How To Use the Google Template System

(as of 27 February 2006)

Motivation

A template system can be used to separate output formatting specifications, which govern the appearance and location of output text and data elements, from the executable logic which prepares the data and makes decisions about what appears in the output.

Template systems lie along a continuum of power versus separation. "Powerful" constructs like variable assignment or conditional statements make it easy to modify the look of an application within the template system exclusively, without having to modify any of the underlying "application logic". They do so, however, at the cost of separation, turning the templates themselves into part of the application logic.

This template system leans strongly towards preserving the separation of logic and presentation. It is intentionally constrained in the features it supports and, as a result, applications tend to require quite a bit of code to instantiate a template. This may not be to everybody's tastes. However, while this design limits the power of the template language, it does not limit the power or flexibility of the template system. This system supports arbitrarily complex text formatting. Many Google applications, including the "main" Google web search, use this system exclusively for formatting output.

Finally, this system is designed with an eye towards efficiency. Template instantiation is very quick, with an eye towards minimizing both memory use and memory fragmentation.

Overview

There are two parts to the Google Template System:

The templates are text files that contain the format specification for the formatted output, i.e, the template language. The data dictionaries contain the mappings from the template elements (markers) embedded in the templates to the data that they will format. Here's a simple template:

   <html><head><title>{{TITLE}}</title>{{META_TAGS}}</head>
   <body>{{BODY}}</body>

Here's a dictionary that one could use to instantiate the template:

   {"TITLE": "Template example",
    "BODY": "This is a simple template example.\nIt's boring",
    "DATE": "11/20/2005"}

If we instantiated the template with this dictionary, here's the output we would get:

   <html><head><title>Template example</title></head>
   <body>This is a simple template example.
It's boring</body></html>

{{TITLE}} and {{{BODY}} are template elements, also called markers. In the dictionary, TITLE, BODY, and DATE are dictionary names, and the values associated with each one, such as 11/20/2005, are dictionary values.

A few points are clear even from this simple example:

  1. Dictionary keys and values are strings; the Google template system is not typed.
  2. Dictionary values come already formatted. It was up to the application code to decide how to format the value for DATE, and to insert the date into the dictionary already formatted.
  3. Not all dictionary values must be used by a templete. DATE is entirely ignored.
  4. Not all template elements may exist in the dictionary. In this example, {{META_TAGS}} is not found in the dictionary. This is perfectly legal; missing variable markers evaluate to the empty string.

Templates

The template language has four types of markers:

  1. VARIABLE markers, which are replaced by text based on dictionary values. All markers in the above example are variable markers. Variable markers look like this: {{VARIABLE}}
  2. START SECTION and END SECTION markers, which delimit sections which may appear zero, one, or N times in the output. The number of times a section appears is determined by the data dictionaries, as explained below. Each time a section is expanded, it uses a different dictionary, so that the output values may be different from one iteration of a section expansion to another. Note that the specification of how sections expand is entirely dependent on the dictionary, as set up by the application; there is no way to specify a repeat rate in the template language itself. Section markers look like this: {{#SECTION_NAME}}...{{/SECTION_NAME}}
  3. TEMPLATE-INCLUDE markers, which designate other templates to be expanded and inserted at the location where the marker appears. These are treated much like sections -- one may think of them as sections whose content is specified in a different file instead of inline -- and just like sections, can be expanded zero, one or N times in the output, each with a different dictionary. Template-include markers look like this: {{>FILENAME}}
  4. COMMENT markers, which may annotate the template structure but drop completely out of the expanded output. Comment markers look like this: {{! comment lives here -- cool, no?}}

These marker types each have their own namespace. For readability, however, it is best to not overuse a single name.

Anything found in a template of the form {{...}} is interpreted as a template marker. All other text is considered formatting text and is output verbatim at template expansion time. Formatting text may consist of HTML tags, XML tags, linefeeds and other spacing characters, constant text, etc.

