> It is also hard to beat the performance of the STL

If you want to re-create the STL, maybe. But you can make custom data structures tailored to your task at hand instead.

For example, instead of a std::map or std::unordered_map that allocates and initializes each node separately, you could preallocate some of them in a big chunk of memory, hand them out via a bump allocator scheme, and later free them all at once. Instead of a std::sort algorithm, you could use a bucket sort if it's possible in your situation, to improve your asymptotics from O(n log n) to O(n). Etc, etc.

There's plenty of papers you can find online where people beat the STL: http://www.cs.cmu.edu/~dga/papers/cuckoo-eurosys14.pdf It's usually because there's some use case or underlying assumption they can make that the STL can't, because the STL is designed to be the best possible one-size-fits-all-approach, which is even further constrained by things like ABI history. So it's always a discouraging sign to hear a software engineer talking about such things as though they're holy, since that'd be like hiring a tailor who uses spandex.