Because of the possibilty that extension module writers have not written their code to take into consideration it being used from multiple sub interpreters, the safest approach is to force all WSGI applications to run within the same application group, with that preferably being the first interpreter instance created by Python.

To force a specific WSGI application to be run within the very first Python sub interpreter created when Python is initialised, the WSGIApplicationGroup directive should be used and the group set to ‘%{GLOBAL}’.

WSGIApplicationGroup %{GLOBAL}

If it is not feasible to force all WSGI applications to run in the same interpreter, then daemon mode of mod_wsgi should be used to assign different WSGI applications to their own daemon processes. Each would then be made to run in the first Python sub interpreter instance within their respective processes.

I did try WSGIApplicationGroup %{GLOBAL} first, but (assuming I implemented it correctly), the problem remained. So I tried WSGI daemon mode and the process has proved stable.

So, I decided to try something kinda crazy: subclassing django.contrib.auth.models.AnonymousUser  and overriding the “anonymity”.  AnonymousUser really takes advantage of Python duck-typing by mimicking the django.contrib.auth.models.User without being a Django model itself, and so isn’t tied to Django’s database.  Here’s what I came up with:

from django.contrib.auth.models import AnonymousUser

class DetachedUser(AnonymousUser):
    """
    Implements a user-like object for user authentication
    when linkage of a backend user identity to a Django user
    is undesirable or unnecessary.
    """

    # mark as not hashable -- see also __eq__() and __ne__(), below
    __hash__ = None 

    # is_active might matter in some contexts, so override
    is_active = True 

    #
    # AnonymousUser sets username to '' and id to None, so we need to at least
    # override those values.
    #
    def __init__(self, username):
        self.username = username
        self.id = username

    def __unicode__(self):
        return self.username

    #
    # is_anonymous() and is_authenticated() are key to distinguishing truly
    # anonymous/unauthenticated users from known/authenticated ones.
    #
    def is_anonymous(self):
        return False

    def is_authenticated(self):
        return True

    #
    # __eq__ and __ne__ are related to hashing, so be consistent with __hash__, above.
    #
    def __eq__(self, other):
        return NotImplemented

    def __ne__(self, other):
        return NotImplemented

    #
    # Some django.contrib.auth code may call this method,
    # e.g, to update the last login time
    #
    def save(self):
        pass

Now I can code the get_user and authenticate methods of my custom authentication backend to return DetachedUser objects instead of Django users. So far, so good.

Example Architecture

• Front end web server (www.example.com): Apache 2.2 + mod_proxy, mod_proxy_balancer, mod_ssl.
• Back end application servers (apps-01.example.com, apps-02.example.com): Apache 2.2 + mod_wsgi, mod_ssl; Python 2.6; Django 1.3.1.
• Backend database server.
• Additional requirements: Remote user authentication; SSL and non-SSL proxies.

Let’s start with the application servers and deal with the front end later.

Application Servers

Obviously both app servers will be configured the same way.  How to keep them in sync will be discussed briefly.

Django Settings Module

In order for Django to properly create fully-qualified URLs for the front-end client, you must set:

USE_X_FORWARDED_HOST = True

This setting, new in Django 1.3.1, affects the get_host() and build_absolute_uri() methods of django.http.HttpRequest.  If not set, Django will use the value of the HTTP_HOST or SERVER_NAME variables, which are most likely set to the host name of the app server, not the front end.

If you’re using Django’s RemoteUserMiddleware and RemoteUserBackend for authentication, you will need to replace RemoteUserMiddleware with a custom subclass:

from django.contrib.auth.middleware import RemoteUserMiddleware

class ProxyRemoteUserMiddleware(RemoteUserMiddleware):
    header = 'HTTP_REMOTE_USER'

Then update your settings:

MIDDLEWARE_CLASSES = (
    'path.to.ProxyRemoteUserMiddleware',
    )

(It is possible to avoid this by setting REMOTE_USER on the app web server to the value of HTTP_REMOTE_USER, but here I will assume a default setup.)

If you’re using Django’s “sites” framework, you will probably want to set SITE_ID to correspond to the front-end site.  And if your WSGIScriptAlias path differs from the proxied path on the front-end server (not covered in detail here), you may have to use FORCE_SCRIPT_NAME (check the docs).

