Just to prevent rampant over-generalization in an article whose subject is the topic of bias: We've implemented various special-purpose sorts at least three times in the last five years. Here's one example: https://github.com/efficient/cuckoofilter/blob/master/src/pa...

Yes, this is all in a very high-performance (sometimes insanely so) context, but it does happen. Most of them were like this - unrolled special-purpose versions derived from a sorting network. Some were for GPU.

While this is for the most part true, questions regarding sorting algorithms and data structures (debug this left-rotate function that operates on binary trees) still come up during interviews.

I mostly agree, however there are exceptions that prove the rule. Some engineers are working on standard libraries. (someone has to write that code!) Some people are taking advantage of their data set to write special-purpose sorting algorithms that blow generic algorithms out of the water. (guilty!) Some people are putting stuff together in interesting ways that requires them to understand and sometimes even re-implement the standard algorithms to take advantage of internal data structures or other interesting effects. (It's bad practice, but knowing the order a map will iterate things in can be helpful if you don't expect it to change -- and make sure to have a unit test that proves it!)

You need them for those whiteboarding tech interviews.