akg-factor-bridge/test_plan_snapshot.py

215 lines
10 KiB
Python
Raw Normal View History

"""接口改读当日全量快照(2026-09-10,台账 057):截取规则、名次重编、回落、字段一致。
要解决的问题:计划接口每次请求都现场装配整池,实测三十三秒。而计划是日频的、用的全是
昨收与昨夜数据,出计划时本来就落了一份全量快照(不裁剪、不设主题限额),接口一直没读它。
离线,不连库不起服务。跑法:python3 test_plan_snapshot.py,预期最后一行是 ALL OK。
真实快照存在时会额外做一轮逐行比对,不存在就跳过那一轮并说明。
"""
import glob
import json
import os
import sys
import types
for _n in ("pandas", "pymysql", "psycopg"):
if _n not in sys.modules:
try:
__import__(_n)
except Exception: # noqa: BLE001
_m = types.ModuleType(_n)
if _n == "pandas":
_m.DataFrame = type("DataFrame", (), {})
sys.modules[_n] = _m
import config # noqa: E402
import plan # noqa: E402
FAILED = []
def t(name, cond, extra=""):
print((" ok " if cond else " FAIL ") + name + ((" " + str(extra)) if not cond else ""))
if not cond:
FAILED.append(name)
def _row(code, theme, rank):
return {"code": code, "rank": rank, "score": 100.0 - rank,
"evidence": ({"theme": theme} if theme else None)}
def test_pick_rules():
print("[截取规则]")
rows = [_row("A", "光伏", 1), _row("B", "光伏", 2), _row("C", "光伏", 3),
_row("D", "军工", 4), _row("E", None, 5)]
t("不设限就按顺序取", [r["code"] for r in plan.pick_rows(rows, 3, 0)] == ["A", "B", "C"])
got = plan.pick_rows(rows, 4, 2)
t("每个主题最多两条,第三条光伏被跳过",
[r["code"] for r in got] == ["A", "B", "D", "E"], [r["code"] for r in got])
t("没有主题的行不归到同一个桶",
[r["code"] for r in plan.pick_rows(
[_row("X", None, 1), _row("Y", None, 2), _row("Z", "光伏", 3)], 3, 1)] == ["X", "Y", "Z"])
选股计划排序轴换成公司质地,传导降为雷达门与带内次级键(台账 063) 按 docs/选股计划调整方案_2026-09-17.md 落地。三份读数(09-02 §1.8、08-28 周报、09-11 抽查)一致:传导越强超额越负,传导分的公式奖励的正是被证明为负的"未动补涨"。所以把传导从排序主轴换成雷达门加带内次级键,公司质地成为排序主轴。改动全在计划装配层,四个开关全设旧值即逐字回旧。 一、排序轴换质地(plan.py)。新增 _rerank_by_axis 按(质地带、被指向、组内分、原始分、代码)词典序重排、_rerank 按 config.PLAN_RANK_AXIS 分派;组内分从原始分反解、档位不进次序;被指向以候选卡门槛一 gates.pointed 为准。调用点先用 _rerank_by_quality 算旧序拿 rank_old 再算新序。_pick 与 pick_rows 加 uncapped_no_theme:质地轴下无传导行免主题限额、不计数,否则主轴被限额推翻。_row 加 quality_band、rank_axis、pointed、rank_old 四键,score/score_adj/tier 值与含义不变。返回字典质地轴下加 rank_axis、rank_rule、quality_band_counts。encoding 追加一句。 二、候选卡内序与质地门(card.py)。新增 quality_band 函数(好三中二差零,无评析与证据不足与超期一百二十天归一),plan.py 与 card.py 共用它。judge 加 stale_days 参数、返回加 quality_band;质地门"差降关注"改用质地带为零并带上一百二十天守卫(超期报告不再降档)。sort_key 在判决与涨幅之间插入质地带、受 CARD_SORT_BY_QUALITY 控,关掉逐字回旧。三门槛一确认三风险一字未动。 三、入池与评析名单(pool.py、review_targets.py)。pool 把选行抽成纯函数 select_plan_rows:rank 源取主榜新序里被指向的前 POOL_TOP、_pool_source 记 rank_top,备用开关 POOL_RANK_KEEP_OLD 并进旧序强传导前若干、去重、受 POOL_MAX 兜底;tier 与 candidate 逐字保留。