An SSD does not solve the problem of page fault chasing, it just makes it slightly less bad. This is fundamentally a software architecture problem.

This is solved with latency-hiding I/O schedulers, which don’t rely on cache locality for throughput.

Hi, I'm the author of the article. The sequential access pattern of IVF makes prefetching and large block sequential reads much easier, whereas it's almost impossible for HNSW to achieve efficient prefetching.

Even SSD won't be fast enough for most indexes due to the random access nature. I've seen more than 1M iops on a huge nvme disk when use DiskANN index

Data farms are all about cost per GB. Spinning media is a fraction of the cost per.