL = "https://order.my.rakuten.co.jp/purchase-history/?ord
_ORDER_LIST_LANDED_URL = "https://order.my.rakuten.co.jp/purchase-history/order-list"
class _FakeBrowser:
"""Browser 替身:只实现 _ensure_browser_alive 用到的 is_connected()
没有它,`_ensure_context_ready` 会判定「浏览器没起来」并尝试真的 launch 一次
Chromium。connected 可以翻成 False 来模拟掉线。
"""
def __init__(self, connected: bool = True):
self.connected = connected
def is_connected(self) -> bool:
return self.connected
def _build_site(tmp_path: Path, context: _FakeContext, auth: _FakeAuthSession) -> SiteInteractor:
"""装配一个不依赖 Playwright 的 SiteInteractor
auth_state_dir 指向空目录:_refresh_context_if_stale 见不到 storage_state
auth_state_dir 指向空目录:_ensure_context_ready 见不到 storage_state
文件就直接返回,不会试图重建 context(重建需要真实 browser)。
"""
from app.shared.config import Settings
@@ -588,6 +603,7 @@ def _build_site(tmp_path: Path, context: _FakeContext, auth: _FakeAuthSession) -
auth_session=auth, # type: ignore[arg-type]
settings=Settings(auth_state_dir=str(tmp_path / "auth"), evidence_dir=str(tmp_path)),
)
site._browser = _FakeBrowser() # type: ignore[assignment]
site._context = context # type: ignore[assignment]
return site
@@ -980,13 +996,16 @@ async def test_check_order_status_still_returns_only_status(tmp_path):
async def test_submit_order_without_checkout_page_raises():
"""浏览器活着但没有留存确认页 → OrderOperationError(不是掉线那类错误)"""
site = SiteInteractor(auth_session=None, settings=None) # type: ignore[arg-type]
site._browser = _FakeBrowser() # type: ignore[assignment]
with pytest.raises(OrderOperationError):
await site.submit_order(_make_task())
async def test_pay_without_checkout_page_raises():
site = SiteInteractor(auth_session=None, settings=None) # type: ignore[arg-type]
site._browser = _FakeBrowser() # type: ignore[assignment]
with pytest.raises(OrderOperationError):
await site.pay(_make_task(), "ord-1")
@@ -1314,15 +1333,17 @@ class _FakeLoggedInAuth:
def _build_clear_cart_site(*, delete_buttons: int, count_body: str, html: str) -> SiteInteractor:
"""构造 clear_cart 可离线跑起来的 SiteInteractor(fake page + fake count API)"""
site = SiteInteractor(auth_session=_FakeLoggedInAuth(), settings=None) # type: ignore[arg-type]
site._browser = _FakeBrowser() # type: ignore[assignment]
site._context = _FakeClearCartContext(
page=_FakeClearCartPage(delete_buttons=delete_buttons, html=html),
count_body=count_body,
) # type: ignore[assignment]
# settings=None,走不了真实的 mtime 检查;探活那半段由 _FakeBrowser 覆盖
async def _noop_refresh() -> None:
return None
site._refresh_context_if_stale = _noop_refresh # type: ignore[assignment]
site._ensure_context_ready = _noop_refresh # type: ignore[assignment]
return site
@@ -1390,3 +1411,210 @@ async def test_clear_cart_html_capture_failure_keeps_clear_result():
# ---- helper ----
# ---- 浏览器掉线:任务边界就地重建,中途掉线转 BrowserDeadError ----
#
# 真实的 Chromium 崩溃(容器 /dev/shm 不足、OOM kill、seccomp 挡 sandbox)没法在
# 离线测试里制造,这里用 _FakeBrowser.connected 翻成 False 模拟「is_connected()
# 返回 False」这个唯一的可观测信号,覆盖的是 SiteInteractor 对该信号的反应。
def _build_dead_browser_site(tmp_path: Path) -> SiteInteractor:
"""浏览器已掉线的 SiteInteractor:_browser 在位但 is_connected() 为 False"""
from app.shared.config import Settings
site = SiteInteractor(
auth_session=_FakeLoggedInAuth(), # type: ignore[arg-type]
settings=Settings(auth_state_dir=str(tmp_path / "auth"), evidence_dir=str(tmp_path)),
)
site._browser = _FakeBrowser(connected=False) # type: ignore[assignment]
return site
def test_browser_alive_reflects_is_connected(tmp_path):
"""browser_alive:没起过 / 已掉线 / 正常,三种情况分别为 False/False/True"""
site = SiteInteractor(auth_session=None, settings=None) # type: ignore[arg-type]
assert site.browser_alive is False # 还没 start()
site._browser = _FakeBrowser(connected=False) # type: ignore[assignment]
assert site.browser_alive is False
site._browser = _FakeBrowser(connected=True) # type: ignore[assignment]
assert site.browser_alive is True
def test_browser_alive_survives_is_connected_raising():
"""is_connected() 自己抛错(驱动已经没了)也要当掉线处理,不能把异常漏出去"""
class _ExplodingBrowser:
def is_connected(self) -> bool:
raise RuntimeError("driver connection closed")
