django-failover provides automatic failover for services used by your Django application. For example, if you are running a master and slave database, you might configure the slave to failover to the master. django-failover will monitor the slave and, if it goes down, will modify your settings to point the slave connection at the master. When the slave comes back online, django-failover will restore the settings and re-point the slave connection at the slave.

You can configure django-failover to perform monitoring, failover, and recovery via a middleware class. You can also specify exception classes (such as socket errors) that, when raised and logged using Python's logging module, will trigger monitoring. django-failover provides a log handler that you can hook up to your logger(s) for this purpose.

django-failover logs when a service goes down and when it comes back up. It also periodically re-logs a service outage if the service remains down.

Technical Note

Monitoring, failover, and recovery occur on a per-process basis. That is, django-failover does not store information about which services are down, which are up, and when they were last monitored in any global location such as a database or cache. Rather, each process running your Django application discovers outages for itself. Ping frequency settings (described below) only apply to a given process; your service may be pinged more often than your settings stipulate if multiple processes are running your Django application. This design is intended to keep django-failover lightweight and to prevent minor network hiccups from triggering failover globally across large-scale applications.

django-failover is best suited for environments where your webserver is serving multiple requests per process. A service outage will trigger failover within the process, the process will continue serving requests while in failover, and then will resume serving requests normally once the service comes back up. If you are working in a development environment where your webserver is reloading your code for each request (using a setting such as MaxRequestsPerChild in Apache), you should configure your webserver not to reload code while experimenting with django-failover. Otherwise, django-failover will start each request assuming all services are operational.

Requirements

django-failover has been tested with Django 1.3 and Python 2.6.

Installation

You can install django-failover with

pip install django-failover

or

easy_install django-failover

This will add failover to your Python path. Add 'failover' to your INSTALLED_APPS if you want to run django-failover's test suite.

Service classes

To use django-failover, you define and register service classes, each class corresponding to a service used by your Django application, such as a database, a cache, a message broker, etc. Each service class should inherit from failover.services.Service and override it as necessary. The base Service class provides the following attributes, methods, and default behavior:

Service.ping()

This method pings the service to see if it is available. Must raise an exception if the service is down, otherwise the service is assumed to be up. Default behavior: attempts to make a socket connection using Service.SOCKET_HOST (default None) and Service.SOCKET_PORT (default None). Depending on your service class, you should either define values for SOCKET_HOST and SOCKET_PORT, or override the ping method entirely to check the service in some other way.

Service.failover()

This method performs failover if the service is down. Default behavior: modifies django.conf.settings, setting name=value for each name/value pair in Service.FAILOVER_SETTINGS (which is an empty dictionary by default).

Service.recover()

This method performs recovery when a service comes back up. Default behavior: restores the settings that were overwritten during failover.

Service.MONITORING_PING_FREQUENCY

Determines the maximum frequency (in seconds) with which django-failover will ping the service during normal monitoring. This means that, when monitoring is triggered, django-failover will not ping the service unless at least MONITORING_PING_FREQUENCY seconds have passed since the last ping. Default: settings.FAILOVER_MONITORING_PING_FREQUENCY (or 120 if undefined).

Service.OUTAGE_PING_FREQUENCY

Like MONITORING_PING_FREQUENCY, but applies when the service is down and is being monitored for recovery. Default: settings.FAILOVER_OUTAGE_PING_FREQUENCY (or 30 if undefined).

Service.ERROR_PING_FREQUENCY

Like MONITORING_PING_FREQUENCY, but applies when monitoring is triggered by an exception. Default: settings.FAILOVER_ERROR_PING_FREQUENCY (or 5 if undefined).

