Thats not at all true though?

If you have a small dataset, yeah, memcpy will outperform a lot of indirect pointer lookups. But that doesn't stay true once you're memcpying around megabytes of data. The trick with indirect data structures on modern hardware is to tune the size of internal nodes and leaf nodes to make the cache misses worth it. For example, Binary trees are insanely inefficient on modern hardware because the internal nodes have a size of 2. If you give them a size of 64 or something, they perform much better. (Ie, make a b-tree). Likewise, a lot of bad tree implementations put just a single item in the leaf nodes. Its much better to have leaves store blocks of 128 items or something. And use memcpy to move data around within the block when needed.

This gets you the best of both worlds.

I spent about 18 months optimising a text based CRDT library (diamond types). We published a paper on it. By default, we store the editing history on disk. When you open a document, we reconstruct the document from scratch from a series of edits. After awhile, actually applying the stream of edits to a text document became the largest performance cost. Ropes were hugely useful. There's a stack of optimisations we made there to make that replay another 10x faster or so on top of most rope implementations. Using a linear data structure? Forget it. For nontrivial workloads, you 100% want indirect data structures. But you've gotta tune them for modern hardware.

My comment is an observation about how this gets tried every few years in major libraries and is usually reverted. I Agree, there are use cases where these are better. But the pattern tends to be to revert to simpler data structures.