• The Django server (yet to be named)
• The client daemon (backtracd)
• The backup library itself (backuplib)

As I mentioned in the previous post, the Django server will be based around Celery. This will offer me the ability to schedule a task for a certain date and time in the future, which is a huge part of the implementation taken care of.

One thing I am yet to figure out, however, is how the server will communicate with the client daemons. At the moment I'm thinking XML-RPC, but it seems to me that it will end up a sort of mish-mash of different technologies - as the Django server will be exposing a JSON API (probably via Piston). I have also considered running Django on the clients too, so everything can use JSON/Piston, but this seems at the moment like an unnecessary burden.

Another thing to consider is the data that is to be transmitted between the nodes. Should I use a bunch of fields in a JSON object to represent the success/failure of a job, and how much data was backed up, etc.? Another option is to pickle a Result object of some sort - representing a result as a Python instance to be interpreted by the server. These are the things I love to think about when designing a system, and I'll certainly have plenty to write about in the inevitable report.

Something I can work on in the meantime, however, is backuplib - the client library that will actually copy the data. At the moment, it's just a wrapper around rsync/tar. That probably won't change, but it will need to get a bit smarter in the way it handles failures (and restores, of course) if it's going to make for the basis a decent backup system.

Backtrac is a pretty old project, and one whose name I am likely to change once I'm working on it. It was started in my first semester of university last year, when I didn't have enough work to keep me busy. I didn't know enough about Django back then, though, so it was left unfinished and only half-working.

Now I think I know enough to pull this off, and it's a chance to do something I really enjoy - while getting marks for it in the process! If everything goes to plan I'll be supervised by Dr. Julie Greensmith, which I'm really looking forward to.

I'll be using an app called Celery to handle the background tasks, which will be simple "pings" to the client machines, telling them to start a backup. Celery uses AMQP, provided by a broker such as RabbitMQ (which is written in Erlang). It's usually used for processing intensive computational tasks, but can also be put to use for jobs that need executing on a periodic basis, like backups.

I can tell you're getting too excited, so I'll stop. Rest assured, though, I will be writing about the discoveries I make along the way to finishing this ambitious project.

Beginnings

Firstly, I use virtualenv to keep the Python environments for the different projects separate, but that's a topic in itself. This isn't a requirement, but it helps if you have dependencies that you don't want installed system-wide, or projects that use different versions of a particular package.

You can read more about virtualenv and Django here.

The directory structure for my webapps directory (which lives inside /home/) looks something like this:

webapps/
|-- apache
|   |-- coral.conf
|   |-- coral.wsgi
|   |-- myuni.conf
|   `-- myuni.wsgi
|-- coral
|   |-- env
|   `-- repository
|-- myuni
|   |-- env
|   `-- repository

The two projects shown here are Coral Issue Tracker and MyUni. The directory for each project contains the virtual env for that project, and the repository. Alongside the project folders is a directory called apache, which holds two files for each project: the WSGI file (<project-name>.wsgi) and the Apache configuration file (<project-name>.conf).

Apache Configuration

Then, in the main Apache configuration, I use a directive that includes all these files automatically:

Include    /home/webapps/apache/*.conf

The individual project configuration files then contain a set of simple directives almost straight from the Django deployment documentation. For example, the MyUni configuration looks like this:

WSGIScriptAlias  /myuni        /home/webapps/apache/myuni.wsgi
Alias            /myuni/docs   /home/webapps/myuni/repository/docs/_build/html
Alias            /myuni/media  /home/webapps/myuni/repository/myuni/static

WSGI Files

The WSGI files are also pretty simple, though there are a couple of things worth explaining. Here is the WSGI file for the MyUni project:

import os
import sys
import site
 
os.environ['PYTHON_EGG_CACHE'] = '/tmp/myuni/egg_cache'
 
root_dir = '/home/webapps/myuni/env'
 
# with mod_wsgi >= 2.4, this line will add this path in front of the python path
site.addsitedir(os.path.join(root_dir, 'lib/python2.6/site-packages'))
 
sys.path.append('/home/webapps/myuni/repository')
 
os.environ['DJANGO_SETTINGS_MODULE'] = 'myuni.settings'
 
import django.core.handlers.wsgi
 
application = django.core.handlers.wsgi.WSGIHandler()

The site.addsitedir() call in the WSGI file adds the library path in the virtual environment to the python path, by inspecting it to find all the packages that are installed. I then append the path to the project's repository to the path also, as is standard practice in this type of deployment.

