Complex algorithms over simple data can have big performance payoffs, remove obstacles, and even simplify things.
For instance, binary search over a sorted array (as opposed to a BinaryTree object):
• simplifies merging (it's just concat & sort)
• no pointers, which makes serialisation easier, but also...
• bypasses the need for serialisation
• an array can be on disk, or in memory, or both (via mmap)
• the data can be bigger than your ram
• allows for 'cold starts': just point your program at the file/mapping and tell it to run, no 'loading into ram' step.
• it's also cache-oblivious
Another example is huffman coding. If you studied it at uni, you probably saw it as a Tree-based algorithm with the associated O(n logn) time complexity. I was unaware that there was an in-place, array-based method of constructing a Huffman tree in linear time.
Of course 99% of the time, I'm doing back-end microservices and just using standard collection data structures. But if I ever had a 'big data' task at work, I would strongly prefer to be able to do it on a single machine with a big disk locally rather than buy into whatever the current flavour of MapReduce is.