Data Dictionaries

A data dictionary is a map from keys to values. The keys are always strings, each string representing either a variable, a section, or a template-include file. (Comments are not stored in the data dictionary!) These values correspond to the name of the associated template marker: a section {{#FOO}} in the template text is matched to the key "FOO" in the dictionary, if it exists. Note the case must match as well.

The value associated with a key differs according to key type. The value associated with a variable is simple: it's the value for that variable. (Niggly implementation note: this value may not have internal NUL -- that is, \0 -- characters. It's a TODO to allow this one day.)

The value associated with a section is more complicated, and somewhat recursive: it's a list of data dictionaries. Come template-expansion time, the section is expanded once for each dictionary in the list, so if there are two dictionaries in the list, then the section text will occur in the output twice. The first time, all variables/etc. in the section will be evaluated taking into account the first dictionary. The second time, all variables/etc. will be evaluated taking into account the second dictionary. (See below for a definition of "taking into account.")

A template-include is a special type of section, so the associated value is the same: a list of dictionaries. Template-includes also have one other, mandatory associated piece of information: the filename of the template to include. This filename may be specified either as an absolute path, or as a relative path. (In the latter case, the path is taken relative to the template_root, as set by the application.)

The application program is responsible for building this data dictionary, including all nesting. It then applies this dictionary to a single template to produce formatted output.

Expanding a Template

A program using Google Templates typically reads in templates at load time. During the course of program execution, the program will repeatedly perform the following two steps: first, instantiate a data dictionary, and second, apply the dictionary to the template to produce output.

The template system applies a dictionary to a template by finding all template markers in the template, and replacing them with the appropriate dictionary values. It matches template markers to dictionary keys in the obvious way. For instance, a template marker {{FOO}} matches the dictionary key FOO. {{FOO:html_escape}} matches FOO as well. The marker {{#BAR}} matches the dictionary key BAR, as does the marker {{/BAR}}. The marker {{>BAZ}} matches the dictionary key BAZ. (And of course, the marker {{! comment}} doesn't match any dictionary key at all.)

Template-variables can also have modifiers. In that case, the template-system starts by finding the appropriate value for that variable in the dictionary, just like normal. Then it applies each modifier to the variable, left to right. Finally, it emits the modified value to the output.

If no dictionary key is found for a given template marker, then the template marker is ignored: if a variable, it expands to the empty string; if a section or include-template, the section or include-template is expanded zero times.

All names are case sensitive. Names -- that is, variable keys and, as a result, template markers -- must be made of (7-bit ascii) alphanumeric characters and the underscore. The commment marker, which does not map to dictionary keys, may contain any chararacters whatsoever except }, the close-curly brace. It's a syntax error for any template marker to violate this rule.

Outside of the template markers, templates may contain any text whatsoever, including (single) curly braces and NUL characters.

Variable Modifiers

Recall that variables look like this: {{VARNAME}}. We actually allow a more generic form: the variable name may be followed by one or more modifiers. A modifier is a filter that's applied at template-expand time, that munges the value of the variable before it's output. For instance, consider a template that looks like this:

   <html><body>{{NAME:html_escape}}</body></html>

This asks the template system to apply the built-in html_escape modifier when expanding {{NAME}}. If you set NAME in your dictionary to be Jim & Bob, what will actually be emitted in the template is Jim &amp; Bob.

You can chain modifiers together. This template first html-escapes NAME, and then javascript-escapes that result:

   <html><body>{{NAME:html_escape:javascript_escape}}</body></html>

Modifiers typically have a long, descriptive name and also a one-letter abbreviation. So this example is equivalent to the previous one:

   <html><body>{{NAME:h:j}}</body></html>

Only the modifiers built in to the template system may be used: It is not possible at the time for users to register modifiers. Here are the modifiers that are supported:

long nameshort namedescription
:html_escape:h html-escapes the variable before output (eg & -> &amp)
:javascript_escape:j javascript-escapes the variable before output (eg " -> \")

Details on Dictionary Lookup

The dictionary structure is a tree: there's a 'main' dictionary, and then sub-dictionaries for each section or include-template. Even with all this complexity, the lookup rules are mostly straightforward: when looking up a marker -- be it a variable, section, or include-template marker -- the system looks in the currently applicable dictionary. If it's found there, great. If not, and the parent dictionary is not an include-template, it continues the look in the parent dictionary, and possibly the grandparent, etc. That is, lookup has static scoping: you look in your dictionary and any parent dictionary that is associated with the same template-file. As soon as continuing the lookup would require you to jump to a new template-file (which is what include-template would do), we stop the lookup.

