Fun and weirdness with SSDs

Cool stuff, and great data.

The behavior with SSDs is much more complicated, and heavily depends on the I/O pattern, which can have significant impact on query timings.

The other way to say this is that we’ve been really spoiled by SSDs. Spinning media have something like a 30x difference in throughput between 32kB random and forward sequential (unbuffered). 4kB is even worse, at over 200x. So while the 1 - 2x here is meaningful for query performance it’s also a huge improvement.

As a systems design rule of thumb, I think of modern SSDs saturating their throughput for any (decently concurrent) workload over 32kB. That seems roughly in line with the data here (but there’s a good opportunity to optimize beyond that).