最近我们的Django项目供Java Sofa应用进行tr调用时, 经常会出现一个异常: django.db.utils.OperationalError: (2006, 'MySQL server has gone away'). 本文记录了分析, 本地重现与解决此问题的全过程.

原因分析:

Django在1.6引入长链接(Persistent connections)的概念, 可以在一个HTTP请求中一直用同一个连接对数据库进行读写操作.
但我们的应用对数据库的操作太不频繁了, 两次操作数据库的间隔大于MySQL配置的超时时间(默认为8个小时), 导致下一次操作数据库时的connection过期失效.

Our databases have a 300-second (5-minute) timeout on inactive connections. That means, if you open a connection to the database, and then you don't do anything with it for 5 minutes, then the server will disconnect, and the next time you try to execute a query, it will fail.

重现问题:

设置mysql `wait_timeout`为10s

在macOS上的mysql配置文件路径: /usr/local/etc/my.cnf

# Default Homebrew MySQL server config
[mysqld]
# Only allow connections from localhost
bind-address = 127.0.0.1
wait_timeout = 10
interactive_timeout = 10

重启mysql:

➜  ~ brew services restart mysql
Stopping `mysql`... (might take a while)
==> Successfully stopped `mysql` (label: homebrew.mxcl.mysql)
==> Successfully started `mysql` (label: homebrew.mxcl.mysql)

检查wait_timeout的值是否已被更新.

mysql> show variables like '%wait_timeout%';
+--------------------------+----------+
| Variable_name            | Value    |
+--------------------------+----------+
| innodb_lock_wait_timeout | 50       |
| lock_wait_timeout        | 31536000 |
| wait_timeout             | 10       |
+--------------------------+----------+
3 rows in set (0.00 sec)

重现exception:

>>> XXX.objects.exists()
True
>>> import time
>>> time.sleep(15)
>>> XXX.objects.exists()
True
>>> XXX.objects.exists()
...
django.db.utils.OperationalError: (2013, 'Lost connection to MySQL server during query')
>>> XXX.objects.exists()
...
django.db.utils.OperationalError: (2006, 'MySQL server has gone away')

有意思的一个点是, sleep 10s 之后, 第一次操作数据库, 会出现(2013, 'Lost connection to MySQL server during query’)异常. 之后再操作数据库, 才会抛出(2006, 'MySQL server has gone away’)异常.

解决问题:

第一个最暴力的方法就是增加mysql的wait_timeout让mysql不要太快放弃连接. 感觉不太靠谱, 因为不能杜绝这种Exception的发生.

第二个办法就是手动把connection直接关闭:

>>> Alarm.objects.exists()
True
>>> from django.db import connection
>>> connection.close()
>>> time.sleep(10)
>>> Alarm.objects.exists()
True
>>>

发现不会出现(2006, 'MySQL server has gone away’)异常了, 但总感觉还是不够优雅.
最终决定在客户端(Django), 设置超时时间(CONN_MAX_AGE: 5)比mysql服务端(wait_timeout = 10)小:

DATABASES = {
	'default': {
		'ENGINE': 'django.db.backends.mysql',
		'CONN_MAX_AGE': 5,
		<other params here>
	}
}

但很奇怪没有生效??? 看了源代码, 发现只有在request_started(HTTP request)和request_finished的时候, 在close_if_unusable_or_obsolete才用到CONN_MAX_AGE并去验证时间关闭connection.

具体代码见: python3.6/site-packages/django/db/__init__.py#64

# Register an event to reset transaction state and close connections past
# their lifetime.
def close_old_connections(**kwargs):
	for conn in connections.all():
		conn.close_if_unusable_or_obsolete()


signals.request_started.connect(close_old_connections)
signals.request_finished.connect(close_old_connections)

而我的代码是处理一个任务而不是HTTP请求, 所以不会触发这个signal. 于是我写了一个装饰器, 在任务的开始和结束的时候, 关闭所有数据库连接.

from django.db import connections


# ref: django.db.close_old_connections
def close_old_connections():
    for conn in connections.all():
        conn.close_if_unusable_or_obsolete()


def handle_db_connections(func):
    def func_wrapper(request):
        close_old_connections()
        result = func(request)
        close_old_connections()

        return result

    return func_wrapper

# ------割-------
@handle_db_connections
def process_trsbrain_request(request):
    ...

ps. CONN_MAX_AGE默认其实为0, 意味着默认在http请求和结束时会关闭所有数据库连接.

其他:

django.db中connection和connections的区别???

connection对应的是默认数据库的连接, 用代码表示就是connections[DEFAULT_DB_ALIAS]
connections对应的是setting.DATABASES中所有数据库的connection

原因分析:

重现问题:

设置mysql wait_timeout为10s

重现exception:

解决问题:

其他:

ref:

设置mysql `wait_timeout`为10s