What it says. Cuz it skips.
Note this is a toy project and is not meant for production usage...yet(?). It's primary purpose will be as part of a database internals teaching project.
The implementation of the algorithms in v1 adheres somewhat faithfully to the algorithms as laid out in the original paper by Pugh.
Uses a geometric distribution for determining if a new key is part of a level (fancy for saying we flip a coin). The geometric distrubution actually defaults to p = 0.25 but this is configurable.
The initial versions of this will be a non-concurrent version, requiring an external lock for
concurrent writes. The hope is to iteratively add concurrency features with Arc
/RwLock
first and
then lock-free methods following.
Rc
/RefCell
construction or using
unsafe
/raw pointers much like the standard library's linked list implementation