strategy_context_entry 加 quality_band、rank_axis、rank_old 三键。review_targets 新增纯函数 pick_main_targets:both 取原始分前一百与新序前五十的并集、去重按新序,score 逐字回旧;评析覆盖与排序轴解耦。 四、周报对照(plan_review.py)。第一节加主榜新序前二十与旧序前二十两份名单(rank_old 缺的旧快照跳过);档位分组外加档位乘质地带交叉分组,看传导在质地档内还剩多少信息量。plan_reconcile 应答加 rank_axis: transmission 一键,说明逐票对账的"主榜第 N 名"仍按原始分序,文案不动。 五、单测、开关、文档。新 test_plan_rank_axis.py 十例(含四开关设旧值逐字回旧、四带序、带内次级键、确定性、超期守卫、入池 rank、评析并集、卡内序、主题限额例外、备用开关),本地 ALL OK。test_plan_quality_rank.py 接线匹配串改成新调用点,test_plan_snapshot.py 加一条质地轴主题限额对照用例。config 加 PLAN_RANK_AXIS、CARD_SORT_BY_QUALITY、REVIEW_TARGETS_AXIS 三个开关与备用开关 POOL_RANK_KEEP_OLD,POOL_SOURCE 默认改 rank。.env.example、README 计划字段、台账 063、入池对接说明、下游对接说明一并改。 没有改的是什么、为什么不合并。分数编码 factors.py 一字未动——它注册在外部平台当因子,改它会污染平台的因子历史;次序改动全部放在装配层。api.py 不碰。三门槛一确认三风险不动,质地门保持"差降关注"。持仓管理系统按上游名次排是协同项,在另一仓库单独审批,不阻塞本次上线(它的看多分桶已把好公司排在前面)。 实测读数。开发机 python3 test_plan_rank_axis.py 与 python3 test_plan_quality_rank.py 均 ALL OK;PLAN_RANK_AXIS=transmission CARD_SORT_BY_QUALITY=0 POOL_SOURCE=tier REVIEW_TARGETS_AXIS=score 下 test_plan_rank_axis.py 仍 ALL OK(逐字回旧那一例断言与 09-14 旧序逐字相同)。test_card、test_pool_logic、test_plan_snapshot 全绿。既有的 test_review_targets、test_plan_review_window、test_plan_logic_state、test_plan_cache、test_plan_reviews_live 因缺 psycopg/fastapi 或本机 Python 3.9 不支持联合类型注解而只能在部署机跑,非本次引入。 Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
2026-09-17 10:28:23 +08:00
# 质地轴对照(2026-09-17):同样两只无主题票,行带 rank_axis=quality 时免主题限额,两只都进
t("质地轴下无主题行不受主题限额(对照上一条)",
[r["code"] for r in plan.pick_rows(
[dict(_row("X", None, 1), rank_axis="quality"),
dict(_row("Y", None, 2), rank_axis="quality"),
dict(_row("Z", "光伏", 3), rank_axis="quality")], 3, 1)] == ["X", "Y", "Z"])
t("要的条数比总行数多就全给", len(plan.pick_rows(rows, 99, 0)) == 5)
t("要零条时先给一条(照抄原实现的边界:先放进去再判断够没够,不是我们要它这样)",
len(plan.pick_rows(rows, 0, 0)) == 1)
t("要一条就给一条", [r["code"] for r in plan.pick_rows(rows, 1, 0)] == ["A"])
def test_replay_against_real_snapshot():
print("[拿真实快照逐行比对 —— 这条是这次改动成立的根据]")
fs = sorted(glob.glob(os.path.join(config.PLAN_SNAPSHOT_DIR, "plan_*.json")))
if not fs:
print(" -- 本机没有快照文件,跳过这一轮(在部署机上跑才有)")
return
p = fs[-1]
snap = json.load(open(p, encoding="utf-8"))
sp = snap.get("shown_params") or {}
if not sp or not isinstance(snap.get("main"), list):
print(" -- 最新快照没有全量段或落盘参数,跳过:" + os.path.basename(p))
return
print(" 用 %s(落盘参数 %s)" % (os.path.basename(p), sp))
for seg, shown_key, n_key in (("main", "main_shown", "top"),