The only other thing worth mentioning here, is the PYTHON_EGG_CACHE variable that I add to the environment. This is required if you are using Python modules that are packaged up as eggs, so that Apache can uncompress them in a directory to which it can write (/tmp/ on Ubuntu is 777 by default). If you are not using eggs, you can skip this line out altogether. Alternatively, you can uncompress the eggs first:

$ unzip [package].egg

You can read more about this issue, and the workaround, here.

Summary

That's pretty much all there is to it. This setup also makes it pretty easy to deploy new versions of my projects, for which I have started using Fabric. I'll save that for another post though.

If you do things differently, then I'd love to hear about how - so please leave a comment!

I didn't want to change the name of the project settings file to settings_local, however, as it would mean changing the WSGI file on every server that the project runs on.

The local_settings Trick

It's a well-known trick to use a file called local_settings.py or something similar, with a piece of code at the bottom of the main settings file:

try:
    from local_settings import *
except:
    pass

to override the value of settings variables. This will not work, however, if you wish to extend a settings variable (for example, adding an app to INSTALLED_APPS). For this, I have found that the following ugly hack seems to do the job.

The Ugly Hack

Replace the snippet at the bottom of the main settings file with the following code

try:
    LOCAL_SETTINGS
except NameError:
    try:
        from local_settings import *
    except ImportError:
        pass

What this is doing is effectively checking for the presence of a variable called LOCAL_SETTINGS. The local_settings file then contains this code:

LOCAL_SETTINGS = True
from settings import *

This means that the local_settings file will only be imported once, and it has all the variables in the main settings file available to extend at will. For example, to enable the Django debug toolbar:

MIDDLEWARE_CLASSES += ('debug_toolbar.middleware.DebugToolbarMiddleware',)
INTERNAL_IPS = ('127.0.0.1',)
INSTALLED_APPS += ('debug_toolbar',)

The code is ugly (and results in the main settings file being parsed twice) but it works!

The model needed to represent the notion of a university module, which can be taught over a number of semesters and/or years, by different people, and with different students each time round. Some information remained common to each of these modules however, such as the code, name, and type of the module (single semester or full-year).

One solution that crossed my mind was to use a sort of template, and copy the common information over each time the module was taught. This would mean duplicating the common fields over a template model and the module's model, and copying the data every time a new "Module" object gets created. This just didn't sit well with me though, as I really don't like denomalising data when it's not absolutely necessary.

The solution I have settled on (for now, anyway) works something like this: I use the model template idea, but with a foreign key from the Module to it's template. Then, to access the fields on the template directly from the Module model, I use a combination of the Python property and a smart lambda function.

Here's a simplified version of the model to illustrate what I'm talking about:

class ModuleDefinition(models.Model):
    code = models.CharField(max_length=6, unique=True)
    name = models.CharField(max_length=200)
    credits = models.IntegerField(max_length=4)
    level = models.IntegerField(max_length=2)
    type = models.CharField(max_length=30, choices=MODULE_TYPE_CHOICES)
 
class Module(models.Model):
    code = property(lambda s: s._get_definition_property('code'))
    name = property(lambda s: s._get_definition_property('name'))
    credits = property(lambda s: s._get_definition_property('credits'))
    level = property(lambda s: s._get_definition_property('level'))
 
    definition = models.ForeignKey('ModuleDefinition')
    semester = models.CharField(max_length=20, choices=SEMESTER_CHOICES)
    year = models.CharField(max_length=20, choices=YEAR_CHOICES)
    convener = models.ForeignKey('auth.User', related_name='modules_convened')
    students = models.ManyToManyField('auth.User', blank=True, related_name='modules_taken')
 
    def _get_definition_property(self, property):
        return self.definition.__getattribute__(property)