Django Application Modules and Templates

If your code or templates contain references to REMOTE_ADDR, REMOTE_USER or other server variables (via HttpRequest.META) affected by proxies, you will probably have to change them.  If you’re using Django’s RemoteUserMiddleware or the ProxyRemoteUserMiddleware subclass shown above, you should probably code with request.user.username instead of request.META['REMOTE_USER']; otherwise, you’ll want to reference HTTP_REMOTE_USER.  REMOTE_ADDR will be set to the IP address of the app server, not the proxy front-end; instead you will have to use HTTP_X_FORWARDED_FOR, which can have multiple comma-separated values.

Django Projects and Python Environments

Since we’ve got two app servers, each will have its own Python environment (created with virtualenv) and Django project.  In my setup I decided to serve the Django MEDIA_ROOT from network storage mounted at the same point on each server to avoid synchronization issues.  Otherwise, it seems OK to keep each instance separate (YMMV).  I use Fabric for ensuring that the Python environments and Django projects stay in sync across the two servers.  The precise way you do this syncing depends on your preferences, the available tools, etc.

Apache Configuration

The Apache config on each app server follows the normal Django/WSGI pattern, so I’ll skip the details here.  Note that while it is possible for WSGIScriptAlias path on the app server to differ from the proxied path on the front-end web server (which we’ll get to), this introduces some additional complexities which we will avoid here.  Some issues can be handled on the reverse proxy (front-end) server by Apache directives such as ProxyPassReverse and ProxyPassReverseCookiePath, but you may also need to use Django’s FORCE_SCRIPT_PATH setting in your project settings module.

Front-end Server

At this point you should have working Django projects on each app server under both SSL and non-SSL virtual hosts.  Now we’re going to set up the reverse proxy and load balancing on the front-end server.

Let’s assume your apps are served under the path /webapps on both port 80 and port 443 (SSL) virtual hosts.

Then, you can add to your port 80 virtual host:

<Proxy balancer://django-http>
    BalancerMember http://apps-01.example.com/webapps route=http-1
    BalancerMember http://apps-02.example.com/webapps route=http-2
</Proxy>

<Location /webapps>
    ProxyPass balancer://django-http stickysession=sessionid
    ProxyPassReverse http://apps-01.example.com/webapps
    ProxyPassReverse http://apps-02.example.com/webapps
    ProxyPassReverseCookieDomain apps-01.example.com www.example.com
    ProxyPassReverseCookieDomain apps-02.example.com www.example.com
</Location>

And to your SSL virtual host on port 443:

<Proxy balancer://django-https>
    BalancerMember https://apps-01.example.com/webapps route=https-1
    BalancerMember https://apps-02.example.com/webapps route=https-2
</Proxy>

<Location /webapps>
    ProxyPass balancer://django-https stickysession=sessionid
    ProxyPassReverse https://apps-01.example.com/webapps
    ProxyPassReverse https://apps-02.example.com/webapps
    ProxyPassReverseCookieDomain apps-01.example.com www.example.com
    ProxyPassReverseCookieDomain apps-02.example.com www.example.com
</Location>

This isn’t the only way to do it, of course, and you may have different requirements, but I’ve tried to cover the basics.

• Don’t reverse engineer the system.  By all means, read the code, just don’t try to outsmart it.
• If you can help it, don’t use methods or attributes marked as private (yes, I have a beef with Django’s Model._meta).
• Don’t use a third-party product for something you can do relatively easily without it.  Stop and ask yourself if the coding you may save is worth the headache of maintaining the product through framework upgrades, etc.
• Keep an eye on framework developments.  Watch the roadmap.  Read the release notes!  I’ve been pleasantly surprised a number of times to find out about new (or existing) features by re-reading documentation.
• Participate in the core community.  Everybody benefits from improvements to the core framework.  This is not to say that every good product should ultimately aim for inclusion in the framework, but with a strong group of core developers new features are subjected to critical analysis, rigorous testing and benefit from wide usage and long term support.

I had recently made changes to the get_absolute_url() methods on a couple of application models (Django 1.3/Python 2.4).  For various reasons, the admin UI for the app is on an intranet site, which is a different host from the public site (represented by a different Django Site object).  The changes involved calling a new function that uses the current site object to determine the appropriate domain for constructing a fully-qualified URL for get_absolute_url().  As originally implemented, I set a variable at the module level in models.py:

BASE_PUBLIC_URL = 'http://%s/path/to/app' % get_public_domain()

where get_public_domain() imports the Site model class from django.contrib.sites, calls Site.objects.get_current() and returns the appropriate public domain.  The get_absolute_url() methods then used BASE_PUBLIC_URL in constructing the final URL.

