Pretty incredible. More than 10 times faster than the substring parser. This really does show that the performance gains to be had by working on UTF-8 can be truly substantial. This is the difference of being able to process more than 50 megabytes of logs per second, or being able to process a measly 4 megabytes of logs per second.

So this is pretty huge. We have found that our parser type is so general and so composable that we are not only able to parse at the low UTF-8 level in order to get huge performance gains, but we can also fluidly move between abstraction levels allowing us to choose the right balance of correctness and speed. This is absolutely incredible, and honestly it’s not something we’ve really ever seen in other parsing libraries.

And as hard as it may be to believe, it gets even better. There is even one more change we can make to our parser library that unlocks the final level of performance, and this will bring the performance of our parsers within a very slim margin of hand-rolled, ad hoc, imperative parsers, which are widely believed to be the fast way to write parsers even if they are a bit messy.

To see where this last performance gain could be hiding, let’s run one of our benchmarks in the time profile instrument and see what it exposes to us.