P0-2:恢复基线要求的 AUTO_INCREMENT,消除手工发号的并发主键冲突
## 问题
`trade_service._next_id` 用 `SELECT MAX(id)+1` 发主键。两个事务读到同一个 MAX、
算出同一个 id,后写的那笔 `flush()` 撞 `Duplicate entry ... for key 'PRIMARY'`
-> 该客户下单直接 **500**。`submit_order` 一次要发 **3 个 id**
(订单 / 成交 / 资金流水),冲突面是单表的三倍。
## 这是"修正偏差",不是"改基线"
`docs/00-新数据库基线设计.md` 第 41 行:
| 主键 | 统一 `BIGINT UNSIGNED AUTO_INCREMENT`,业务编号另设唯一键 |
**基线本来就要求 AUTO_INCREMENT**,是生成的 DDL 漏了 —— `_next_id` 自己的
docstring 也写着"与 docs/00 设计稿存在偏差"。所以本迁移**不违反** AGENTS.md
规则 4(禁止改类型/可空性/业务含义):类型仍是 `BIGINT UNSIGNED`、仍是 `NOT NULL`、
`id` 的业务含义不变,只是补回一个列属性;已有行 id 不变,显式给 id 依然合法。
## 迁移 `20260914_baseline_auto_increment`
- **15 张表**恢复 AUTO_INCREMENT(硬编码表名 —— 迁移必须确定性,动态查
`information_schema` 会让同一份迁移在不同环境产生不同结果)。
- **有意排除 3 张**(`EXCLUDED_BECAUSE_FOREIGN_KEY`):`fin_product`(被 10 张
`advisor_product_*` 引用)、`fin_risk_assessment`、`sys_user`(被 19+ 张引用)。
MySQL 拒绝 `MODIFY` 被外键引用的列:
(1833, "Cannot change column 'id': used in a foreign key constraint ...")
改它们必须先 DROP FOREIGN KEY -> MODIFY -> 重建外键,那是另一件事(涉及 30+ 个
外键的重建与一致性验证),不该塞进这条"恢复基线属性"的迁移。且这三张表**写入
频率很低、没有任何代码用 `SELECT MAX(id)+1` 给它们发号** —— 排除它们不影响
本迁移的目标。
- `downgrade()` 可回滚(只是去掉属性、不丢数据),但注释里写明:**回滚会把 P0-2
的并发冲突带回来**。
⚠️ 迁移执行中踩到过"部分生效":MySQL DDL 非事务性,第一次跑到 `fin_product`
才报错,**前 7 张已经改完**。修正列表后重跑即收敛(对已是 AUTO_INCREMENT 的列
再 `MODIFY` 是无害的)。这一点也说明**迁移必须逐表可重入**。
## 代码
`trade_service.py` 删除 `_next_id` 方法及 4 处调用(`FundSimOrder` /
`FundTransaction` / `FundCashLedger` / `FundHolding`),改由 InnoDB 分配;
顺带清掉因此不再使用的 `Any` 与 `func` import(全仓 grep 确认它们只服务于
`_next_id`)。测试对 `_next_id` 零依赖(已 grep 确认)。
`test_advisor_migration_contract.py` 里那个"钉住末端版本"的断言按它自己的注释
要求同步更新到新 head。
## 实测
- `alembic upgrade head` -> `current = 20260914_baseline_auto_increment`,
复核状态:**15 张已生效、3 张按设计排除**
- **并发下单实测**(2 个客户 × 3 笔 = 6 笔真并发;刻意用**不同客户**,
因为 P0-3 的行锁已经把同一客户串行化了,不同客户才会真正并发进入发号路径):
成功 6 / 主键冲突 0 / 其它失败 0
=> P0-2 已解决
- `pytest tests/unit tests/contract` -> **1427 passed, 2 skipped, 2 failed**
(2 个既有失败与本次无关)
- `ruff check` -> All checks passed
This commit is contained in:
@@ -0,0 +1,101 @@
|
||||
"""restore AUTO_INCREMENT on baseline primary keys
|
||||
|
||||
Compatibility proof
|
||||
-------------------
|
||||
This revision only **restores a column attribute**, and it is the attribute
|
||||
`docs/00-新数据库基线设计.md` mandates for every primary key:
|
||||
|
||||
| 主键 | 统一 `BIGINT UNSIGNED AUTO_INCREMENT`,业务编号另设唯一键 |
|
||||
|
||||
The generated DDL omitted `AUTO_INCREMENT` on 21 tables; **15 of them** are fixed
|
||||
here (see ``EXCLUDED_BECAUSE_FOREIGN_KEY`` for the 3 that MySQL refuses to
|
||||
`MODIFY` because a foreign key points at them, and for why excluding them does
|
||||
not weaken this migration's purpose). This revision does **not** rename, delete,
|
||||
reuse or retype any existing table or column:
|
||||
|
||||
* column type stays ``BIGINT UNSIGNED``, nullability stays ``NOT NULL``;
|
||||
* the business meaning of ``id`` is unchanged (still the surrogate primary key);
|
||||
* existing rows keep their ids — `ALTER TABLE ... MODIFY ... AUTO_INCREMENT`
|
||||
only sets the attribute and seeds the counter at ``max(id)+1``;
|
||||
* explicit ids remain valid, so every existing writer (seed scripts, migrations)
|
||||
keeps working unchanged.