Service.outage_middleware_class

You can assign a middleware class to this attribute and that middleware will run when the service goes down (assuming you're using django-failover's middleware). The middleware may define process_request and/or process_view methods. This could be useful if you wanted to close down some sections of your site during the outage. Default: None.

django-failover provides a few ready-to-use service classes.

failover.services.cache.Memcached

This service class will ping your Memcached node(s) and let you know if any are down. Requires python-memcached client. Doesn't perform any failover, as the cache fails silently and thus automatically fails over to the database. This service class is useful for notifying you of Memcached outages.

failover.services.celery.Celery

Can be used with django-celery. Pings your message broker (using settings.BROKER_HOST and settings.BROKER_PORT). If the message broker is down, sets CELERY_ALWAYS_EAGER to True, meaning your celery tasks will execute locally during the outage.

failover.services.db.Database

Fails over from one of your database connections in settings.DATABASES to another. By default, fails over from a connection with alias "slave" to a connection with alias "default". To failover using different db aliases, subclass Database and define DB_ALIAS and FAILOVER_DB_ALIAS on the subclass.

failover.services.db.MySQL

A subclass of Database that sets sensible timeouts when pinging in order to reduce false positives and avoid hanging. Database connection timeout settings are backend-specific; if you use a backend other than MySQL and want to set timeouts, subclass Database and override the set_timeout method.

Registering service classes

You can register a service class in several ways. The easiest way is to define FAILOVER_SERVICES in your settings file. FAILOVER_SERVICES should be a tuple of service classes in dot notation:

FAILOVER_SERVICES = (
    "failover.services.cache.Memcached",
    "failover.services.celery.Celery",
    "myproject.myapp.MyServiceClass",
)

Alternatively, import failover.monitor.ServiceMonitor and call:

ServiceMonitor.register(MyServiceClass)

for each service class you want to register. You can also use this as a decorator:

@ServiceMonitor.register
class MyServiceClass(Service):

Middleware configuration

Add failover.middleware.FailoverMiddleware to your MIDDLEWARE_CLASSES, at or near the beginning:

MIDDLEWARE_CLASSES = (
    "failover.middleware.FailoverMiddleware",
    "django.middleware.common.CommonMiddleware",
    ...
)

This will trigger monitoring, failover, and recovery before each request (taking into account your ping frequency settings).

Logging configuration

To trigger monitoring when certain exception classes are raised, add failover.log.FailoverHandler to one of your loggers. To capture django-failover's own logging, add a handler to the "failover" logger. Service outages are logged as CRITICAL and recoveries are logged as INFO, so you probably want to set your logger and handler levels to INFO so you will know when services recover.

LOGGING = {
    'version': 1,
    'handlers': {
        'mail_admins': {
            'level': 'INFO',
            'class': 'django.utils.log.AdminEmailHandler',
        },
        'failover': {
            'level': 'ERROR',
            'class': 'failover.log.FailoverHandler',
    },
    'loggers': {
        'django.request': {
            'handlers': ['mail_admins', 'failover'],
            'level': 'ERROR',
        },
        'failover': {
            'handlers': ['mail_admins'],
            'level': 'INFO',
        },
    },
}

By default, FailoverHandler will trigger monitoring for the following exception classes:

socket.error
django.db.DatabaseError
urllib2.URLError

To add to this list, define FAILOVER_OUTAGE_EXTRA_EXCEPTION_CLASSES in settings. To override the list, define FAILOVER_OUTAGE_EXCEPTION_CLASSES.

Monitoring decorators

You can also trigger monitoring using several decorators django-failover provides. Decorating a function with failover.decorators.monitor will trigger monitoring prior to the function being called. Wrapping the function with failover.decorators.recover_from_outages will monitor services that are currently down, but won't monitor services that are operating normally. recover_from_outages is useful in conjunction with exception-based monitoring, as the latter provides no trigger for checking when a service recovers. If you are only using django-failover's middleware, you probably don't need either of these decorators.

Additional Settings

FAILOVER_OUTAGE_LOGGING_FREQUENCY

Controls how often, after a service outage is discovered, the outage will be re-logged if the service remains down. Default: 3600 seconds (1 hour).

FAILOVER_MONITORING_PING_FREQUENCY

Controls the ping frequency during normal monitoring on service classes that don't define their own value. Default: 120 seconds.

FAILOVER_OUTAGE_PING_FREQUENCY

Controls the ping frequency during outages on service classes that don't define their own value. Default: 30 seconds.

FAILOVER_ERROR_PING_FREQUENCY

Controls the ping frequency for exception-triggered monitoring on service classes that don't define their own value. Default: 5 seconds.