* Howard Chu:
We can provide plenty more documentation on LMDB performance and
reliability if desired.
Can you cope with incompletely written pages (e.g., only the first 512
bytes of a page is written) or write reordering between fsyncs?
Berkeley DB doesn't deal with torn writes, either, but it can deal
with write reordering. The latter is what makes the checkpoints so
expensive, but I don't think there's a way around that. You can
spread out the writes (PostgreSQL tries to do that, but it is not
entirely successful because of kernel API limitations), but then you
risk doing additional writes which are not strictly required for
durability.