django-pancake

By Adrian Holovaty <[email protected]>

Library for "flattening" Django templates.

Run make_pancakes.py with an output directory name, and it will fill that directory with a "flat" version of each template -- a pancake. A pancake is a template in which:

• Template inheritance is fully expanded. {% extends %} and {% block %} tags are gone, and the parent templates are fully folded in.
• {% include %} tags are gone, with their contents fully folded in.
• Template comments are removed: {% comment %} syntax and {# short syntax #}.

Pancakes behave exactly the same as the original templates. But they might render more quickly, because they avoid the overhead of template inheritance and template includes at rendering time. (Whether they do or don't actually render more quickly depends on other factors such as whether you're using template caching.)

Think of it as "denormalizing" your templates, as a database administrator might denormalize SQL tables for performance. You give up the DRY principle but make up for it in faster application speed.

You can also use pancakes as a learning tool. You can examine a pancake to see how all of your template blocks fit together in the resulting "compiled" template.

Pros and cons

Pros:

• Might make run-time template rendering faster. In some cases, the rendering is significantly faster. See "How fast is the speed improvement?" below.
• Can help you understand how a complex template inheritance structure works. You can examine a pancake to see how all of the template blocks fit together.
• Not a huge commitment. Give it a shot and see whether it makes things faster for you.

Cons:

• Might not actually result in a significant speed increase.
• Makes your deployment more complex, as you now have to manage generated templates and run django-pancake to generate them whenever you deploy.
• May require you to change the way you write templates, specifically by removing dynamic {% include %} and {% extends %} tags. See "Limitations" below.
• (Philosophical.) Django really should do this in memory rather than compiling to templates on the filesystem. See "Related projects" below.

How fast is the speed improvement?

It depends on what you're doing in templates, and it depends on whether you have template caching enabled.

If you're just using basic template inheritance and includes, you should expect a tiny/negligible performance improvement -- on the order of 10 milliseconds for a single template render. In this case, it may not be worth the added complexity it introduces to your deployment environment.

But if you're doing crazy things -- say, having a template with a template tag that loads other templates, within a loop, with each subtemplate being in an inheritance structure three levels deep -- then you might see a significant benefit that makes it worth the complexity.

At EveryBlock, I found it sped a certain type of page up by 200 milliseconds, which is pretty great. But that was on my laptop, and unfortunately, the speed increase went away when the code was deployed onto the production servers (which were running the cached template loader). So django-pancake basically had no positive effect for us, and I disabled it. But, who knows, it might help somebody else.

Usage

1. Generate the pancakes. Pass it the directory that contains your source templates and the directory you want pancakes to be generated in:

   python make_pancakes.py /path/to/source/directory /path/to/pancake/directory
   

2. Point Django at the pancake directory:

   TEMPLATE_DIRS = [
       '/path/to/pancake/directory',
   ]
   

Limitations

If you want django-pancake to work with your templates, make sure your templates do the following:

• Avoid using block.super in anything but a standalone variable. This is OK:

  {{ block.super }}
  

  But these statements are not:

  {% if block.super %}
  {{ block.super|lower }}
  {% some_other_tag block.super %}
  

  If you use block.super in one of these prohibited ways, django-pancake will not detect it and will generate your templates as if everything is OK. But you'll likely get odd behavior when the template is rendered. The problem is that block is no longer a variable in the pancakes, so it'll be evaluated as False.

• Avoid dynamic {% extends %} tags -- that is, when the parent template name is a variable. Example: {% extends my_template_name %} (note the lack of quotes around my_template_name). If you do this, django-pancake will raise a PancakeFail exception.

• Likewise, avoid dynamic {% include %} tags. Example: {% include some_include %}. If you do this, django-pancake will raise a PancakeFail exception.

• Don't use the "only" keyword in {% include %} tags. If you do, django-pancake won't raise an exception, but it'll merely output the same {% include %} tag, so you don't get the benefit of flattening.

Other notes

Obviously, pancakes are very redundant -- each of them includes all of the markup from the base template(s), etc. You shouldn't check pancakes into revision control or hand-edit them. They're purely for performance and should be handled as any automatically generated code.

Related projects

Ideally, django-pancake wouldn't exist, and Django would do this "flattening" itself, along with providing a more robust template compilation step. That's a significantly harder problem, and we're working on it. The goal is for django-pancake not to have to exist.

But, in the meantime, here are some related projects that have gone down that road:

• templatetk (https://github.com/mitsuhiko/templatetk/)
• django-template-preprocessor (https://github.com/citylive/django-template-preprocessor/)