全量修复:预测系统正确性、安全性与部署问题

P0 严重问题修复:
- 修复 form_slice/stats_slice 主客身份反转(历史比赛视角错误)
- 修复 understat.py httpx 未导入导致的 NameError
- 修复 LLM 解析失败时静默产生假成功预测(0-0 平局+置信度0.5)

预测路径修复:
- multi-agent 路径增加 backtest cutoff 透传,回测防泄漏生效
- H2H 切片汇总统计改为从当前主队视角计数
- 预测唯一约束增加 mode+run_type 维度,防止回测覆盖实盘预测

伤停管线修复:
- IntegrityError 后不再整批回滚丢数据(改用逐条 flush)
- return_date 正确解析并写入
- retrieved_at 比较统一用 date() 避免当天数据不可见
- 唯一索引改为 partial unique index(排除 NULL 重复)
- HTTP 缓存 TTL 从 7 天改为 6 小时

安全与连接管理:
- /api/v1/predict 增加内存滑动窗口限流(10次/分钟/IP)
- 预测路由改用短 session 模式,LLM 调用期间不持有 DB 连接

Docker 部署修复:
- 修复 .dockerignore 排除 *.md 导致 COPY README.md 失败
- 容器内 DATABASE_URL 使用 postgres 服务名(非 localhost)
- 启动时自动执行 alembic upgrade head
- 前端改用多阶段构建(Dockerfile.frontend)

新增测试(5个文件,24+用例):
- test_p0_home_away.py: 主客身份反转回归测试
- test_p0_parse_failure.py: LLM 解析失败回归测试
- test_multi_agent_cutoff.py: multi-agent cutoff 透传测试
- test_h2h_perspective.py: H2H 视角测试
- test_injuries_pipeline.py: 伤停管线 5 项修复测试
- test_predict_protection.py: 限流+短 session 测试
- test_prediction_unique_constraint.py: 唯一约束测试

迁移:
- 0012_injuries_partial_unique_and_return_date.py
- 0013_predictions_unique_constraint_mode_run_type.py
This commit is contained in:
Profeto Agent
2026-09-19 06:43:55 +00:00
parent 11efe91ce9
commit bee330f31f
27 changed files with 1666 additions and 137 deletions
+33
View File
@@ -2,6 +2,7 @@
from __future__ import annotations
from collections.abc import AsyncIterator
from contextlib import asynccontextmanager
from sqlalchemy.ext.asyncio import AsyncSession, async_sessionmaker, create_async_engine
from sqlalchemy.orm import DeclarativeBase
@@ -54,6 +55,38 @@ async def get_db_read() -> AsyncIterator[AsyncSession]:
await session.close()
@asynccontextmanager
async def short_read():
"""短生命周期 read session: 用于非路由上下文(如后台任务、手动调用)。
用法:
async with short_read() as session:
m = await session.get(Match, match_id)
# session 已关闭,连接已释放
"""
async with AsyncSessionLocal() as session:
yield session
@asynccontextmanager
async def short_write():
"""短生命周期 write session: 提交后立即释放。
用法:
async with short_write() as session:
session.add(pred)
await session.commit()
# session 已关闭,连接已释放
"""
async with AsyncSessionLocal() as session:
try:
yield session
await session.commit()
except Exception:
await session.rollback()
raise
async def init_db() -> None:
"""验证数据库连接(不建表)。
+21 -4
View File
@@ -16,6 +16,7 @@ from sqlalchemy import (
String,
Text,
UniqueConstraint,
and_,
func,
)
from sqlalchemy.dialects.postgresql import JSONB
@@ -164,7 +165,19 @@ class Injury(Base):
team: Mapped["Team | None"] = relationship()
__table_args__ = (
Index("ix_injuries_player_fixture", "player_id", "fixture_id", "injury_type", unique=True),
# Fix 4: partial unique index — 只在 player_id 和 fixture_id 都非空时强制唯一
# PostgreSQL 中 NULL != NULL,普通唯一索引无法防止 NULL 重复
Index(
"ix_injuries_player_fixture",
"player_id",
"fixture_id",
"injury_type",
unique=True,
postgresql_where=and_(
player_id.is_not(None),
fixture_id.is_not(None),
),
),
Index("ix_injuries_team_date", "team_id", "injury_date"),
)
@@ -194,6 +207,8 @@ class Prediction(Base):
agent_outputs: Mapped[dict | None] = mapped_column(JSONB)
# 预测状态: success / failed / degraded
status: Mapped[str] = mapped_column(String(20), nullable=False, default="success")
# Fix: run_type 区分实盘(live)与回测(backtest),避免回测覆盖实盘预测
run_type: Mapped[str] = mapped_column(String(10), nullable=False, default="live")
# 时间语义:区分比赛时间、预测创建时间、数据截止时间
match_kickoff_at: Mapped[datetime | None] = mapped_column(DateTime(timezone=True))
prediction_created_at: Mapped[datetime] = mapped_column(DateTime(timezone=True), default=_utcnow)
@@ -207,10 +222,11 @@ class Prediction(Base):
match: Mapped[Match] = relationship(back_populates="predictions")
__table_args__ = (
# P1-6: 数据库级唯一约束,防止同一 match+provider+model 产生重复预测
# Fix: 唯一约束增加 mode + run_type,允许 live 与 backtest 共存
# 防止回测覆盖未结算的实盘预测(后续 settle 会污染评估数据)
UniqueConstraint(
"match_id", "provider", "model",
name="uq_predictions_match_provider_model",
"match_id", "provider", "model", "mode", "run_type",
name="uq_predictions_match_provider_model_mode_run_type",
),
Index("ix_predictions_match", "match_id"),
Index("ix_predictions_provider_model", "provider", "model"),
@@ -223,6 +239,7 @@ class Prediction(Base):
CheckConstraint("pred_1x2 IN ('1', 'X', '2')", name="ck_pred_1x2_enum"),
CheckConstraint("mode IN ('single', 'multi')", name="ck_mode_enum"),
CheckConstraint("status IN ('success', 'failed', 'degraded')", name="ck_status_enum"),
CheckConstraint("run_type IN ('live', 'backtest')", name="ck_run_type_enum"),
)