(In a nutshell, pack some hot data with a lot of cold data on many large drives, then put a Flash-based cache in front of them to get long tail performance predictability back.)
There was also a talk about the low level storage service and the performance isolation work that allows it to mix batch and latency-sensitive traffic on the same drive, but it doesn't seem to have been recorded: http://www.pdl.cmu.edu/SDI/2012/101112.html
Gory details are in the patents, 9781054, 9262093 and 8612990, which I'm not linking directly, because your lawyers might not approve. There's even a follow up, 10257111. It's so new, from two days ago, that Google Patents can't find it, while Justia can.
You might not need to speculate much about how it works, it's probably implemented as described by Google themselves in slides 22ff. here:
http://www.pdsw.org/pdsw-discs17/slides/PDSW-DISCS-Google-Ke...
(In a nutshell, pack some hot data with a lot of cold data on many large drives, then put a Flash-based cache in front of them to get long tail performance predictability back.)