In this model I've called the template a "definition", but the idea is the same. The _get_definition_property() method allows me to get a property from the parent definition programmatically, which is used in the lambda functions to add the fields from the definition to the model itself, as read-only properties. This allows me to access them as if they were a field stored on the model itself (i.e. module.code, module.name, etc.) which is handy in the templates where module.definition.code would get very old, very fast.

I'd be really interested to hear how others have tackled the same issue, as I'm sure it's not just restricted to university modules. Maybe soon I'll be making a post about how I model a timetable in Django...after all, how hard can it be!?

Oh, and if you're wondering, this code is for the MyUni project that is occupying me recently. I'm restricting myself to just thinking about how the models could work (and maybe writing little bit of code) until Rob and Ben come back to university.

The project has gone through several iterations since then, never really reaching any form of completeness. This time round, we are using proper version control (thanks to Mercurial and BitBucket), and trying to adhere to Django best practices as much as possible.

Straight from the repository website:

MyUni is a portal-style web application for Universities, offering students a central location to access all the relevant information about their course, modules, assignments, and staff members.

The project came about because of a passionate hatred for the WebCT system that our university is using, and a general feeling that things could be (and should be) better.

Django 1.2 will bring some really exciting new features, which I am sure would make developing and using MyUni a better experience. One of the things I'm most looking forward to using is the messages framework. I plan to use it for a bunch of things, possibly including form error messages...maybe.

Now, if only the release date would stay in the same place for more than a week!

I've been working with Django more and more lately, and I've written my first "commercial" application using my new favourite framework - a booking system for taught causes at the City Council. I'm really enjoying writing web applications with Django, and I'm sure this blog will start to resemble a web-developer's in the near future.

Also, I'm really looking forward to this year at university. As I understand it, there's a lot more work to be done, but the software engineering group project should be fun - as long as my "randomly chosen" team are happy with us using Python!

Something strange happens to me every time exams come around. I seem to pick up new projects, and just run with them. This time, I've become involved with a small group of people at university, writing a portal-style information system for universities. I suppose most people call this behavior procrastination, but I'm quite deeply in denial about that.

Ever since my post about the Backtrac Backup System, I've been really enjoying using Django. Something about it just makes developing for the web, well, exciting. That can only be good, right? I am the designated server administrator for this latest project, mostly due to the fact that I am the only one with a server to administer, and some of the things I've learned so far seem worthy of a mention here.  Firstly, we as developers were - how can I put it - stepping on each other toes somewhat. The project at this point had no version control, so we were just editing a bunch of files over SFTP. Obviously, some sort of Source Control Management was in order. I did some research, and decided that Git was a nice, modern alternative to the ever-popular SVN. It also meant that my server was constantly backed up by everyone on the team - but that's just a bonus!

So, I installed Git, and started a repository. A lot of effort went into learning how the system works, and more importantly, how to make it work for us. Directed Acyclic Graphs thankfully made some sense to me, so I could just about understance the documentation. I wrote some custom hooks, and a C Program to syncronise the web-server. I was happy, and absolutely certain that this was the solution to all our woes. I was mistaken. Git just didn't work the way I had hoped. The custom hooks were throwing permission errors all over the place, and my development team (read: my friend Rob Miles) was locked out of the repository. We made the decision yesterday to scrap Git, and go back to the previous system of editing the files over SFTP. We are always in constant communication when developing for the project, so it's not too big a deal, but I feel that I failed as an adminisrator. You see, as an admin your job isn't just to play with cool toys and loud servers - your primary purpose is to give the users what they need to work, and that is most certainly not what I achieved.

I'm glad I took the time to learn Git - and I'm sure it will help me later in life, in some way or another. It's just unfortunate that it didn't work out the way I had hoped for our project.