[Fixed]-Django Multiple Databases Fallback to Master if Slave is down

19👍

You are on the right track with using a router. I’m assuming the fact that your two db definitions are identical is just a typo.

(FYI, I’m going to refer to the database hierarchy using the more sensitive master->follower)

In your db_for_read() functions, you can check for connectivity to your follower. This might incur a bit more overhead, but that’s the cost of having auto-failover for a database. An example database definition would be:

DATABASES = {
'follower': {
        'ENGINE': 'django.db.backends.mysql',
        'NAME': 'follower',
        'USER': 'root',
        'HOST': '54.34.65.24',
        'PORT': '3306',
    },
'default': {
        'ENGINE': 'django.db.backends.mysql',
        'NAME': 'application',
        'USER': 'root',
        'HOST': '54.34.65.23',
        'PORT': '3306',
    },
}  

You can test the connection with a quick try/except like this example. A router using this that does what you need would look like:

from django.conf import settings
import socket


def test_connection_to_db(database_name):
    try:
        db_definition = getattr(settings, 'DATABASES')[database_name]
        s = socket.create_connection((db_definition['HOST'], db_definition['PORT']), 5)
        s.close()
        return True
    except (AttributeError, socket.timeout) as e:
        return False


class FailoverRouter(object):
    """A router that defaults reads to the follower but provides a failover back to the default"""

    def db_for_read(self, model, **hints):
        if test_connection_to_db('follower'):
            return 'follower'
        return 'default'

    def db_for_write(self, model, **hints):
        "Point all writes to the default db"
        return 'default'

    def allow_syncdb(self, db, model):
        "Make sure only the default db allows syncdb"
        return db == 'default'

This will still syncdb in master like you want. Also, you could make the logic for both db_for_read() and db_for_write() more complicated (like pick the follower db only for certain models that are queried for your reports.

I don’t know what overhead this test_connection() will cause for every read, since that will depend on the MySQL server and the timeout. Perhaps a better architecture is to cache these reports using memcached, or just work out the issues with the slave ever going down and update your database definitions in settings first.

Leave a comment