Some related comments, with respect to the article:
- Modern database models are one process per physical core regardless of how many sessions, queries, connections, etc that there are. These cores own the part of the data they work with. This has several advantages on modern computing architectures. It also makes for a pretty elegant implementation.
- Related to the first point, new database kernels are increasingly "shared nothing" even within a single server. As in, physical resources will be cut up between cores / processes and rarely shared. It is like running a cluster within a machine. Again, this has some significant performance advantages on modern machines.
- The key advantage of SSDs, which was not true for spinning disk, is that you can usually guarantee effective disk I/O bandwidth is always significantly greater than the full-duplex network bandwidth to the server no matter what the workload. This has an interesting technical implication: if the I/O scheduler is correctly designed and implemented, an in-memory database engine should never be faster than a disk-backed database. In-memory was only an optimization that made sense for spinning disk; with SSD, if you aren't compute-bound, you can always saturate the network (and if you are compute-bound, in-memory does not help).