lambda fairy

The fastest template engine in the West

Lately I’ve been working on Maud, an HTML template library for Rust. One of the features that make it special is that it works at compile time. That is, your templates are compiled to plain Rust code, and type-checked and optimized with the rest of your app.

Among other benefits, this design makes the library fast. Like, really, really fast. To get an idea of how fast it is, take a look at this graph:

Graph of render times for different template engines

That’s right. Maud is 69 times faster than Handlebars.1

I know what you’re thinking:

For the sake of professionalism, I will only answer the second question.

Maud is fast, because it does as little work as possible at runtime. For example, the following template:

expands to this Rust code:

In other words, the resulting code does little more than building a string.

Compare this to other—more dynamic—template engines, which may encode the input data as JSON, or look up the name of the template in a global registry, or do a variety of other things. These engines are effectively language interpreters, with all the pros and cons of working that way. And as the benchmark above shows, one of these cons is reduced performance.

That’s not to say that these dynamic approaches aren’t useful. Speed isn’t the only factor in choosing a template engine; and to be fair, when an average request takes hundreds of milliseconds already, a microsecond difference doesn’t matter that much. Other engines also let you edit a template without re-compiling the app, and their syntax can feel more familiar to users of other languages.

What I do want to show, then, is how design decisions that seem minor at first can have a big impact down the road. I’ve added a bunch of optimizations over the years (some of them stolen from Horrorshow), but none of them have affected performance that much. The largest difference is in static vs dynamic; the rest is details.

  1. The benchmark code can be found on GitHub.

(Posted on .)