("observe", "observe_shown", "obs_top")):
got = plan.pick_rows(snap[seg], sp[n_key], sp["theme_cap"])
shown = snap[shown_key]
t("%s 截出来的代码序列与当时下发的完全一致" % seg,
[r["code"] for r in got] == [r["code"] for r in shown],
([r["code"] for r in got][:6], [r["code"] for r in shown][:6]))
# 名次是唯一该不一样的字段:全量快照里存的是全量排名,下发时要重编成连续名次
diff = {k for a, b in zip(got, shown) for k in set(a) | set(b) if a.get(k) != b.get(k)}
t("%s 除名次外逐字段一致" % seg, diff <= {"rank"}, diff)
renum = [{**r, "rank": i} for i, r in enumerate(got, 1)]
t("%s 名次重编后与当时下发的一模一样" % seg, renum == shown)
def test_from_snapshot(tmpdir=None):
print("[from_snapshot 的输出与回落]")
fs = sorted(glob.glob(os.path.join(config.PLAN_SNAPSHOT_DIR, "plan_*.json")))
if not fs:
print(" -- 本机没有快照文件,跳过这一轮")
return
ds = os.path.basename(fs[-1])[len("plan_"):-len(".json")]
d = plan.from_snapshot(ds, top=5, obs_top=3, theme_cap=2)
t("名次是连续的 1 到 n", [r["rank"] for r in d["main"]] == list(range(1, len(d["main"]) + 1)))
t("条数不超过请求的", len(d["main"]) <= 5 and len(d["observe"]) <= 3)
t("主题限额按本次请求生效,不是落盘那次的", d["theme_cap"] == 2)
t("标明了这份是从快照来的", d.get("plan_source") == "snapshot")
t("带上快照的生成时刻,一眼看得出多新", bool(d.get("snapshot_generated_at")))
for k in ("main_shown", "observe_shown", "shown_params"):
t("不把落盘那次的裁剪结果带出去(%s)" % k, k not in d)
for k in ("market", "regime"):
t("环境字段来自指定快照,不在查询时覆盖(%s)" % k,
d.get(k) == plan.plan_versions.load(config.PLAN_SNAPSHOT_DIR, ds).get(k))
for k in ("date", "generated_at", "plan_version", "counts", "candidates", "watch",
"segments_pointed", "card_counts", "card_params", "changes", "encoding"):
t("顶层保留 %s" % k, k in d)
# 全量计数不随裁剪变化:counts 说的是整池有多少,不是这次给了几条
t("全量计数原样保留", (d.get("counts") or {}).get("main", 0) >= len(d["main"]))
# 认不出的日期、对不上的日期一律抛异常,交给调用方回落
try:
plan.from_snapshot("1999-01-01", 5, 3, 2)
t("快照不存在要抛异常", False)
except Exception as e: # noqa: BLE001
t("快照不存在要抛异常", isinstance(e, (FileNotFoundError, RuntimeError)), type(e).__name__)
def test_freshness_guard():
print("[新鲜度守卫 —— 旧快照不许蒙混过关]")
import datetime as _dt
now = _dt.datetime.now()
import version
cur = ""
try:
cur = str(version.git_short_rev() or "")
except Exception: # noqa: BLE001
cur = ""
def snap(gen, ver=None):
return {"generated_at": gen, "plan_version": cur if ver is None else ver}
# 时间判据是「距今多久」,不是「晚于今天几点」——容器跑在世界协调时上,
# 拿本地日期算界会差八小时,把早上七点十分出的快照判成过期(2026-09-10 实测踩到)
ok, why = plan.snapshot_fresh_enough(
snap((now - _dt.timedelta(hours=2)).isoformat(timespec="seconds")))
