I use splay trees for their indirect utility. That is, the splay tree is a trivial way to calculate minimal, incremental updates to an append-only log. With most (all?) other b-tree structures, you would need to rewrite an unsustainable (or otherwise unpredictable) amount of tree data to the append-only log each time it is modified. The magic is that the splay operation redefines the root node every time you access the tree, so you just have to keep track of all the nodes you visited in each transaction batch and then write that sub-tree to the end of the log.

>paying attention to memory layout and cache line size is far, far more important.

Totally agree. For scenarios where I know I am going to have fewer than ~100k of something (assume an array of structs), I typically just do a linear scan over memory. If its a bunch of objects in memory or a cold access to storage, the trees become far more practical.