锁定表会导致 django 的保存点问题

Question

我正在尝试在 Django 应用程序的 MariaDB 数据库中构建有向无环图 (DAG)。因为这是非循环的，所以我需要验证任何添加的元素（顶点/边）不会在图中创建循环。

许多客户端会尝试在一天中同时添加元素，但是这些循环检查需要是原子的，所以我认为在添加/更新元素时​​需要使用一些锁。 Django 似乎没有提供类似的东西，所以我尝试使用原始的LOCK TABLES/UNLOCK TABLES 查询。这是我用来执行此操作的代码...

def lock_tables():                                                                    
    cursor = get_connection(DEFAULT_DB_ALIAS).cursor()                            

    tables = [                                                                    
        'vertex',                                                  
        'edge'                                                                                             
    ]                                                                             

    lock_query = ', '.join(                                                
        "{} {}".format(table, 'WRITE') for table in tables                        
    )                                                                             

    query = 'LOCK TABLES {}'.format(lock_query)                            
    cursor.execute(query)                                                         


def unlock_tables():                                                                  
    cursor = get_connection(DEFAULT_DB_ALIAS).cursor()                            
    cursor.execute('UNLOCK TABLES')

然后在我模式的save 方法中......

@transaction.atomic()
def save(self, *args, **kwargs):

    print("---INSIDE MODEL SAVE")

    try:
        print("---LOCKING TABLES")
        lock_tables()
        print("---LOCKED TABLES")

        super().save(*args, **kwargs)

        # TODO: Add Cycle check here
    except Exception as ex:
        print("---EXCEPTION THROWN INSIDE SAVE: {}".format(ex))
        raise
    finally:
        print("---UNLOCKING TABLES")
        unlock_tables()
        print("---UNLOCKED TABLES")   

然而，关于锁定和解锁这些表的一些事情会弄乱使用django.db.transaction.atomic 创建的保存点...有时当 Django 尝试退出 atomic 上下文时，它会尝试回滚到它已经存在的保存点发布。

这里有一些我试图捕捉问题的日志，Executing Query 行来自django.db.backends.mysql.base，STARTING/EXITING ATOMIC 行来自django.db.transactions.atomic __enter__/__exit__ 方法，#### 之后的注释是我在事后添加的 cmets 试图解释我的想法。

---STARTING ATOMIC  #### Atomic context wrapping my serializer's create method
Executing query: 'SAVEPOINT `s139667621889792_x1`' - args: None

---STARTING ATOMIC  #### Atomic context wrapping my model's save method
Executing query: 'SAVEPOINT `s139667621889792_x2`' - args: None

---INSIDE MODEL SAVE
---LOCKING TABLES
Executing query: 'LOCK TABLES vertex WRITE, edge WRITE
---LOCKED TABLES

---STARTING ATOMIC  #### I think Django must wrap some queries in an atomic block, but this doesnt even create a savepoint
Executing query: 'INSERT INTO `edge`...
---EXITING ATOMIC

#### WHERE MY CYCLE CHECK CODE WOULD RUN - not implemented yet

---UNLOCKING TABLES
Executing query: 'UNLOCK TABLES' - args: None
---UNLOCKED TABLES

---EXITING ATOMIC
Executing query: 'RELEASE SAVEPOINT `s139667621889792_x2`' - args: None
Executing query: 'ROLLBACK TO SAVEPOINT `s139667621889792_x2`' - args: None   ### WHAT I BELIEVE TO BE THE OFFENDING QUERY

---EXITING ATOMIC
Executing query: 'ROLLBACK TO SAVEPOINT `s139667621889792_x1`' - args: None

Traceback (most recent call last):
  File ".../site-packages/django/db/backends/utils.py", line 83, in _execute
    return self.cursor.execute(sql)
  File ".../site-packages/django/db/backends/mysql/base.py", line 72, in execute
    return self.cursor.execute(query, args)
  File ".../site-packages/pymysql/cursors.py", line 170, in execute
    result = self._query(query)
  File ".../site-packages/pymysql/cursors.py", line 328, in _query
    conn.query(q)
  File ".../site-packages/pymysql/connections.py", line 516, in query
    self._affected_rows = self._read_query_result(unbuffered=unbuffered)
  File ".../site-packages/pymysql/connections.py", line 727, in _read_query_result
    result.read()
  File ".../site-packages/pymysql/connections.py", line 1066, in read
    first_packet = self.connection._read_packet()
  File ".../site-packages/pymysql/connections.py", line 683, in _read_packet
    packet.check_error()
  File ".../site-packages/pymysql/protocol.py", line 220, in check_error
    err.raise_mysql_exception(self._data)
  File ".../site-packages/pymysql/err.py", line 109, in raise_mysql_exception
    raise errorclass(errno, errval)
pymysql.err.InternalError: (1305, 'SAVEPOINT s139667621889792_x2 does not exist')

如上所示，django 尝试回滚到它已经释放的保存点。如果我删除对锁定/解锁表的调用，此代码可以完美运行，但是我不能再保证我的循环检查是原子的。

以前有没有人遇到过这个问题，或者有任何关于如何深入挖掘原因的提示？

编辑：我读得越多，我就越认为我想要的行为是不可能的。根据MySQL docs on locks，似乎在您获得表锁定时提交了事务。这打破了我的用例，因为如果我的周期检查失败，我希望回滚事务。

Answer 1

任何反循环算法都依赖于在执行检查时表不发生变化。正确的？执行周期检查需要多长时间？您每天需要多少张支票？

假设您有足够的时间来完成所有这些工作，那么请考虑一下：

SELECT GET_LOCK('cycle_check');  -- (you may want timeout)
BEGIN;
INSERT new item in graph
perform cycle check
if ... COMMIT else ROLLBACK
SELECT RELEASE_LOCK('cycle_check');

请注意，这种锁定机制与导致LOCK TABLES 无用的特性不同。

为了防止在循环检查期间读取，您还需要：

SELECT GET_LOCK('cycle_check');
SELECT ...;
SELECT RELEASE_LOCK('cycle_check');

（旁注：GET_LOCK 是执行锁定的“正确”方式极为罕见。请不要将其扩展到任意其他情况。）