fix: gate orchestrator materialize leak + cancel-drain on abort (algo #6)
Codex 质量审 2 Critical: - C001: 候选物化移入 try、成功一个登记一个,第 N 个题型物化失败时 finally 仍清理前 N-1 个已建目录,不泄漏 - C002: gather 首异常(护栏 raise)后显式取消其余任务并排水,确保 finally 删除候选目录时无在飞任务访问、事件循环无 pending task 警告; 护栏中止整轮语义不变 - 新增 2 测试:部分物化失败清理 / 护栏 raise 取消收束不悬挂(wait_for 5s) Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
This commit is contained in:
+34
-23
@@ -1272,34 +1272,45 @@ async def validate_skills_concurrent(
|
||||
_validate_gate_specs(specs)
|
||||
base_skills_dir = workspace_dir / "skills" / base_skills_version
|
||||
runs = [_GateRun.from_spec(s) for s in specs]
|
||||
cand_dirs = {
|
||||
r.spec.task_type: materialize_candidate_skill(
|
||||
workspace_dir, base_skills_version, r.spec.target_file, r.spec.candidate_content
|
||||
)
|
||||
for r in runs
|
||||
}
|
||||
cand_dirs: dict[str, Path] = {}
|
||||
slots_gate = _QuestionSlots(concurrency)
|
||||
try:
|
||||
coros = [
|
||||
_run_unit_arm(
|
||||
r,
|
||||
rank,
|
||||
arm,
|
||||
slots_gate,
|
||||
run_inference,
|
||||
log,
|
||||
baseline_cache,
|
||||
prompts_version,
|
||||
base_skills_dir,
|
||||
cand_dirs[r.spec.task_type],
|
||||
gate_params,
|
||||
gate_guard_err,
|
||||
# 成功一个登记一个:第 N 个题型物化抛 OSError 时,已登记的前 N-1 个
|
||||
# 目录仍由 finally 统一清理,不泄漏(Codex 质量审 C001)。
|
||||
for r in runs:
|
||||
cand_dirs[r.spec.task_type] = materialize_candidate_skill(
|
||||
workspace_dir, base_skills_version, r.spec.target_file, r.spec.candidate_content
|
||||
)
|
||||
tasks = [
|
||||
asyncio.ensure_future(
|
||||
_run_unit_arm(
|
||||
r,
|
||||
rank,
|
||||
arm,
|
||||
slots_gate,
|
||||
run_inference,
|
||||
log,
|
||||
baseline_cache,
|
||||
prompts_version,
|
||||
base_skills_dir,
|
||||
cand_dirs[r.spec.task_type],
|
||||
gate_params,
|
||||
gate_guard_err,
|
||||
)
|
||||
)
|
||||
for r, rank, arm in _build_launch_order(runs)
|
||||
]
|
||||
# gather 任一任务 raise(INFRA 护栏)即向上传播中止整轮,与现行"护栏
|
||||
# 中止训练"语义一致;finally 仍清理候选目录。
|
||||
await asyncio.gather(*coros)
|
||||
# 护栏 raise 中止整轮的语义不变(Codex 质量审 C002):首异常先取消其余
|
||||
# 任务并排水(return_exceptions 吞取消回报),确保外层 finally 删除候选
|
||||
# 目录时已无在飞任务访问该目录、事件循环收尾无 pending task 警告;
|
||||
# _run_unit_arm 的题槽获取自带取消回滚,cancel 安全。
|
||||
try:
|
||||
await asyncio.gather(*tasks)
|
||||
except BaseException:
|
||||
for t in tasks:
|
||||
t.cancel()
|
||||
await asyncio.gather(*tasks, return_exceptions=True)
|
||||
raise
|
||||
finally:
|
||||
_cleanup_candidate_dirs(cand_dirs)
|
||||
|
||||
|
||||
Reference in New Issue
Block a user