Files
rakuten-api/tests/test_worker_client_telemetry.py
q792602257andClaude Opus 5 8403b9b586 test(observability): 补抓取客户端与网关客户端的埋点测试
上一提交(8381896)的配套测试,钉住两个失败可见性问题:

- 抓取侧回归:失败分支原先写的是无参 `span.record_exception()`,而 exception
  是必填位置参数——每次抓取失败都会在记异常时抛 TypeError,把真正的原因(反爬
  阻断 / 解析失败)整个顶替掉,失败页面快照也永远落不下来(抛错在 snapshot
  之前)。测试断言业务异常原样抛出、stage 能区分 fetch/parse/delegate、
  parse 阶段确实落了快照且带得上 URL
- 网关侧:success=false code=6002 在 HTTP 层是 200,断言 span 记下了信封结论;
  以及空转长轮询在 suppressed() 里确实不产 span,抑制之外照常产(防止空断言
  是因为埋点根本没生效)
- 两个文件都各带一条成功路径对照,证明 ERROR 状态不是恒真

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
2026-08-28 16:08:35 +08:00

159 lines
6.3 KiB
Python

"""GatewayClient 埋点测试:信封里的失败必须在链路上看得见
要钉住的问题:**网关的失败在响应信封里,不在 HTTP 状态码上**。一次
`success=false, code=6002`(租约无效)的回报,httpx 自动 instrumentation 只看到
一个完成了的请求——而且那个 CLIENT span 在 `request()` 返回时就结束了,此时信封
还没解,所以它永远不可能带上这次调用的结论。链路里于是只剩「调用发生过」。
另一半是抑制:worker 的空转长轮询在 `suppressed()` 里调 `_request`,而
`suppress_instrumentation` 只被 instrumentation 库尊重,手工建的 span 不看它。
手工埋点若不配合 `span_unless_suppressed`,孤立 trace 会从这个口子重新灌回来。
全部用 httpx.MockTransport 拦截,不触达真实网关。
"""
from __future__ import annotations
import httpx
import pytest
from opentelemetry.sdk.trace import TracerProvider
from opentelemetry.sdk.trace.export import SimpleSpanProcessor
from opentelemetry.sdk.trace.export.in_memory_span_exporter import InMemorySpanExporter
from opentelemetry.trace import SpanKind, StatusCode
from app.shared.config import Settings
from app.shared.errors import AppError
from app.shared.task_state import OrderState
from app.shared.telemetry import suppressed
from app.trading.worker import client as worker_client
from app.trading.worker.client import GatewayClient
@pytest.fixture
def spans(monkeypatch) -> InMemorySpanExporter:
"""把 client 模块级 tracer 换成写内存的(不碰全局 provider)"""
exporter = InMemorySpanExporter()
provider = TracerProvider()
provider.add_span_processor(SimpleSpanProcessor(exporter))
monkeypatch.setattr(worker_client, "tracer", provider.get_tracer("test"))
return exporter
def _client(handler) -> GatewayClient:
"""构建 GatewayClient 并把它的 httpx 客户端换成 MockTransport 版本"""
client = GatewayClient(
"https://gateway.example", "token", settings=Settings(_env_file=None)
)
client._client = httpx.AsyncClient(
base_url="https://gateway.example",
transport=httpx.MockTransport(handler),
)
return client
def _envelope(*, success: bool, code: int, msg: str, data=None, status: int = 200):
return httpx.Response(
status,
json={"success": success, "msg": msg, "data": data, "code": code},
)
async def test_envelope_failure_is_recorded_on_span(spans):
"""success=false 要把 span 置 ERROR 并落错误码——HTTP 层看这次是 200"""
gateway = _client(
lambda _r: _envelope(
success=False, code=6002, msg="租约无效:不是持有者", status=200
)
)
try:
with pytest.raises(AppError) as excinfo:
await gateway.report(
"task-1", "worker-1", state=OrderState.IN_CART, terminal=False
)
assert excinfo.value.err_code == 6002
finally:
await gateway.aclose()
span = spans.get_finished_spans()[0]
assert span.kind is SpanKind.CLIENT
assert span.status.status_code is StatusCode.ERROR
# 这次调用的结论——HTTP 状态码是 200,光看它分不出成败
assert span.attributes["api.success"] is False
assert span.attributes["api.code"] == 6002
assert span.attributes["api.status_code"] == 200
assert "租约无效" in span.attributes["api.msg"]
assert span.attributes["gateway.path"] == "/api/orders/task-1/report"
assert span.attributes["gateway.method"] == "POST"
async def test_successful_envelope_leaves_span_ok(spans):
"""success=true 不置 ERROR;作为对照说明上面的断言不是恒真"""
gateway = _client(
lambda _r: _envelope(success=True, code=0, msg="success", data={"recorded": True})
)
try:
await gateway.report("task-1", "worker-1", state=OrderState.IN_CART)
finally:
await gateway.aclose()
span = spans.get_finished_spans()[0]
assert span.status.status_code is not StatusCode.ERROR
assert span.attributes["api.success"] is True
assert "api.code" not in span.attributes
async def test_non_json_body_is_recorded_as_error(spans):
"""网关回了非 JSON(网关挂了/被反代拦了):记成可重试的 3001"""
gateway = _client(lambda _r: httpx.Response(502, text="<html>502 Bad Gateway</html>"))
try:
with pytest.raises(AppError) as excinfo:
await gateway.renew("task-1", "worker-1")
assert excinfo.value.err_code == 3001
finally:
await gateway.aclose()
span = spans.get_finished_spans()[0]
assert span.status.status_code is StatusCode.ERROR
assert span.attributes["error.code"] == 3001
# 这类失败是可重试的,与信封里的业务拒绝(retryable=False)要能区分开
assert span.attributes["error.retryable"] is True
assert span.attributes["gateway.status_code"] == 502
async def test_idle_long_poll_produces_no_span(spans):
"""回归:空转长轮询在 suppressed() 里不能产生 span
worker 每 30 秒 lease 一次、绝大多数返回空。手工埋点若不看抑制标记,这些
调用会各自成为一条孤立 trace 把观测后台刷满——正是 suppressed() 要解决的问题。
"""
gateway = _client(lambda _r: _envelope(success=True, code=0, msg="success", data=None))
try:
with suppressed():
assert await gateway.lease("worker-1", wait=0) is None
assert spans.get_finished_spans() == ()
# 抑制之外照常埋点,证明上面的空断言不是因为埋点根本没生效
assert await gateway.lease("worker-1", wait=0) is None
finally:
await gateway.aclose()
assert [s.attributes["gateway.path"] for s in spans.get_finished_spans()] == [
"/api/orders/lease"
]
async def test_leased_task_call_is_still_traced(spans):
"""领到任务后的调用不受抑制影响:它们挂在任务根 span 底下,是要看的那部分"""
gateway = _client(
lambda _r: _envelope(
success=True, code=0, msg="success",
data={"task_id": "task-1", "site": "rakuten", "lease_count": 1},
)
)
try:
task = await gateway.lease("worker-1", wait=0)
assert task is not None and task.task_id == "task-1"
finally:
await gateway.aclose()
assert len(spans.get_finished_spans()) == 1