in a different order than they began, and blow up a bad assert.

separate the difference between the oldest referenced xid we can use for
implicit promotion vs simple garbage collection.

run full garbage collection when a leafentry has > 1 committed entry.

node and notices that it is over 32mb

osx doesn't support __thread well so scoped_malloc breaks it.  Rather than
fix it, we implement a correct but poorly-performing implementation for
now that will be used on osx machines.

fixes #160

Refs Tokutek/ft-index#159 Calculate klpair size correctly when message application deletes a leafentry

Refs Tokutek/ft-index#159 Calculate klpair size correctly when garbage collection deletes a leafentry

Garbage collection was done on the clones and did not always make it to in-memory.
Also, closing a db and reopening no longer (necessarily) flushes the nodes out of memory; we close the entire environment when we want to ensure we read it.

partial eviction

thread-local buffer when possible.

frwlock

refs Tokutek/mongo#886, change get_last_key to use a callback instead of copying data, so that we don't have
mongo code freeing data that fractal tree allocated.

Add db->get_last_key() and test.
Add test

partition

eviction callback

serialization. Assume the fifo is about 1kb larger than the read buffer
(in practice, the fifo is 5 bytes smaller than the read buffer).

create_no_array() so that deserialization does not malloc/free needlessly.

logger

clock behavior more consistent with the rest of the code.

fixes #122