ironoxidizer
Fastest Template Engine
2020-06-23
Preface: I realize most of the following template engines are template libraries, but referring to them all as engines makes things less confusing for readers and they both fulfill similar roles anyway.
Preface 2: I’ve recently discovered Sailfish and I’m considering testing it out. I’ll make a blog post about it in the future.
I’ve recently been working on lemmy-lite and I’ve had a few people ask me why I chose to use Maud as the template engine when there are more popular engines even within the Rust community. I referred them to this dated benchmark from 2016 and realized I should make a quick blog post with updated 2020 performance. Here are the results.
Methodology
Hardware
CPU: i7 6700
RAM: 16GB 2133MHz DDR4
Drive: Samsung 860 EVO 480GB
Software
linux lts 5.4
alpine 3.12
musl 1.1
llvm 3.6.2
cargo 1.45 nightly
rustc 1.45 nightly
serde 1.0.113
serde_json 1.0.55
rustc-serialize 0.3.24
askama 0.9.0
handlebars 3.1.0
horrorshow 0.8.3
liquid 0.20.1
maud 0.22.0
tera 1.3.1
Commands
git clone https://github.com/lambda-fairy/maud.git
cd maud/benchmarks
cargo bench
Results
Benchmark was done 5 times, values are times in nanoseconds to complete each iteration
| Run 1 | Run 2 | Run 3 | Run 4 | Run 5 | Median | Relative | |
|---|---|---|---|---|---|---|---|
| askama | 899 | 1010 | 923 | 1032 | 1008 | 1008 | 1.47 |
| handlebars | 16652 | 18398 | 17976 | 17997 | 17899 | 17976 | 26.20 |
| horrorshow | 798 | 888 | 893 | 859 | 892 | 888 | 1.29 |
| liquid | 12823 | 15052 | 14948 | 15143 | 15333 | 15052 | 21.94 |
| maud | 599 | 700 | 686 | 690 | 659 | 686 | 1.00 |
| tera | 8115 | 9227 | 9311 | 9346 | 8974 | 9227 | 13.45 |
Producing this median graph
Conclusion
Maud is 23% faster than the next fastest (horrorshow) and 26 times faster than handlebars; the most popular template engine in Rust according to crates.io downloads.
I know some of you may be thinking that a few nanoseconds doesn’t matter in a world where requests take hundreds of milliseconds anyways, but consider how many more simultaneous client connections could be handled on the same machine given how fast and efficient Maud is. This makes it ideal in situations where processing is limited but bandwidth is not which happens to be the case for systems like the Raspberry Pi 4 with gigabit networking.
That being said, speed isn’t everything, and Maud has it shortcomings as well (which I might cover in a future blog post), but for the time being, I’m planning on sticking with Maud for lemmy-lite and future projects where templating is needed. I’ve been enjoying the library and the confidence it gives me knowing there isn’t a faster solution out there that I could otherwise be using.