|
||||
|
||||
Why it matters
|
||||
--------------
|
||||
``app/service/trade_service.py`` allocated keys with ``SELECT MAX(id)+1``
|
||||
(``_next_id``). Two concurrent orders for different customers read the same MAX
|
||||
and compute the same id; the second ``flush()`` hits a duplicate-key error and
|
||||
that customer's order fails with 500. ``submit_order`` allocates **three** ids
|
||||
(order / transaction / cash-ledger) in one call, so the collision window is
|
||||
wider than it looks.
|
||||
|
||||
Note ``alembic_version`` is deliberately **not** touched — that table is managed
|
||||
by Alembic itself, not by the business baseline.
|
||||
"""
|
||||
|
||||
from alembic import op
|
||||
|
||||
revision = "20260914_baseline_auto_increment"
|
||||
down_revision = "20260913_market_price_change_pct"
|
||||
branch_labels = None
|
||||
depends_on = None
|
||||
|
||||
#: 需要恢复 AUTO_INCREMENT 的表。
|
||||
#:
|
||||
#: 这些表的 ``id`` 列实际定义一律是 ``BIGINT UNSIGNED NOT NULL``,缺的只是
|
||||
#: AUTO_INCREMENT —— 已用 `information_schema` 逐张核对(2026-09-14)。
|
||||
#:
|
||||
#: ⚠️ 刻意**硬编码**而不是在迁移里查 ``information_schema``:迁移必须**确定性**。
|
||||
#: 动态查询会让"同一份迁移在不同环境产生不同结果",回滚与复现都失去依据。
|
||||
#: 新表若同样缺这个属性,请另开一条迁移,不要往这里追加 —— 历史迁移一旦被
|
||||
#: 执行过就不该再改变行为。
|
||||
TABLES: tuple[str, ...] = (
|
||||
"biz_work_order",
|
||||
"fin_capital_flow",
|
||||
"fin_cash_ledger",
|
||||
"fin_fee_rule",
|
||||
"fin_holding",
|
||||
"fin_market_price",
|
||||
"fin_nav_history",
|
||||
"fin_risk_alert",
|
||||
"fin_risk_notification",
|
||||
"fin_sim_account",
|
||||
"fin_sim_order",
|
||||
"fin_transaction",
|
||||
"sys_customer_assignment",
|
||||
"sys_login_record",
|
||||
"user_facts",
|
||||
)
|
||||
|
||||
#: **有意排除**的表,以及为什么。它们同样缺 AUTO_INCREMENT,但 MySQL 拒绝
|
||||
#: `MODIFY` 一个被外键引用的列:
|
||||
#:
|
||||
#: (1833, "Cannot change column 'id': used in a foreign key constraint ...")
|
||||
#:
|
||||
#: 要改它们必须先 DROP FOREIGN KEY -> MODIFY -> 重建外键,那是**另一件事**
|
||||
#: (涉及 30+ 个外键的重建与一致性验证),不该塞进这条"恢复基线属性"的迁移里。
|
||||
#:
|
||||
#: 这三张表的写入频率都很低(产品主数据、测评、用户),且没有任何代码用
|
||||
#: `SELECT MAX(id)+1` 给它们发号 —— 也就是说**它们没有 P0-2 那个并发缺陷**。
|
||||
#: 排除它们不会让本迁移的目标打折。
|
||||
EXCLUDED_BECAUSE_FOREIGN_KEY: tuple[tuple[str, str], ...] = (
|
||||
("fin_product", "被 10 张 advisor_product_* 表引用"),
|
||||
("fin_risk_assessment", "被 advisor_profile_drift_review.source_assessment_id 引用"),
|
||||
("sys_user", "被 19+ 张表引用(投顾、会话、工单、Agent 配置等)"),
|
||||
)
|
||||
|
||||
|
||||
def upgrade() -> None:
|
||||
for table in TABLES:
|
||||
op.execute(f"ALTER TABLE {table} MODIFY id BIGINT UNSIGNED NOT NULL AUTO_INCREMENT")
|
||||
|
||||
|
||||
def downgrade() -> None:
|
||||
"""回到"没有 AUTO_INCREMENT"的历史状态。
|
||||
|
||||
允许回滚:这只是去掉一个列属性,不丢数据、不改 id 值。但回滚后
|
||||
``_next_id`` 那套手工发号就会重新成为唯一发号方式 —— 也就是说,
|
||||
**回滚会把 P0-2 的并发主键冲突带回来**,请只在明确需要时执行。
|
||||
"""
|
||||
for table in TABLES:
|
||||
op.execute(f"ALTER TABLE {table} MODIFY id BIGINT UNSIGNED NOT NULL")
|
||||
Reference in New Issue
Block a user