咨询：Airflow MySQL 2013 Lost connection 报错

[yearn21c]

Hi @pingzh，
打扰您，目前我们生产环境 Airflow 频繁报 MySQL 2013 错误：Lost connection to MySQL server during query，多发生在凌晨批量任务调度阶段。
日志：
MySQLdb.OperationalError: (2013, 'Lost connection to MySQL server during query')
sqlalchemy.exc.OperationalError: (MySQLdb.OperationalError) (2013, 'Lost connection to MySQL server during query')

[latent-9]

A (2013, 'Lost connection to MySQL server during query') against the Airflow metadata database is almost always a pooled connection gone stale, not a network fault. MySQL closes idle connections after wait_timeout, but SQLAlchemy keeps them in its pool and can hand one back mid-query, which surfaces as 2013. The nightly-only timing fits this: connections sit idle through the quiet window, the server reaps them, and the batch then reuses a dead one.

Two settings cover the common cases:

[database]
sql_alchemy_pool_pre_ping = True
sql_alchemy_pool_recycle = 1800

pool_pre_ping makes SQLAlchemy test a connection before use and reconnect transparently if it is dead. pool_recycle should be lower than MySQL's wait_timeout so connections refresh before the server drops them (check SHOW VARIABLES LIKE 'wait_timeout').

If it still shows up on large runs, raise MySQL's max_allowed_packet, since big serialized DAGs or XComs can exceed it and appear as a mid-query disconnect. net_read_timeout is worth a look too. Start with pool_pre_ping, which clears the majority of 2013 errors on the metadata DB.