This worked fine in the normal application contexts, i.e., the admin site and the public site.  However, a custom management command which updates app data from an external source raised an ImportError on the relevant model.  The significant part of the traceback was as follows:

  File "/path/to/pyenv/lib/python2.4/site-packages/django/contrib/sites/models.py", line 25, in get_current
    current_site = self.get(pk=sid)

Ad hoc testing showed the ImportError to be a red herring — and, in any case, it didn’t square with the fact that the rest of the app (minus the admin command) was functional.

Now, the management command doesn’t actually call get_absolute_url(), so I figured that maybe the solution was to wrap the base public URL in a memoized function, so that the current site object is accessed lazily:

@memoize
def get_base_public_url():
    from path.to.mymodule import get_public_domain
    return 'http://%s/path/to/app' % get_public_domain()

That did the trick.  I’m still not sure exactly why the sites framework barfed, and it doesn’t seem worth digging for …

I suspect there may be some moaning around the deprecation of function-based generic views in favor of class-based views.   Class-based views make sense, but it looks like there will be a little pain in the transition, partly because the keyword arguments for Django’s built-in generic view functions don’t map exactly to the generic view class keyword arguments.  It would have been nice to provide a little smoother transition there.  Also, the class-based view documentation is rather dense because you have to refer to the mixin classes that compose the actual generic view class you want to use.  I’m sure it will get easier with time, but it does feel like a jump in complexity that could make generic views more difficult for new users.    For example, where I did this before in a URLconf module:

from django.conf.urls.defaults import *
from django.views.generic.simple import direct_to_template

urlpatterns = patterns('',
    (r'^status/$', direct_to_template,
     {'template': 'sitetest/status.txt', 'mimetype': 'text/plain'}),
)

I now have to do something like this:

from django.conf.urls.defaults import *
from django.views.generic.base import TemplateView

class PlainTextTemplateView(TemplateView):
    """A plain text generic template view."""
    def render_to_response(self, context, **kwargs):
        return super(PlainTextTemplateView, self).render_to_response(
            context, content_type='text/plain', **kwargs
            )

urlpatterns = patterns('',
    (r'^status/$', PlainTextTemplateView.as_view(template_name='sitetest/status.txt')),
)

While class-based views may have been a step in the right direction for the framework, I wonder how it will play out.

ldaps://example.com
ou=CompanyPeople,dc=example,dc=com

In this case, however, the users spanned two different top-level containers or “domains”:

ou=CompanyPeople,dc=example,dc=com
ou=OtherPeople,dc=example,dc=com

So, I tried setting the base DN to the top level:

dc=example,dc=com

but authentication failed with this ugly error in the log:

[ldap_search_ext_s() for user failed][Operations error]

It took some hunting, but I finally found that if you want to query the Active Directory “Global Catalog” (GC) via LDAP, you have to use port 3268 or 3269 (LDAPS) instead of the usual default port 389 or 636. So, the working URL and base DN are:

ldaps://example.com:3269
dc=example,dc=com

Django – The de facto standard for Python web application development.  I’ve learned a lot from studying its code.  Includes a library of useful utilities (django.utils) that can used outside of web application contexts (e.g., check out django.utils.datastructures.SortedDict).

sphinx – Also a de facto standard in the Python universe.  It’s made me appreciate reStructured Text and improve my code documentation practices.  Ironically I find its own documentation rather hard to use.

virtualenv – How did we manage without it?

pip – Better package management than easy_install.

ipython – Worth it for the command history alone.

decorator – Almost essential for writing decorators, especially if you’re on Python < 2.5.

Fabric – A great addition to the developer’s or sysadmin’s toolkit.

lxml – For XML processing, I almost never use Python’s builtin XML libraries.

pycurl – Brings the power of libcurl into Python, filling gaps left by urllib/urllib2 and httplib (e.g., multiple asynchronous requests, multipart form data).

xlrd, xlwt — Good API for MS Excel processing.  Unfortunately, no support (yet) for Excel 2007 XML format.

simplejson – The standard JSON library for Python < 2.6.

py.test — Anything that makes writing and running unit tests easier is very good.

unittest2 — Makes available to Python 2.4-2.6 the significant enhancements made to the standard unittest module in Python 2.7.

I also want to thank Christof Gohlke for his “Unofficial Windows Binaries” site, since up-to-date versions of lxml and pycurl would be difficult to use on Windows without his builds.

And finally, there are essential libraries that I depend on without normally using directly: MySQL-Python, pysqlite (stuck on CentOS 5/Python 2.4), python-ldap, docutils, setuptools.