I've found my implementation to be CPU intensive: creating Python objects from the raw pages is expensive. That's why bulk inserts and iterations are much faster than insert/get in a loop.
Maybe someone on HN knows a Python serialization library beyond pickle that would allow to describe how the data is laid out and take care of the rest. It looks like struct is not flexible enough for this usage.
And and can you use keys that aren't builtin values?
You might also mention that, if replacing large values that use overflow pages, the file has the potential to grow without bounds as it looks like overflow pages are not collected?
Or FusionIO ioDrive2 Duo, which is above 900K IOPS for both reads and writes.