t("两小时前出的、版本对得上的快照可用", ok, why)
ok2, why2 = plan.snapshot_fresh_enough(
snap((now - _dt.timedelta(hours=30)).isoformat(timespec="seconds")))
t("三十小时前的快照不可用", not ok2 and "距今" in why2, why2)
ok_edge, _ = plan.snapshot_fresh_enough(
snap((now - _dt.timedelta(hours=15.5)).isoformat(timespec="seconds")))
t("十五个半小时刚好在界内", ok_edge)
ok_edge2, _ = plan.snapshot_fresh_enough(
snap((now - _dt.timedelta(hours=16.5)).isoformat(timespec="seconds")))
t("十六个半小时超界", not ok_edge2)
ok_f, why_f = plan.snapshot_fresh_enough(
snap((now + _dt.timedelta(hours=5)).isoformat(timespec="seconds")))
t("生成时刻在未来的不可用(多半是两边时区不一致)", not ok_f and "未来" in why_f, why_f)
if cur:
ok3, why3 = plan.snapshot_fresh_enough(
snap((now - _dt.timedelta(hours=2)).isoformat(timespec="seconds"), "deadbee"))
t("代码部署不让有效期内已发布的旧计划失效", ok3, why3)
ok4, _ = plan.snapshot_fresh_enough(snap(""))
t("没有生成时刻的快照不可用", not ok4)
ok5, _ = plan.snapshot_fresh_enough(snap("不是时间"))
t("生成时刻认不出的快照不可用", not ok5)
t("守卫用的是专门的异常类型,调用方接得住", issubclass(plan.StaleSnapshot, RuntimeError))
# 新鲜度只管「今天该用哪份」。复盘要看 09-08 那天的计划,拿的就该是那天那份旧快照
import glob as _g, os as _o
fs = sorted(_g.glob(_o.path.join(config.PLAN_SNAPSHOT_DIR, "plan_*.json")))
if fs:
ds = _o.path.basename(fs[-1])[len("plan_"):-len(".json")]
try:
plan.from_snapshot(ds, 5, 3, 2)
t("显式指定日期不受今天的新鲜度约束(复盘要拿历史那份)", True)
except plan.StaleSnapshot as e:
t("显式指定日期不受今天的新鲜度约束(复盘要拿历史那份)", False, e)
def test_api_falls_back():
import ast
from types import SimpleNamespace
print('[缺历史快照不能用当前研究补造]')
src = open(os.path.join(os.path.dirname(os.path.abspath(__file__)), 'api.py'), encoding='utf-8').read()
fn = next(n for n in ast.parse(src).body if isinstance(n, ast.FunctionDef) and n.name == '_plan_build')
def missing(*a, **k): raise FileNotFoundError('无历史版本')
def forbidden(*a, **k): raise AssertionError('查询禁止重装配')
env = {'plan': SimpleNamespace(from_snapshot=missing, collect=forbidden)}
exec(compile(ast.Module(body=[fn], type_ignores=[]), 'api.py', 'exec'), env)
try:
env['_plan_build']('2020-01-02', 20, 10, 5)
t('缺档必须明确失败', False)
except FileNotFoundError:
t('缺档失败,没有回落当前数据', True)
def main():
test_pick_rules()
test_replay_against_real_snapshot()
test_from_snapshot()
test_freshness_guard()
test_api_falls_back()
print("-" * 70)
if FAILED:
print("FAILED %d: %s" % (len(FAILED), "; ".join(FAILED)))
return 1
print("ALL OK — 计划读快照:截取规则 / 与真实快照逐行一致 / 输出与回落 / 接口接线")
return 0
if __name__ == "__main__":
sys.exit(main())