For instance, for a template that says {{#RESULTS}}{{RESULTNUM}}. {{>ONE_RESULT}}{{#RESULTS}}, "ONE_RESULT" is looked for in the "RESULTS" dictionary, and if not found there, is looked for in the main, top-level dictionary. Likewise, the variable "RESULTNUM" is looked for first in the "RESULTS" dictionary, then in the main dictionary if necessary. However, "ONE_RESULT" will not do equivalent cascading lookups. In fact, it will have no parent dictionaries at all, because it's a different template file and thus in a different scope.

Because of these scoping rules, it's perfectly reasonable to set all variables that are needed in a given template file, in the top-level dictionary for that template. In fact, the ShowSection() function is provided to support just this idiom. To avoid confusion in such a usage mode, it's strongly encouraged that you give unique names to all sections and include-templates in a single template file. (It's no problem, given the template scoping rules, for a single section or include-template name to be repeated across different template files.)

There's a single special case: the global variable dictionary. Every dictionary inherits its initial set of values from the global dictionary. Clients can set variables in the global dictionary just like they can in normal template dictionaries they create.

The system initializes the global dictionary with a few useful values for your convenience. All system variables are prefixed with BI, to emphasize they are "built in" variables.

As is usual for inheritence, if a user explicitly assigns a value to these variable-names in its own dictionary, this overrides the inherited value. So, dict->SetValue("BI_SPACE", "&nbsp;") causes BI_SPACE to have the value &nbsp;, rather than <space>, when expanding dict.

Note that only variables can be inherited from the global dictionary, not section dictionaries or include-file dictionaries.

A couple of small implementation notes: global inheritence is "last chance", so if a section's parent dictionary redefined BI_SPACE, say, the section dictionary inherits the parent-dict value, not the global-dict value. Second, variable inheritence happens at expand time, not at dictionary-create time. So if you create a section dictionary, and then afterwards set a variable in its parent dictionary (or in the global dictionary), the section will inherit that variable value, if it doesn't define the value itself.

Writing Application Code To Use Templates

Most application code concerns filling a template dictionary, but there is also code for loading templates themselves from disk. A final category of code lets you inspect and control the template system.

The code below assumes the default configuration option of putting all template code in namespace google.

Loading A Template

The main routine to load a template is google::Template::GetTemplate(), defined in template.h. This is a static, factory method, that loads a template from either disk or from an internal template cache, and returns a pointer to a Template object. Besides a filename to load from, this routine takes a 'strip' argument which defines how to expand whitespace found in a template file. It can have one of the following values:

This factory method returns NULL if the template cannot be found, or if there is a syntax error trying to load it.

Besides loading templates, the application can also ask the template system to reload a template, via template->ReloadIfChanged(). (You can also reload all templates at once via google::Template::ReloadAllIfChanged().) ReloadIfChanged() looks on disk, and if it notices the template file has changed since the last load, it will reload the template from disk, replacing the old contents. Actually, the reload is done lazily: ReloadIfChanged just sets a bit that causes the template to be reloaded next time GetTemplate is called.

