Performance characteristics of the underlying hardware dictate the software abstractions that make sense. In today's machines, we still have differential cost for sequential, aligned, rightly sized block io operations vs random io operations of random sizes. And we have a hierarchy of storages with different latency, performance and costs – cpu caches, ram – and ram, ssd, spinning disks – and these via local attached vs in disaggregated and distributed clusters.
So, if you want absolutely optimal performance, you still have to care about your column sizes and order of columns in your records and how your records are keyed and how your tables involved in join operations are organized and what kinds of new queries are likely in your application. This matters for both very small scale (embedded devices) and very large scale systems. For mid-scale systems with plenty of latency and cost margins, it matters a lot less than it used. Hence, we have the emergence of nosql in the last decade and distributed sql in this decade.
