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
+9 -1
View File
@@ -591,6 +591,10 @@ class RunManager:
error_node: str | None = None
deadline: asyncio.Timeout | None = None
# 停在审批上的那一刻:这一段的执行算到这里为止。后面写审批记录、发 run.interrupted 是收尾,
# 等待从 run.interrupted 算起——以前执行一直计到 finally,收尾那一截执行和等待两边都算,
# active + wait 比墙钟还长(机器一忙就差出几十毫秒)
paused: float | None = None
async with self._semaphore:
try:
async with SessionLocal() as session:
@@ -655,8 +659,11 @@ class RunManager:
if mode == "custom":
await self._handle_custom(run_id, chunk)
elif mode == "updates":
# LangGraph 这一轮的节点都跑完了才发出中断:此刻引擎已经停下
at = time.perf_counter()
if await self._handle_updates(run_id, chunk):
interrupted = True
paused = at if paused is None else paused
snapshot = await app.aget_state(config)
final_state = dict(snapshot.values or {})
@@ -710,7 +717,8 @@ class RunManager:
error_node = terminal[1].get("node_id")
logger.exception("run %s 失败", run_id)
finally:
elapsed = int((time.perf_counter() - started) * 1000)
stop = paused if status == "interrupted" and paused is not None else time.perf_counter()
elapsed = int((stop - started) * 1000)
self._finalizing.add(run_id)
try:
await self._finalize(run_id, status, error, final_state, elapsed,