Creating A Template Dictionary

The class google::TemplateDictionary is used for all template dictionary operations. new google::TemplateDictionary(name) is used to create a new top-level dictionary. dict->AddSectionDictionary(name) and dict->AddIncludeDictionary(name) are used to create sub-dictionaries for sections or include-files. After creating a dictionary, the application should call one or more functions for each marker in the template. As an example, consider the following template:

<html><body> {{! This page has no head section.}}
{{#CHANGE_USER}}
<A HREF="/login">Click here</A> if you are not {{USERNAME}}<br>
{{/CHANGE_USER}}

Last five searches:<ol>
{{#PREV_SEARCHES}
<li> {{PREV_SEARCH}}
{{/PREV_SEARCH}}
</ol>

{{>RESULT_TEMPLATE}}

{{FOOTER}}
</body></html>

To instantiate the template, the user should call a function to set up FOOTER, and a function to say what to do for the sections CHANGE_USER and PREV_SEARCHES, and for the include-template RESULT_TEMPLATE. Quite likely, the application will also want to create a sub-dictionary for CHANGE_USER, and in that sub-dictionary call a function to set up USERNAME. There will also be sub-dictionaries for PREV_SEARCHES, each of which will need to set PREV_SEARCH. Only when this is all set up will the application be able to apply the dictionary to the template to get output.

The appropriate function to call for a given template marker depends on its type.

Variables

For variables, the only interesting action is to set the variable's value. For most variables, the right method to call is dict->SetValue(name, value). (The name and value can be specified as strings in a variety of ways: C++ strings, char *'s, or char *'s plus length.) You can also call google::TemplateDictionary::SetGlobalValue(name, value) -- no TemplateDictionary instance needed here -- to set a variable that can be used by all templates in an applications. This is quite rare.

In addition to SetValue(), there are a few helper routines to help setting values of a few special forms.

Example:

   google::TemplateDictionary* dict = new google::TemplateDictionary("var example");
   dict->SetValue("FOOTER", "Aren't these great results?");
   class StarEscape { string operator()(const string& in) const { return string("*") + in + string("*"); } };
   dict->SetEscapedValue("USERNAME", username, StarEscape());

Note that the template itself can also specify escaping via variable modifiers! It's very possible for you to escape the value when setting it in the dictionary, and then have the template escape it again when outputting, so be careful you escape only as much as you need to.

Sections

Sections are used in two ways in templates. One is to expand some text multiple times. This is how PREV_SEARCHES is used in the example above. In this case we'll have one small sub-dictionary for each of the five previous searches the user did. To do this, call AddSectionDictionary(section_name) to create the sub-dictionary. It returns a TemplateDictionary* that you can use to fill the sub-dictionary.

The other use of sections is to conditionally show or hide a block of text at template-expand time. This is how CHANGE_USER is used in the example template: if the user is logged in, we show the section with the user's username, otherwise we choose not to show the section.

This second case is a special case of the first, and the "standard" way to show a section is to expand it exactly one time, by calling AddSectionDictionary() once, and then setting USERNAME in the sub-dictionary.

However, the hide/show idiom is so common there are a few convenience methods to make it simpler. The first takes advantage of the fact sections inherit variables from their parent: you set USERNAME in the parent dictionary, rather than a section sub-dictionary, and then call ShowSection(), which adds a single, empty dictionary for that section. This causes the section to be shown once, and to inherit all its variable values from its parent.

A second convenience method is written for the particular case we have with USERNAME: if the user's username is non-empty, we wish to show the section with USERNAME set to the username, otherwise we wish to hide the section and show neither USERNAME nor the text around it. The method SetValueAndShowSection(name, value, section_name) does exactly that: if value is non-empty, add a single single dictionary to section_name and call section_dict->AddValue(name, value). There's also SetEscapedValueAndShowSection(name, value, escape_functor, section_name), which lets you escape value.

