Last July I was looking around for some Rust projects to contribute to in order to sharpen my teeth on an existing codebase. I ran into a Rust crate called datetime. After I heard that Rust’s Regex was supposedly very fast - I though - a crate with such a prominent name1 and it does not yet support parsing dates?
So I went ahead and added the functionality to parse iso8601 with regular expressions. With a little help of http://regexr.com/ I got it done. While I dug around in the internals of datetime I also found and fixed a few minor issues, hence it’s original maintainer already had good experiences excepting my pull requests :D
After my Regex parser was accepted I read about yet another implementation of iso8601 in Rust, using a parser combinator called nom instead of regular expressions. At the time I just though “Oh well - but I just got done, and Rust’s Regex is fast.”
2 months later I played around with Rust’s benchmarking. Because I heard many good things about nom I decided to compare chrono’s2 and datetimes parsing performance with the implementation presented by Jan-Erik Rediger’s implementation using nom. The Results were devastating:
Running target/release/dateparser_benchmarks-71747040d4b5a03b running 4 tests test chrono_bench::parse_iso8601 ... bench: 866 ns/iter (+/- 3) test datetime_bench::parse_iso8601 ... bench: 82,033 ns/iter (+/- 847) test nomdate_bench::parse_iso8601 ... bench: 260 ns/iter (+/- 2) test result: ok. 0 passed; 0 failed; 0 ignored; 4 measured
Too my defense: I hadn’t spent a second on optimizations, and the datetime implementation was still the most standard conform out of the three.
A little later, without my participation, this benchmark got a little bit of attention on reddit. With help of Piotr Czarnecki it was possible to deduce that the instantiation of Regex objects is the most expensive part:
... test datetime_regex_pure_bench::create_regex ... bench: 41,473 ns/iter (+/- 1,217) test datetime_regex_pure_bench::apply_regex ... bench: 1,484 ns/iter (+/- 137) ...
That doesn’t help much though, because when you parse a single string you still have to instantiate the Regex each time, and pulling it out of the implementation like Piotr did would just not make for a neat API. And even then, the nom version was still twice 3x as fast.
Rolling up the sleeves
So the next time when I had a little time on my hands, I though I’d give it a try and put in Jan-Erik’s nom parser into datetime instead of mine.3 After all, I kinda owed ogham as much, after leaving him with such a slow implementation :D
Turns out, the nom parser was still a bit incomplete since it was never meant for release as a crate. So I decided to contribute to both crates:
- make iso8601 a releasable general purpose library that exposes a simple API
- remove regex parsing from datetime all together
The important differentiation is that iso8601 is no datetime library, it only knows how to parse those darn strings, while datetime itself does not care about iso standards, but understands what a valid date is. There is preliminary validity checking in iso8601, but that may even be dropped later on, if it has a performance advantage. Though the nom version became slightly slower through this 4.
After sending pull requests to two separate crate owners I am glad to announce, datetime now has a faster and more complete parsing date-string-parsing implementation than chrono, thanks to @badboy and @ogham.
This may be a very small contribution to the Rust crates ecosystem, but I had the chance to improve my Regex skills, could play around with Rusts benchmarking features, got a good look at how fast and safe parsers are really build and got a little bit of github-cred along the way. Let’s call that a day.
currently the most popular date/time crate ↩
Meanwhile Datetime’s maintainer Benjamin Sago decided to tackle the same issue by using lazy static, At this point, sorry Ben for missing this, you had to pay the price. Ogham: “This merge sucks :p” ↩
I was never able to reproduce those 260ns numbers, even without modifications. ↩