[est31]

Another anecdote: A few months ago, I was adding a bunch of features to an open source library. Before changing, I studied the library's code to find out where to put the change best. During that study, I found an instance where the code performed a slow operation, but I didn't attempt to change it because it was out of scope for my change, and I didn't want to introduce bugs [1].

A little bit after I have made the change, an user filed a bug [2] to the library about slow behavior when you passed a large amount of data to the library. They were using the public release instead of git master, so my code wasn't at fault thankfully. Due to the knowledge I attained from doing the change, I could quickly confirm that precisely this slow part was cause for the performance slowdown, and was able to file a PR to reduce the library's complexity from quadratic to linear [3].

It wasn't very complicated from the algorithmic point of view, I only had to create a lookup structure, but the lookup structure had to be created in a certain way so that the library still behaved the same way as tested by the test suite. It also had to support things like duplicates as the original implementation also supported them, as a very important feature in fact (toml arrays).

[1]: https://github.com/alexcrichton/toml-rs/pull/333

[2]: https://github.com/alexcrichton/toml-rs/issues/342

[3]: https://github.com/alexcrichton/toml-rs/pull/349