Example:

   using google::TemplateDictionary;
   TemplateDictionary* dict = new TemplateDictionary("section example");   
   const char* username = GetUsername();   // returns "" for no user
   if (username[0] != '\0') {
      TemplateDictionary* sub_dict = dict->AddSectionDictionary("CHANGE_USER");
      sub_dict->SetValue("USERNAME", username);
   } else {
      // don't need to do anything; we want a hidden section, which is the default
   }

   // Instead of the above 'if' statement, we could have done this:
   if (username[0] != '\0') {
      dict->ShowSection("CHANGE_USER");       // adds a single, empty dictionary
      dict->SetValue("USERNAME", username);   // take advantage of inheritence
   } else {
      // don't need to do anything; we want a hidden section, which is the default
   }

   // Or we could have done this:
   dict->SetValueAndShowSection("USERNAME", username, "CHANGE_USER");

   // Moving on...
   GetPrevSearches(prev_searches, &num_prev_searches);
   if (num_prev_searches > 0) {
      for (int i = 0; i < num_prev_searches; ++i) {
         TemplateDictionary* sub_dict = dict->AddSectionDictionary("CHANGE_USER");
         sub_dict->SetEscapedValue("PREV_SEARCH", prev_searches[i],
                                   TemplateDictionary::html_escape);
      }
   }

Template-includes

Template-include markers are much like section markers, so SetIncludeDictionary(name) acts, not surprisingly, exactly like SetSectionDictionary(name). However, since variable inheritence doesn't work across include boundaries, there is no template-include equivalent to ShowSection() or SetValueAndShowSection().

One difference bewteen template-includes and sections is that for a sub-dictionary that you create via SetIncludeDictionary(), you must call subdict->SetFilename() to indicate the name of the template to include. If you do not set this, the sub-dictionary will be ignored. The filename may be absolute, or relative, in which case it's relative to template_root.

Example:

   using google::TemplateDictionary;
   TemplateDictionary* dict = new TemplateDictionary("include example");   
   GetResults(results, &num_results);
   for (int i = 0; i < num_results; ++i) {
      TemplateDictionary* sub_dict = dict->AddIncludeDictionary("RESULT_TEMPLATE");
      sub_dict->SetFilename("results.tpl");
      FillResultsTemplate(sub_dict, results[i]);
   }

In practice, it's much more likely that FillResultsTemplate() will be the one to call SetFilename(). Note that it's not an error to call SetFilename() on a dictionary even if the dictionary is not being used for a template-include; in that case, the function is a no-op, but is perhaps still useful as self-documenting code.

Expanding a Template

Once you have a template and a template dictionary, it's simplicity itself to expand the template with those dictionary values, putting the output in a string:

   google::Template* tpl = google::Template::GetTemplate(<filename>, google::STRIP_WHITESPACE);
   google::TemplateDictionary dict("debug-name");
   FillDictionary(&dict, ...);
   string output;
   tpl->Expand(&output, &dict);
   // output now holds the expanded template

The expanded template is written to the string output. If output was not empty before calling Expand(), the expanded template is appended to the end of output.

Getting a Template From a String Rather Than a File

The TemplateFromString class, in template_from_string.h, is an alternative to the Template class when you really want your template to be built in to the executable rather than read from a file. It's a drop-in replacement, that takes an extra argument which is the template contents.

Prefer Template to TemplateFromString, for several reasons. For one, updating the template requires merely a data push, rather than pushing the new executable. Also, you can load the new template without needing to restart the binary. It also makes it easier for non-programmers to modify the template. Finally, string-templates cannot be included by other templates, since {{>include}} takes a filename.

One reason to use TemplateFromString is if you are in an environment where having data files could be dangerous -- for instance, you work on a disk that is usually full, or need the template to work even in the face of disk I/O errors.

This package comes with a script, template-converter, that takes a template file as input and emits a C++ code snippet (an .h file) that defines a string with those template contents. This makes it easy to start by using a normal, file-based template, and then switch to template-from-string later if you so desire.

Working Effectively with Templates

Registering Template Strings

Both dictionary keys and template filenames are strings. Instead of using raw strings, we encourage you to use a bit of machinery to help protect against various types of errors.

For dictionary keys, you can use the make_tpl_varnames_h tool to create static string variables to use instead of a string constant. This will protect against typos, as the make_tpl_varnames_h documentation describes.

