修正审批恢复的运行耗时:停在审批上之后的收尾不再同时算进执行和等待

- 每一段的执行时长以前一直计到收尾结束,等待时长又从 run.interrupted 算起。停在审批上之后
  写审批记录、发事件这一截被两边都算,执行 + 等待比墙钟还长:平时只差几毫秒,机器一忙就差出几十毫秒
  (全量并行测试里约每 20 遍失败一次的那条就是它)
- 现在这一段的执行算到引擎发出中断的那一刻为止(LangGraph 这一轮的节点都跑完了才发出中断),
  之后的收尾不算执行;执行、等待、墙钟由构造保证对得上账,不再靠容差
- 新测试把中断后的收尾人为拖慢 0.4 秒,稳定复现以前的重复计算;全量测试并行连跑 26 遍全过
This commit is contained in:
ThearchyHelios
2026-09-29 16:50:31 +08:00
parent e859c13433
commit dbae680abf
2 changed files with 40 additions and 1 deletions
+31
View File
@@ -123,6 +123,37 @@ async def test_resumed_run_keeps_its_start_and_sums_every_segment(slow_transform
}
async def test_bookkeeping_after_the_pause_is_not_counted_twice(slow_transforms, monkeypatch):
"""停在审批上之后的收尾(写审批记录、发 run.interrupted)不算执行时长。
以前这一段的执行一直计到 _drive 的 finally,等待又从 run.interrupted 算起:收尾那一截两边都算,
active + wait 比墙钟还长。平时只差几毫秒,机器一忙就超出容差(全量并行测试里偶发)。
这里把收尾拖慢 0.4 秒,不靠机器忙不忙碰运气。
"""
original = run_manager._handle_updates
async def slow(run_id, chunk):
hit = await original(run_id, chunk)
if hit:
await asyncio.sleep(0.4) # 审批卡已经发出去了,引擎也不在执行
return hit
monkeypatch.setattr(run_manager, "_handle_updates", slow)
run = await run_manager.start(graph=APPROVAL_GRAPH, input_payload={"question": "q"})
paused = await _wait(run.id, ("interrupted", "failed"))
assert paused.status == "interrupted", paused.error
await asyncio.sleep(WAIT)
await run_manager.resume(run.id, {"approved": True})
done = await _wait(run.id, ("succeeded", "failed"))
assert done.status == "succeeded", done.error
usage = done.usage
# 执行、等待、墙钟严格对得上账:取整各丢不到 1 毫秒
assert usage["wall_ms"] >= usage["active_ms"] + usage["wait_ms"] - 2, usage
# 0.4 秒的收尾没有算进执行(两段各 SEGMENT 秒,其余开销远小于 0.4 秒)
assert usage["active_ms"] < 2 * SEGMENT * 1000 + 400, usage
@pytest.fixture
def fails_once(monkeypatch):
original = compiler.RUNNERS[NodeType.TRANSFORM]