site = SiteInteractor(auth_session=None, settings=None) # type: ignore[arg-type]
site._browser = _ExplodingBrowser() # type: ignore[assignment]
assert site.browser_alive is False
def test_browser_status_distinguishes_not_started_from_dead():
"""/health 要能分清「还没启动」和「起过但死了」——只有后者算 degraded"""
site = SiteInteractor(auth_session=None, settings=None) # type: ignore[arg-type]
not_started = site.browser_status()
assert not_started["started"] is False
assert not_started["alive"] is False
site._browser = _FakeBrowser(connected=False) # type: ignore[assignment]
dead = site.browser_status()
assert dead["started"] is True
assert dead["alive"] is False
site._browser = _FakeBrowser(connected=True) # type: ignore[assignment]
assert site.browser_status() == {
"started": True,
"alive": True,
"pending_checkout_tasks": [],
"detail": "连接正常",
}
async def test_ensure_browser_alive_relaunches_on_task_boundary(tmp_path):
"""任务边界探到掉线 → 重放一遍启动流程,调用方无感知"""
site = _build_dead_browser_site(tmp_path)
relaunched = []
async def _fake_launch() -> None:
relaunched.append(True)
site._browser = _FakeBrowser(connected=True) # type: ignore[assignment]
site._launch = _fake_launch # type: ignore[assignment]
await site._ensure_browser_alive()
assert relaunched == [True]
assert site.browser_alive is True
async def test_ensure_browser_alive_is_noop_when_connected(tmp_path):
"""浏览器好好的就不能重建——重建一次要几秒,还会丢掉当前 context 的 cookie"""
site = _build_dead_browser_site(tmp_path)
site._browser = _FakeBrowser(connected=True) # type: ignore[assignment]
async def _fail_launch() -> None:
raise AssertionError("浏览器连接正常时不应重建")
site._launch = _fail_launch # type: ignore[assignment]
await site._ensure_browser_alive()
async def test_ensure_browser_alive_drops_orphaned_checkout_pages(tmp_path):
"""重建前丢掉残留的确认页:那些 Page 已随浏览器一起没了,留着必然报错"""
site = _build_dead_browser_site(tmp_path)
site._checkout_pages["t-stuck"] = object() # type: ignore[assignment]
async def _fake_launch() -> None:
site._browser = _FakeBrowser(connected=True) # type: ignore[assignment]
site._launch = _fake_launch # type: ignore[assignment]
await site._ensure_browser_alive()
assert site._checkout_pages == {}
async def test_ensure_browser_alive_raises_browser_dead_when_relaunch_fails(tmp_path):
"""重建也起不来 → BrowserDeadError(AppError,runner 接得住),并把半成品收干净"""
site = _build_dead_browser_site(tmp_path)
async def _broken_launch() -> None:
raise RuntimeError("chromium 起不来")
site._launch = _broken_launch # type: ignore[assignment]
with pytest.raises(BrowserDeadError):
await site._ensure_browser_alive()
assert site._browser is None # 已 teardown,下次调用从头再试
async def test_new_page_wraps_target_closed_into_app_error():
"""中途掉线的裸 TargetClosedError 必须被包成 AppError
这是整条链最关键的一环:不包的话异常会穿过 runner 的 except AppError,
落到主循环那个只记日志的兜底里——任务一次都不上报,网关要干等租约过期。
"""
class _DeadContext:
async def new_page(self):
raise RuntimeError("Target page, context or browser has been closed")
site = SiteInteractor(auth_session=None, settings=None) # type: ignore[arg-type]
site._context = _DeadContext() # type: ignore[assignment]
with pytest.raises(BrowserDeadError) as excinfo:
await site._new_page()
from app.shared.errors import AppError
assert isinstance(excinfo.value, AppError)
assert excinfo.value.err_code == 5006
async def test_request_keeps_original_error_when_browser_is_alive():
"""浏览器活着时的请求失败是正常业务失败,不能被误标成掉线"""
class _FlakyRequest:
async def get(self, url: str, **kwargs):
raise RuntimeError("站点 502")
class _Context:
request = _FlakyRequest()
site = SiteInteractor(auth_session=None, settings=None) # type: ignore[arg-type]
site._browser = _FakeBrowser(connected=True) # type: ignore[assignment]
site._context = _Context() # type: ignore[assignment]
with pytest.raises(RuntimeError, match="站点 502"):
await site._request("get", "https://example.test/")
async def test_request_converts_to_browser_dead_when_browser_is_gone():
"""同样的请求失败,浏览器确实没了时才改写成 BrowserDeadError"""
class _FlakyRequest:
async def get(self, url: str, **kwargs):
raise RuntimeError("Target closed")
class _Context:
request = _FlakyRequest()