For template filenames that a program uses -- including sub-templates -- we suggest the following idiom:

   #include "example.tpl.varnames.h"   // defines 1 string per dictionary key
   RegisterTemplateFilename(EXAMPLE_FN, "example.tpl");   // defines template
   ...
   google::Template* tpl = google::Template::GetTemplate(EXAMPLE_FN, ...);
   ...
   include_dict->SetFilename(EXAMPLE_FN);

By registering the filename, you can query the template system to detect syntax errors, reload-status, and so forth.

Managing Templates

The following functions affect the global state of the template system.

There are some administrative tools that can help with tweaking template performance and debugging template problems. The following functions work on registered templates.

The following functions help with debugging, by allowing you to examine the template dictionaries and expanded templates in more detail.

Finally, ClearCache() removes all template objects from the cache used by GetTemplate(). Typically, this is only used in environments that check for memory leaks: calling this at the end of the program will clean up all memory that the template system uses.

Template and Threads

All static methods on Template and TemplateDictionary objects are threadsafe: you can safely call google::TemplateDictionary::SetGlobalValue() without needing to worry about locking.

Non-static methods are not thread-safe. It is not safe for two threads to assign values to the same template-dictionary without doing their own locking. Note that this is expected to be quite rare: usually only one thread will care about a given template-dictionary.

For Template objects, the most common idiom is that a template is loaded via GetTemplate(), and after that only const methods like Expand() are called on the template. With such usage, it's safe to use the same Template object in multiple threads without locking. Be careful, however, if you also call functions like ReloadIfChanged().

Development Tools

This package includes several tools to make it easier to use write and use templates.

make_tpl_varnames_h: Template Syntax Checker and Header File Generator

make_tpl_varnames_h is a "lint" style syntax checker and header file generator. It takes the names of template files as command line arguments and loads each file into a Template object by retrieving the file via the Template factory method. The loading of the file does pure syntax checking and reports such errors as mis-matched section start/end markers, mis-matched open/close double-curly braces, such as "{{VAR}", or invalid characters in template variables/names/comments.

If the template passes the syntax check, by default the utility then creates a header file for use in the executable code that fills the dictionary for the template. If the developer includes this header file, then constants in the header file may be referenced in the dictionary building function, rather than hard-coding strings as variable and section names. By using these constants, the compiler can notify the developer of spelling errors and mismatched names. Here's an example of how this is used, and how it helps prevent errors:

   const char * const kosr_RESULT_NUMBER = "RESULT_NUMBER";  // script output
   dict.SetValue("RESSULT_NUMBER", "4");    // typo is silently missed
   dict.SetValue(kosr_RESSULT_NUMBER, "4");   // compiler catches typo

Each constant is named as follows:

As an example, the section name "RESULT_NUMBER" in the file one_search_result_post20020815.tpl would be given the constant name kosr_RESULT_NUMBER and would appear in the header file as const char * const kosr_RESULT_NUMBER = "RESULT_NUMBER"; -- as in the example above.

By default, the header file is produced in the current directory. An alternate output directory may be specified by the command line flag --header_dir.

The name of the generated header file is the same as the name of the template file with an extension added to the name. By default, that extension is .varnames.h. In the above example, the header file containing the constant declarations would be named one_search_result_post20020815.tpl.varnames.h. An alternate extension may be provided via the command line flag --outputfile_suffix.

Important command line flags:

For a full list of command line flags, run make_tpl_varnames_h --help.

template-converter: convert a template to a C++ string

The TemplateFromString class lets you load a template from a string instead of a file. Applications may prefer this option to reduce the dependencies of the executable, or use it in environments where data files are not practical. In such cases, template-converter can be used as a template "compiler", letting the developer write a template file as a data file in the normal way, and then "compiling" it to a C++ string to be included in the executable.

Usage is template-converter <template filename>. C++ code is output is to stdout; it can be stored in a .h file or included directly into a C++ file. Perl must be installed to use this script.



Craig Silverstein
27 February 2006