Approach: Building a toy template engine in Python. | Alex Michael

03 Apr 2013

TL;DR Let's build a simplistic template engine and explore how it works under the hood – in case you were ever wondering.

If you wish to dive straight into the code, then Github is your friend.

Language design

Our language is pretty basic. We will have two types of tags, variables and blocks.

<!-- variables are surrounded by `{{` and `}}` --><div>{{my_var}}</div><!-- blocks are surrounded by `{%` and `%}` -->
{% each items %}<div>{{it}}</div>
{% end %}

Most blocks need to be closed as in the example and the closing tag is just the word end.

Our template engine will be able to handle basic loops and conditionals. We will also add support for callables in blocks – personally, I find it handy being able to call arbitrary Python functions in my templates.

Loops

Loops allow for iterations over collections or iterable objects.

{% each people %}<div>{{it.name}}</div>
{% end %}
{% each [1, 2, 3] %}<div>{{it}}</div>
{% end %}
{% each records %}<div>{{..name}}</div>
{% end %}

In the example above, people is the collection and it refers to the current item in the iteration. Dotted paths in names will resolve to nested dictionary attributes. Using ’..’ we can access names in the parent context.

Conditionals

Conditionals need no explanation. Our language will support if and else constructs, and the following operators: ==, <=, >=, !=, is, >, <.

{% if num > 5 %}<div>more than 5</div>
{% else %}<div>less than or equal to 5</div>
{% end %}

Callables

Callables can be passed via the template context and get called with positional or keyword arguments in the template. Call blocks do not need to be closed.

<!-- supports positional arguments... --><divclass='date'>{% call prettify date_created %}</div><!-- ...and keyword arguments --><div>{% call log 'here' verbosity='debug' %}</div>

Theory

Before delving into the details of how our engine will compile and render the templates we need to talk a bit about how we will represent a compiled template in memory.

Compilers use Abstract Syntax Trees to represent the structure of a computer program. ASTs are the product of lexical analysis on the source code. An AST has a lot of advantages over source code since it does not contain unnecessary textual elements such as delimiters; furthermore, nodes in the tree can be enhanced with attributes without altering the actual code.

We will parse and analyse the template and build such a tree to represent the compiled template. To render the template we will walk the tree, pass the correct context into each node and spit out HTML.

Tokenizing the template

The first step in parsing the template is splitting the contents into fragments. Each fragment can either be arbitrary HTML or a tag. To split the contents we will use regular expressions and the split() function.

VAR_TOKEN_START='{{'VAR_TOKEN_END='}}'BLOCK_TOKEN_START='{%'BLOCK_TOKEN_END='%}'TOK_REGEX=re.compile(r"(%s.*?%s|%s.*?%s)"%(VAR_TOKEN_START,VAR_TOKEN_END,BLOCK_TOKEN_START,BLOCK_TOKEN_END))

Let’s analyze TOK_REGEX a bit. We can see that there’s an option between a variable tag and a block tag. That makes sense – we want to split on either variables or blocks. We surround the option with capturing parenthesis to include the matching text in the resulting tokens. The ? inside the options is for non-greedy repetition. We want our regex to be lazy and stop on first match, so, for example, we can extract variables inside blocks. This is a very nice explanation of how to control the greediness of regular expressions.

The following example shows our regular expression in action:

>>>TOK_REGEX.split('{% each vars %}<i>{{it}}</i>{% endeach %}')['{% each vars %}','<i>','{{it}}','</i>','{% endeach %}']

We will encapsulate each fragment of text in a Fragment object. This object will determine the fragment type and prepare the fragment for consumption by the compile function. Fragments can be one of four types:

VAR_FRAGMENT=0OPEN_BLOCK_FRAGMENT=1CLOSE_BLOCK_FRAGMENT=2TEXT_FRAGMENT=3

Building the AST

Once we’ve tokenized the template content it’s time to go through each fragment and build the syntax tree. We will use a Node class to serve as the base class for tree nodes, and then create concrete subclasses for each possible node type. A subclass should provide implementations for process_fragment() and render(). process_fragment() is used to further parse the fragment contents and store necessary attributes on the Node object. render() is responsible for converting the node to HTML using the provided context.

A subclass could optionally implement enter_scope() and exit_scope() which are hooks that get called by the compiler during compilation and should perform further initialization and cleanup. enter_scope() is called when the node creates a new scope (more on this later), and exit_scope() is called when the node’s scope is about to get popped off the scope stack.

This is our base Node class:

class_Node(object):def__init__(self,fragment=None):self.children=[]self.creates_scope=Falseself.process_fragment(fragment)defprocess_fragment(self,fragment):passdefenter_scope(self):passdefrender(self,context):passdefexit_scope(self):passdefrender_children(self,context,children=None):ifchildrenisNone:children=self.childrendefrender_child(child):child_html=child.render(context)return''ifnotchild_htmlelsestr(child_html)return''.join(map(render_child,children))

As an example of a concrete subclass, this is a Variable node:

class_Variable(_Node):defprocess_fragment(self,fragment):self.name=fragmentdefrender(self,context):returnresolve_in_context(self.name,context)

In order to determine the node type (and hence instantiate the correct node class) we will look at the fragment type and text. Text and Variable fragments directly translate to text nodes and variable nodes. Block fragments need a bit more processing – their type is denoted by the block command which is the first word in the fragment text. For example, the fragment:

is a block node of ‘each’ type since the block command is each.

A node can also create a scope. During compiling, we keep track of the current scope and new nodes are added as children of that scope. Once we encounter a correct closing tag we close the scope, pop it off the scope stack and set the top of the stack as the new current scope.

defcompile(self):root=_Root()scope_stack=[root]forfragmentinself.each_fragment():ifnotscope_stack:raiseTemplateError('nesting issues')parent_scope=scope_stack[-1]iffragment.type==CLOSE_BLOCK_FRAGMENT:parent_scope.exit_scope()scope_stack.pop()continuenew_node=self.create_node(fragment)ifnew_node:parent_scope.children.append(new_node)ifnew_node.creates_scope:scope_stack.append(new_node)new_node.enter_scope()returnroot

Rendering

The last step in the pipeline is to render the AST to HTML. We visit all the nodes in the AST and call render() passing as a parameter the context given to the template. During rendering we will need to deduce whether we are dealing with literals or context variable names that need to be resolved. For this, we will use ast.literal_eval() which can safely evaluate strings containing Python code:

defeval_expression(expr):try:return'literal',ast.literal_eval(expr)exceptValueError,SyntaxError:return'name',expr

If we are dealing with a context variable name instead of a literal then we need to resolve it by searching for its value in the context. We need to take care of dotted names and names that reference the parent context. Here’s our resolving function which is the final piece of the puzzle:

defresolve(name,context):ifname.startswith('..'):context=context.get('..',{})name=name[2:]try:fortokinname.split('.'):context=context[tok]returncontextexceptKeyError:raiseTemplateContextError(name)

Conclusion

I hope this academic exercise gave you a taste of how the internals of a templating engine might work. This is far and away from production quality but it can serve as the basis of better things.

You can find the full code on Github, and you can discuss this further on Hacker News.

Huge thanks to Nassos Hadjipapas, Alex Loizou, Panagiotis Papageorgiou and Gearoid O’Rourke for reviewing.