site = SiteInteractor(auth_session=None, settings=None) # type: ignore[arg-type]
site._browser = _FakeBrowser(connected=False) # type: ignore[assignment]
site._context = _Context() # type: ignore[assignment]
with pytest.raises(BrowserDeadError):
await site._request("get", "https://example.test/")
async def test_submit_order_reports_browser_dead_before_blaming_selectors(tmp_path):
"""浏览器死时 submit_order 要报掉线,而不是「找不到确认按钮」
留存的 Page 是个死对象(不是 None),不先探活的话 selector 轮询会全部超时,
最后把「浏览器崩了」误诊成「站点改版了」——这两个结论的处置方式完全不同。
"""
site = _build_dead_browser_site(tmp_path)
site._checkout_pages["t1"] = object() # type: ignore[assignment]
with pytest.raises(BrowserDeadError):
await site.submit_order(_make_task(task_id="t1"))
async def test_pay_reports_browser_dead(tmp_path):
"""pay 之前 submit_order 已经真的下过单了,掉线必须当场抛出去交人工"""
site = _build_dead_browser_site(tmp_path)
site._checkout_pages["t1"] = object() # type: ignore[assignment]
with pytest.raises(BrowserDeadError):
await site.pay(_make_task(task_id="t1"), "ord-1")
+30
View File
@@ -85,6 +85,7 @@ class StubSiteInteractor:
self.clear_calls = 0
self.remove_calls: list[str] = []
self.fail_with: Exception | None = None
self.browser_alive = True
async def add_to_cart_payload(
self,
@@ -129,6 +130,15 @@ class StubSiteInteractor:
raise self.fail_with
return {"removed": True, "item_id": item_id}
def browser_status(self) -> dict:
"""/health 读的浏览器状态;browser_alive 可翻成 False 模拟掉线"""
return {
"started": True,
"alive": self.browser_alive,
"pending_checkout_tasks": [],
"detail": "连接正常" if self.browser_alive else "浏览器已掉线",
}
async def close(self) -> None:
"""lifespan 收尾会调用;桩没有真实浏览器要关"""
@@ -168,6 +178,26 @@ def test_health_needs_no_token(client):
body = response.json()
assert body["data"]["status"] == "ok"
assert set(body["data"]["auth"]) == {"rakuten"}
assert body["data"]["browser"]["alive"] is True
def test_health_returns_503_when_browser_is_dead(client, stub_site):
"""浏览器掉线 → 503,让容器 HEALTHCHECK 探到并触发重启
进程还活着、端口还通,但这个服务的所有站点操作都要靠那一个浏览器,它没了以后
每一单都会失败。返回 200 的话 HEALTHCHECK 永远绿灯,缺口就一直挂在那儿。
重启是安全的:网关侧任务绝不自动重投(docs/order-gateway.md §5)。
"""
stub_site.browser_alive = False
response = client.get("/health")
assert response.status_code == 503
body = response.json()
assert body["success"] is False
assert body["code"] == 5006 # 与 BrowserDeadError 同码
assert body["data"]["status"] == "degraded"
assert body["data"]["browser"]["alive"] is False
def test_health_does_not_probe_the_site(client, stub):
+36 -1
View File
@@ -20,7 +20,7 @@ from typing import Any
import pytest
from app.shared.errors import CheckoutBlockedError, OrderGuardError
from app.shared.errors import BrowserDeadError, CheckoutBlockedError, OrderGuardError
from app.shared.task_state import OrderState, TaskStatus
from app.trading.worker import verify
from app.trading.worker.evidence import EvidenceStore
@@ -352,6 +352,41 @@ async def test_checkout_blocked_becomes_needs_human(
assert "风控" in terminal["detail"]
# ---- 浏览器掉线 → needs_human(不是 failed)----
async def test_browser_dead_becomes_needs_human(runner: WorkerRunner):
"""执行途中浏览器掉线 → BrowserDeadError → 转 needs_human
两件事同时被这个用例守着:
1. BrowserDeadError 是 AppError 的子类,能被 _execute_with_renewal 接住并**上报**。
不是 AppError 的话它会一路飘到主循环的兜底日志里,任务一次都不上报,网关侧
要干等整个 lease_ttl 才被 sweep 置 stale。
2. 结论是 needs_human 而不是 failed——`except BrowserDeadError` 必须排在
`except AppError` 前面,顺序反了这个断言就会挂。浏览器是在动作中途没的,
站点侧生效与否无从判断,不能给上游一个「明确失败」的结论。
"""
async def _noop(task): # noqa: ANN001
return None
async def _browser_died(task): # noqa: ANN001
raise BrowserDeadError("pay:浏览器已掉线,本单是否已提交无法判断,需人工核对订单列表")
runner._site.add_to_cart = _noop # type: ignore[assignment]
runner._site.verify_cart = _noop # type: ignore[assignment]
runner._site.enter_checkout = _browser_died # type: ignore[assignment]
_set_site_clear(runner._site, {"removed_count": 0, "cart_count": 0})
await runner.handle(_make_task(task_id="t1"))
gateway: FakeGateway = runner._gateway_for_test # type: ignore[attr-defined]
terminal = gateway.last_terminal_report()
assert terminal["terminal_status"] == TaskStatus.NEEDS_HUMAN
assert "掉线" in terminal["detail"]
# ---- ラクマ 不在交易范围 → needs_human ----